app.py

import gradio as gr
import os
import random
import spaces
import torch
from transformers import MllamaForConditionalGeneration, AutoProcessor
from OmniGen import OmniGenPipeline
from PIL import Image

from huggingface_hub import login
Llama32V_HFtoken = os.getenv("Llama32V")
login(Llama32V_HFtoken)

pipe = OmniGenPipeline.from_pretrained("Shitao/OmniGen-v1")
model_id = "meta-llama/Llama-3.2-11B-Vision-Instruct"
model = MllamaForConditionalGeneration.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")
processor = AutoProcessor.from_pretrained(model_id)


@spaces.GPU()
def predict_clothing(images):
    messages = [{"role": "user", "content":
                 [
                     {"type": "image"},
                     {"type": "text", "text": "Define this clothing in 1-3 words, always include the colour of the clothing too. Your response should be only the definition."}
                 ]}
               ]
    input_text = processor.apply_chat_template(messages, add_generation_prompt=True)

    output_texts = []
    for img_path in images:
        image = Image.open(img_path)
        print(type(image))
        inputs = processor(image, input_text, add_special_tokens=False, return_tensors="pt").to(model.device)
        with torch.no_grad():
            output = model.generate(**inputs, max_new_tokens=32)

        
        output_reponse = str(processor.decode(output[0])).split('\n')[-1] 
        output_texts.append(output_reponse[:-11])   # without .<|eot_id|>
        
    print(output_texts)
    return output_texts
    

@spaces.GPU(duration=180)
def generate_image(category, img1, img2, img3, height, width, img_guidance_scale, inference_steps, seed, separate_cfg_infer, offload_model,
                   use_input_image_size_as_output, max_input_image_size, randomize_seed, guidance_scale=2.6):

    print()
    input_images = [img1, img2, img3]
    # Delete None
    input_images = [img for img in input_images if img is not None]
    if len(input_images) == 0:
        input_images = None

    wears = predict_clothing(input_images[1:])
    if len(wears)==1:
        dress = wears[0]
        text = f"""<img><|image_1|></img> Replace the {category}'s clothes with a {dress}. The {dress} is in <img><|image_2|></img>."""
    elif len(wears)==2:
        topwear, bottomwear = wears[0], wears[1]
        text = f"""<img><|image_1|></img> Replace the {category}'s topwear with a {topwear} and the bottomwear with a {bottomwear}.
                    The {topwear} is in <img><|image_2|></img>. The {bottomwear} is in <img><|image_3|></img>."""
    else:
        input_images = None

        
    if randomize_seed:
        seed = random.randint(0, 10000000)

    print(text)
    output = pipe(prompt=text, input_images=input_images, height=height, width=width, guidance_scale=guidance_scale,
                  img_guidance_scale=img_guidance_scale, num_inference_steps=inference_steps, separate_cfg_infer=separate_cfg_infer, 
                  use_kv_cache=True, offload_kv_cache=True, offload_model=offload_model, 
                  use_input_image_size_as_output=use_input_image_size_as_output, seed=seed, max_input_image_size=max_input_image_size,)
    img = output[0]
    return img

    
def get_example():
    case = [
            [   "./imgs/test_cases/icl1.jpg",
                "./imgs/test_cases/icl2.jpg",
                "./imgs/test_cases/icl3.jpg",
                224,
                224,
                1.6,
                1,
                768,
                False,
                False,
                2.5
            ],
    ]
    return case

def run_for_examples(img1, img2, img3, height, width, img_guidance_scale, seed, max_input_image_size, randomize_seed, use_input_image_size_as_output,):    
    # Check the internal configuration of the function
    inference_steps = 50
    separate_cfg_infer = True
    offload_model = False

    text = "According to the following examples, generate an output for the input.\nInput: <img><|image_1|></img>\nOutput: <img><|image_2|></img>\n\nInput: <img><|image_3|></img>\nOutput:"
    return generate_image(img1, img2, img3, height, width, img_guidance_scale, inference_steps, seed, separate_cfg_infer,
                          offload_model, use_input_image_size_as_output, max_input_image_size, randomize_seed)


description = """
### Usage:
- First upload your own image as the first image, also tagged 'Person'
- Then upload you 'Top-wear' and 'Bottom-wear' images
- If its a single dress, and/or you don't have a Topwear and Bottomwear as separate images upload that single image under 'Topwear'

### Tips:
- For out-of-memory or time cost, you can set `offload_model=True` or refer to [./docs/inference.md#requiremented-resources](https://github.com/VectorSpaceLab/OmniGen/blob/main/docs/inference.md#requiremented-resources) to select a appropriate setting.
- If inference time is too long when inputting multiple images, please try to reduce the `max_input_image_size`. For more details please refer to [./docs/inference.md#requiremented-resources](https://github.com/VectorSpaceLab/OmniGen/blob/main/docs/inference.md#requiremented-resources).

**Please note that HF Spaces often encounter errors due to GPU quota or other limitations, so please try lowering the image sizes and inference steps to manage the generation.**
"""

Credits = """**Credits**

Made using [OmniGen](https://huggingface.co/Shitao/OmniGen-v1): Unified Image Generation [paper](https://arxiv.org/abs/2409.11340) [code](https://github.com/VectorSpaceLab/OmniGen)
"""


# Gradio 
with gr.Blocks() as demo:
    gr.Markdown("# Virtual Try-On ✨")
    gr.Markdown(description)
    with gr.Row():
        with gr.Row(equal_height=True):
            # input images
            image_input_1 = gr.Image(label="Person", type="filepath")
            image_input_2 = gr.Image(label="Top-wear", type="filepath")
            image_input_3 = gr.Image(label="Bottom-wear", type="filepath")

    with gr.Row():
        with gr.Column():
            category = gr.Radio(["man", "woman", "boy", "girl"], label="Category", info="Choose one category from the following")
            
            # sliders            
            height_input = gr.Slider(label="Height", minimum=128, maximum=1024, value=768, step=16)
            width_input = gr.Slider(label="Width", minimum=128, maximum=1024, value=432, step=16)
    
            # guidance_scale_input = gr.Slider(label="Guidance Scale", minimum=1.0, maximum=5.0, value=2.5, step=0.1)
    
            num_inference_steps = gr.Slider(label="Inference Steps", minimum=1, maximum=128, value=32, step=1)
    
            seed_input = gr.Slider(label="Seed", minimum=0, maximum=2147483647, value=42, step=1)
            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)

        with gr.Column():
            max_input_image_size = gr.Slider(label="max_input_image_size", minimum=128, maximum=2048, value=768, step=16)
            
            img_guidance_scale_input = gr.Slider(label="img_guidance_scale", minimum=1.0, maximum=2.0, value=1.6, step=0.1)
            
            separate_cfg_infer = gr.Checkbox(
                label="separate_cfg_infer", info="Whether to use separate inference process for different guidance. This will reduce the memory cost.", value=True,)
            
            offload_model = gr.Checkbox(
                label="offload_model", info="Offload model to CPU, which will significantly reduce the memory cost but slow down the generation speed. You can cancel separate_cfg_infer and set offload_model=True. If both separate_cfg_infer and offload_model are True, further reduce the memory, but slowest generation", value=False,)
            
            use_input_image_size_as_output = gr.Checkbox(
                label="use_input_image_size_as_output", info="Automatically adjust the output image size to be same as input image size. For editing and controlnet task, it can make sure the output image has the same size as input image leading to better performance", value=False,)

    with gr.Row():
        # generate
        generate_button = gr.Button("Try On!")
            
    with gr.Row():
        # output image
        output_image = gr.Image(label="Output Image")

    # click
    generate_button.click(
        generate_image,
        inputs=[category, image_input_1, image_input_2, image_input_3, height_input, width_input, img_guidance_scale_input, num_inference_steps,
                seed_input, separate_cfg_infer, offload_model, use_input_image_size_as_output, max_input_image_size, randomize_seed,],
        outputs=output_image,
    )

    gr.Examples(
        examples=get_example(),
        fn=run_for_examples,
        inputs=[image_input_1, image_input_2, image_input_3, height_input, width_input, img_guidance_scale_input, seed_input,
                max_input_image_size, randomize_seed, use_input_image_size_as_output],
        outputs=output_image,
    )

    gr.Markdown(Credits)

# launch
demo.launch()