update README

README.md

@@ -58,7 +58,7 @@ HunyuanVideo-1.5 is a video generation model that delivers top-tier quality with
 
 ## 🔥🔥🔥 News
 * 🚀 Dec 05, 2025: **New Release**: We now release the [480p I2V step-distilled model](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_i2v_step_distilled), which generates videos in 8 or 12 steps (recommended)! On RTX 4090, end-to-end generation time is reduced by 75%, and a single RTX 4090 can generate videos within 75 seconds. The step-distilled model maintains comparable quality to the original model while achieving significant speedup. See [Step Distillation Comparison](./assets/step_distillation_comparison.md) for detailed quality comparisons. For even faster generation, you can also try 4 steps (faster speed with slightly reduced quality). **To enable the step-distilled model, run `generate.py` with the `--enable_step_distill` parameter.** See [Usage](#-usage) for detailed usage instructions. 🔥🔥🔥🆕
-* 📚 Training code is coming soon. HunyuanVideo-1.5 is trained using the Muon optimizer, which we have open-sourced in the in [Training](#-training) section. **If you would like to continue training our model or fine-tune it with LoRA, please use the Muon optimizer.**
+* 📚 Dec 05, 2025: **Training Code Released**: We now open-source the training code for HunyuanVideo-1.5! The training script (`train.py`) provides a full training pipeline with support for distributed training, FSDP, context parallel, gradient checkpointing, and more. HunyuanVideo-1.5 is trained using the Muon optimizer, which we have open-sourced in the [Training](#-training) section. **If you would like to continue training our model or fine-tune it with LoRA, please use the Muon optimizer.** See [Training](#-training) section for detailed usage instructions. 🔥🔥🔥🆕
 * 🎉 **Diffusers Support**: HunyuanVideo-1.5 is now available on Hugging Face Diffusers! Check out [Diffusers collection](https://huggingface.co/collections/hunyuanvideo-community/hunyuanvideo-15) for easy integration. 🔥🔥🔥🆕
 * 🚀 Nov 27, 2025: We now support cache inference (deepcache, teacache, taylorcache), achieving significant speedup! Pull the latest code to try it. 🔥🔥🔥🆕 
 * 🚀 Nov 24, 2025: We now support deepcache inference.
@@ -106,7 +106,7 @@ If you develop/use HunyuanVideo-1.5 in your projects, welcome to let us know.
 - [🛠️ Dependencies and Installation](#️-dependencies-and-installation)
 - [🧱 Download Pretrained Models](#-download-pretrained-models)
 - [📝 Prompt Guide](#-prompt-guide)
-- [🔑 Usage](#-usage)
+- [🔑 Inference](#-inference)
   - [Inference with Source Code](#inference-with-source-code)
   - [Usage with Diffusers](#usage-with-diffusers)
   - [Prompt Enhancement](#prompt-enhancement)
@@ -114,7 +114,6 @@ If you develop/use HunyuanVideo-1.5 in your projects, welcome to let us know.
   - [Image to Video](#image-to-video)
   - [Command Line Arguments](#command-line-arguments)
   - [Optimal Inference Configurations](#optimal-inference-configurations)
-- [🧱 Models Cards](#-models-cards)
 - [🎓 Training](#-training)
 - [🎬 More Examples](#-more-examples)
 - [📊 Evaluation](#-evaluation)
@@ -205,6 +204,23 @@ pip install -i https://mirrors.tencent.com/pypi/simple/ --upgrade tencentcloud-s
 
 Download the pretrained models before generating videos. Detailed instructions are available at [checkpoints-download.md](checkpoints-download.md).
 
+### Model Cards
+|ModelName| Download                     |
+|-|---------------------------| 
+|HunyuanVideo-1.5-480P-T2V|[480P-T2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_t2v) |
+|HunyuanVideo-1.5-480P-I2V |[480P-I2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_i2v) |
+|HunyuanVideo-1.5-480P-T2V-cfg-distill | [480P-T2V-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_t2v_distilled) |
+|HunyuanVideo-1.5-480P-I2V-cfg-distill |[480P-I2V-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_i2v_distilled) |
+|HunyuanVideo-1.5-480P-I2V-step-distill |[480P-I2V-step-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_i2v_step_distilled) |
+|HunyuanVideo-1.5-720P-T2V|[720P-T2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_t2v) |
+|HunyuanVideo-1.5-720P-I2V |[720P-I2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_i2v) |
+|HunyuanVideo-1.5-720P-T2V-cfg-distill| Comming soon |
+|HunyuanVideo-1.5-720P-I2V-cfg-distill |[720P-I2V-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_i2v_distilled) |
+|HunyuanVideo-1.5-720P-T2V-sparse-cfg-distill| Comming soon |
+|HunyuanVideo-1.5-720P-I2V-sparse-cfg-distill |[720P-I2V-sparse-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_i2v_distilled_sparse) |
+|HunyuanVideo-1.5-720P-sr-step-distill |[720P-sr](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_sr_distilled) |
+|HunyuanVideo-1.5-1080P-sr-step-distill |[1080P-sr](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/1080p_sr_distilled) |
+
 ## 📝 Prompt Guide
 ### Prompt Writing Handbook
 Prompt enhancement plays a crucial role in enabling our model to generate high-quality videos. By writing longer and more detailed prompts, the generated video will be significantly improved. We encourage you to craft comprehensive and descriptive prompts to achieve the best possible video quality. we recommend community partners consulting our official guide on how to write effective prompts. 
@@ -214,7 +230,7 @@ Prompt enhancement plays a crucial role in enabling our model to generate high-q
 ### System Prompts for Automatic Prompt Enhancement
 For users seeking to optimize prompts for other large models, it is recommended to consult the definition of `t2v_rewrite_system_prompt` in the file `hyvideo/utils/rewrite/t2v_prompt.py` to guide text-to-video rewriting. Similarly, for image-to-video rewriting, refer to the definition of `i2v_rewrite_system_prompt` in `hyvideo/utils/rewrite/i2v_prompt.py`.
 
-## 🔑 Usage
+## 🔑 Inference
 
 ### Inference with Source Code
 
@@ -251,24 +267,28 @@ export I2V_REWRITE_MODEL_NAME="<your_model_name>"
 
 PROMPT='A girl holding a paper with words "Hello, world!"'
 
-IMAGE_PATH=none # Optional, none or <image path> to enable i2v mode
+IMAGE_PATH=/path/to/image.png # Optional, none or <image path> to enable i2v mode
 SEED=1
 ASPECT_RATIO=16:9
 RESOLUTION=480p
 OUTPUT_PATH=./outputs/output.mp4
+MODEL_PATH=./ckpts # Path to pretrained model
 
-# Configuration
-REWRITE=true # Enable prompt rewriting. Please ensure rewrite vLLM server is deployed and configured.
+# Configuration for faster inference
 N_INFERENCE_GPU=8 # Parallel inference GPU count
 CFG_DISTILLED=true # Inference with CFG distilled model, 2x speedup
-ENABLE_STEP_DISTILL=true # Enable step distilled model for 480p I2V, recommended 8 or 12 steps, 75% speedup on RTX 4090
-SPARSE_ATTN=false # Inference with sparse attention (only 720p models are equipped with sparse attention). Please ensure flex-block-attn is installed
 SAGE_ATTN=true # Inference with SageAttention
+SPARSE_ATTN=false # Inference with sparse attention (only 720p models are equipped with sparse attention). Please ensure flex-block-attn is installed
 OVERLAP_GROUP_OFFLOADING=true # Only valid when group offloading is enabled, significantly increases CPU memory usage but speeds up inference
 ENABLE_CACHE=true # Enable feature cache during inference. Significantly speeds up inference.
 CACHE_TYPE=deepcache # Support: deepcache, teacache, taylorcache
+ENABLE_STEP_DISTILL=true # Enable step distilled model for 480p I2V, recommended 8 or 12 steps, up to 6x speedup
+
+
+# Configuration for better quality
+REWRITE=true # Enable prompt rewriting. Please ensure rewrite vLLM server is deployed and configured.
 ENABLE_SR=true # Enable super resolution
-MODEL_PATH=ckpts # Path to pretrained model
+
 
 torchrun --nproc_per_node=$N_INFERENCE_GPU generate.py \
   --prompt "$PROMPT" \
@@ -412,64 +432,87 @@ with attention_backend("_flash_3_hub"): # or `"flash_hub"` if you are not on H10
 For more details, please visit [HunyuanVideo-1.5 Diffusers Collection](https://huggingface.co/collections/hunyuanvideo-community/hunyuanvideo-15).
 
 
-## 🧱 Models Cards
-|ModelName| Download                     |
-|-|---------------------------| 
-|HunyuanVideo-1.5-480P-T2V|[480P-T2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_t2v) |
-|HunyuanVideo-1.5-480P-I2V |[480P-I2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_i2v) |
-|HunyuanVideo-1.5-480P-T2V-cfg-distill | [480P-T2V-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_t2v_distilled) |
-|HunyuanVideo-1.5-480P-I2V-cfg-distill |[480P-I2V-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_i2v_distilled) |
-|HunyuanVideo-1.5-480P-I2V-step-distill |[480P-I2V-step-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_i2v_step_distilled) |
-|HunyuanVideo-1.5-720P-T2V|[720P-T2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_t2v) |
-|HunyuanVideo-1.5-720P-I2V |[720P-I2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_i2v) |
-|HunyuanVideo-1.5-720P-T2V-cfg-distill| Comming soon |
-|HunyuanVideo-1.5-720P-I2V-cfg-distill |[720P-I2V-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_i2v_distilled) |
-|HunyuanVideo-1.5-720P-T2V-sparse-cfg-distill| Comming soon |
-|HunyuanVideo-1.5-720P-I2V-sparse-cfg-distill |[720P-I2V-sparse-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_i2v_distilled_sparse) |
-|HunyuanVideo-1.5-720P-sr-step-distill |[720P-sr](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_sr_distilled) |
-|HunyuanVideo-1.5-1080P-sr-step-distill |[1080P-sr](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/1080p_sr_distilled) |
+## 🎓 Training
 
+HunyuanVideo-1.5 is trained using the **Muon optimizer**, which accelerates convergence and improves training stability. The Muon optimizer combines momentum-based updates with Newton-Schulz orthogonalization for efficient optimization of large-scale video generation models.
 
+### Quick Start
 
-## 🎓 Training
+The training script (`train.py`) provides a complete training pipeline for HunyuanVideo-1.5. Here's how to use it:
 
-> 💡 Training code is coming soon. We will release the complete training pipeline in the future.
+#### 1. Implement Your DataLoader
 
-HunyuanVideo-1.5 is trained using the **Muon optimizer**, which accelerates convergence and improves training stability. The Muon optimizer combines momentum-based updates with Newton-Schulz orthogonalization for efficient optimization of large-scale video generation models.
+Replace the `create_dummy_dataloader()` function in `train.py` with your own implementation. Your dataloader should return batches with the following format:
 
-### Creating a Muon Optimizer
+- **Required fields:**
+  - `"pixel_values"`: `torch.Tensor` - Video: `[B, C, F, H, W]` or Image: `[B, C, H, W]`
+    - Note: For video data, temporal dimension F must be `4n+1` (e.g., 1, 5, 9, 13, 17, ...)
+  - `"text"`: `List[str]` - Text prompts for each sample
+  - `"data_type"`: `str` - `"video"` or `"image"`
 
-Here's how to create a Muon optimizer for your model:
+- **Optional fields (for performance optimization):**
+  - `"latents"`: Pre-encoded VAE latents (skips VAE encoding for faster training)
+  - `"byt5_text_ids"` and `"byt5_text_mask"`: Pre-tokenized byT5 inputs
 
-```python
-from hyvideo.optim.muon import get_muon_optimizer
-
-# Create Muon optimizer for your model
-optimizer = get_muon_optimizer(
-    model=your_model,
-    lr=lr,                      # Learning rate
-    weight_decay=weight_decay,  # Weight decay
-    momentum=momentum,          # Momentum coefficient
-    adamw_betas=adamw_betas,   # AdamW betas for 1D parameters
-    adamw_eps=adamw_eps        # AdamW epsilon
-)
+See the `create_dummy_dataloader()` function in `train.py` for detailed batch format documentation.
+
+#### 2. Run Training
+
+**Single GPU:**
+```bash
+python train.py --pretrained_model_root <path_to_pretrained_model> [other args]
 ```
 
-> 📝 **To be continued**: More training details and the complete training pipeline will be released soon. Stay tuned!
+**Multi-GPU:**
+```bash
+N=8
+torchrun --nproc_per_node=$N train.py --pretrained_model_root <path_to_pretrained_model> [other args]
+```
 
-## 🎬 More Examples
-|Features|Demo1|Demo2|
-|------|------|------|
-|Strong Instruction Following|<video src="https://github.com/user-attachments/assets/fdc3c27b-69f5-46a1-b707-0b57510fa32f" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```一名哀伤的黑发中国女子凝望天空，复古胶片风格烘托出怀旧戏剧氛围``` </details> <details><summary>📋 Show rewrite prompt</summary> ```俯视角度，一位有着深色，略带凌乱的长卷发的年轻中国女性，佩戴着闪耀的珍珠项链和圆形金色耳环，她凌乱的头发被风吹散，她微微抬头，望向天空，神情十分哀伤，眼中含着泪水。嘴唇涂着红色口红。背景是带有华丽红色花纹的图案。画面呈现复古电影风格，色调低饱和，带着轻微柔焦，烘托情绪氛围，质感仿佛20世纪90年代的经典胶片风格，营造出怀旧且富有戏剧性的感觉。``` </details>|<video src="https://github.com/user-attachments/assets/3fcb42cc-cdd3-4651-86a6-645a858561c4" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```建筑蓝图上的线条化为实体，瞬间生长出一个完整的复古工业风办公空间。``` </details> <details><summary>📋 Show rewrite prompt</summary> ```一座空旷的现代阁楼里，有一张铺展在地板中央的建筑蓝图。忽然间，图纸上的线条泛起微光，仿佛被某种无形的力量唤醒。紧接着，那些发光的线条开始向上延伸，从平面中挣脱，勾勒出立体的轮廓——就像在空中进行一场无声的3D打印。随后，奇迹在加速发生：极简的橡木办公桌、优雅的伊姆斯风格皮质椅、高挑的工业风金属书架，还有几盏爱迪生灯泡，以光纹为骨架迅速“生长”出来。转瞬间，线条被真实的材质填充——木材的温润、皮革的质感、金属的冷静，都在眨眼间完整呈现。最终，所有家具稳固落地，蓝图的光芒悄然褪去。一个完整的办公空间，就这样从二维的图纸中诞生。``` </details>|
-|Smooth Motion Generation|<video src="https://github.com/user-attachments/assets/447847f0-490a-45f9-a86d-a67ab1ff4231" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```A DJ is immersed in his musical world. He wears a pair of professional, matte-black headphones, revealing a focused expression. He wears a black bomber jacket, zipped open to reveal a T-shirt underneath. His upper body sways back and forth rhythmically to the throbbing electronic beats, his head moving with precise movement. The mixing console in front of him serves as the primary source of light. In the distance, the cool white glow of several stadium floodlights casts a deep, dark haze across the vast field, casting long shadows across the emerald green grass, creating a stark contrast to the brightly lit area surrounding the DJ booth. His hands danced swiftly and precisely across the equipment. The entire scene was filled with high-tech dynamics and the solitary creative passion. Against the backdrop of the vast and silent night stadium, it created an atmosphere of high focus, energy, and a slightly surreal feeling.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```slowly advancing medium shot, shot from a level angle, focuses on the center of an empty football field, where a DJ is immersed in his musical world. He wears a pair of professional, matte-black headphones, one earcup slightly removed, revealing a focused expression and a brow beaded with sweat from his intense concentration. He wears a black bomber jacket, zipped open to reveal a T-shirt underneath. His upper body sways back and forth rhythmically to the throbbing electronic beats, his head moving with precise movement. The mixing console in front of him serves as the primary source of light. In the distance, the cool white glow of several stadium floodlights casts a deep, dark haze across the vast field, casting long shadows across the emerald green grass, creating a stark contrast to the brightly lit area surrounding the DJ booth. His hands danced swiftly and precisely across the equipment, one hand steadily pushing and pulling a long volume fader, while the fingers of the other nimbly jumped between the illuminated knobs and pads, sometimes decisively cutting a bass line, sometimes triggering an echo effect. The entire scene was filled with high-tech dynamics and the solitary creative passion. Against the backdrop of the vast and silent night stadium, it created an atmosphere of high focus, energy, and a slightly surreal feeling.``` </details>|<video src="https://github.com/user-attachments/assets/49057fe8-a102-4fd7-bd92-e9561abb9f45" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```A figure skater performs a rapid, graceful Biellmann spin, captured from all angles.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```The video captures a figure skater performing a Biellmann spin on ice. The subject is a female skater in a glittering costume. Initially, she spins on one leg. Then, she reaches back and pulls her free leg up. Next, she spins rapidly, becoming a blur of motion, with ice shavings spraying from her skate blade. The background is an ice rink with blurred advertising boards. The camera circles around the subject to capture the spin from all angles. The lighting is spotlit, creating lens flares and sparkles on her costume. The overall video presents a graceful artistic sports style.``` </details>|
-|Cinematic Aesthetics|<video src="https://github.com/user-attachments/assets/4098cf72-357d-4b81-97df-6752064ce0c3" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```固定镜头,焦点在图片里的挂钟上，镜头轻微摇晃营造手持摄影感，​wjw,filmphotos,Film Grain,Reversal film photography，Wong Kar-wai movies,cinematic photography, HK film style,neon lighting, in the style of Wong Kar Wai film``` </details> <details><summary>📋 Show rewrite prompt</summary> ```Handheld lens shooting, the camera focuses on the wall clock hanging on the green-toned wall, shaking slightly. The second hand sweeps steadily across the clock face, and the shadow of the clock cast on the wall shifts subtly with the movement of the lens.``` </details>|<video src="https://github.com/user-attachments/assets/2b4575e5-79f1-4011-bed0-e8380198f7c9" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```The leaves of calamus shine in the sunlight, dotted with dewdrops that trickle down to the ground with the breeze.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```A macro shot focuses on long, slender calamus leaves, rendered in a cinematic photography realistic style. The main leaf, a vibrant, deep green, is positioned diagonally across the frame. Its surface is covered in tiny, glistening spherical dewdrops that catch and refract the bright morning sunlight, creating sparkling highlights. Initially, a larger, perfectly round dewdrop clings to the upper section of the leaf, its surface tension holding it in place. Then, as the leaf sways almost imperceptibly, the dewdrop begins to slowly dislodge. Next, it starts to trickle down the central vein of the leaf, its shape elongating slightly as it moves, leaving a subtle, glistening wet trail in its path. Finally, it reaches the pointed tip of the leaf, hangs for a brief moment, and falls out of the bottom of the frame. In the background, other leaves and blades of grass are softly blurred, creating a beautiful bokeh effect with soft, out-of-focus circles of light. The environment is bathed in the warm, golden glow of early morning sunlight, which streams in from behind the leaves, backlighting them and causing their wet edges to shine brilliantly. The overall impression is one of serene, natural beauty, captured in a highly realistic and detailed manner. This is a macro shot. The camera tilts down very slowly, following the path of the main dewdrop as it travels down the leaf. The lighting is soft and natural, with strong backlighting to create a radiant, glowing effect on the dewdrops and leaf edges, characteristic of professional nature photography. The atmosphere is peaceful and serene. The overall video presents a cinematic photography realistic style.``` </details>|
-|Text Rendering|<video src="https://github.com/user-attachments/assets/7c964fc5-c27e-4bd0-bf3f-eb8fca2caef6" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```赛博朋克风格的夜晚街角，一个巨大的招牌上， “Hunyuan Video 1.5”的霓虹灯管轮廓已经安装好。镜头推进，霓虹灯从“H”开始，伴随着‘滋滋’的电流声，每个字母依次亮起粉紫色的光芒，直到全部点亮，照亮了潮湿的街道。赛博朋克，城市美学``` </details> <details><summary>📋 Show rewrite prompt</summary> ```On a wet street corner in a cyberpunk city at night, a large neon sign reading "Hunyuan Video 1.5" lights up sequentially, illuminating the dark, rainy environment with a pinkish-purple glow. he scene is a dark, rain-slicked street corner in a futuristic, cinematic cyberpunk city. Mounted on the metallic, weathered facade of a building is a massive, unlit neon sign. The sign's glass tube framework clearly spells out the words "Hunyuan Video 1.5". Initially, the street is dimly lit, with ambient light from distant skyscrapers creating shimmering reflections on the wet asphalt below. Then, the camera zooms in slowly toward the sign. As it moves, a low electrical sizzling sound begins. In the background, the dense urban landscape of the cyberpunk metropolis is visible through a light atmospheric haze, with towering structures adorned with their own flickering advertisements. A complex web of cables and pipes crisscrosses between the buildings. The shot is at a low angle, looking up at the sign to emphasize its grand scale. The lighting is high-contrast and dramatic, dominated by the neon glow which creates sharp, specular reflections and deep shadows. The atmosphere is moody and tech-noir. The overall video presents a cinematic photography realistic style.,``` </details>|<video src="https://github.com/user-attachments/assets/73e8b741-baec-4a40-9d36-a1435172ab64" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```一张铺开的中国宣纸上，浓墨滴入水中，晕染出壮丽的山水画轮廓。山峰、云雾、孤舟在墨色中自然形成。随后，这些水墨元素巧妙地流动、重组，在画面的留白处汇聚成"Hunyuan Video 1.5"的书法字体。优雅，诗意，文化底蕴``` </details> <details><summary>📋 Show rewrite prompt</summary> ```A drop of black ink blooms on wet Chinese Xuan paper, forming a landscape painting before the ink elements fluidly reassemble into the calligraphic text "Hunyuan Video 1.5". On a flat, laid-out sheet of off-white Chinese Xuan paper with a subtle, fibrous texture, the scene unfolds. Initially, a single, concentrated drop of deep black ink falls into a clear, wet area at the center of the paper. Then, the ink instantly begins to bloom outwards in intricate, flowing tendrils of varying shades from jet-black to smoky grey. As it spreads, the ink wash naturally and rapidly forms the silhouette of a majestic mountain range with sharp, defined peaks. Next, softer, diluted grey tones billow around the mountains, creating layers of atmospheric mist and clouds, while a simple, dark stroke materializes as a lone boat on a tranquil, watery expanse at the base. As the landscape is formed, the ink elements—the lines of the mountains, wisps of cloud, and the shape of the boat—begin to deconstruct, dissolving into flowing streams of liquid ink. Finally, these streams move gracefully across the paper's empty white space, converging and elegantly reorganizing to form the text "Hunyuan Video 1.5" in a fluid, semi-cursive calligraphic style. The background is the minimalist expanse of the Xuan paper itself, its texture providing a subtle depth. The entire process is lit by soft, even, diffused light from above, which enhances the rich tonal variations of the ink and the delicate texture of the paper without creating harsh shadows. Bird's-eye view. The camera is positioned directly above the subject, capturing the entire process. The camera remains static. The aesthetic is a high-quality, dynamic Chinese ink wash animation style, perfectly simulating the real-world physics of ink spreading on wet paper. The entire sheet of paper and the final text are kept fully within the frame. Poetic, elegant, artistic. The overall video presents a dynamic Chinese ink wash animation style.``` </details>|
-|Physics Compliance|<video src="https://github.com/user-attachments/assets/f1d74e48-cc03-415d-b75f-f7186a4fb41d" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```In a sleek museum gallery, a woman pauses before a gilded oil painting. The painted man inside slowly comes alive, lifting a bottle and pouring real wine straight from the canvas into her glass. Surrounded by stylish art critics moving naturally through the hall, she accepts the pour with calm elegance, as if the impossible were routine. ``` </details> <details><summary>📋 Show rewrite prompt</summary> ```In a sleek museum gallery, a woman receives a glass of wine poured directly from an animated oil painting. A sophisticated woman with dark hair tied back elegantly stands in the mid-ground. She is wearing a simple, black silk sleeveless dress and holds a clear, crystal wine glass in her right hand. She is positioned before a large, baroque-style oil painting in an ornate, gilded frame. Inside the painting, an aristocratic man with a mustache, dressed in a dark velvet doublet with a white lace collar, is depicted. His form is defined by visible, impasto oil brushstrokes. Initially, the woman watches the painting with calm poise. Then, the painted man's arm slowly animates, his painted texture retained as he lifts a dark bottle. Next, a photorealistic stream of red wine emerges directly from the flat canvas surface, arcing through the air and splashing gently into the real crystal glass she holds. She remains perfectly still, accepting the impossible pour with a subtle, knowing smile. The setting is a modern art gallery with high white walls and polished dark concrete floors that reflect the ambient light. Focused track lighting from the high ceiling casts a warm, dramatic spotlight on the woman and the painting, creating soft shadows. In the background, two other gallery patrons, a man and a woman in stylish, modern attire, stroll slowly from right to left, their figures slightly blurred by a shallow depth of field, moving naturally through the hall. The shot is at an eye-level angle with the woman. The camera remains static, capturing the surreal event in a steady medium shot. The lighting is high-contrast and dramatic, reminiscent of a cinematic photography realistic style, using soft side lighting to accentuate the woman's features and the texture of the painting. The mood is surreal, elegant, and mysterious. The overall video presents a cinematic photography realistic style.``` </details>|<video src="https://github.com/user-attachments/assets/07bcce06-ff4f-4688-8c60-c02f600635ea" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```An intact soda can is slowly crushed by a hand.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```In a medium close-up, a hand slowly crushes an intact red and white soda can on a wooden table. A male hand with visible, realistic skin texture is wrapped firmly around the middle of an intact, pristine red and white aluminum soda can. The can, covered in glistening condensation droplets, rests on a dark, polished wooden surface. The cinematic realism captures every minute detail of the scene. Initially, the hand's grip is steady, with the can's cylindrical shape perfectly preserved. Then, the fingers begin to tighten slowly, the knuckles whitening slightly from the exertion. Next, the smooth aluminum surface starts to buckle under the controlled pressure, a sharp crease forming vertically down its side as the metallic sheen distorts. As the hand continues its deliberate squeeze, the can collapses inward progressively, the vibrant red paint wrinkling as the metal structure crumples. Finally, the can is left significantly crushed, its form now an irregular, crumpled shape held tightly in the fist. The scene takes place on a dark, polished wooden tabletop that catches soft, diffuse reflections. The grain of the wood is faintly discernible, adding a layer of texture to the foreground. The background is completely out of focus, rendered as a soft, dark, and non-descript blur, which isolates the main action and enhances the photorealistic quality of the shot. The shot is a medium close-up, presented in a cinematic photography realistic style. The camera remains static at a slightly high angle, looking down to provide a clear and unobstructed view of the can's deformation. Soft side lighting creates high contrast, sculpting the muscles and tendons of the hand while casting specular highlights on the metallic can and the water droplets. The atmosphere is focused and intense. The overall video presents a cinematic photography realistic style.``` </details>|
-|Camera Movement|<video src="https://github.com/user-attachments/assets/6deacbfe-4cca-48d7-a2be-cb638a3e01cb" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```圣诞节的家中，小女孩靠着妈妈听妈妈读书，背景是下着雪的窗外，镜头缓慢下移，一只可爱的长毛小白猫戴着圣诞帽趴在��暖的地摊上``` </details> <details><summary>📋 Show rewrite prompt</summary> ```In a cozy home on Christmas, a young girl leans against her mother as they read a book, and the camera moves down to reveal a fluffy white cat in a Santa hat resting on a warm rug. In a warmly lit living room on a snowy Christmas evening, a young mother and her little daughter are sitting together on a comfortable sofa. The mother, with a gentle expression and wearing a cream-colored knitted sweater, holds an open storybook with colorful illustrations. Her daughter, a small girl with brown hair in pigtails and a red pajama set, leans her head affectionately on her mother's shoulder, her eyes fixed on the book. On the floor below them, a fluffy, long-haired white cat is curled up on a plush, beige wool rug. The cat wears a tiny red and white Santa hat perched between its ears. Initially, the shot focuses on the mother and daughter, capturing their quiet, shared moment. The mother’s finger gently rests on the page of the book. Then, the camera slowly moves downward, gliding past the book and their laps. Finally, the camera settles at a low angle, bringing the adorable white cat into sharp focus as the primary subject. The cat's chest gently rises and falls with each breath, its eyes peacefully closed. Through a large window in the background, large, soft snowflakes can be seen falling silently against the dark blue twilight sky, creating a peaceful and serene backdrop. Faint, out-of-focus golden Christmas lights twinkle in the corner of the room, adding to the warm, festive atmosphere. The scene is imbued with a sense of comfort and holiday warmth, creating a beautiful cinematic photography realistic image. The camera slowly moves downward. The shot uses soft, warm interior lighting that casts gentle shadows, creating a high-contrast, cinematic look. A shallow depth of field keeps the focus on the subjects while beautifully blurring the background elements. The mood is heartwarming, peaceful, and festive. The overall video presents a cinematic photography realistic style.``` </details>|<video src="https://github.com/user-attachments/assets/8e72ed0f-f8ac-445b-97e5-eb4b16fbc121" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```The hiker begins walking forward along the trail, causing the water bottle to swing rhythmically with each step. The camera gradually pulls back and rises to reveal a vast desert landscape stretching out ahead.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```The hiker begins walking forward along the trail, causing the water bottle to swing rhythmically with each step. The camera gradually pulls back and rises to reveal a vast desert landscape stretching out ahead, while the sun position shifts from afternoon to dusk, casting increasingly longer shadows across the terrain as the figure becomes smaller in the frame.``` </details>|
-|Multi-Style Support|<video src="https://github.com/user-attachments/assets/65b2c5a5-e6ba-43be-9462-a98b03b675f1" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```Have the cake man begin to take chunks out of himself and eat it.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```The cake man sits on the chair, with his hands resting on his knees. Then, he slowly raises his right hand and breaks off a piece of cake from his left shoulder. Next, he brings the piece of cake to his mouth and begins to chew. At the same time, his eyes widen slightly, and his mouth parts gently. After that, he raises his right hand again, breaks off another piece of cake from his right arm, and repeats the action of bringing it to his mouth to chew.``` </details>|<video src="https://github.com/user-attachments/assets/de5f7480-b79c-4fc1-b345-c5880a3b5f9e" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```A little girl, carrying a colorful handbag, skips through the garden.  The video uses claymation style.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```A little girl with a colorful handbag skips through a whimsical claymation garden. In a vibrant garden constructed entirely from clay, a young girl, meticulously crafted in a claymation style, skips joyfully. She has chunky, sculpted yellow clay hair tied in pigtails that bounce with a slight stiffness, simple black button eyes, and a wide, permanently etched smile. She wears a simple pink clay dress with a white collar. In her left hand, she carries a small handbag molded from bright red and blue clay, which swings in a slightly jerky arc as she moves. Initially, the girl lifts her right leg high, her body momentarily suspended in a classic stop-motion pose. Then, she hops forward, landing lightly as her left leg swings through for the next skip. Her arms move in an exaggerated, back-and-forth rhythm, characteristic of stop-motion animation. Her movements are intentionally not perfectly fluid, highlighting the frame-by-frame nature of the claymation technique. The garden around her is a whimsical, textured world. In the foreground and mid-ground, oversized flowers with swirled purple and orange petals stand on thick green stems. The ground is a textured mat of green clay, showing subtle fingerprints and tool marks that add to the handmade charm. In the background, a pale blue clay backdrop features a simplified, smiling sun molded from yellow clay. The shot is at an eye-level angle with the main subject. The camera follows the subject, moving smoothly to the right to keep her in the frame. The lighting is bright and even, casting soft shadows that emphasize the rounded, three-dimensional forms of the clay models. The overall video presents a charming and detailed claymation style.``` </details>|
-|High Image-Video Consistency|<img src="https://github.com/user-attachments/assets/3bc8e55d-c211-454e-8067-128c0e215eb6"> <video src="https://github.com/user-attachments/assets/3e6b7ee9-ec66-4e46-a446-801b1c1a1c81" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```女孩放下书，站起身，转身向屋内走去。镜头拉远。``` </details> <details><summary>📋 Show rewrite prompt</summary> ```女孩合上手中的书，将书放在身侧的窗台上。随后，她缓缓站起身，转身向屋内走去，身影逐渐没入门后的阴影中。镜头缓缓拉远，露出更多被绿植覆盖的屋檐和墙体。``` </details>|<img src="https://github.com/user-attachments/assets/7657ce60-90b5-4fdc-b713-0eaa55829b09"> <video src="https://github.com/user-attachments/assets/9ca24021-2353-40d5-8a4d-0f8e67d51826" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```女人手上的鸟亲了女人一口``` </details> <details><summary>📋 Show rewrite prompt</summary> ```女人手臂上的白色鹦鹉缓缓转过头，将喙轻轻触碰女人的脸颊，随后收回头部。女人嘴角微微上扬，目光温柔地注视着鹦鹉。背景中的绿植保持静止。``` </details>|
+**Example:**
+```bash
+torchrun --nproc_per_node=8 train.py \
+  --pretrained_model_root ./ckpts \
+  --learning_rate 1e-5 \
+  --batch_size 1 \
+  --max_steps 10000 \
+  --output_dir ./outputs \
+  --enable_fsdp \
+  --enable_gradient_checkpointing \
+  --sp_size 8
+```
 
+#### 3. Key Training Parameters
 
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `--pretrained_model_root` | Path to pretrained model (required) | - |
+| `--learning_rate` | Learning rate | 1e-5 |
+| `--batch_size` | Batch size | 1 |
+| `--max_steps` | Maximum training steps | 10000 |
+| `--warmup_steps` | Warmup steps | 500 |
+| `--gradient_accumulation_steps` | Gradient accumulation steps | 1 |
+| `--enable_fsdp` | Enable FSDP for distributed training | true |
+| `--enable_gradient_checkpointing` | Enable gradient checkpointing | true |
+| `--sp_size` | Sequence parallelism size (must divide world_size) | 8 |
+| `--i2v_prob` | Probability of i2v task for video data | 0.3 |
+| `--use_muon` | Use Muon optimizer | true |
+| `--resume_from_checkpoint` | Resume from checkpoint directory | None |
+
+#### 4. Monitor Training
+
+- Checkpoints are saved to `output_dir` at intervals specified by `--save_interval`
+- Validation videos are generated at intervals specified by `--validation_interval`
+- Training logs are printed to console at intervals specified by `--log_interval`
+
+#### 5. Resume Training
+
+Use `--resume_from_checkpoint <checkpoint_dir>` to resume from a saved checkpoint:
+```bash
+python train.py \
+  --pretrained_model_root <path> \
+  --resume_from_checkpoint ./outputs/checkpoint-1000
+```
 
 
 ## 📊 Evaluation
@@ -512,6 +555,20 @@ We report the total inference time for 50 diffusion steps for HunyuanVideo 1.5 b
 <img src="./assets/speed.png" alt="" width="100%">
 </div> 
 
+## 🎬 More Examples
+|Features|Demo1|Demo2|
+|------|------|------|
+|Strong Instruction Following|<video src="https://github.com/user-attachments/assets/fdc3c27b-69f5-46a1-b707-0b57510fa32f" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```一名哀伤的黑发中国女子凝望天空，复古胶片风格烘托出怀旧戏剧氛围``` </details> <details><summary>📋 Show rewrite prompt</summary> ```俯视角度，一位有着深色，略带凌乱的长卷发的年轻中国女性，佩戴着闪耀的珍珠项链和圆形金色耳环，她凌乱的头发被风吹散，她微微抬头，望向天空，神情十分哀伤，眼中含着泪水。嘴唇涂着红色口红。背景是带有华丽红色花纹的图案。画面呈现复古电影风格，色调低饱和，带着轻微柔焦，烘托情绪氛围，质感仿佛20世纪90年代的经典胶片风格，营造出怀旧且富有戏剧性的感觉。``` </details>|<video src="https://github.com/user-attachments/assets/3fcb42cc-cdd3-4651-86a6-645a858561c4" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```建筑蓝图上的线条化为实体，瞬间生长出一个完整的复古工业风办公空间。``` </details> <details><summary>📋 Show rewrite prompt</summary> ```一座空旷的现代阁楼里，有一张铺展在地板中央的建筑蓝图。忽然间，图纸上的线条泛起微光，仿佛被某种无形的力量唤醒。紧接着，那些发光的线条开始向上延伸，从平面中挣脱，勾勒出立体的轮廓——就像在空中进行一场无声的3D打印。随后，奇迹在加速发生：极简的橡木办公桌、优雅的伊姆斯风格皮质椅、高挑的工业风金属书架，还有几盏爱迪生灯泡，以光纹为骨架迅速“生长”出来。转瞬间，线条被真实的材质填充——木材的温润、皮革的质感、金属的冷静，都在眨眼间完整呈现。最终，所有家具稳固落地，蓝图的光芒悄然褪去。一个完整的办公空间，就这样从二维的图纸中诞生。``` </details>|
+|Smooth Motion Generation|<video src="https://github.com/user-attachments/assets/447847f0-490a-45f9-a86d-a67ab1ff4231" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```A DJ is immersed in his musical world. He wears a pair of professional, matte-black headphones, revealing a focused expression. He wears a black bomber jacket, zipped open to reveal a T-shirt underneath. His upper body sways back and forth rhythmically to the throbbing electronic beats, his head moving with precise movement. The mixing console in front of him serves as the primary source of light. In the distance, the cool white glow of several stadium floodlights casts a deep, dark haze across the vast field, casting long shadows across the emerald green grass, creating a stark contrast to the brightly lit area surrounding the DJ booth. His hands danced swiftly and precisely across the equipment. The entire scene was filled with high-tech dynamics and the solitary creative passion. Against the backdrop of the vast and silent night stadium, it created an atmosphere of high focus, energy, and a slightly surreal feeling.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```slowly advancing medium shot, shot from a level angle, focuses on the center of an empty football field, where a DJ is immersed in his musical world. He wears a pair of professional, matte-black headphones, one earcup slightly removed, revealing a focused expression and a brow beaded with sweat from his intense concentration. He wears a black bomber jacket, zipped open to reveal a T-shirt underneath. His upper body sways back and forth rhythmically to the throbbing electronic beats, his head moving with precise movement. The mixing console in front of him serves as the primary source of light. In the distance, the cool white glow of several stadium floodlights casts a deep, dark haze across the vast field, casting long shadows across the emerald green grass, creating a stark contrast to the brightly lit area surrounding the DJ booth. His hands danced swiftly and precisely across the equipment, one hand steadily pushing and pulling a long volume fader, while the fingers of the other nimbly jumped between the illuminated knobs and pads, sometimes decisively cutting a bass line, sometimes triggering an echo effect. The entire scene was filled with high-tech dynamics and the solitary creative passion. Against the backdrop of the vast and silent night stadium, it created an atmosphere of high focus, energy, and a slightly surreal feeling.``` </details>|<video src="https://github.com/user-attachments/assets/49057fe8-a102-4fd7-bd92-e9561abb9f45" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```A figure skater performs a rapid, graceful Biellmann spin, captured from all angles.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```The video captures a figure skater performing a Biellmann spin on ice. The subject is a female skater in a glittering costume. Initially, she spins on one leg. Then, she reaches back and pulls her free leg up. Next, she spins rapidly, becoming a blur of motion, with ice shavings spraying from her skate blade. The background is an ice rink with blurred advertising boards. The camera circles around the subject to capture the spin from all angles. The lighting is spotlit, creating lens flares and sparkles on her costume. The overall video presents a graceful artistic sports style.``` </details>|
+|Cinematic Aesthetics|<video src="https://github.com/user-attachments/assets/4098cf72-357d-4b81-97df-6752064ce0c3" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```固定镜头,焦点在图片里的挂钟上，镜头轻微摇晃营造手持摄影感，​wjw,filmphotos,Film Grain,Reversal film photography，Wong Kar-wai movies,cinematic photography, HK film style,neon lighting, in the style of Wong Kar Wai film``` </details> <details><summary>📋 Show rewrite prompt</summary> ```Handheld lens shooting, the camera focuses on the wall clock hanging on the green-toned wall, shaking slightly. The second hand sweeps steadily across the clock face, and the shadow of the clock cast on the wall shifts subtly with the movement of the lens.``` </details>|<video src="https://github.com/user-attachments/assets/2b4575e5-79f1-4011-bed0-e8380198f7c9" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```The leaves of calamus shine in the sunlight, dotted with dewdrops that trickle down to the ground with the breeze.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```A macro shot focuses on long, slender calamus leaves, rendered in a cinematic photography realistic style. The main leaf, a vibrant, deep green, is positioned diagonally across the frame. Its surface is covered in tiny, glistening spherical dewdrops that catch and refract the bright morning sunlight, creating sparkling highlights. Initially, a larger, perfectly round dewdrop clings to the upper section of the leaf, its surface tension holding it in place. Then, as the leaf sways almost imperceptibly, the dewdrop begins to slowly dislodge. Next, it starts to trickle down the central vein of the leaf, its shape elongating slightly as it moves, leaving a subtle, glistening wet trail in its path. Finally, it reaches the pointed tip of the leaf, hangs for a brief moment, and falls out of the bottom of the frame. In the background, other leaves and blades of grass are softly blurred, creating a beautiful bokeh effect with soft, out-of-focus circles of light. The environment is bathed in the warm, golden glow of early morning sunlight, which streams in from behind the leaves, backlighting them and causing their wet edges to shine brilliantly. The overall impression is one of serene, natural beauty, captured in a highly realistic and detailed manner. This is a macro shot. The camera tilts down very slowly, following the path of the main dewdrop as it travels down the leaf. The lighting is soft and natural, with strong backlighting to create a radiant, glowing effect on the dewdrops and leaf edges, characteristic of professional nature photography. The atmosphere is peaceful and serene. The overall video presents a cinematic photography realistic style.``` </details>|
+|Text Rendering|<video src="https://github.com/user-attachments/assets/7c964fc5-c27e-4bd0-bf3f-eb8fca2caef6" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```赛博朋克风格的夜晚街角，一个巨大的招牌上， “Hunyuan Video 1.5”的霓虹灯管轮廓已经安装好。镜头推进，霓虹灯从“H”开始，伴随着‘滋滋’的电流声，每个字母依次亮起粉紫色的光芒，直到全部点亮，照亮了潮湿的街道。赛博朋克，城市美学``` </details> <details><summary>📋 Show rewrite prompt</summary> ```On a wet street corner in a cyberpunk city at night, a large neon sign reading "Hunyuan Video 1.5" lights up sequentially, illuminating the dark, rainy environment with a pinkish-purple glow. he scene is a dark, rain-slicked street corner in a futuristic, cinematic cyberpunk city. Mounted on the metallic, weathered facade of a building is a massive, unlit neon sign. The sign's glass tube framework clearly spells out the words "Hunyuan Video 1.5". Initially, the street is dimly lit, with ambient light from distant skyscrapers creating shimmering reflections on the wet asphalt below. Then, the camera zooms in slowly toward the sign. As it moves, a low electrical sizzling sound begins. In the background, the dense urban landscape of the cyberpunk metropolis is visible through a light atmospheric haze, with towering structures adorned with their own flickering advertisements. A complex web of cables and pipes crisscrosses between the buildings. The shot is at a low angle, looking up at the sign to emphasize its grand scale. The lighting is high-contrast and dramatic, dominated by the neon glow which creates sharp, specular reflections and deep shadows. The atmosphere is moody and tech-noir. The overall video presents a cinematic photography realistic style.,``` </details>|<video src="https://github.com/user-attachments/assets/73e8b741-baec-4a40-9d36-a1435172ab64" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```一张铺开的中国宣纸上，浓墨滴入水中，晕染出壮丽的山水画轮廓。山峰、云雾、孤舟在墨色中自然形成。随后，这些水墨元素巧妙地流动、重组，在画面的留白处汇聚成"Hunyuan Video 1.5"的书法字体。优雅，诗意，文化底蕴``` </details> <details><summary>📋 Show rewrite prompt</summary> ```A drop of black ink blooms on wet Chinese Xuan paper, forming a landscape painting before the ink elements fluidly reassemble into the calligraphic text "Hunyuan Video 1.5". On a flat, laid-out sheet of off-white Chinese Xuan paper with a subtle, fibrous texture, the scene unfolds. Initially, a single, concentrated drop of deep black ink falls into a clear, wet area at the center of the paper. Then, the ink instantly begins to bloom outwards in intricate, flowing tendrils of varying shades from jet-black to smoky grey. As it spreads, the ink wash naturally and rapidly forms the silhouette of a majestic mountain range with sharp, defined peaks. Next, softer, diluted grey tones billow around the mountains, creating layers of atmospheric mist and clouds, while a simple, dark stroke materializes as a lone boat on a tranquil, watery expanse at the base. As the landscape is formed, the ink elements—the lines of the mountains, wisps of cloud, and the shape of the boat—begin to deconstruct, dissolving into flowing streams of liquid ink. Finally, these streams move gracefully across the paper's empty white space, converging and elegantly reorganizing to form the text "Hunyuan Video 1.5" in a fluid, semi-cursive calligraphic style. The background is the minimalist expanse of the Xuan paper itself, its texture providing a subtle depth. The entire process is lit by soft, even, diffused light from above, which enhances the rich tonal variations of the ink and the delicate texture of the paper without creating harsh shadows. Bird's-eye view. The camera is positioned directly above the subject, capturing the entire process. The camera remains static. The aesthetic is a high-quality, dynamic Chinese ink wash animation style, perfectly simulating the real-world physics of ink spreading on wet paper. The entire sheet of paper and the final text are kept fully within the frame. Poetic, elegant, artistic. The overall video presents a dynamic Chinese ink wash animation style.``` </details>|
+|Physics Compliance|<video src="https://github.com/user-attachments/assets/f1d74e48-cc03-415d-b75f-f7186a4fb41d" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```In a sleek museum gallery, a woman pauses before a gilded oil painting. The painted man inside slowly comes alive, lifting a bottle and pouring real wine straight from the canvas into her glass. Surrounded by stylish art critics moving naturally through the hall, she accepts the pour with calm elegance, as if the impossible were routine. ``` </details> <details><summary>📋 Show rewrite prompt</summary> ```In a sleek museum gallery, a woman receives a glass of wine poured directly from an animated oil painting. A sophisticated woman with dark hair tied back elegantly stands in the mid-ground. She is wearing a simple, black silk sleeveless dress and holds a clear, crystal wine glass in her right hand. She is positioned before a large, baroque-style oil painting in an ornate, gilded frame. Inside the painting, an aristocratic man with a mustache, dressed in a dark velvet doublet with a white lace collar, is depicted. His form is defined by visible, impasto oil brushstrokes. Initially, the woman watches the painting with calm poise. Then, the painted man's arm slowly animates, his painted texture retained as he lifts a dark bottle. Next, a photorealistic stream of red wine emerges directly from the flat canvas surface, arcing through the air and splashing gently into the real crystal glass she holds. She remains perfectly still, accepting the impossible pour with a subtle, knowing smile. The setting is a modern art gallery with high white walls and polished dark concrete floors that reflect the ambient light. Focused track lighting from the high ceiling casts a warm, dramatic spotlight on the woman and the painting, creating soft shadows. In the background, two other gallery patrons, a man and a woman in stylish, modern attire, stroll slowly from right to left, their figures slightly blurred by a shallow depth of field, moving naturally through the hall. The shot is at an eye-level angle with the woman. The camera remains static, capturing the surreal event in a steady medium shot. The lighting is high-contrast and dramatic, reminiscent of a cinematic photography realistic style, using soft side lighting to accentuate the woman's features and the texture of the painting. The mood is surreal, elegant, and mysterious. The overall video presents a cinematic photography realistic style.``` </details>|<video src="https://github.com/user-attachments/assets/07bcce06-ff4f-4688-8c60-c02f600635ea" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```An intact soda can is slowly crushed by a hand.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```In a medium close-up, a hand slowly crushes an intact red and white soda can on a wooden table. A male hand with visible, realistic skin texture is wrapped firmly around the middle of an intact, pristine red and white aluminum soda can. The can, covered in glistening condensation droplets, rests on a dark, polished wooden surface. The cinematic realism captures every minute detail of the scene. Initially, the hand's grip is steady, with the can's cylindrical shape perfectly preserved. Then, the fingers begin to tighten slowly, the knuckles whitening slightly from the exertion. Next, the smooth aluminum surface starts to buckle under the controlled pressure, a sharp crease forming vertically down its side as the metallic sheen distorts. As the hand continues its deliberate squeeze, the can collapses inward progressively, the vibrant red paint wrinkling as the metal structure crumples. Finally, the can is left significantly crushed, its form now an irregular, crumpled shape held tightly in the fist. The scene takes place on a dark, polished wooden tabletop that catches soft, diffuse reflections. The grain of the wood is faintly discernible, adding a layer of texture to the foreground. The background is completely out of focus, rendered as a soft, dark, and non-descript blur, which isolates the main action and enhances the photorealistic quality of the shot. The shot is a medium close-up, presented in a cinematic photography realistic style. The camera remains static at a slightly high angle, looking down to provide a clear and unobstructed view of the can's deformation. Soft side lighting creates high contrast, sculpting the muscles and tendons of the hand while casting specular highlights on the metallic can and the water droplets. The atmosphere is focused and intense. The overall video presents a cinematic photography realistic style.``` </details>|
+|Camera Movement|<video src="https://github.com/user-attachments/assets/6deacbfe-4cca-48d7-a2be-cb638a3e01cb" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```圣诞节的家中，小女孩靠着妈妈听妈妈读书，背景是下着雪的窗外，镜头缓慢下移，一只可爱的长毛小白猫戴着圣诞帽趴在温暖的地摊上``` </details> <details><summary>📋 Show rewrite prompt</summary> ```In a cozy home on Christmas, a young girl leans against her mother as they read a book, and the camera moves down to reveal a fluffy white cat in a Santa hat resting on a warm rug. In a warmly lit living room on a snowy Christmas evening, a young mother and her little daughter are sitting together on a comfortable sofa. The mother, with a gentle expression and wearing a cream-colored knitted sweater, holds an open storybook with colorful illustrations. Her daughter, a small girl with brown hair in pigtails and a red pajama set, leans her head affectionately on her mother's shoulder, her eyes fixed on the book. On the floor below them, a fluffy, long-haired white cat is curled up on a plush, beige wool rug. The cat wears a tiny red and white Santa hat perched between its ears. Initially, the shot focuses on the mother and daughter, capturing their quiet, shared moment. The mother’s finger gently rests on the page of the book. Then, the camera slowly moves downward, gliding past the book and their laps. Finally, the camera settles at a low angle, bringing the adorable white cat into sharp focus as the primary subject. The cat's chest gently rises and falls with each breath, its eyes peacefully closed. Through a large window in the background, large, soft snowflakes can be seen falling silently against the dark blue twilight sky, creating a peaceful and serene backdrop. Faint, out-of-focus golden Christmas lights twinkle in the corner of the room, adding to the warm, festive atmosphere. The scene is imbued with a sense of comfort and holiday warmth, creating a beautiful cinematic photography realistic image. The camera slowly moves downward. The shot uses soft, warm interior lighting that casts gentle shadows, creating a high-contrast, cinematic look. A shallow depth of field keeps the focus on the subjects while beautifully blurring the background elements. The mood is heartwarming, peaceful, and festive. The overall video presents a cinematic photography realistic style.``` </details>|<video src="https://github.com/user-attachments/assets/8e72ed0f-f8ac-445b-97e5-eb4b16fbc121" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```The hiker begins walking forward along the trail, causing the water bottle to swing rhythmically with each step. The camera gradually pulls back and rises to reveal a vast desert landscape stretching out ahead.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```The hiker begins walking forward along the trail, causing the water bottle to swing rhythmically with each step. The camera gradually pulls back and rises to reveal a vast desert landscape stretching out ahead, while the sun position shifts from afternoon to dusk, casting increasingly longer shadows across the terrain as the figure becomes smaller in the frame.``` </details>|
+|Multi-Style Support|<video src="https://github.com/user-attachments/assets/65b2c5a5-e6ba-43be-9462-a98b03b675f1" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```Have the cake man begin to take chunks out of himself and eat it.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```The cake man sits on the chair, with his hands resting on his knees. Then, he slowly raises his right hand and breaks off a piece of cake from his left shoulder. Next, he brings the piece of cake to his mouth and begins to chew. At the same time, his eyes widen slightly, and his mouth parts gently. After that, he raises his right hand again, breaks off another piece of cake from his right arm, and repeats the action of bringing it to his mouth to chew.``` </details>|<video src="https://github.com/user-attachments/assets/de5f7480-b79c-4fc1-b345-c5880a3b5f9e" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```A little girl, carrying a colorful handbag, skips through the garden.  The video uses claymation style.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```A little girl with a colorful handbag skips through a whimsical claymation garden. In a vibrant garden constructed entirely from clay, a young girl, meticulously crafted in a claymation style, skips joyfully. She has chunky, sculpted yellow clay hair tied in pigtails that bounce with a slight stiffness, simple black button eyes, and a wide, permanently etched smile. She wears a simple pink clay dress with a white collar. In her left hand, she carries a small handbag molded from bright red and blue clay, which swings in a slightly jerky arc as she moves. Initially, the girl lifts her right leg high, her body momentarily suspended in a classic stop-motion pose. Then, she hops forward, landing lightly as her left leg swings through for the next skip. Her arms move in an exaggerated, back-and-forth rhythm, characteristic of stop-motion animation. Her movements are intentionally not perfectly fluid, highlighting the frame-by-frame nature of the claymation technique. The garden around her is a whimsical, textured world. In the foreground and mid-ground, oversized flowers with swirled purple and orange petals stand on thick green stems. The ground is a textured mat of green clay, showing subtle fingerprints and tool marks that add to the handmade charm. In the background, a pale blue clay backdrop features a simplified, smiling sun molded from yellow clay. The shot is at an eye-level angle with the main subject. The camera follows the subject, moving smoothly to the right to keep her in the frame. The lighting is bright and even, casting soft shadows that emphasize the rounded, three-dimensional forms of the clay models. The overall video presents a charming and detailed claymation style.``` </details>|
+|High Image-Video Consistency|<img src="https://github.com/user-attachments/assets/3bc8e55d-c211-454e-8067-128c0e215eb6"> <video src="https://github.com/user-attachments/assets/3e6b7ee9-ec66-4e46-a446-801b1c1a1c81" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```女孩放下书，站起身，转身向屋内走去。镜头拉远。``` </details> <details><summary>📋 Show rewrite prompt</summary> ```女孩合上手中的书，将书放在身侧的窗台上。随后，她缓缓站起身，转身向屋内走去，身影逐渐没入门后的阴影中。镜头缓缓拉远，露出更多被绿植覆盖的屋檐和墙体。``` </details>|<img src="https://github.com/user-attachments/assets/7657ce60-90b5-4fdc-b713-0eaa55829b09"> <video src="https://github.com/user-attachments/assets/9ca24021-2353-40d5-8a4d-0f8e67d51826" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```女人手上的鸟亲了女人一口``` </details> <details><summary>📋 Show rewrite prompt</summary> ```女人手臂上的白色鹦鹉缓缓转过头，将喙轻轻触碰女人的脸颊，随后收回头部。女人嘴角微微上扬，目光温柔地注视着鹦鹉。背景中的绿植保持静止。``` </details>|
+
+
+
 
 ## 📚 Citation
 

README_CN.md

@@ -41,7 +41,7 @@ HunyuanVideo-1.5作为一款轻量级视频生成模型，仅需83亿参数即
 
 ## 🔥🔥🔥 最新动态
 * 🚀 Dec 05, 2025: **新模型发布**：我们现已发布 [480p I2V 步数蒸馏模型](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_i2v_step_distilled)，建议使用 8 或 12 步生成视频！在 RTX 4090 上，端到端生成耗时减少 75%，单卡 RTX 4090 可在 75 秒内生成视频。步数蒸馏模型在保持与原模型相当质量的同时实现了显著的加速。详细的质量对比请参见[步数蒸馏对比文档](./assets/step_distillation_comparison.md)。如需更快的生成速度，您也可以尝试使用4步推理（速度更快，质量略有下降）。**启用步数蒸馏模型，请运行 `generate.py` 并使用 `--enable_step_distill` 参数。** 详细的使用说明请参见[使用方法](#-使用方法)。 🔥🔥🔥🆕
-* 📚 训练代码即将发布。HunyuanVideo-1.5 使用 Muon 优化器进行训练，我们在[Training](#-training) 部分开源。**如果您希望继续训练我们的模型，或使用 LoRA 进行微调，请使用 Muon 优化器。**
+* 📚 Dec 05, 2025: **训练代码已发布**：我们现已开源 HunyuanVideo-1.5 的完整训练代码！训练脚本（`train.py`）提供了完整的训练流程，支持分布式训练、FSDP、context parallel、梯度检查点等功能。HunyuanVideo-1.5 使用 Muon 优化器进行训练，我们在[训练](#-训练)部分已开源。**如果您希望继续训练我们的模型，或使用 LoRA 进行微调，请使用 Muon 优化器。** 详细使用说明请参见[训练](#-训练)部分。 🔥🔥🔥🆕
 * 🎉 **Diffusers 支持**：HunyuanVideo-1.5 现已支持 Hugging Face Diffusers！查看我们的 [Diffusers 集合](https://huggingface.co/collections/hunyuanvideo-community/hunyuanvideo-15) 以便轻松集成。 🔥🔥🔥🆕
 * 🚀 Nov 27, 2025: 我们现已支持 cache 推理（deepcache, teacache, taylorcache），可极大加速推理！请 pull 最新代码体验。 🔥🔥🔥🆕 
 * 🚀 Nov 24, 2025: 我们现已支持 deepcache 推理。
@@ -89,12 +89,11 @@ HunyuanVideo-1.5作为一款轻量级视频生成模型，仅需83亿参数即
 - [🛠️ 依赖安装](#️-依赖安装)
 - [🧱 下载预训练模型](#-下载预训练模型)
 - [📝 提示词指南](#-提示词指南)
-- [🔑 使用方法](#-使用方法)
+- [🔑 推理](#-推理)
   - [使用源代码推理](#使用源代码推理)
   - [使用 Diffusers](#使用-diffusers)
   - [命令行参数](#命令行参数)
   - [最优推理配置](#最优推理配置)
-- [🧱 模型卡片](#-模型卡片)
 - [🎓 训练](#-训练)
 - [🎬 更多示例](#-更多示例)
 - [📊 性能评估](#-性能评估)
@@ -186,6 +185,23 @@ pip install -i https://mirrors.tencent.com/pypi/simple/ --upgrade tencentcloud-s
 
 在生成视频之前，请先下载预训练模型。详细说明请参考 [checkpoints-download.md](checkpoints-download.md)。
 
+### 模型卡片
+|模型名称| 下载链接                     |
+|-|---------------------------| 
+|HunyuanVideo-1.5-480P-T2V|[480P-T2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_t2v) |
+|HunyuanVideo-1.5-480P-I2V |[480P-I2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_i2v) |
+|HunyuanVideo-1.5-480P-T2V-cfg-distill | [480P-T2V-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_t2v_distilled) |
+|HunyuanVideo-1.5-480P-I2V-cfg-distill |[480P-I2V-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_i2v_distilled) |
+|HunyuanVideo-1.5-480P-I2V-step-distill |[480P-I2V-step-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_i2v_step_distilled) |
+|HunyuanVideo-1.5-720P-T2V|[720P-T2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_t2v) |
+|HunyuanVideo-1.5-720P-I2V |[720P-I2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_i2v) |
+|HunyuanVideo-1.5-720P-T2V-cfg-distill| Comming soon |
+|HunyuanVideo-1.5-720P-I2V-cfg-distill |[720P-I2V-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_i2v_distilled) |
+|HunyuanVideo-1.5-720P-T2V-sparse-cfg-distill| Comming soon |
+|HunyuanVideo-1.5-720P-I2V-sparse-cfg-distill |[720P-I2V-sparse-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_i2v_distilled_sparse) |
+|HunyuanVideo-1.5-720P-sr-step-distill |[720P-sr](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_sr_distilled) |
+|HunyuanVideo-1.5-1080P-sr-step-distill |[1080P-sr](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/1080p_sr_distilled) |
+
 ## 📝 提示词指南
 ### 提示词撰写手册
 提示词增强在我们的模型生成高质量视频方面起着至关重要的作用。通过撰写更长、更详细的提示词，生成的视频质量将得到显著改善。我们鼓励您编写全面且描述性��提示词，以获得最佳的视频质量。我们建议社区伙伴参考我们的官方指南，了解如何撰写有效的提示词。
@@ -198,10 +214,11 @@ pip install -i https://mirrors.tencent.com/pypi/simple/ --upgrade tencentcloud-s
 对于希望为其他大模型优化提示词的用户，建议参考文件 `hyvideo/utils/rewrite/t2v_prompt.py` 中 `t2v_rewrite_system_prompt` 的定义来指导文生视频的提示词重写。同样，对于图生视频重写，请参考 `hyvideo/utils/rewrite/i2v_prompt.py` 中 `i2v_rewrite_system_prompt` 的定义。
 
 
-## 🔑 使用方法
+## 🔑 推理
 
 ### 使用源代码推理
 
+
 对于提示词重写，我们推荐使用 Gemini 或通过 vLLM 部署的大模型。当前代码库仅支持兼容 vLLM 接口的模型，如果您希望使用 Gemini，需自行实现相关接口调用。
 
 对于 vLLM 接口的模型，需要注意 T2V 和 I2V 推荐使用不同的模型和环境变量：
@@ -235,24 +252,27 @@ export I2V_REWRITE_MODEL_NAME="<your_model_name>"
 
 PROMPT='A girl holding a paper with words "Hello, world!"'
 
-IMAGE_PATH=none # 可选，none 或 <图像路径> 以启用 i2v 模式
+IMAGE_PATH=/path/to/image.png # 可选，none 或 <图像路径> 以启用 i2v 模式
 SEED=1
 ASPECT_RATIO=16:9
 RESOLUTION=480p
 OUTPUT_PATH=./outputs/output.mp4
+MODEL_PATH=./ckpts # 预训练模型路径
 
-# 配置
-REWRITE=true # 启用提示词重写。请确保 rewrite vLLM server 已部署和配置。
+# 加速推理配置
 N_INFERENCE_GPU=8 # 并行推理 GPU 数量
 CFG_DISTILLED=true # 使用 CFG 蒸馏模型进行推理，2倍加速
-ENABLE_STEP_DISTILL=true # 启用 480p I2V 步数蒸馏模型，推荐 8 或 12 步，在 RTX 4090 上可提速 75%
-SPARSE_ATTN=false # 使用稀疏注意力进行推理（仅 720p 模型配备了稀疏注意力）。请确保 flex-block-attn 已安装
 SAGE_ATTN=true # 使用 SageAttention 进行推理
+SPARSE_ATTN=false # 使用稀疏注意力进行推理（仅 720p 模型配备了稀疏注意力）。请确保 flex-block-attn 已安装
 OVERLAP_GROUP_OFFLOADING=true # 仅在组卸载启用时有效，会显著增加 CPU 内存占用，但能够提速
 ENABLE_CACHE=true # 启用特征缓存进行推理。显著提升推理速度
 CACHE_TYPE=deepcache # 支持：deepcache, teacache, taylorcache
+ENABLE_STEP_DISTILL=true # 启用 480p I2V 步数蒸馏模型，推荐 8 或 12 步，最高可达 6 倍加速
+
+
+# 提升质量配置
+REWRITE=true # 启用提示词重写。请确保 rewrite vLLM server 已部署和配置。
 ENABLE_SR=true # 启用超分辨率
-MODEL_PATH=ckpts # 预训练模型路径
 
 torchrun --nproc_per_node=$N_INFERENCE_GPU generate.py \
   --prompt "$PROMPT" \
@@ -395,62 +415,87 @@ with attention_backend("_flash_3_hub"): # 如果您不在 H100/H800 上，可以
 更多详情，请访问 [HunyuanVideo-1.5 Diffusers 集合](https://huggingface.co/collections/hunyuanvideo-community/hunyuanvideo-15)。
 
 
-## 🧱 模型卡片
-|模型名称| 下载链接                     |
-|-|---------------------------| 
-|HunyuanVideo-1.5-480P-T2V|[480P-T2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_t2v) |
-|HunyuanVideo-1.5-480P-I2V |[480P-I2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_i2v) |
-|HunyuanVideo-1.5-480P-T2V-cfg-distill | [480P-T2V-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_t2v_distilled) |
-|HunyuanVideo-1.5-480P-I2V-cfg-distill |[480P-I2V-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_i2v_distilled) |
-|HunyuanVideo-1.5-480P-I2V-step-distill |[480P-I2V-step-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/480p_i2v_step_distilled) |
-|HunyuanVideo-1.5-720P-T2V|[720P-T2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_t2v) |
-|HunyuanVideo-1.5-720P-I2V |[720P-I2V](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_i2v) |
-|HunyuanVideo-1.5-720P-T2V-cfg-distill| Comming soon |
-|HunyuanVideo-1.5-720P-I2V-cfg-distill |[720P-I2V-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_i2v_distilled) |
-|HunyuanVideo-1.5-720P-T2V-sparse-cfg-distill| Comming soon |
-|HunyuanVideo-1.5-720P-I2V-sparse-cfg-distill |[720P-I2V-sparse-cfg-distill](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_i2v_distilled_sparse) |
-|HunyuanVideo-1.5-720P-sr-step-distill |[720P-sr](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/720p_sr_distilled) |
-|HunyuanVideo-1.5-1080P-sr-step-distill |[1080P-sr](https://huggingface.co/tencent/HunyuanVideo-1.5/tree/main/transformer/1080p_sr_distilled) |
+## 🎓 训练
 
+HunyuanVideo-1.5 使用 **Muon 优化器**进行训练，该优化器能够加速收敛并提高训练稳定性。Muon 优化器结合了基于动量的更新和 Newton-Schulz 正交化方法，可高效优化大规模视频生成模型。
 
+### 快速开始
 
-## 🎓 训练
+训练脚本（`train.py`）为 HunyuanVideo-1.5 提供了完整的训练流程。使用方法如下：
 
-> 💡 训练代码即将发布。我们将在未来发布完整的训练流程。
+#### 1. 实现您的数据加载器
 
-HunyuanVideo-1.5 使用 **Muon 优化器**进行训练，该优化器能够加速收敛并提高训练稳定性。Muon 优化器结合了基于动量的更新和 Newton-Schulz 正交化方法，可高效优化大规模视频生成模型。
+替换 `train.py` 中的 `create_dummy_dataloader()` 函数，实现您自己的数据加载器。数据加载器应返回以下格式的批次：
 
-### 创建 Muon 优化器
+- **必需字段：**
+  - `"pixel_values"`: `torch.Tensor` - 视频：`[B, C, F, H, W]` 或图像：`[B, C, H, W]`
+    - 注意：对于视频数据，时间维度 F 必须是 `4n+1`（例如：1, 5, 9, 13, 17, ...）
+  - `"text"`: `List[str]` - 每个样本的文本提示词
+  - `"data_type"`: `str` - `"video"` 或 `"image"`
 
-以下是如何为您的模型创建 Muon 优化器：
+- **可选字段（用于性能优化）：**
+  - `"latents"`: 预编码的 VAE 潜在表示（跳过 VAE 编码以加速训练）
+  - `"byt5_text_ids"` 和 `"byt5_text_mask"`: 预分词的 byT5 输入
 
-```python
-from hyvideo.optim.muon import get_muon_optimizer
-
-# 为您的模型创建 Muon 优化器
-optimizer = get_muon_optimizer(
-    model=your_model,
-    lr=lr,                      # 学习率
-    weight_decay=weight_decay,  # 权重衰减
-    momentum=momentum,          # 动量系数
-    adamw_betas=adamw_betas,   # 1D 参数的 AdamW betas
-    adamw_eps=adamw_eps        # AdamW epsilon
-)
+详细的批次格式文档请参见 `train.py` 中的 `create_dummy_dataloader()` 函数。
+
+#### 2. 运行训练
+
+**单 GPU：**
+```bash
+python train.py --pretrained_model_root <预训练模型路径> [其他参数]
 ```
 
-> 📝 **未完待续**：更多训练细节和完整的训练流程即将发布，敬请期待！
+**多 GPU：**
+```bash
+N=8
+torchrun --nproc_per_node=$N train.py --pretrained_model_root <预训练模型路径> [其他参数]
+```
 
-## 🎬 更多示例
-|特性|示例1|示例2|
-|------|------|------|
-|指令跟随能力|<video src="https://github.com/user-attachments/assets/fdc3c27b-69f5-46a1-b707-0b57510fa32f" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```一名哀伤的黑发中国女子凝望天空，复古胶片风格烘托出怀旧戏剧氛围``` </details> <details><summary>📋 Show rewrite prompt</summary> ```俯视角度，一位有着深色，略带凌乱的长卷发的年轻中国女性，佩戴着闪耀的珍珠项链和圆形金色耳环，她凌乱的头发被风吹散，她微微抬头，望向天空，神情十分哀伤，眼中含着泪水。嘴唇涂着红色口红。背景是带有华丽红色花纹的图案。画面呈现复古电影风格，色调低饱和，带着轻微柔焦，烘托情绪氛围，质感仿佛20世纪90年代的经典胶片风格，营造出怀旧且富有戏剧性的感觉。``` </details>|<video src="https://github.com/user-attachments/assets/3fcb42cc-cdd3-4651-86a6-645a858561c4" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```建筑蓝图上的线条化为实体，瞬间生长出一个完整的复古工业风办公空间。``` </details> <details><summary>📋 Show rewrite prompt</summary> ```一座空旷的现代阁楼里，有一张铺展在地板中央的建筑蓝图。忽然间，图纸上的线条泛起微光，仿佛被某种无形的力量唤醒。紧接着，那些发光的线条开始向上延伸，从平面中挣脱，勾勒出立体的轮廓——就像在空中进行一场无声的3D打印。随后，奇迹在加速发生：极简的橡木办公桌、优雅的伊姆斯风格皮质椅、高挑的工业风金属书架，还有几盏爱迪生灯泡，以光纹为骨架迅速“生长”出来。转瞬间，线条被真实的材质填充——木材的温润、皮革的质感、金属的冷静，都在眨眼间完整呈现。最终，所有家具稳固落地，蓝图的光芒悄然褪去。一个完整的办公空间，就这样从二维的图纸中诞生。``` </details>|
-|流畅运动生成|<video src="https://github.com/user-attachments/assets/447847f0-490a-45f9-a86d-a67ab1ff4231" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```A DJ is immersed in his musical world. He wears a pair of professional, matte-black headphones, revealing a focused expression. He wears a black bomber jacket, zipped open to reveal a T-shirt underneath. His upper body sways back and forth rhythmically to the throbbing electronic beats, his head moving with precise movement. The mixing console in front of him serves as the primary source of light. In the distance, the cool white glow of several stadium floodlights casts a deep, dark haze across the vast field, casting long shadows across the emerald green grass, creating a stark contrast to the brightly lit area surrounding the DJ booth. His hands danced swiftly and precisely across the equipment. The entire scene was filled with high-tech dynamics and the solitary creative passion. Against the backdrop of the vast and silent night stadium, it created an atmosphere of high focus, energy, and a slightly surreal feeling.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```slowly advancing medium shot, shot from a level angle, focuses on the center of an empty football field, where a DJ is immersed in his musical world. He wears a pair of professional, matte-black headphones, one earcup slightly removed, revealing a focused expression and a brow beaded with sweat from his intense concentration. He wears a black bomber jacket, zipped open to reveal a T-shirt underneath. His upper body sways back and forth rhythmically to the throbbing electronic beats, his head moving with precise movement. The mixing console in front of him serves as the primary source of light. In the distance, the cool white glow of several stadium floodlights casts a deep, dark haze across the vast field, casting long shadows across the emerald green grass, creating a stark contrast to the brightly lit area surrounding the DJ booth. His hands danced swiftly and precisely across the equipment, one hand steadily pushing and pulling a long volume fader, while the fingers of the other nimbly jumped between the illuminated knobs and pads, sometimes decisively cutting a bass line, sometimes triggering an echo effect. The entire scene was filled with high-tech dynamics and the solitary creative passion. Against the backdrop of the vast and silent night stadium, it created an atmosphere of high focus, energy, and a slightly surreal feeling.``` </details>|<video src="https://github.com/user-attachments/assets/49057fe8-a102-4fd7-bd92-e9561abb9f45" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```A figure skater performs a rapid, graceful Biellmann spin, captured from all angles.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```The video captures a figure skater performing a Biellmann spin on ice. The subject is a female skater in a glittering costume. Initially, she spins on one leg. Then, she reaches back and pulls her free leg up. Next, she spins rapidly, becoming a blur of motion, with ice shavings spraying from her skate blade. The background is an ice rink with blurred advertising boards. The camera circles around the subject to capture the spin from all angles. The lighting is spotlit, creating lens flares and sparkles on her costume. The overall video presents a graceful artistic sports style.``` </details>|
-|电影级美学|<video src="https://github.com/user-attachments/assets/4098cf72-357d-4b81-97df-6752064ce0c3" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```固定镜头,焦点在图片里的挂钟上，镜头轻微摇晃营造手持摄影感，​wjw,filmphotos,Film Grain,Reversal film photography，Wong Kar-wai movies,cinematic photography, HK film style,neon lighting, in the style of Wong Kar Wai film``` </details> <details><summary>📋 Show rewrite prompt</summary> ```Handheld lens shooting, the camera focuses on the wall clock hanging on the green-toned wall, shaking slightly. The second hand sweeps steadily across the clock face, and the shadow of the clock cast on the wall shifts subtly with the movement of the lens.``` </details>|<video src="https://github.com/user-attachments/assets/2b4575e5-79f1-4011-bed0-e8380198f7c9" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```The leaves of calamus shine in the sunlight, dotted with dewdrops that trickle down to the ground with the breeze.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```A macro shot focuses on long, slender calamus leaves, rendered in a cinematic photography realistic style. The main leaf, a vibrant, deep green, is positioned diagonally across the frame. Its surface is covered in tiny, glistening spherical dewdrops that catch and refract the bright morning sunlight, creating sparkling highlights. Initially, a larger, perfectly round dewdrop clings to the upper section of the leaf, its surface tension holding it in place. Then, as the leaf sways almost imperceptibly, the dewdrop begins to slowly dislodge. Next, it starts to trickle down the central vein of the leaf, its shape elongating slightly as it moves, leaving a subtle, glistening wet trail in its path. Finally, it reaches the pointed tip of the leaf, hangs for a brief moment, and falls out of the bottom of the frame. In the background, other leaves and blades of grass are softly blurred, creating a beautiful bokeh effect with soft, out-of-focus circles of light. The environment is bathed in the warm, golden glow of early morning sunlight, which streams in from behind the leaves, backlighting them and causing their wet edges to shine brilliantly. The overall impression is one of serene, natural beauty, captured in a highly realistic and detailed manner. This is a macro shot. The camera tilts down very slowly, following the path of the main dewdrop as it travels down the leaf. The lighting is soft and natural, with strong backlighting to create a radiant, glowing effect on the dewdrops and leaf edges, characteristic of professional nature photography. The atmosphere is peaceful and serene. The overall video presents a cinematic photography realistic style.``` </details>|
-|文字渲染|<video src="https://github.com/user-attachments/assets/7c964fc5-c27e-4bd0-bf3f-eb8fca2caef6" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```赛博朋克风格的夜晚街角，一个巨大的招牌上， “Hunyuan Video 1.5”的霓虹灯管轮廓已经安装好。镜头推进，霓虹灯从“H”开始，伴随着‘滋滋’的电流声，每个字母依次亮起粉紫色的光芒，直到全部点亮，照亮了潮湿的街道。赛博朋克，城市美学``` </details> <details><summary>📋 Show rewrite prompt</summary> ```On a wet street corner in a cyberpunk city at night, a large neon sign reading "Hunyuan Video 1.5" lights up sequentially, illuminating the dark, rainy environment with a pinkish-purple glow. he scene is a dark, rain-slicked street corner in a futuristic, cinematic cyberpunk city. Mounted on the metallic, weathered facade of a building is a massive, unlit neon sign. The sign's glass tube framework clearly spells out the words "Hunyuan Video 1.5". Initially, the street is dimly lit, with ambient light from distant skyscrapers creating shimmering reflections on the wet asphalt below. Then, the camera zooms in slowly toward the sign. As it moves, a low electrical sizzling sound begins. In the background, the dense urban landscape of the cyberpunk metropolis is visible through a light atmospheric haze, with towering structures adorned with their own flickering advertisements. A complex web of cables and pipes crisscrosses between the buildings. The shot is at a low angle, looking up at the sign to emphasize its grand scale. The lighting is high-contrast and dramatic, dominated by the neon glow which creates sharp, specular reflections and deep shadows. The atmosphere is moody and tech-noir. The overall video presents a cinematic photography realistic style.,``` </details>|<video src="https://github.com/user-attachments/assets/73e8b741-baec-4a40-9d36-a1435172ab64" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```一张铺开的中国宣纸上，浓墨滴入水中，晕染出壮丽的山水画轮廓。山峰、云雾、孤舟在墨色中自然形成。随后，这些水墨元素巧妙地流动、重组，在画面的留白处汇聚成"Hunyuan Video 1.5"的书法字体。优雅，诗意，文化底蕴``` </details> <details><summary>📋 Show rewrite prompt</summary> ```A drop of black ink blooms on wet Chinese Xuan paper, forming a landscape painting before the ink elements fluidly reassemble into the calligraphic text "Hunyuan Video 1.5". On a flat, laid-out sheet of off-white Chinese Xuan paper with a subtle, fibrous texture, the scene unfolds. Initially, a single, concentrated drop of deep black ink falls into a clear, wet area at the center of the paper. Then, the ink instantly begins to bloom outwards in intricate, flowing tendrils of varying shades from jet-black to smoky grey. As it spreads, the ink wash naturally and rapidly forms the silhouette of a majestic mountain range with sharp, defined peaks. Next, softer, diluted grey tones billow around the mountains, creating layers of atmospheric mist and clouds, while a simple, dark stroke materializes as a lone boat on a tranquil, watery expanse at the base. As the landscape is formed, the ink elements—the lines of the mountains, wisps of cloud, and the shape of the boat—begin to deconstruct, dissolving into flowing streams of liquid ink. Finally, these streams move gracefully across the paper's empty white space, converging and elegantly reorganizing to form the text "Hunyuan Video 1.5" in a fluid, semi-cursive calligraphic style. The background is the minimalist expanse of the Xuan paper itself, its texture providing a subtle depth. The entire process is lit by soft, even, diffused light from above, which enhances the rich tonal variations of the ink and the delicate texture of the paper without creating harsh shadows. Bird's-eye view. The camera is positioned directly above the subject, capturing the entire process. The camera remains static. The aesthetic is a high-quality, dynamic Chinese ink wash animation style, perfectly simulating the real-world physics of ink spreading on wet paper. The entire sheet of paper and the final text are kept fully within the frame. Poetic, elegant, artistic. The overall video presents a dynamic Chinese ink wash animation style.``` </details>|
-|物理合理性|<video src="https://github.com/user-attachments/assets/f1d74e48-cc03-415d-b75f-f7186a4fb41d" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```In a sleek museum gallery, a woman pauses before a gilded oil painting. The painted man inside slowly comes alive, lifting a bottle and pouring real wine straight from the canvas into her glass. Surrounded by stylish art critics moving naturally through the hall, she accepts the pour with calm elegance, as if the impossible were routine. ``` </details> <details><summary>📋 Show rewrite prompt</summary> ```In a sleek museum gallery, a woman receives a glass of wine poured directly from an animated oil painting. A sophisticated woman with dark hair tied back elegantly stands in the mid-ground. She is wearing a simple, black silk sleeveless dress and holds a clear, crystal wine glass in her right hand. She is positioned before a large, baroque-style oil painting in an ornate, gilded frame. Inside the painting, an aristocratic man with a mustache, dressed in a dark velvet doublet with a white lace collar, is depicted. His form is defined by visible, impasto oil brushstrokes. Initially, the woman watches the painting with calm poise. Then, the painted man's arm slowly animates, his painted texture retained as he lifts a dark bottle. Next, a photorealistic stream of red wine emerges directly from the flat canvas surface, arcing through the air and splashing gently into the real crystal glass she holds. She remains perfectly still, accepting the impossible pour with a subtle, knowing smile. The setting is a modern art gallery with high white walls and polished dark concrete floors that reflect the ambient light. Focused track lighting from the high ceiling casts a warm, dramatic spotlight on the woman and the painting, creating soft shadows. In the background, two other gallery patrons, a man and a woman in stylish, modern attire, stroll slowly from right to left, their figures slightly blurred by a shallow depth of field, moving naturally through the hall. The shot is at an eye-level angle with the woman. The camera remains static, capturing the surreal event in a steady medium shot. The lighting is high-contrast and dramatic, reminiscent of a cinematic photography realistic style, using soft side lighting to accentuate the woman's features and the texture of the painting. The mood is surreal, elegant, and mysterious. The overall video presents a cinematic photography realistic style.``` </details>|<video src="https://github.com/user-attachments/assets/07bcce06-ff4f-4688-8c60-c02f600635ea" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```An intact soda can is slowly crushed by a hand.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```In a medium close-up, a hand slowly crushes an intact red and white soda can on a wooden table. A male hand with visible, realistic skin texture is wrapped firmly around the middle of an intact, pristine red and white aluminum soda can. The can, covered in glistening condensation droplets, rests on a dark, polished wooden surface. The cinematic realism captures every minute detail of the scene. Initially, the hand's grip is steady, with the can's cylindrical shape perfectly preserved. Then, the fingers begin to tighten slowly, the knuckles whitening slightly from the exertion. Next, the smooth aluminum surface starts to buckle under the controlled pressure, a sharp crease forming vertically down its side as the metallic sheen distorts. As the hand continues its deliberate squeeze, the can collapses inward progressively, the vibrant red paint wrinkling as the metal structure crumples. Finally, the can is left significantly crushed, its form now an irregular, crumpled shape held tightly in the fist. The scene takes place on a dark, polished wooden tabletop that catches soft, diffuse reflections. The grain of the wood is faintly discernible, adding a layer of texture to the foreground. The background is completely out of focus, rendered as a soft, dark, and non-descript blur, which isolates the main action and enhances the photorealistic quality of the shot. The shot is a medium close-up, presented in a cinematic photography realistic style. The camera remains static at a slightly high angle, looking down to provide a clear and unobstructed view of the can's deformation. Soft side lighting creates high contrast, sculpting the muscles and tendons of the hand while casting specular highlights on the metallic can and the water droplets. The atmosphere is focused and intense. The overall video presents a cinematic photography realistic style.``` </details>|
-|摄像机运动|<video src="https://github.com/user-attachments/assets/6deacbfe-4cca-48d7-a2be-cb638a3e01cb" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```圣诞节的家中，小女孩靠着妈妈听妈妈读书，背景是下着雪的窗外，镜头缓慢下移，一只可爱的长毛小白猫戴着圣诞帽趴在温暖的地摊上``` </details> <details><summary>📋 Show rewrite prompt</summary> ```In a cozy home on Christmas, a young girl leans against her mother as they read a book, and the camera moves down to reveal a fluffy white cat in a Santa hat resting on a warm rug. In a warmly lit living room on a snowy Christmas evening, a young mother and her little daughter are sitting together on a comfortable sofa. The mother, with a gentle expression and wearing a cream-colored knitted sweater, holds an open storybook with colorful illustrations. Her daughter, a small girl with brown hair in pigtails and a red pajama set, leans her head affectionately on her mother's shoulder, her eyes fixed on the book. On the floor below them, a fluffy, long-haired white cat is curled up on a plush, beige wool rug. The cat wears a tiny red and white Santa hat perched between its ears. Initially, the shot focuses on the mother and daughter, capturing their quiet, shared moment. The mother’s finger gently rests on the page of the book. Then, the camera slowly moves downward, gliding past the book and their laps. Finally, the camera settles at a low angle, bringing the adorable white cat into sharp focus as the primary subject. The cat's chest gently rises and falls with each breath, its eyes peacefully closed. Through a large window in the background, large, soft snowflakes can be seen falling silently against the dark blue twilight sky, creating a peaceful and serene backdrop. Faint, out-of-focus golden Christmas lights twinkle in the corner of the room, adding to the warm, festive atmosphere. The scene is imbued with a sense of comfort and holiday warmth, creating a beautiful cinematic photography realistic image. The camera slowly moves downward. The shot uses soft, warm interior lighting that casts gentle shadows, creating a high-contrast, cinematic look. A shallow depth of field keeps the focus on the subjects while beautifully blurring the background elements. The mood is heartwarming, peaceful, and festive. The overall video presents a cinematic photography realistic style.``` </details>|<video src="https://github.com/user-attachments/assets/8e72ed0f-f8ac-445b-97e5-eb4b16fbc121" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```The hiker begins walking forward along the trail, causing the water bottle to swing rhythmically with each step. The camera gradually pulls back and rises to reveal a vast desert landscape stretching out ahead.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```The hiker begins walking forward along the trail, causing the water bottle to swing rhythmically with each step. The camera gradually pulls back and rises to reveal a vast desert landscape stretching out ahead, while the sun position shifts from afternoon to dusk, casting increasingly longer shadows across the terrain as the figure becomes smaller in the frame.``` </details>|
-|多风格支持|<video src="https://github.com/user-attachments/assets/65b2c5a5-e6ba-43be-9462-a98b03b675f1" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```Have the cake man begin to take chunks out of himself and eat it.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```The cake man sits on the chair, with his hands resting on his knees. Then, he slowly raises his right hand and breaks off a piece of cake from his left shoulder. Next, he brings the piece of cake to his mouth and begins to chew. At the same time, his eyes widen slightly, and his mouth parts gently. After that, he raises his right hand again, breaks off another piece of cake from his right arm, and repeats the action of bringing it to his mouth to chew.``` </details>|<video src="https://github.com/user-attachments/assets/de5f7480-b79c-4fc1-b345-c5880a3b5f9e" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```A little girl, carrying a colorful handbag, skips through the garden.  The video uses claymation style.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```A little girl with a colorful handbag skips through a whimsical claymation garden. In a vibrant garden constructed entirely from clay, a young girl, meticulously crafted in a claymation style, skips joyfully. She has chunky, sculpted yellow clay hair tied in pigtails that bounce with a slight stiffness, simple black button eyes, and a wide, permanently etched smile. She wears a simple pink clay dress with a white collar. In her left hand, she carries a small handbag molded from bright red and blue clay, which swings in a slightly jerky arc as she moves. Initially, the girl lifts her right leg high, her body momentarily suspended in a classic stop-motion pose. Then, she hops forward, landing lightly as her left leg swings through for the next skip. Her arms move in an exaggerated, back-and-forth rhythm, characteristic of stop-motion animation. Her movements are intentionally not perfectly fluid, highlighting the frame-by-frame nature of the claymation technique. The garden around her is a whimsical, textured world. In the foreground and mid-ground, oversized flowers with swirled purple and orange petals stand on thick green stems. The ground is a textured mat of green clay, showing subtle fingerprints and tool marks that add to the handmade charm. In the background, a pale blue clay backdrop features a simplified, smiling sun molded from yellow clay. The shot is at an eye-level angle with the main subject. The camera follows the subject, moving smoothly to the right to keep her in the frame. The lighting is bright and even, casting soft shadows that emphasize the rounded, three-dimensional forms of the clay models. The overall video presents a charming and detailed claymation style.``` </details>|
-|高图视一致性|<img src="https://github.com/user-attachments/assets/3bc8e55d-c211-454e-8067-128c0e215eb6"> <video src="https://github.com/user-attachments/assets/3e6b7ee9-ec66-4e46-a446-801b1c1a1c81" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```女孩放下书，站起身，转身向屋内走去。镜头拉远。``` </details> <details><summary>📋 Show rewrite prompt</summary> ```女孩合上手中的书，将书放在身侧的窗台上。随后，她缓缓站起身，转身向屋内走去，身影逐渐没入门后的阴影中。镜头缓缓拉远，露出更多被绿植覆盖的屋檐和墙体。``` </details>|<img src="https://github.com/user-attachments/assets/7657ce60-90b5-4fdc-b713-0eaa55829b09"> <video src="https://github.com/user-attachments/assets/9ca24021-2353-40d5-8a4d-0f8e67d51826" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```女人手上的鸟亲了女人一口``` </details> <details><summary>📋 Show rewrite prompt</summary> ```女人手臂上的白色鹦鹉缓缓转过头，将喙轻轻触碰女人的脸颊，随后收回头部。女人嘴角微微上扬，目光温柔地注视着鹦鹉。背景中的绿植保持静止。``` </details>|
+**示例：**
+```bash
+torchrun --nproc_per_node=8 train.py \
+  --pretrained_model_root ./ckpts \
+  --learning_rate 1e-5 \
+  --batch_size 1 \
+  --max_steps 10000 \
+  --output_dir ./outputs \
+  --enable_fsdp \
+  --enable_gradient_checkpointing \
+  --sp_size 8
+```
+
+#### 3. 关键训练参数
+
+| 参数 | 描述 | 默认值 |
+|-----------|-------------|---------|
+| `--pretrained_model_root` | 预训练模型路径（必需） | - |
+| `--learning_rate` | 学习率 | 1e-5 |
+| `--batch_size` | 批次大小 | 1 |
+| `--max_steps` | 最大训练步数 | 10000 |
+| `--warmup_steps` | 预热步数 | 500 |
+| `--gradient_accumulation_steps` | 梯度累积步数 | 1 |
+| `--enable_fsdp` | 启用 FSDP 进行分布式训练 | true |
+| `--enable_gradient_checkpointing` | 启用梯度检查点 | true |
+| `--sp_size` | 序列并行大小（必须能整除 world_size） | 8 |
+| `--i2v_prob` | 视频数据使用 i2v 任务的概率 | 0.3 |
+| `--use_muon` | 使用 Muon 优化器 | true |
+| `--resume_from_checkpoint` | 从检查点目录恢复训练 | None |
+
+#### 4. 监控训练
+
+- 检查点按 `--save_interval` 指定的间隔保存到 `output_dir`
+- 验证视频按 `--validation_interval` 指定的间隔生成
+- 训练日志按 `--log_interval` 指定的间隔打印到控制台
+
+#### 5. 恢复训练
+
+使用 `--resume_from_checkpoint <检查点目录>` 从保存的检查点恢复训练：
+```bash
+python train.py \
+  --pretrained_model_root <路径> \
+  --resume_from_checkpoint ./outputs/checkpoint-1000
+```
 
 
 ## 📊 性能评估
@@ -492,6 +537,18 @@ GSB（Good/Same/Bad）评估法被广泛用于基于整体视频感知质量来
 <img src="./assets/speed.png" alt="" width="100%">
 </div> 
 
+## 🎬 更多示例
+|特性|示例1|示例2|
+|------|------|------|
+|指令跟随能力|<video src="https://github.com/user-attachments/assets/fdc3c27b-69f5-46a1-b707-0b57510fa32f" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```一名哀伤的黑发中国女子凝望天空，复古胶片风格烘托出怀旧戏剧氛围``` </details> <details><summary>📋 Show rewrite prompt</summary> ```俯视角度，一位有着深色，略带凌乱的长卷发的年轻中国女性，佩戴着闪耀的珍珠项链和圆形金色耳环，她凌乱的头发被风吹散，她微微抬头，望向天空，神情十分哀伤，眼中含着泪水。嘴唇涂着红色口红。背景是带有华丽红色花纹的图案。画面呈现复古电影风格，色调低饱和，带着轻微柔焦，烘托情绪氛围，质感仿佛20世纪90年代的经典胶片风格，营造出怀旧且富有戏剧性的感觉。``` </details>|<video src="https://github.com/user-attachments/assets/3fcb42cc-cdd3-4651-86a6-645a858561c4" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```建筑蓝图上的线条化为实体，瞬间生长出一个完整的复古工业风办公空间。``` </details> <details><summary>📋 Show rewrite prompt</summary> ```一座空旷的现代阁楼里，有一张铺展在地板中央的建筑蓝图。忽然间，图纸上的线条泛起微光，仿佛被某种无形的力量唤醒。紧接着，那些发光的线条开始向上延伸，从平面中挣脱，勾勒出立体的轮廓——就像在空中进行一场无声的3D打印。随后，奇迹在加速发生：极简的橡木办公桌、优雅的伊姆斯风格皮质椅、高挑的工业风金属书架，还有几盏爱迪生灯泡，以光纹为骨架迅速“生长”出来。转瞬间，线条被真实的材质填充——木材的温润、皮革的质感、金属的冷静，都在眨眼间完整呈现。最终，所有家具稳固落地，蓝图的光芒悄然褪去��一个完整的办公空间，就这样从二维的图纸中诞生。``` </details>|
+|流畅运动生成|<video src="https://github.com/user-attachments/assets/447847f0-490a-45f9-a86d-a67ab1ff4231" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```A DJ is immersed in his musical world. He wears a pair of professional, matte-black headphones, revealing a focused expression. He wears a black bomber jacket, zipped open to reveal a T-shirt underneath. His upper body sways back and forth rhythmically to the throbbing electronic beats, his head moving with precise movement. The mixing console in front of him serves as the primary source of light. In the distance, the cool white glow of several stadium floodlights casts a deep, dark haze across the vast field, casting long shadows across the emerald green grass, creating a stark contrast to the brightly lit area surrounding the DJ booth. His hands danced swiftly and precisely across the equipment. The entire scene was filled with high-tech dynamics and the solitary creative passion. Against the backdrop of the vast and silent night stadium, it created an atmosphere of high focus, energy, and a slightly surreal feeling.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```slowly advancing medium shot, shot from a level angle, focuses on the center of an empty football field, where a DJ is immersed in his musical world. He wears a pair of professional, matte-black headphones, one earcup slightly removed, revealing a focused expression and a brow beaded with sweat from his intense concentration. He wears a black bomber jacket, zipped open to reveal a T-shirt underneath. His upper body sways back and forth rhythmically to the throbbing electronic beats, his head moving with precise movement. The mixing console in front of him serves as the primary source of light. In the distance, the cool white glow of several stadium floodlights casts a deep, dark haze across the vast field, casting long shadows across the emerald green grass, creating a stark contrast to the brightly lit area surrounding the DJ booth. His hands danced swiftly and precisely across the equipment, one hand steadily pushing and pulling a long volume fader, while the fingers of the other nimbly jumped between the illuminated knobs and pads, sometimes decisively cutting a bass line, sometimes triggering an echo effect. The entire scene was filled with high-tech dynamics and the solitary creative passion. Against the backdrop of the vast and silent night stadium, it created an atmosphere of high focus, energy, and a slightly surreal feeling.``` </details>|<video src="https://github.com/user-attachments/assets/49057fe8-a102-4fd7-bd92-e9561abb9f45" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```A figure skater performs a rapid, graceful Biellmann spin, captured from all angles.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```The video captures a figure skater performing a Biellmann spin on ice. The subject is a female skater in a glittering costume. Initially, she spins on one leg. Then, she reaches back and pulls her free leg up. Next, she spins rapidly, becoming a blur of motion, with ice shavings spraying from her skate blade. The background is an ice rink with blurred advertising boards. The camera circles around the subject to capture the spin from all angles. The lighting is spotlit, creating lens flares and sparkles on her costume. The overall video presents a graceful artistic sports style.``` </details>|
+|电影级美学|<video src="https://github.com/user-attachments/assets/4098cf72-357d-4b81-97df-6752064ce0c3" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```固定镜头,焦点在图片里的挂钟上，镜头轻微摇晃营造手持摄影感，​wjw,filmphotos,Film Grain,Reversal film photography，Wong Kar-wai movies,cinematic photography, HK film style,neon lighting, in the style of Wong Kar Wai film``` </details> <details><summary>📋 Show rewrite prompt</summary> ```Handheld lens shooting, the camera focuses on the wall clock hanging on the green-toned wall, shaking slightly. The second hand sweeps steadily across the clock face, and the shadow of the clock cast on the wall shifts subtly with the movement of the lens.``` </details>|<video src="https://github.com/user-attachments/assets/2b4575e5-79f1-4011-bed0-e8380198f7c9" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```The leaves of calamus shine in the sunlight, dotted with dewdrops that trickle down to the ground with the breeze.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```A macro shot focuses on long, slender calamus leaves, rendered in a cinematic photography realistic style. The main leaf, a vibrant, deep green, is positioned diagonally across the frame. Its surface is covered in tiny, glistening spherical dewdrops that catch and refract the bright morning sunlight, creating sparkling highlights. Initially, a larger, perfectly round dewdrop clings to the upper section of the leaf, its surface tension holding it in place. Then, as the leaf sways almost imperceptibly, the dewdrop begins to slowly dislodge. Next, it starts to trickle down the central vein of the leaf, its shape elongating slightly as it moves, leaving a subtle, glistening wet trail in its path. Finally, it reaches the pointed tip of the leaf, hangs for a brief moment, and falls out of the bottom of the frame. In the background, other leaves and blades of grass are softly blurred, creating a beautiful bokeh effect with soft, out-of-focus circles of light. The environment is bathed in the warm, golden glow of early morning sunlight, which streams in from behind the leaves, backlighting them and causing their wet edges to shine brilliantly. The overall impression is one of serene, natural beauty, captured in a highly realistic and detailed manner. This is a macro shot. The camera tilts down very slowly, following the path of the main dewdrop as it travels down the leaf. The lighting is soft and natural, with strong backlighting to create a radiant, glowing effect on the dewdrops and leaf edges, characteristic of professional nature photography. The atmosphere is peaceful and serene. The overall video presents a cinematic photography realistic style.``` </details>|
+|文字渲染|<video src="https://github.com/user-attachments/assets/7c964fc5-c27e-4bd0-bf3f-eb8fca2caef6" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```赛博朋克风格的夜晚街角，一个巨大的招牌上， “Hunyuan Video 1.5”的霓虹灯管轮廓已经安装好。镜头推进，霓虹灯从“H”开始，伴随着‘滋滋’的电流声，每个字母依次亮起粉紫色的光芒，直到全部点亮，照亮了潮湿的街道。赛博朋克，城市美学``` </details> <details><summary>📋 Show rewrite prompt</summary> ```On a wet street corner in a cyberpunk city at night, a large neon sign reading "Hunyuan Video 1.5" lights up sequentially, illuminating the dark, rainy environment with a pinkish-purple glow. he scene is a dark, rain-slicked street corner in a futuristic, cinematic cyberpunk city. Mounted on the metallic, weathered facade of a building is a massive, unlit neon sign. The sign's glass tube framework clearly spells out the words "Hunyuan Video 1.5". Initially, the street is dimly lit, with ambient light from distant skyscrapers creating shimmering reflections on the wet asphalt below. Then, the camera zooms in slowly toward the sign. As it moves, a low electrical sizzling sound begins. In the background, the dense urban landscape of the cyberpunk metropolis is visible through a light atmospheric haze, with towering structures adorned with their own flickering advertisements. A complex web of cables and pipes crisscrosses between the buildings. The shot is at a low angle, looking up at the sign to emphasize its grand scale. The lighting is high-contrast and dramatic, dominated by the neon glow which creates sharp, specular reflections and deep shadows. The atmosphere is moody and tech-noir. The overall video presents a cinematic photography realistic style.,``` </details>|<video src="https://github.com/user-attachments/assets/73e8b741-baec-4a40-9d36-a1435172ab64" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```一张铺开的中国宣纸上，浓墨滴入水中，晕染出壮丽的山水画轮廓。山峰、云雾、孤舟在墨色中自然形成。随后，这些水墨元素巧妙地流动、重组，在画面的留白处汇聚成"Hunyuan Video 1.5"的书法字体。优雅，诗意，文化底蕴``` </details> <details><summary>📋 Show rewrite prompt</summary> ```A drop of black ink blooms on wet Chinese Xuan paper, forming a landscape painting before the ink elements fluidly reassemble into the calligraphic text "Hunyuan Video 1.5". On a flat, laid-out sheet of off-white Chinese Xuan paper with a subtle, fibrous texture, the scene unfolds. Initially, a single, concentrated drop of deep black ink falls into a clear, wet area at the center of the paper. Then, the ink instantly begins to bloom outwards in intricate, flowing tendrils of varying shades from jet-black to smoky grey. As it spreads, the ink wash naturally and rapidly forms the silhouette of a majestic mountain range with sharp, defined peaks. Next, softer, diluted grey tones billow around the mountains, creating layers of atmospheric mist and clouds, while a simple, dark stroke materializes as a lone boat on a tranquil, watery expanse at the base. As the landscape is formed, the ink elements—the lines of the mountains, wisps of cloud, and the shape of the boat—begin to deconstruct, dissolving into flowing streams of liquid ink. Finally, these streams move gracefully across the paper's empty white space, converging and elegantly reorganizing to form the text "Hunyuan Video 1.5" in a fluid, semi-cursive calligraphic style. The background is the minimalist expanse of the Xuan paper itself, its texture providing a subtle depth. The entire process is lit by soft, even, diffused light from above, which enhances the rich tonal variations of the ink and the delicate texture of the paper without creating harsh shadows. Bird's-eye view. The camera is positioned directly above the subject, capturing the entire process. The camera remains static. The aesthetic is a high-quality, dynamic Chinese ink wash animation style, perfectly simulating the real-world physics of ink spreading on wet paper. The entire sheet of paper and the final text are kept fully within the frame. Poetic, elegant, artistic. The overall video presents a dynamic Chinese ink wash animation style.``` </details>|
+|物理合理性|<video src="https://github.com/user-attachments/assets/f1d74e48-cc03-415d-b75f-f7186a4fb41d" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```In a sleek museum gallery, a woman pauses before a gilded oil painting. The painted man inside slowly comes alive, lifting a bottle and pouring real wine straight from the canvas into her glass. Surrounded by stylish art critics moving naturally through the hall, she accepts the pour with calm elegance, as if the impossible were routine. ``` </details> <details><summary>📋 Show rewrite prompt</summary> ```In a sleek museum gallery, a woman receives a glass of wine poured directly from an animated oil painting. A sophisticated woman with dark hair tied back elegantly stands in the mid-ground. She is wearing a simple, black silk sleeveless dress and holds a clear, crystal wine glass in her right hand. She is positioned before a large, baroque-style oil painting in an ornate, gilded frame. Inside the painting, an aristocratic man with a mustache, dressed in a dark velvet doublet with a white lace collar, is depicted. His form is defined by visible, impasto oil brushstrokes. Initially, the woman watches the painting with calm poise. Then, the painted man's arm slowly animates, his painted texture retained as he lifts a dark bottle. Next, a photorealistic stream of red wine emerges directly from the flat canvas surface, arcing through the air and splashing gently into the real crystal glass she holds. She remains perfectly still, accepting the impossible pour with a subtle, knowing smile. The setting is a modern art gallery with high white walls and polished dark concrete floors that reflect the ambient light. Focused track lighting from the high ceiling casts a warm, dramatic spotlight on the woman and the painting, creating soft shadows. In the background, two other gallery patrons, a man and a woman in stylish, modern attire, stroll slowly from right to left, their figures slightly blurred by a shallow depth of field, moving naturally through the hall. The shot is at an eye-level angle with the woman. The camera remains static, capturing the surreal event in a steady medium shot. The lighting is high-contrast and dramatic, reminiscent of a cinematic photography realistic style, using soft side lighting to accentuate the woman's features and the texture of the painting. The mood is surreal, elegant, and mysterious. The overall video presents a cinematic photography realistic style.``` </details>|<video src="https://github.com/user-attachments/assets/07bcce06-ff4f-4688-8c60-c02f600635ea" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```An intact soda can is slowly crushed by a hand.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```In a medium close-up, a hand slowly crushes an intact red and white soda can on a wooden table. A male hand with visible, realistic skin texture is wrapped firmly around the middle of an intact, pristine red and white aluminum soda can. The can, covered in glistening condensation droplets, rests on a dark, polished wooden surface. The cinematic realism captures every minute detail of the scene. Initially, the hand's grip is steady, with the can's cylindrical shape perfectly preserved. Then, the fingers begin to tighten slowly, the knuckles whitening slightly from the exertion. Next, the smooth aluminum surface starts to buckle under the controlled pressure, a sharp crease forming vertically down its side as the metallic sheen distorts. As the hand continues its deliberate squeeze, the can collapses inward progressively, the vibrant red paint wrinkling as the metal structure crumples. Finally, the can is left significantly crushed, its form now an irregular, crumpled shape held tightly in the fist. The scene takes place on a dark, polished wooden tabletop that catches soft, diffuse reflections. The grain of the wood is faintly discernible, adding a layer of texture to the foreground. The background is completely out of focus, rendered as a soft, dark, and non-descript blur, which isolates the main action and enhances the photorealistic quality of the shot. The shot is a medium close-up, presented in a cinematic photography realistic style. The camera remains static at a slightly high angle, looking down to provide a clear and unobstructed view of the can's deformation. Soft side lighting creates high contrast, sculpting the muscles and tendons of the hand while casting specular highlights on the metallic can and the water droplets. The atmosphere is focused and intense. The overall video presents a cinematic photography realistic style.``` </details>|
+|摄像机运动|<video src="https://github.com/user-attachments/assets/6deacbfe-4cca-48d7-a2be-cb638a3e01cb" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```圣诞节的家中，小女孩靠着妈妈听妈妈读书，背景是下着雪的窗外，镜头缓慢下移，一只可爱的长毛小白猫戴着圣诞帽趴在温暖的地摊上``` </details> <details><summary>📋 Show rewrite prompt</summary> ```In a cozy home on Christmas, a young girl leans against her mother as they read a book, and the camera moves down to reveal a fluffy white cat in a Santa hat resting on a warm rug. In a warmly lit living room on a snowy Christmas evening, a young mother and her little daughter are sitting together on a comfortable sofa. The mother, with a gentle expression and wearing a cream-colored knitted sweater, holds an open storybook with colorful illustrations. Her daughter, a small girl with brown hair in pigtails and a red pajama set, leans her head affectionately on her mother's shoulder, her eyes fixed on the book. On the floor below them, a fluffy, long-haired white cat is curled up on a plush, beige wool rug. The cat wears a tiny red and white Santa hat perched between its ears. Initially, the shot focuses on the mother and daughter, capturing their quiet, shared moment. The mother’s finger gently rests on the page of the book. Then, the camera slowly moves downward, gliding past the book and their laps. Finally, the camera settles at a low angle, bringing the adorable white cat into sharp focus as the primary subject. The cat's chest gently rises and falls with each breath, its eyes peacefully closed. Through a large window in the background, large, soft snowflakes can be seen falling silently against the dark blue twilight sky, creating a peaceful and serene backdrop. Faint, out-of-focus golden Christmas lights twinkle in the corner of the room, adding to the warm, festive atmosphere. The scene is imbued with a sense of comfort and holiday warmth, creating a beautiful cinematic photography realistic image. The camera slowly moves downward. The shot uses soft, warm interior lighting that casts gentle shadows, creating a high-contrast, cinematic look. A shallow depth of field keeps the focus on the subjects while beautifully blurring the background elements. The mood is heartwarming, peaceful, and festive. The overall video presents a cinematic photography realistic style.``` </details>|<video src="https://github.com/user-attachments/assets/8e72ed0f-f8ac-445b-97e5-eb4b16fbc121" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```The hiker begins walking forward along the trail, causing the water bottle to swing rhythmically with each step. The camera gradually pulls back and rises to reveal a vast desert landscape stretching out ahead.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```The hiker begins walking forward along the trail, causing the water bottle to swing rhythmically with each step. The camera gradually pulls back and rises to reveal a vast desert landscape stretching out ahead, while the sun position shifts from afternoon to dusk, casting increasingly longer shadows across the terrain as the figure becomes smaller in the frame.``` </details>|
+|多风格支持|<video src="https://github.com/user-attachments/assets/65b2c5a5-e6ba-43be-9462-a98b03b675f1" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```Have the cake man begin to take chunks out of himself and eat it.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```The cake man sits on the chair, with his hands resting on his knees. Then, he slowly raises his right hand and breaks off a piece of cake from his left shoulder. Next, he brings the piece of cake to his mouth and begins to chew. At the same time, his eyes widen slightly, and his mouth parts gently. After that, he raises his right hand again, breaks off another piece of cake from his right arm, and repeats the action of bringing it to his mouth to chew.``` </details>|<video src="https://github.com/user-attachments/assets/de5f7480-b79c-4fc1-b345-c5880a3b5f9e" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```A little girl, carrying a colorful handbag, skips through the garden.  The video uses claymation style.``` </details> <details><summary>📋 Show rewrite prompt</summary> ```A little girl with a colorful handbag skips through a whimsical claymation garden. In a vibrant garden constructed entirely from clay, a young girl, meticulously crafted in a claymation style, skips joyfully. She has chunky, sculpted yellow clay hair tied in pigtails that bounce with a slight stiffness, simple black button eyes, and a wide, permanently etched smile. She wears a simple pink clay dress with a white collar. In her left hand, she carries a small handbag molded from bright red and blue clay, which swings in a slightly jerky arc as she moves. Initially, the girl lifts her right leg high, her body momentarily suspended in a classic stop-motion pose. Then, she hops forward, landing lightly as her left leg swings through for the next skip. Her arms move in an exaggerated, back-and-forth rhythm, characteristic of stop-motion animation. Her movements are intentionally not perfectly fluid, highlighting the frame-by-frame nature of the claymation technique. The garden around her is a whimsical, textured world. In the foreground and mid-ground, oversized flowers with swirled purple and orange petals stand on thick green stems. The ground is a textured mat of green clay, showing subtle fingerprints and tool marks that add to the handmade charm. In the background, a pale blue clay backdrop features a simplified, smiling sun molded from yellow clay. The shot is at an eye-level angle with the main subject. The camera follows the subject, moving smoothly to the right to keep her in the frame. The lighting is bright and even, casting soft shadows that emphasize the rounded, three-dimensional forms of the clay models. The overall video presents a charming and detailed claymation style.``` </details>|
+|高图视一致性|<img src="https://github.com/user-attachments/assets/3bc8e55d-c211-454e-8067-128c0e215eb6"> <video src="https://github.com/user-attachments/assets/3e6b7ee9-ec66-4e46-a446-801b1c1a1c81" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```女孩放下书，站起身，转身向屋内走去。镜头拉远。``` </details> <details><summary>📋 Show rewrite prompt</summary> ```女孩合上手中的书，将书放在身侧的窗台上。随后，她缓缓站起身，转身向屋内走去，身影逐渐没入门后的阴影中。镜头缓缓拉远，露出更多被绿植覆盖的屋檐和墙体。``` </details>|<img src="https://github.com/user-attachments/assets/7657ce60-90b5-4fdc-b713-0eaa55829b09"> <video src="https://github.com/user-attachments/assets/9ca24021-2353-40d5-8a4d-0f8e67d51826" width="600"> </video> <details><summary>📋 Show input prompt</summary> ```女人手上的鸟亲了女人一口``` </details> <details><summary>📋 Show rewrite prompt</summary> ```女人手臂上的白色鹦鹉缓缓转过头，将喙轻轻触碰女人的脸颊，随后收回头部。女人嘴角微微上扬，目光温柔地注视着鹦鹉。背景中的绿植保持静止。``` </details>|
+
 
 ## 📚 引用
 ```bibtex

assets/step_distillation_comparison.md

@@ -0,0 +1,29 @@
+# Step Distillation Comparison
+
+This document provides detailed quality comparisons between the original 480p I2V model and the step-distilled model.
+
+## Overview
+
+The step-distilled model reduces inference steps from 50 to 8 (or 12 steps recommended) while maintaining comparable visual quality to the original model. On RTX 4090, this achieves up to 75% reduction in end-to-end generation time, enabling a single RTX 4090 to generate videos within 75 seconds. This document showcases side-by-side comparisons to demonstrate that the distillation process does not significantly degrade output quality. For even faster generation, you can also try 4 steps, which provides faster speed with slightly reduced quality.
+
+## Comparison Results
+
+The following table shows side-by-side comparisons between the original 480p I2V model (50 steps) and the step-distilled model (8 steps). The comparisons demonstrate that the step-distilled model maintains comparable visual quality while achieving significant speedup.
+
+<div align="center">
+<video src="https://github.com/user-attachments/assets/8ac11d99-1e8e-4b73-a7b4-e1a990c830ec" width="100%"></video>
+</div>
+
+
+
+
+
+
+
+## Usage Notes
+
+- **8 or 12 steps**: Recommended default setting, provides the best balance between speed and quality
+- **4 steps**: Faster generation with slightly reduced quality, suitable for rapid prototyping
+
+Detailed usage instructions can be found in [Usage](https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5/blob/main/README.md#-usage).
+