app.py

# app.py
import gradio as gr
from src.model_loader import load_model
from src.video_utils import process_video_for_internvl3
from src.ar_prompts import generate_conversation_questions

tokenizer, model = load_model()

def evaluate_ar_multi_turn(video):
    pixel_values, num_patches_list, image_prefix = process_video_for_internvl3(video)
    conversation = generate_conversation_questions(include_descriptions=True)

    history = None
    visible_outputs = []

    for i, question in enumerate(conversation):
        prompt = image_prefix + question if i == 0 else question

        output, history = model.chat(
            tokenizer,
            pixel_values,
            prompt,
            generation_config={"max_new_tokens": 1024},
            num_patches_list=num_patches_list,
            history=history,
            return_history=True
        )

        # 仅保留评测和拓展部分的回答（即从第3轮开始）
        if i >= 2:
            visible_outputs.append(output)

    # 多个输出拼接成文本显示
    return "\n\n".join(visible_outputs)

gr.Interface(
    fn=evaluate_ar_multi_turn,
    inputs=gr.Video(label="Upload your AR video"),
    outputs="text",
    title="InternVL3 AR Evaluation (Multi-turn)",
    description="Upload a short AR video clip. The model will sample frames and conduct a multi-turn dialogue to assess occlusion/rendering/placement/lighting."
).launch()