app.py

import gradio as gr
import torch
from diffusers import AutoPipelineForInpainting
from PIL import Image
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BlipForConditionalGeneration,
    BlipProcessor,
    OwlViTForObjectDetection,
    OwlViTProcessor,
    SamModel,
    SamProcessor,
)


def delete_model(model):
    model.to("cpu")
    del model
    torch.cuda.empty_cache()


def run_language_model(edit_prompt, device):
    language_model_id = "Qwen/Qwen1.5-0.5B-Chat"
    language_model = AutoModelForCausalLM.from_pretrained(
        language_model_id, device_map="auto"
    )
    tokenizer = AutoTokenizer.from_pretrained(language_model_id)
    messages = [
        {
            "role": "system",
            "content": "Follow the examples and return the expected output",
        },
        {"role": "user", "content": "swap mountain and lion"},  # example 1
        {"role": "assistant", "content": "mountain, lion"},  # example 1
        {"role": "user", "content": "change the dog with cat"},  # example 2
        {"role": "assistant", "content": "dog, cat"},  # example 2
        {"role": "user", "content": "replace the human with a boat"},  # example 3
        {"role": "assistant", "content": "human, boat"},  # example 3
        {"role": "user", "content": edit_prompt},
    ]
    text = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
    model_inputs = tokenizer([text], return_tensors="pt").to(device)
    generated_ids = language_model.generate(model_inputs.input_ids, max_new_tokens=512)
    generated_ids = [
        output_ids[len(input_ids) :]
        for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
    ]
    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
    to_replace, replace_with = response.split(", ")

    delete_model(language_model)
    return (to_replace, replace_with)


def run_image_captioner(image, device):
    caption_model_id = "Salesforce/blip-image-captioning-base"
    caption_model = BlipForConditionalGeneration.from_pretrained(caption_model_id).to(
        device
    )
    caption_processor = BlipProcessor.from_pretrained(caption_model_id)
    inputs = caption_processor(image, return_tensors="pt").to(device)
    with torch.no_grad():
        outputs = caption_model.generate(**inputs, max_new_tokens=200)
    caption = caption_processor.decode(outputs[0], skip_special_tokens=True)

    delete_model(caption_model)
    return caption


def run_segmentation(image, object_to_segment, device):
    # OWL-ViT for object detection
    owl_vit_model_id = "google/owlvit-base-patch32"
    processor = OwlViTProcessor.from_pretrained(owl_vit_model_id)
    od_model = OwlViTForObjectDetection.from_pretrained(owl_vit_model_id).to(device)
    text_queries = [object_to_segment]
    inputs = processor(text=text_queries, images=image, return_tensors="pt").to(device)
    with torch.no_grad():
        outputs = od_model(**inputs)
        target_sizes = torch.tensor([image.size]).to(device)
        results = processor.post_process_object_detection(
            outputs, threshold=0.1, target_sizes=target_sizes
        )[0]

    boxes = results["boxes"].tolist()

    delete_model(od_model)

    # SAM for image segmentation
    sam_model_id = "facebook/sam-vit-base"
    seg_model = SamModel.from_pretrained(sam_model_id).to(device)
    processor = SamProcessor.from_pretrained(sam_model_id)
    input_boxes = [boxes]
    inputs = processor(image, input_boxes=input_boxes, return_tensors="pt").to(device)
    with torch.no_grad():
        outputs = seg_model(**inputs)
    masks = processor.image_processor.post_process_masks(
        outputs.pred_masks.cpu(),
        inputs["original_sizes"].cpu(),
        inputs["reshaped_input_sizes"].cpu(),
    )

    delete_model(seg_model)
    return masks


def run_inpainting(image, replaced_caption, masks, device):
    pipeline = AutoPipelineForInpainting.from_pretrained(
        "diffusers/stable-diffusion-xl-1.0-inpainting-0.1",
        torch_dtype=torch.float16,
        variant="fp16",
    ).to(device)

    prompt = replaced_caption
    negative_prompt = """lowres, bad anatomy, bad hands,
    text, error, missing fingers, extra digit, fewer digits,
    cropped, worst quality, low quality"""

    output = pipeline(
        prompt=prompt,
        image=image,
        mask_image=Image.fromarray(masks[0][0][0, :, :].numpy()),
        negative_prompt=negative_prompt,
        guidance_scale=7.5,
        strength=0.6,
    ).images[0]

    delete_model(pipeline)
    return output


def run_open_gen_fill(image, edit_prompt):
    device = "cuda" if torch.cuda.is_available() else "cpu"

    # Resize the image to (512, 512)
    image = image.resize((512, 512))

    # Run the langauge model to extract the objects to be swapped from
    # the edit prompt
    to_replace, replace_with = run_language_model(
        edit_prompt=edit_prompt, device=device
    )

    # Caption the input image
    caption = run_image_captioner(image, device=device)

    # Replace the object in the caption with the new object
    replaced_caption = caption.replace(to_replace, replace_with)

    # Segment the `to_replace` object from the input image
    masks = run_segmentation(image, to_replace, device=device)

    # Diffusion pipeline for inpainting
    return run_inpainting(
        image=image, replaced_caption=replaced_caption, masks=masks, device=device
    )


def setup_gradio_interface():
    block = gr.Blocks()

    with block:
        gr.Markdown("<h1><center>Open Generative Fill<h1><center>")

        with gr.Row():
            with gr.Column():
                input_image_placeholder = gr.Image(type="pil", label="Input Image")
                edit_prompt_placeholder = gr.Textbox(label="Enter the editing prompt")
                run_button_placeholder = gr.Button(value="Run")

            with gr.Column():
                output_image_placeholder = gr.Image(type="pil", label="Output Image")

        run_button_placeholder.click(
            fn=lambda image, edit_prompt: run_open_gen_fill(
                image=image,
                edit_prompt=edit_prompt,
            ),
            inputs=[input_image_placeholder, edit_prompt_placeholder],
            outputs=[output_image_placeholder],
        )

    return block


if __name__ == "__main__":
    gradio_interface = setup_gradio_interface()
    gradio_interface.queue(max_size=5)
    gradio_interface.launch(share=False, show_api=False, show_error=True)