app.py

import os
import torch
import gradio as gr
from PIL import Image
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import spaces

# Configuration
MODEL_ID = "Qwen/Qwen2.5-VL-3B-Instruct"
device = "cuda" if torch.cuda.is_available() else "cpu"

# Load Processor
processor = AutoProcessor.from_pretrained(MODEL_ID)

# Load Model
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    MODEL_ID, 
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
).eval()
print("Model loaded.")

@spaces.GPU
def process_images(image_files, instruction):
    """
    Process a batch of images sequentially.
    Yields the updated results list as each image is processed.
    """
    if not image_files:
        yield "No images uploaded."
        return

    results = []
    
    for idx, img_file in enumerate(image_files):
        try:
            # We assume it is a path to the file passed from gradio
            img_path = img_file.name if hasattr(img_file, 'name') else img_file
            
            # Use Qwen-VL specific conversational format
            messages = [
                {
                    "role": "user",
                    "content": [
                        {"type": "image", "image": img_path},
                        {"type": "text", "text": instruction},
                    ],
                }
            ]
            
            # Preparation for inference
            text = processor.apply_chat_template(
                messages, tokenize=False, add_generation_prompt=True
            )
            image_inputs, video_inputs = process_vision_info(messages)
            
            inputs = processor(
                text=[text],
                images=image_inputs,
                videos=video_inputs,
                padding=True,
                return_tensors="pt",
            )
            # Move inputs to the same device as the model
            inputs = inputs.to(model.device)
            
            # Generate output
            generated_ids = model.generate(**inputs, max_new_tokens=256)
            
            # Trim the generated ids to only contain the new tokens
            generated_ids_trimmed = [
                out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
            ]
            
            output_text = processor.batch_decode(
                generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
            )[0]
            
            results.append(f"### Image {idx + 1}\n**Caption:** {output_text}\n")
            
            # Yield accumulated results so user sees progress
            yield "\n---\n".join(results)
            
        except Exception as e:
            results.append(f"### Image {idx + 1}\n**Error processing image:** {str(e)}\n")
            yield "\n---\n".join(results)

# Gradio Interface Construction
with gr.Blocks(title="Batch Image Captioning") as demo:
    gr.Markdown("# 🖼️ Batch Image Captioning with Qwen2.5-VL")
    gr.Markdown(
        "Upload multiple images and provide an instruction prompt. The system uses "
        "[Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct) "
        "to generate descriptions sequentially. Designed to run smoothly on Hugging Face ZeroGPU."
    )
    
    with gr.Row():
        with gr.Column(scale=1):
            input_images = gr.File(
                label="Upload Images", 
                file_count="multiple", 
                file_types=["image"], 
                type="filepath" # returns temp paths
            )
            
            # Default instruction panel
            instruction_textbox = gr.Textbox(
                label="Instructions", 
                placeholder="Describe this image in detail...",
                value="Provide a detailed, highly descriptive caption for this image focusing on lighting, composition, and subjects.",
                lines=3
            )
            
            submit_btn = gr.Button("Generate Captions", variant="primary")
            
        with gr.Column(scale=1):
            output_text = gr.Markdown("Captions will appear here...", label="Results")

    submit_btn.click(
        fn=process_images,
        inputs=[input_images, instruction_textbox],
        outputs=output_text
    )

if __name__ == "__main__":
    demo.launch()