Upload 11 files

app.py

@@ -0,0 +1,220 @@
+from typing import TypedDict
+
+import diffusers.image_processor
+import gradio as gr
+import pillow_heif
+import spaces
+import torch
+from PIL import Image
+from transformers import SegformerImageProcessor, AutoModelForSemanticSegmentation
+from pipeline import TryOffAnyone
+import numpy as np
+
+
+pillow_heif.register_heif_opener()
+pillow_heif.register_avif_opener()
+
+torch.set_float32_matmul_precision("high")
+torch.backends.cuda.matmul.allow_tf32 = True
+
+TITLE = """
+# Try Off Anyone
+
+## Important
+
+1. Choose an example image or upload your own
+
+[[arxiv:2412.08573]](https://arxiv.org/abs/2412.08573)
+[[github:ixarchakos/try-off-anyone]](https://github.com/ixarchakos/try-off-anyone)
+"""
+
+DEVICE = torch.device("cuda" if torch.cuda.is_available() else "mps")
+DTYPE = torch.bfloat16 if DEVICE == 'cuda' else torch.float32
+
+pipeline_tryoff = TryOffAnyone(
+    device=DEVICE,
+    dtype=DTYPE,
+)
+mask_processor = diffusers.image_processor.VaeImageProcessor(
+    vae_scale_factor=8,
+    do_normalize=False,
+    do_binarize=True,
+    do_convert_grayscale=True,
+)
+vae_processor = diffusers.image_processor.VaeImageProcessor(
+    vae_scale_factor=8,
+)
+
+
+def mask_generation(image, processor, model, category):
+    inputs = processor(images=image, return_tensors="pt")
+    outputs = model(**inputs)
+    logits = outputs.logits.cpu()
+
+    upsampled_logits = torch.nn.functional.interpolate(
+        logits,
+        size=image.size[::-1],
+        mode="bilinear",
+        align_corners=False,
+    )
+
+    predicted_mask = upsampled_logits.argmax(dim=1).squeeze().cpu().numpy()
+    if category == "Tops":
+        predicted_mask_1 = predicted_mask == 4
+        predicted_mask_2 = predicted_mask == 7
+    elif category == "Bottoms":
+        predicted_mask_1 = predicted_mask == 5
+        predicted_mask_2 = predicted_mask == 6
+    else:
+        raise NotImplementedError
+
+    predicted_mask = predicted_mask_1 + predicted_mask_2
+    mask_image = Image.fromarray((predicted_mask * 255).astype(np.uint8))
+    return mask_image
+
+
+class ImageData(TypedDict):
+    background: Image.Image
+    composite: Image.Image
+    layers: list[Image.Image]
+
+
+@spaces.GPU
+def process(
+    image_data: ImageData,
+    image_width: int,
+    image_height: int,
+    num_inference_steps: int,
+    condition_scale: float,
+    seed: int,
+) -> Image.Image:
+    assert image_width > 0
+    assert image_height > 0
+    assert num_inference_steps > 0
+    assert condition_scale > 0
+    assert seed >= 0
+
+    # extract image and mask from image_data
+    image = image_data["background"]
+    processor = SegformerImageProcessor.from_pretrained("sayeed99/segformer_b3_clothes")
+    model = AutoModelForSemanticSegmentation.from_pretrained("sayeed99/segformer_b3_clothes")
+    model.to('cpu')
+
+    # preprocess image
+    image = image.convert("RGB").resize((image_width, image_height))
+    mask = mask_generation(image, processor, model, "Tops")
+    image_preprocessed = vae_processor.preprocess(
+        image=image,
+        width=image_width,
+        height=image_height,
+    )[0]
+
+    # preprocess mask
+    mask = mask.resize((image_width, image_height))
+    mask_preprocessed = mask_processor.preprocess(  # pyright: ignore[reportUnknownMemberType]
+        image=mask,
+        width=image_width,
+        height=image_height,
+    )[0]
+
+    # generate the TryOff image
+    gen = torch.Generator(device=DEVICE).manual_seed(seed)
+    tryoff_image = pipeline_tryoff(
+        image_preprocessed,
+        mask_preprocessed,
+        inference_steps=num_inference_steps,
+        scale=condition_scale,
+        generator=gen,
+    )[0]
+
+    return tryoff_image
+
+
+with gr.Blocks() as demo:
+    gr.Markdown(TITLE)
+
+    with gr.Row():
+        with gr.Column():
+            input_image = gr.ImageMask(
+                label="Input Image",
+                height=1024,
+                type="pil",
+                interactive=True,
+            )
+            run_button = gr.Button(
+                value="Extract Clothing",
+            )
+            gr.Examples(
+                examples=[
+                    ["examples/model_1.jpg"],
+                    ["examples/model_2.jpg"],
+                    ["examples/model_3.jpg"],
+                    ["examples/model_4.jpg"],
+                    ["examples/model_5.jpg"],
+                    ["examples/model_6.jpg"],
+                    ["examples/model_7.jpg"],
+                    ["examples/model_8.jpg"],
+                    ["examples/model_9.jpg"],
+                ],
+                inputs=[input_image],
+            )
+        with gr.Column():
+            output_image = gr.Image(
+                label="TryOff result",
+                height=1024,
+                image_mode="RGB",
+                type="pil",
+            )
+
+    with gr.Accordion("Advanced Settings", open=False):
+        seed = gr.Slider(
+            label="Seed",
+            minimum=36,
+            maximum=36,
+            value=36,
+            step=1,
+        )
+        scale = gr.Slider(
+            label="Scale",
+            minimum=2.5,
+            maximum=2.5,
+            value=2.5,
+            step=0,
+        )
+        num_inference_steps = gr.Slider(
+            label="Number of inference steps",
+            minimum=50,
+            maximum=50,
+            value=50,
+            step=1,
+        )
+        with gr.Row():
+            image_width = gr.Slider(
+                label="Image Width",
+                minimum=384,
+                maximum=384,
+                value=384,
+                step=8,
+            )
+            image_height = gr.Slider(
+                label="Image Height",
+                minimum=512,
+                maximum=512,
+                value=512,
+                step=8,
+            )
+
+    run_button.click(
+        fn=process,
+        inputs=[
+            input_image,
+            image_width,
+            image_height,
+            num_inference_steps,
+            scale,
+            seed,
+        ],
+        outputs=output_image,
+    )
+
+demo.launch()

pipeline.py

@@ -0,0 +1,159 @@
+# type: ignore
+# Inspired from https://github.com/ixarchakos/try-off-anyone/blob/aa3045453013065573a647e4536922bac696b968/src/model/pipeline.py
+# Inspired from https://github.com/ixarchakos/try-off-anyone/blob/aa3045453013065573a647e4536922bac696b968/src/model/attention.py
+
+import torch
+from accelerate import load_checkpoint_in_model
+from diffusers import AutoencoderKL, DDIMScheduler, UNet2DConditionModel
+from diffusers.models.attention_processor import AttnProcessor
+from diffusers.utils.torch_utils import randn_tensor
+from huggingface_hub import hf_hub_download
+from PIL import Image
+
+
+class Skip(torch.nn.Module):
+    def __init__(self) -> None:
+        super().__init__()
+
+    def __call__(
+        self,
+        attn: torch.Tensor,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor = None,
+        attention_mask: torch.Tensor = None,
+        temb: torch.Tensor = None,
+    ) -> torch.Tensor:
+        return hidden_states
+
+
+def fine_tuned_modules(unet: UNet2DConditionModel) -> torch.nn.ModuleList:
+    trainable_modules = torch.nn.ModuleList()
+
+    for blocks in [unet.down_blocks, unet.mid_block, unet.up_blocks]:
+        if hasattr(blocks, "attentions"):
+            trainable_modules.append(blocks.attentions)
+        else:
+            for block in blocks:
+                if hasattr(block, "attentions"):
+                    trainable_modules.append(block.attentions)
+
+    return trainable_modules
+
+
+def skip_cross_attentions(unet: UNet2DConditionModel) -> dict[str, AttnProcessor | Skip]:
+    attn_processors = {
+        name: unet.attn_processors[name] if name.endswith("attn1.processor") else Skip()
+        for name in unet.attn_processors.keys()
+    }
+    return attn_processors
+
+
+def encode(image: torch.Tensor, vae: AutoencoderKL) -> torch.Tensor:
+    image = image.to(memory_format=torch.contiguous_format).float().to(vae.device, dtype=vae.dtype)
+    with torch.no_grad():
+        return vae.encode(image).latent_dist.sample() * vae.config.scaling_factor
+
+
+class TryOffAnyone:
+    def __init__(
+        self,
+        device: torch.device,
+        dtype: torch.dtype,
+        concat_dim: int = -2,
+    ) -> None:
+        self.concat_dim = concat_dim
+        self.device = device
+        self.dtype = dtype
+
+        self.noise_scheduler = DDIMScheduler.from_pretrained(
+            pretrained_model_name_or_path="stable-diffusion-v1-5/stable-diffusion-inpainting",
+            subfolder="scheduler",
+        )
+        self.vae = AutoencoderKL.from_pretrained(
+            pretrained_model_name_or_path="stabilityai/sd-vae-ft-mse",
+        ).to(device, dtype=dtype)
+        self.unet = UNet2DConditionModel.from_pretrained(
+            pretrained_model_name_or_path="stable-diffusion-v1-5/stable-diffusion-inpainting",
+            subfolder="unet",
+            variant="fp16",
+        ).to(device, dtype=dtype)
+
+        self.unet.set_attn_processor(skip_cross_attentions(self.unet))
+        load_checkpoint_in_model(
+            model=fine_tuned_modules(unet=self.unet),
+            checkpoint=hf_hub_download(
+                repo_id="ixarchakos/tryOffAnyone",
+                filename="model.safetensors",
+            ),
+        )
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        image: torch.Tensor,
+        mask: torch.Tensor,
+        inference_steps: int,
+        scale: float,
+        generator: torch.Generator,
+    ) -> list[Image.Image]:
+        image = image.unsqueeze(0).to(self.device, dtype=self.dtype)
+        mask = (mask.unsqueeze(0) > 0.5).to(self.device, dtype=self.dtype)
+        masked_image = image * (mask < 0.5)
+
+        masked_latent = encode(masked_image, self.vae)
+        image_latent = encode(image, self.vae)
+        mask = torch.nn.functional.interpolate(mask, size=masked_latent.shape[-2:], mode="nearest")
+
+        masked_latent_concat = torch.cat([masked_latent, image_latent], dim=self.concat_dim)
+        mask_concat = torch.cat([mask, torch.zeros_like(mask)], dim=self.concat_dim)
+
+        latents = randn_tensor(
+            shape=masked_latent_concat.shape,
+            generator=generator,
+            device=self.device,
+            dtype=self.dtype,
+        )
+
+        self.noise_scheduler.set_timesteps(inference_steps, device=self.device)
+        timesteps = self.noise_scheduler.timesteps
+
+        if do_classifier_free_guidance := (scale > 1.0):
+            masked_latent_concat = torch.cat(
+                [
+                    torch.cat([masked_latent, torch.zeros_like(image_latent)], dim=self.concat_dim),
+                    masked_latent_concat,
+                ]
+            )
+
+            mask_concat = torch.cat([mask_concat] * 2)
+
+        extra_step = {"generator": generator, "eta": 1.0}
+        for t in timesteps:
+            input_latents = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+            input_latents = self.noise_scheduler.scale_model_input(input_latents, t)
+
+            input_latents = torch.cat([input_latents, mask_concat, masked_latent_concat], dim=1)
+
+            noise_pred = self.unet(
+                input_latents,
+                t.to(self.device),
+                encoder_hidden_states=None,
+                return_dict=False,
+            )[0]
+
+            if do_classifier_free_guidance:
+                noise_pred_unc, noise_pred_text = noise_pred.chunk(2)
+                noise_pred = noise_pred_unc + scale * (noise_pred_text - noise_pred_unc)
+
+            latents = self.noise_scheduler.step(noise_pred, t, latents, **extra_step).prev_sample
+
+        latents = latents.split(latents.shape[self.concat_dim] // 2, dim=self.concat_dim)[0]
+        latents = 1 / self.vae.config.scaling_factor * latents
+        image = self.vae.decode(latents.to(self.device, dtype=self.dtype)).sample
+        image = (image / 2 + 0.5).clamp(0, 1)
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+        image = (image * 255).round().astype("uint8")
+        image = [Image.fromarray(im) for im in image]
+
+        return image