Integrate FluxPipeline and AutoencoderKL for enhanced image generation; add live preview helper functions and update requirements

app.py

@@ -15,9 +15,16 @@ from trellis.pipelines import TrellisImageTo3DPipeline
 from trellis.representations import Gaussian, MeshExtractResult
 from trellis.utils import render_utils, postprocessing_utils
 from gradio_client import Client
+from diffusers import FluxPipeline, AutoencoderKL
+from live_preview_helpers import flux_pipe_call_that_returns_an_iterable_of_images
 
 llm_client = Client("Qwen/Qwen2.5-72B-Instruct")
-t2i_client = Client("multimodalart/FLUX.1-merged")
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16).to(device)
+good_vae = AutoencoderKL.from_pretrained("black-forest-labs/FLUX.1-dev", subfolder="vae", torch_dtype=torch.bfloat16).to(device)
+pipe.flux_pipe_call_that_returns_an_iterable_of_images = flux_pipe_call_that_returns_an_iterable_of_images.__get__(pipe)
 
 def generate_t2i_prompt(item_name):
     llm_prompt_template = """You are tasked with creating a concise yet highly detailed description of an item to be used for generating an image in a game development pipeline. The image should show the **entire item** with no parts cropped or hidden. The background should always be plain and monocolor, with no focus on it.  
@@ -64,20 +71,33 @@ def preprocess_pil_image(image: Image.Image) -> Tuple[str, Image.Image]:
     processed_image.save(f"{TMP_DIR}/{trial_id}.png")
     return trial_id, processed_image
 
+@spaces.GPU
 def generate_item_image(object_t2i_prompt):
-    img_path = t2i_client.predict(
-		prompt=object_t2i_prompt,
-		seed=0,
-		randomize_seed=True,
-		width=1024,
-		height=1024,
-		guidance_scale=3.5,
-		num_inference_steps=8,
-		api_name="/infer"
-    )[0]
-    image = Image.open(img_path)
+    trial_id = ""
+    for image in pipe.flux_pipe_call_that_returns_an_iterable_of_images(
+        prompt=object_t2i_prompt,
+        guidance_scale=3.5,
+        num_inference_steps=28,
+        width=1024,
+        height=1024,
+        generator=torch.Generator("cpu").manual_seed(0),
+        output_type="pil",
+        good_vae=good_vae,
+    ):
+        yield trial_id, image
+    # img_path = t2i_client.predict(
+	# 	prompt=object_t2i_prompt,
+	# 	seed=0,
+	# 	randomize_seed=True,
+	# 	width=1024,
+	# 	height=1024,
+	# 	guidance_scale=3.5,
+	# 	num_inference_steps=8,
+	# 	api_name="/infer"
+    # )[0]
+    # image = Image.open(img_path)
     trial_id, processed_image = preprocess_pil_image(image)
-    return trial_id, processed_image
+    yield trial_id, processed_image
 
 MAX_SEED = np.iinfo(np.int32).max
 TMP_DIR = "/tmp/Trellis-demo"

live_preview_helpers.py

@@ -0,0 +1,166 @@
+import torch
+import numpy as np
+from diffusers import FluxPipeline, AutoencoderTiny, FlowMatchEulerDiscreteScheduler
+from typing import Any, Dict, List, Optional, Union
+
+# Helper functions
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.16,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+# FLUX pipeline function
+@torch.inference_mode()
+def flux_pipe_call_that_returns_an_iterable_of_images(
+    self,
+    prompt: Union[str, List[str]] = None,
+    prompt_2: Optional[Union[str, List[str]]] = None,
+    height: Optional[int] = None,
+    width: Optional[int] = None,
+    num_inference_steps: int = 28,
+    timesteps: List[int] = None,
+    guidance_scale: float = 3.5,
+    num_images_per_prompt: Optional[int] = 1,
+    generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+    latents: Optional[torch.FloatTensor] = None,
+    prompt_embeds: Optional[torch.FloatTensor] = None,
+    pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+    output_type: Optional[str] = "pil",
+    return_dict: bool = True,
+    joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+    max_sequence_length: int = 512,
+    good_vae: Optional[Any] = None,
+):
+    height = height or self.default_sample_size * self.vae_scale_factor
+    width = width or self.default_sample_size * self.vae_scale_factor
+
+    # 1. Check inputs
+    self.check_inputs(
+        prompt,
+        prompt_2,
+        height,
+        width,
+        prompt_embeds=prompt_embeds,
+        pooled_prompt_embeds=pooled_prompt_embeds,
+        max_sequence_length=max_sequence_length,
+    )
+
+    self._guidance_scale = guidance_scale
+    self._joint_attention_kwargs = joint_attention_kwargs
+    self._interrupt = False
+
+    # 2. Define call parameters
+    batch_size = 1 if isinstance(prompt, str) else len(prompt)
+    device = self._execution_device
+
+    # 3. Encode prompt
+    lora_scale = joint_attention_kwargs.get("scale", None) if joint_attention_kwargs is not None else None
+    prompt_embeds, pooled_prompt_embeds, text_ids = self.encode_prompt(
+        prompt=prompt,
+        prompt_2=prompt_2,
+        prompt_embeds=prompt_embeds,
+        pooled_prompt_embeds=pooled_prompt_embeds,
+        device=device,
+        num_images_per_prompt=num_images_per_prompt,
+        max_sequence_length=max_sequence_length,
+        lora_scale=lora_scale,
+    )
+    # 4. Prepare latent variables
+    num_channels_latents = self.transformer.config.in_channels // 4
+    latents, latent_image_ids = self.prepare_latents(
+        batch_size * num_images_per_prompt,
+        num_channels_latents,
+        height,
+        width,
+        prompt_embeds.dtype,
+        device,
+        generator,
+        latents,
+    )
+    # 5. Prepare timesteps
+    sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+    image_seq_len = latents.shape[1]
+    mu = calculate_shift(
+        image_seq_len,
+        self.scheduler.config.base_image_seq_len,
+        self.scheduler.config.max_image_seq_len,
+        self.scheduler.config.base_shift,
+        self.scheduler.config.max_shift,
+    )
+    timesteps, num_inference_steps = retrieve_timesteps(
+        self.scheduler,
+        num_inference_steps,
+        device,
+        timesteps,
+        sigmas,
+        mu=mu,
+    )
+    self._num_timesteps = len(timesteps)
+
+    # Handle guidance
+    guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32).expand(latents.shape[0]) if self.transformer.config.guidance_embeds else None
+
+    # 6. Denoising loop
+    for i, t in enumerate(timesteps):
+        if self.interrupt:
+            continue
+
+        timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+        noise_pred = self.transformer(
+            hidden_states=latents,
+            timestep=timestep / 1000,
+            guidance=guidance,
+            pooled_projections=pooled_prompt_embeds,
+            encoder_hidden_states=prompt_embeds,
+            txt_ids=text_ids,
+            img_ids=latent_image_ids,
+            joint_attention_kwargs=self.joint_attention_kwargs,
+            return_dict=False,
+        )[0]
+        # Yield intermediate result
+        latents_for_image = self._unpack_latents(latents, height, width, self.vae_scale_factor)
+        latents_for_image = (latents_for_image / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+        image = self.vae.decode(latents_for_image, return_dict=False)[0]
+        yield self.image_processor.postprocess(image, output_type=output_type)[0]
+        
+        latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+        torch.cuda.empty_cache()
+
+    # Final image using good_vae
+    latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
+    latents = (latents / good_vae.config.scaling_factor) + good_vae.config.shift_factor
+    image = good_vae.decode(latents, return_dict=False)[0]
+    self.maybe_free_model_hooks()
+    torch.cuda.empty_cache()
+    yield self.image_processor.postprocess(image, output_type=output_type)[0]

requirements.txt

@@ -1,5 +1,8 @@
 --extra-index-url https://download.pytorch.org/whl/cu121
 
+accelerate
+sentencepiece
+diffusers
 torch==2.4.0
 torchvision==0.19.0
 pillow==10.4.0