app.py

import spaces
import gradio as gr
import torch
from PIL import Image
from transformers import PaliGemmaForConditionalGeneration, PaliGemmaProcessor, pipeline
from transformers import AutoProcessor, AutoModelForCausalLM
import re
import random
import os
from huggingface_hub import snapshot_download
from kolors.pipelines.pipeline_stable_diffusion_xl_chatglm_256 import StableDiffusionXLPipeline
from kolors.models.modeling_chatglm import ChatGLMModel
from kolors.models.tokenization_chatglm import ChatGLMTokenizer
from diffusers import UNet2DConditionModel, AutoencoderKL
from diffusers import EulerDiscreteScheduler
import subprocess

# Initialize models
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16

# Download Kolors model
ckpt_dir = snapshot_download(repo_id="Kwai-Kolors/Kolors")

import subprocess
subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)

# Load Kolors models
text_encoder = ChatGLMModel.from_pretrained(os.path.join(ckpt_dir, 'text_encoder'), torch_dtype=dtype).to(device)
tokenizer = ChatGLMTokenizer.from_pretrained(os.path.join(ckpt_dir, 'text_encoder'))
vae = AutoencoderKL.from_pretrained(os.path.join(ckpt_dir, "vae"), revision=None).to(dtype).to(device)
scheduler = EulerDiscreteScheduler.from_pretrained(os.path.join(ckpt_dir, "scheduler"))
unet = UNet2DConditionModel.from_pretrained(os.path.join(ckpt_dir, "unet"), revision=None).to(dtype).to(device)

kolors_pipe = StableDiffusionXLPipeline(
    vae=vae,
    text_encoder=text_encoder,
    tokenizer=tokenizer,
    unet=unet,
    scheduler=scheduler,
    force_zeros_for_empty_prompt=False
).to(device)

# VLM Captioner
vlm_model = PaliGemmaForConditionalGeneration.from_pretrained("gokaygokay/sd3-long-captioner-v2").to(device).eval()
vlm_processor = PaliGemmaProcessor.from_pretrained("gokaygokay/sd3-long-captioner-v2")

# Initialize Florence model
florence_model = AutoModelForCausalLM.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True).to(device).eval()
florence_processor = AutoProcessor.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True)

# Prompt Enhancer
enhancer_medium = pipeline("summarization", model="gokaygokay/Lamini-Prompt-Enchance", device=device)
enhancer_long = pipeline("summarization", model="gokaygokay/Lamini-Prompt-Enchance-Long", device=device)

MAX_SEED = 2**32 - 1

# Florence caption function
def florence_caption(image):
    # Convert image to PIL if it's not already
    if not isinstance(image, Image.Image):
        image = Image.fromarray(image)
    
    inputs = florence_processor(text="<MORE_DETAILED_CAPTION>", images=image, return_tensors="pt").to(device)
    generated_ids = florence_model.generate(
        input_ids=inputs["input_ids"],
        pixel_values=inputs["pixel_values"],
        max_new_tokens=1024,
        early_stopping=False,
        do_sample=False,
        num_beams=3,
    )
    generated_text = florence_processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
    parsed_answer = florence_processor.post_process_generation(
        generated_text,
        task="<MORE_DETAILED_CAPTION>",
        image_size=(image.width, image.height)
    )
    return parsed_answer["<MORE_DETAILED_CAPTION>"]

# VLM Captioner function
def create_captions_rich(image):
    prompt = "caption en"
    model_inputs = vlm_processor(text=prompt, images=image, return_tensors="pt").to(device)
    input_len = model_inputs["input_ids"].shape[-1]

    with torch.inference_mode():
        generation = vlm_model.generate(**model_inputs, repetition_penalty=1.10, max_new_tokens=256, do_sample=False)
        generation = generation[0][input_len:]
        decoded = vlm_processor.decode(generation, skip_special_tokens=True)

    return modify_caption(decoded)

# Helper function for caption modification
def modify_caption(caption: str) -> str:
    prefix_substrings = [
        ('captured from ', ''),
        ('captured at ', '')
    ]
    pattern = '|'.join([re.escape(opening) for opening, _ in prefix_substrings])
    replacers = {opening: replacer for opening, replacer in prefix_substrings}
    
    def replace_fn(match):
        return replacers[match.group(0)]
    
    return re.sub(pattern, replace_fn, caption, count=1, flags=re.IGNORECASE)

# Prompt Enhancer function
def enhance_prompt(input_prompt, model_choice):
    if model_choice == "Medium":
        result = enhancer_medium("Enhance the description: " + input_prompt)
        enhanced_text = result[0]['summary_text']
        
        pattern = r'^.*?of\s+(.*?(?:\.|$))'
        match = re.match(pattern, enhanced_text, re.IGNORECASE | re.DOTALL)
        
        if match:
            remaining_text = enhanced_text[match.end():].strip()
            modified_sentence = match.group(1).capitalize()
            enhanced_text = modified_sentence + ' ' + remaining_text
    else:  # Long
        result = enhancer_long("Enhance the description: " + input_prompt)
        enhanced_text = result[0]['summary_text']
    
    return enhanced_text

def generate_image(prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps, num_images_per_prompt):
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    
    generator = torch.Generator(device=device).manual_seed(seed)
    
    image = kolors_pipe(
        prompt=prompt, 
        negative_prompt=negative_prompt,
        guidance_scale=guidance_scale, 
        num_inference_steps=num_inference_steps, 
        width=width, 
        height=height,
        num_images_per_prompt=num_images_per_prompt,
        generator=generator
    ).images
    
    return image, seed

@spaces.GPU(duration=200)
def process_workflow(image, text_prompt, vlm_model_choice, use_enhancer, model_choice, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps, num_images_per_prompt):
    if image is not None:
        # Convert image to PIL if it's not already
        if not isinstance(image, Image.Image):
            image = Image.fromarray(image)
        
        if vlm_model_choice == "Long Captioner":
            prompt = create_captions_rich(image)
        else:  # Florence
            prompt = florence_caption(image)
    else:
        prompt = text_prompt
    
    if use_enhancer:
        prompt = enhance_prompt(prompt, model_choice)
    
    generated_image, used_seed = generate_image(prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps, num_images_per_prompt)
    
    return generated_image, prompt, used_seed

custom_css = """
.input-group, .output-group {
    border: 1px solid #e0e0e0;
    border-radius: 10px;
    padding: 20px;
    margin-bottom: 20px;
    background-color: #f9f9f9;
}
.submit-btn {
    background-color: #2980b9 !important;
    color: white !important;
}
.submit-btn:hover {
    background-color: #3498db !important;
}
"""

title = """<h1 align="center">Kolors with VLM Captioner and Prompt Enhancer</h1>
<p><center>
<a href="https://huggingface.co/Kwai-Kolors/Kolors" target="_blank">[Kolors Model]</a>
<a href="https://huggingface.co/microsoft/Florence-2-base" target="_blank">[Florence-2 Model]</a>
<a href="https://huggingface.co/gokaygokay/sd3-long-captioner-v2" target="_blank">[Long Captioner Model]</a>
<a href="https://huggingface.co/gokaygokay/Lamini-Prompt-Enchance-Long" target="_blank">[Prompt Enhancer Long]</a>
<a href="https://huggingface.co/gokaygokay/Lamini-Prompt-Enchance" target="_blank">[Prompt Enhancer Medium]</a>

<p align="center">Create long prompts from images or enhance your short prompts with prompt enhancer</p>
</center></p>
"""

with gr.Blocks(css=custom_css, theme=gr.themes.Soft(primary_hue="blue", secondary_hue="gray")) as demo:
    gr.HTML(title)
    
    with gr.Row():
        with gr.Column(scale=1):
            with gr.Group(elem_classes="input-group"):
                input_image = gr.Image(label="Input Image (VLM Captioner)")
                vlm_model_choice = gr.Radio(["Florence-2", "Long Captioner"], label="VLM Model", value="Florence-2")
            
            with gr.Accordion("Advanced Settings", open=False):
                text_prompt = gr.Textbox(label="Text Prompt (optional, used if no image is uploaded)")
                use_enhancer = gr.Checkbox(label="Use Prompt Enhancer", value=False)
                model_choice = gr.Radio(["Medium", "Long"], label="Enhancer Model", value="Long")
                negative_prompt = gr.Textbox(label="Negative Prompt")
                seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0)
                randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
                width = gr.Slider(label="Width", minimum=512, maximum=2048, step=64, value=1024)
                height = gr.Slider(label="Height", minimum=512, maximum=2048, step=64, value=1024)
                guidance_scale = gr.Slider(label="Guidance Scale", minimum=1.0, maximum=20.0, step=0.5, value=5.0)
                num_inference_steps = gr.Slider(label="Inference Steps", minimum=20, maximum=50, step=1, value=20)
                num_images_per_prompt = gr.Slider(1, 4, 1, step=1, label="Number of images per prompt")
            
            generate_btn = gr.Button("Generate Image", elem_classes="submit-btn")
        
        with gr.Column(scale=1):
            with gr.Group(elem_classes="output-group"):
                output_image = gr.Gallery(label="Result", elem_id="gallery", show_label=False)
                final_prompt = gr.Textbox(label="Final Prompt Used")
                used_seed = gr.Number(label="Seed Used")
    
    generate_btn.click(
        fn=process_workflow,
        inputs=[
            input_image, text_prompt, vlm_model_choice, use_enhancer, model_choice,
            negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps,
            num_images_per_prompt
        ],
        outputs=[output_image, final_prompt, used_seed]
    )

demo.launch(debug=True)