app.py

import gradio as gr
from transformers import PaliGemmaForConditionalGeneration, PaliGemmaProcessor
import spaces
import torch
import re

model = PaliGemmaForConditionalGeneration.from_pretrained("gokaygokay/sd3-long-captioner").to("cuda").eval()
processor = PaliGemmaProcessor.from_pretrained("gokaygokay/sd3-long-captioner")

def modify_caption(caption: str) -> str:
    """
    Removes specific prefixes from captions.
    Args:
        caption (str): A string containing a caption.
    Returns:
        str: The caption with the prefix removed if it was present.
    """
    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)

@spaces.GPU
def create_captions_rich(images):
    captions = []
    prompt = "caption en"
    
    for image in images:
        model_inputs = processor(text=prompt, images=image, return_tensors="pt").to("cuda")
        input_len = model_inputs["input_ids"].shape[-1]
        
        with torch.inference_mode():
            generation = model.generate(**model_inputs, max_new_tokens=256, do_sample=False)
            generation = generation[0][input_len:]
            decoded = processor.decode(generation, skip_special_tokens=True)

            modified_caption = modify_caption(decoded)
            captions.append(modified_caption)
    
    return captions

css = """
  #mkd {
    height: 500px; 
    overflow: auto; 
    border: 16px solid #ccc; 
  }
"""

with gr.Blocks(css=css) as demo:
    gr.HTML("<h1><center>Fine-tuned PaliGemma for SD3 Image Guided Prompt Generation.<center><h1>")
    
    with gr.Tab(label="Image to Prompt for SD3."):
        with gr.Row():
            with gr.Column():
                input_imgs = gr.Image(label="Input Images", type="pil", tool="editor", interactive=True, multiple=True)
                submit_btn = gr.Button(value="Start")
            outputs = gr.Text(label="Prompts", interactive=False)

        submit_btn.click(create_captions_rich, [input_imgs], [outputs])

demo.launch(debug=True)