init project

app.py

@@ -1,142 +1,419 @@
+from typing import Tuple, Optional
+
 import gradio as gr
 import numpy as np
 import random
-#import spaces #[uncomment to use ZeroGPU]
+import spaces
 from diffusers import DiffusionPipeline
+from diffusers import FluxInpaintPipeline
 import torch
+from PIL import Image, ImageFilter
+from huggingface_hub import login
+from huggingface_hub import hf_hub_download, HfFileSystem, ModelCard, snapshot_download
+import copy
+import random
+import time
+import boto3
+from io import BytesIO
+from datetime import datetime
+from diffusers.utils import load_image
+import json
+
+from utils.florence import load_florence_model, run_florence_inference, \
+    FLORENCE_OPEN_VOCABULARY_DETECTION_TASK
+from utils.sam import load_sam_image_model, run_sam_inference
+
 
-device = "cuda" if torch.cuda.is_available() else "cpu"
-model_repo_id = "stabilityai/sdxl-turbo" #Replace to the model you would like to use
 
-if torch.cuda.is_available():
-    torch_dtype = torch.float16
-else:
-    torch_dtype = torch.float32
+HF_TOKEN = os.environ.get("HF_TOKEN")
 
-pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
-pipe = pipe.to(device)
+login(token=HF_TOKEN)
 
 MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 1024
+IMAGE_SIZE = 1024
 
-#@spaces.GPU #[uncomment to use ZeroGPU]
-def infer(prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps, progress=gr.Progress(track_tqdm=True)):
+# init
+dtype = torch.bfloat16
+device = "cuda" if torch.cuda.is_available() else "cpu"
+base_model = "black-forest-labs/FLUX.1-dev"
 
-    if randomize_seed:
-        seed = random.randint(0, MAX_SEED)
-        
-    generator = torch.Generator().manual_seed(seed)
+taef1 = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype).to(device)
+good_vae = AutoencoderKL.from_pretrained(base_model, subfolder="vae", torch_dtype=dtype).to(device)
+pipe = FluxInpaintPipeline.from_pretrained(base_model, torch_dtype=dtype, vae=taef1).to(device)
+
+class calculateDuration:
+    def __init__(self, activity_name=""):
+        self.activity_name = activity_name
+
+    def __enter__(self):
+        self.start_time = time.time()
+        return self
     
-    image = pipe(
-        prompt = prompt, 
-        negative_prompt = negative_prompt,
-        guidance_scale = guidance_scale, 
-        num_inference_steps = num_inference_steps, 
-        width = width, 
-        height = height,
-        generator = generator
-    ).images[0] 
+    def __exit__(self, exc_type, exc_value, traceback):
+        self.end_time = time.time()
+        self.elapsed_time = self.end_time - self.start_time
+        if self.activity_name:
+            print(f"Elapsed time for {self.activity_name}: {self.elapsed_time:.6f} seconds")
+        else:
+            print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
+
+
+def calculate_image_dimensions_for_flux(
+    original_resolution_wh: Tuple[int, int],
+    maximum_dimension: int = IMAGE_SIZE
+) -> Tuple[int, int]:
+    width, height = original_resolution_wh
+
+    if width > height:
+        scaling_factor = maximum_dimension / width
+    else:
+        scaling_factor = maximum_dimension / height
+
+    new_width = int(width * scaling_factor)
+    new_height = int(height * scaling_factor)
+
+    new_width = new_width - (new_width % 32)
+    new_height = new_height - (new_height % 32)
+
+    return new_width, new_height
+
+def is_mask_empty(image: Image.Image) -> bool:
+    gray_img = image.convert("L")
+    pixels = list(gray_img.getdata())
+    return all(pixel == 0 for pixel in pixels)
+
+def process_mask(
+    mask: Image.Image,
+    mask_inflation: Optional[int] = None,
+    mask_blur: Optional[int] = None
+) -> Image.Image:
+    """
+    Inflates and blurs the white regions of a mask.
+    Args:
+        mask (Image.Image): The input mask image.
+        mask_inflation (Optional[int]): The number of pixels to inflate the mask by.
+        mask_blur (Optional[int]): The radius of the Gaussian blur to apply.
+    Returns:
+        Image.Image: The processed mask with inflated and/or blurred regions.
+    """
+    if mask_inflation and mask_inflation > 0:
+        mask_array = np.array(mask)
+        kernel = np.ones((mask_inflation, mask_inflation), np.uint8)
+        mask_array = cv2.dilate(mask_array, kernel, iterations=1)
+        mask = Image.fromarray(mask_array)
+
+    if mask_blur and mask_blur > 0:
+        mask = mask.filter(ImageFilter.GaussianBlur(radius=mask_blur))
+
+    return mask
+
+def upload_image_to_r2(image, account_id, access_key, secret_key, bucket_name):
+    print("upload_image_to_r2", account_id, access_key, secret_key, bucket_name)
+    connectionUrl = f"https://{account_id}.r2.cloudflarestorage.com"
+
+    s3 = boto3.client(
+        's3',
+        endpoint_url=connectionUrl,
+        region_name='auto',
+        aws_access_key_id=access_key,
+        aws_secret_access_key=secret_key
+    )
+
+    current_time = datetime.now().strftime("%Y/%m/%d/%H%M%S")
+    image_file = f"generated_images/{current_time}_{random.randint(0, MAX_SEED)}.png"
+    buffer = BytesIO()
+    image.save(buffer, "PNG")
+    buffer.seek(0)
+    s3.upload_fileobj(buffer, bucket_name, image_file)
+    print("upload finish", image_file)
+    return image_file
+
+
+@spaces.GPU(duration=50)
+def run_flux(
+    image: Image.Image,
+    mask: Image.Image,
+    prompt: str,
+    lora_path: str,
+    lora_weights: str,
+    lora_scale: float,
+    seed_slicer: int,
+    randomize_seed_checkbox: bool,
+    strength_slider: float,
+    num_inference_steps_slider: int,
+    resolution_wh: Tuple[int, int],
+) -> Image.Image:
+    print("Running FLUX...")
+
+    with calculateDuration("load lora"):
+        print("start to load lora", lora_path, lora_weights)
+        pipe.load_lora_weights(lora_path, weight_name=lora_weights)
+
+    width, height = resolution_wh
+    if randomize_seed_checkbox:
+        seed_slicer = random.randint(0, MAX_SEED)
+    generator = torch.Generator().manual_seed(seed_slicer)
+
+    return PIPE(
+        prompt=prompt,
+        image=image,
+        mask_image=mask,
+        width=width,
+        height=height,
+        strength=strength_slider,
+        generator=generator,
+        num_inference_steps=num_inference_steps_slider,
+        max_sequence_length=256,
+        joint_attention_kwargs={"scale": lora_scale},
+    ).images[0]
+
+
+@spaces.GPU(duration=50)
+def genearte_mask(image: Image.Image, masking_prompt_text: str) -> Image.Image:
+    # generate mask by florence & sam
+    print("Generating mask...")
+
+    return 
+
+
+def process(
+    image_url: str,
+    inpainting_prompt_text: str,
+    masking_prompt_text: str,
+    mask_inflation_slider: int,
+    mask_blur_slider: int,
+    seed_slicer: int,
+    randomize_seed_checkbox: bool,
+    strength_slider: float,
+    num_inference_steps_slider: int,
+    lora_path: str,
+    lora_weights: str,
+    lora_scale: str,
+    upload_to_r2: bool,
+    account_id: str,
+    access_key: str,
+    secret_key: str,
+    bucket:str
+):
+    result = {"status": "false", "message": ""}
+    if not image_url:
+        gr.Info("please enter image url for inpaiting")
+        result["message"] = "invalid image url"
+        return None, None, json.dumps(result)
+
+    if not inpainting_prompt_text:
+        gr.Info("Please enter inpainting text prompt.")
+        result["message"] = "invalid inpainting prompt"
+        return None, None, json.dumps(result)
     
-    return image, seed
-
-examples = [
-    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
-    "An astronaut riding a green horse",
-    "A delicious ceviche cheesecake slice",
-]
-
-css="""
-#col-container {
-    margin: 0 auto;
-    max-width: 640px;
-}
-"""
-
-with gr.Blocks(css=css) as demo:
+    if not masking_prompt_text:
+        gr.Info("Please enter masking_prompt_text.")
+        result["message"] = "invalid masking prompt"
+        return None, None, json.dumps(result)
+
+
+    image = load_image(image_url)
+    mask = genearte_mask(image, masking_prompt_text)
     
-    with gr.Column(elem_id="col-container"):
-        gr.Markdown(f"""
-        # Text-to-Image Gradio Template
-        """)
-        
-        with gr.Row():
-            
-            prompt = gr.Text(
-                label="Prompt",
-                show_label=False,
-                max_lines=1,
-                placeholder="Enter your prompt",
-                container=False,
-            )
-            
-            run_button = gr.Button("Run", scale=0)
-        
-        result = gr.Image(label="Result", show_label=False)
+    if not image:
+        gr.Info("Please upload an image.")
+        result["message"] = "can not load image"
+        return None, None, json.dumps(result)
 
-        with gr.Accordion("Advanced Settings", open=False):
-            
-            negative_prompt = gr.Text(
-                label="Negative prompt",
-                max_lines=1,
-                placeholder="Enter a negative prompt",
-                visible=False,
-            )
-            
-            seed = gr.Slider(
-                label="Seed",
-                minimum=0,
-                maximum=MAX_SEED,
-                step=1,
-                value=0,
-            )
+    if is_mask_empty(mask):
+        gr.Info("Please draw a mask or enter a masking prompt.")
+        result["message"] = "can not generate mask"
+        return None, None,  json.dumps(result)
+
+    # generate
+    width, height = calculate_image_dimensions_for_flux(original_resolution_wh=image.size)
+    image = image.resize((width, height), Image.LANCZOS)
+    mask = mask.resize((width, height), Image.LANCZOS)
+    mask = process_mask(mask, mask_inflation=mask_inflation_slider, mask_blur=mask_blur_slider)
+    image = run_flux(
+        image=image,
+        mask=mask,
+        prompt=inpainting_prompt_text,
+        lora_path=lora_path,
+        lora_scale=lora_scale,
+        lora_weights=lora_weights,
+        seed_slicer=seed_slicer,
+        randomize_seed_checkbox=randomize_seed_checkbox,
+        strength_slider=strength_slider,
+        num_inference_steps_slider=num_inference_steps_slider,
+        resolution_wh=(width, height)
+    )
+    if upload_to_r2:
+        url = upload_image_to_r2(image, account_id, access_key, secret_key, bucket)
+        result = {"status": "success", "url": url}
+    else:
+        result = {"status": "success", "message": "Image generated but not uploaded"}
+
+    return image, mask, json.dumps(result)
+
+
+
+with gr.Blocks() as demo:
+    
+    with gr.Row():
+        with gr.Column():
             
-            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+            image_url =  gr.Text(
+                    label="Image url for inpainting",
+                    show_label=False,
+                    max_lines=1,
+                    placeholder="Enter image url for inpainting",
+                    container=False,
+                )
             
-            with gr.Row():
-                
-                width = gr.Slider(
-                    label="Width",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=1024, #Replace with defaults that work for your model
+            masking_prompt_text_component = gr.Text(
+                    label="Masking prompt",
+                    show_label=False,
+                    max_lines=1,
+                    placeholder="Enter text to generate masking",
+                    container=False,
                 )
-                
-                height = gr.Slider(
-                    label="Height",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=1024, #Replace with defaults that work for your model
+
+            inpainting_prompt_text_component = gr.Text(
+                    label="Inpainting prompt",
+                    show_label=False,
+                    max_lines=1,
+                    placeholder="Enter text to generate inpainting",
+                    container=False,
                 )
-            
-            with gr.Row():
-                
-                guidance_scale = gr.Slider(
-                    label="Guidance scale",
-                    minimum=0.0,
-                    maximum=10.0,
+
+            submit_button_component = gr.Button(value='Submit', variant='primary', scale=0)
+
+            with gr.Accordion("Lora Settings", open=True):
+                lora_path = gr.Textbox(
+                    label="Lora model path",
+                    show_label=True,
+                    max_lines=1,
+                    placeholder="Enter your model path",
+                    info="Currently, only LoRA hosted on Hugging Face'model can be loaded properly.",
+                    value="XLabs-AI/flux-RealismLora"
+                )
+                lora_weights = gr.Textbox(
+                    label="Lora weights",
+                    show_label=True,
+                    max_lines=1,
+                    placeholder="Enter your lora weights name",
+                    value="lora.safetensors"
+                )
+                lora_scale = gr.Slider(
+                    label="Lora scale",
+                    show_label=True,
+                    minimum=0,
+                    maximum=1,
                     step=0.1,
-                    value=0.0, #Replace with defaults that work for your model
+                    value=0.9,
                 )
                 
-                num_inference_steps = gr.Slider(
-                    label="Number of inference steps",
-                    minimum=1,
-                    maximum=50,
+                
+
+            with gr.Accordion("Advanced Settings", open=False):
+               
+
+                with gr.Row():
+                    mask_inflation_slider_component = gr.Slider(
+                        label="Mask inflation",
+                        info="Adjusts the amount of mask edge expansion before "
+                             "inpainting.",
+                        minimum=0,
+                        maximum=20,
+                        step=1,
+                        value=5,
+                    )
+
+                    mask_blur_slider_component = gr.Slider(
+                        label="Mask blur",
+                        info="Controls the intensity of the Gaussian blur applied to "
+                             "the mask edges.",
+                        minimum=0,
+                        maximum=20,
+                        step=1,
+                        value=5,
+                    )
+
+                seed_slicer_component = gr.Slider(
+                    label="Seed",
+                    minimum=0,
+                    maximum=MAX_SEED,
                     step=1,
-                    value=2, #Replace with defaults that work for your model
+                    value=42,
                 )
+
+                randomize_seed_checkbox_component = gr.Checkbox(
+                    label="Randomize seed", value=True)
+
+                with gr.Row():
+
+                    strength_slider_component = gr.Slider(
+                        label="Strength",
+                        info="Indicates extent to transform the reference `image`. "
+                             "Must be between 0 and 1. `image` is used as a starting "
+                             "point and more noise is added the higher the `strength`.",
+                        minimum=0,
+                        maximum=1,
+                        step=0.01,
+                        value=0.85,
+                    )
+
+                    num_inference_steps_slider_component = gr.Slider(
+                        label="Number of inference steps",
+                        info="The number of denoising steps. More denoising steps "
+                             "usually lead to a higher quality image at the",
+                        minimum=1,
+                        maximum=50,
+                        step=1,
+                        value=20,
+                    )
+                
+                upload_to_r2 = gr.Checkbox(label="Upload to R2", value=False)
+                account_id = gr.Textbox(label="Account Id", placeholder="Enter R2 account id")
+                access_key = gr.Textbox(label="Access Key", placeholder="Enter R2 access key here")
+                secret_key = gr.Textbox(label="Secret Key", placeholder="Enter R2 secret key here")
+                bucket = gr.Textbox(label="Bucket Name", placeholder="Enter R2 bucket name here")
         
-        gr.Examples(
-            examples = examples,
-            inputs = [prompt]
-        )
-    gr.on(
-        triggers=[run_button.click, prompt.submit],
-        fn = infer,
-        inputs = [prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
-        outputs = [result, seed]
-    )
+                
+        with gr.Column():
 
+            output_image_component = gr.Image(
+                type='pil', image_mode='RGB', label='Generated image', format="png")
+            
+            
+
+            with gr.Accordion("Debug", open=False):
+                output_mask_component = gr.Image(
+                    type='pil', image_mode='RGB', label='Input mask', format="png")
+
+            output_json_component = gr.Textbox()
+
+    submit_button_component.click(
+        fn=process,
+        inputs=[
+            image_url,
+            inpainting_prompt_text_component,
+            masking_prompt_text_component,
+            mask_inflation_slider_component,
+            mask_blur_slider_component,
+            seed_slicer_component,
+            randomize_seed_checkbox_component,
+            strength_slider_component,
+            num_inference_steps_slider_component,
+            lora_path,
+            lora_weights,
+            lora_scale,
+            upload_to_r2,
+            account_id,
+            access_key,
+            secret_key,
+            bucket
+        ],
+        outputs=[
+            output_image_component,
+            output_mask_component,
+            output_json_component
+        ]
+    )
 demo.queue().launch()

configs/sam2_hiera_b+.yaml

@@ -0,0 +1,113 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 112
+      num_heads: 2
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [896, 448, 224, 112]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False

configs/sam2_hiera_l.yaml

@@ -0,0 +1,117 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 144
+      num_heads: 2
+      stages: [2, 6, 36, 4]
+      global_att_blocks: [23, 33, 43]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+      window_spec: [8, 4, 16, 8]
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [1152, 576, 288, 144]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False

configs/sam2_hiera_s.yaml

@@ -0,0 +1,116 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 96
+      num_heads: 1
+      stages: [1, 2, 11, 2]
+      global_att_blocks: [7, 10, 13]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [768, 384, 192, 96]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False

configs/sam2_hiera_t.yaml

@@ -0,0 +1,118 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 96
+      num_heads: 1
+      stages: [1, 2, 7, 2]
+      global_att_blocks: [5, 7, 9]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [768, 384, 192, 96]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  # SAM decoder
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  # HieraT does not currently support compilation, should always be set to False
+  compile_image_encoder: False

requirements.txt

@@ -1,6 +1,17 @@
 accelerate
-diffusers
 invisible_watermark
 torch
 transformers
-xformers
+xformers
+tqdm
+einops
+spaces
+timm
+samv2
+gradio
+supervision
+opencv-python
+pytest
+requests
+git+https://github.com/Gothos/diffusers.git@flux-inpaint
+boto3

utils/florence.py

@@ -0,0 +1,64 @@
+import os
+from typing import Union, Any, Tuple, Dict
+from unittest.mock import patch
+
+import torch
+from PIL import Image
+from transformers import AutoModelForCausalLM, AutoProcessor
+from transformers.dynamic_module_utils import get_imports
+
+# FLORENCE_CHECKPOINT = "microsoft/Florence-2-base"
+FLORENCE_CHECKPOINT = "microsoft/Florence-2-large-ft"
+FLORENCE_OBJECT_DETECTION_TASK = '<OD>'
+FLORENCE_DETAILED_CAPTION_TASK = '<MORE_DETAILED_CAPTION>'
+FLORENCE_CAPTION_TO_PHRASE_GROUNDING_TASK = '<CAPTION_TO_PHRASE_GROUNDING>'
+FLORENCE_OPEN_VOCABULARY_DETECTION_TASK = '<OPEN_VOCABULARY_DETECTION>'
+FLORENCE_DENSE_REGION_CAPTION_TASK = '<DENSE_REGION_CAPTION>'
+
+
+def fixed_get_imports(filename: Union[str, os.PathLike]) -> list[str]:
+    """Work around for https://huggingface.co/microsoft/phi-1_5/discussions/72."""
+    if not str(filename).endswith("/modeling_florence2.py"):
+        return get_imports(filename)
+    imports = get_imports(filename)
+    imports.remove("flash_attn")
+    return imports
+
+
+def load_florence_model(
+    device: torch.device, checkpoint: str = FLORENCE_CHECKPOINT
+) -> Tuple[Any, Any]:
+    with patch("transformers.dynamic_module_utils.get_imports", fixed_get_imports):
+        model = AutoModelForCausalLM.from_pretrained(
+            checkpoint, trust_remote_code=True).to(device).eval()
+        processor = AutoProcessor.from_pretrained(
+            checkpoint, trust_remote_code=True)
+        return model, processor
+
+
+def run_florence_inference(
+    model: Any,
+    processor: Any,
+    device: torch.device,
+    image: Image,
+    task: str,
+    text: str = None
+) -> Tuple[str, Dict]:
+    if text:
+        prompt = task + text
+    else:
+        prompt = task
+    inputs = processor(text=prompt, images=image, return_tensors="pt").to(device)
+    print(inputs)
+    generated_ids = model.generate(
+        input_ids=inputs["input_ids"],
+        pixel_values=inputs["pixel_values"],
+        max_new_tokens=1024,
+        num_beams=3
+    )
+    generated_text = processor.batch_decode(
+        generated_ids, skip_special_tokens=False)[0]
+    response = processor.post_process_generation(
+        generated_text, task=task, image_size=image.size)
+    print(generated_text, response)
+    return generated_text, response

utils/sam.py

@@ -0,0 +1,50 @@
+from typing import Any
+
+import numpy as np
+import supervision as sv
+import torch
+from PIL import Image
+from sam2.build_sam import build_sam2, build_sam2_video_predictor
+from sam2.sam2_image_predictor import SAM2ImagePredictor
+
+# SAM_CHECKPOINT = "checkpoints/sam2_hiera_small.pt"
+# SAM_CONFIG = "sam2_hiera_s.yaml"
+SAM_CHECKPOINT = "checkpoints/sam2_hiera_large.pt"
+SAM_CONFIG = "sam2_hiera_l.yaml"
+
+
+def load_sam_image_model(
+    device: torch.device,
+    config: str = SAM_CONFIG,
+    checkpoint: str = SAM_CHECKPOINT
+) -> SAM2ImagePredictor:
+    model = build_sam2(config, checkpoint, device=device)
+    return SAM2ImagePredictor(sam_model=model)
+
+
+def load_sam_video_model(
+    device: torch.device,
+    config: str = SAM_CONFIG,
+    checkpoint: str = SAM_CHECKPOINT
+) -> Any:
+    return build_sam2_video_predictor(config, checkpoint, device=device)
+
+
+def run_sam_inference(
+    model: Any,
+    image: Image,
+    detections: sv.Detections
+) -> sv.Detections:
+    image = np.array(image.convert("RGB"))
+    model.set_image(image)
+    # from left to right
+    bboxes = detections.xyxy
+    bboxes = sorted(bboxes, key=lambda bbox: bbox[0])
+    mask, score, _ = model.predict(box=bboxes, multimask_output=False)
+
+    # dirty fix; remove this later
+    if len(mask.shape) == 4:
+        mask = np.squeeze(mask)
+
+    detections.mask = mask.astype(bool)
+    return detections