VideoMind-2B

VideoMind is a multi-modal agent framework that enhances video reasoning by emulating human-like processes, such as breaking down tasks, localizing and verifying moments, and synthesizing answers.

🔖 Model Details

• Model type: Multi-modal Large Language Model
• Language(s): English
• License: BSD-3-Clause

🚀 Quick Start

Install the environment

1. Clone the repository from GitHub.

git clone git@github.com:yeliudev/VideoMind.git
cd VideoMind

2. Initialize conda environment.

conda create -n videomind python=3.11 -y
conda activate videomind

3. Install dependencies.

pip install -r requirements.txt

For NPU users, please modify Line 18-25 of requirements.txt.

Quick Inference Demo

The script below showcases how to perform inference with VideoMind's different roles. Please refer to our GitHub Repository for more details about this model.

import torch

from videomind.constants import GROUNDER_PROMPT, PLANNER_PROMPT, VERIFIER_PROMPT
from videomind.dataset.utils import process_vision_info
from videomind.model.builder import build_model
from videomind.utils.io import get_duration
from videomind.utils.parser import parse_span

MODEL_PATH = 'yeliudev/VideoMind-2B'

video_path = '<path-to-video>'
question = '<question>'

# initialize role *grounder*
model, processor = build_model(MODEL_PATH)
device = next(model.parameters()).device

# initialize role *planner*
model.load_adapter(f'{MODEL_PATH}/planner', adapter_name='planner')

# initialize role *verifier*
model.load_adapter(f'{MODEL_PATH}/verifier', adapter_name='verifier')

# ==================== Planner ====================

messages = [{
    'role':
    'user',
    'content': [{
        'type': 'video',
        'video': video_path,
        'min_pixels': 36 * 28 * 28,
        'max_pixels': 64 * 28 * 28,
        'max_frames': 100,
        'fps': 1.0
    }, {
        'type': 'text',
        'text': PLANNER_PROMPT.format(question)
    }]
}]

# preprocess inputs
text = processor.apply_chat_template(messages, add_generation_prompt=True)
images, videos = process_vision_info(messages)
data = processor(text=[text], images=images, videos=videos, return_tensors='pt').to(device)

# switch adapter to *planner*
model.base_model.disable_adapter_layers()
model.base_model.enable_adapter_layers()
model.set_adapter('planner')

# run inference
output_ids = model.generate(**data, do_sample=False, temperature=None, top_p=None, top_k=None, max_new_tokens=256)

# decode output ids
output_ids = output_ids[0, data.input_ids.size(1):-1]
response = processor.decode(output_ids, clean_up_tokenization_spaces=False)

print(f'Planner Response: {response}')

# ==================== Grounder ====================

messages = [{
    'role':
    'user',
    'content': [{
        'type': 'video',
        'video': video_path,
        'min_pixels': 36 * 28 * 28,
        'max_pixels': 64 * 28 * 28,
        'max_frames': 150,
        'fps': 1.0
    }, {
        'type': 'text',
        'text': GROUNDER_PROMPT.format(question)
    }]
}]

# preprocess inputs
text = processor.apply_chat_template(messages, add_generation_prompt=True)
images, videos = process_vision_info(messages)
data = processor(text=[text], images=images, videos=videos, return_tensors='pt').to(device)

# switch adapter to *grounder*
model.base_model.disable_adapter_layers()
model.base_model.enable_adapter_layers()
model.set_adapter('grounder')

# run inference
output_ids = model.generate(**data, do_sample=False, temperature=None, top_p=None, top_k=None, max_new_tokens=256)

# decode output ids
output_ids = output_ids[0, data.input_ids.size(1):-1]
response = processor.decode(output_ids, clean_up_tokenization_spaces=False)

print(f'Grounder Response: {response}')

duration = get_duration(video_path)

# 1. extract timestamps and confidences
blob = model.reg[0].cpu().float()
pred, conf = blob[:, :2] * duration, blob[:, -1].tolist()

# 2. clamp timestamps
pred = pred.clamp(min=0, max=duration)

# 3. sort timestamps
inds = (pred[:, 1] - pred[:, 0] < 0).nonzero()[:, 0]
pred[inds] = pred[inds].roll(1)

# 4. convert timestamps to list
pred = pred.tolist()

print(f'Grounder Regressed Timestamps: {pred}')

# ==================== Verifier ====================

# using top-5 predictions
probs = []
for cand in pred[:5]:
    s0, e0 = parse_span(cand, duration, 2)
    offset = (e0 - s0) / 2
    s1, e1 = parse_span([s0 - offset, e0 + offset], duration)

    # percentage of s0, e0 within s1, e1
    s = (s0 - s1) / (e1 - s1)
    e = (e0 - s1) / (e1 - s1)

    messages = [{
        'role':
        'user',
        'content': [{
            'type': 'video',
            'video': video_path,
            'video_start': s1,
            'video_end': e1,
            'min_pixels': 36 * 28 * 28,
            'max_pixels': 64 * 28 * 28,
            'max_frames': 64,
            'fps': 2.0
        }, {
            'type': 'text',
            'text': VERIFIER_PROMPT.format(question)
        }]
    }]

    text = processor.apply_chat_template(messages, add_generation_prompt=True)
    images, videos = process_vision_info(messages)
    data = processor(text=[text], images=images, videos=videos, return_tensors='pt')

    # ===== insert segment start/end tokens =====
    video_grid_thw = data['video_grid_thw'][0]
    num_frames, window = int(video_grid_thw[0]), int(video_grid_thw[1] * video_grid_thw[2] / 4)
    assert num_frames * window * 4 == data['pixel_values_videos'].size(0)

    pos_s, pos_e = round(s * num_frames), round(e * num_frames)
    pos_s, pos_e = min(max(0, pos_s), num_frames), min(max(0, pos_e), num_frames)
    assert pos_s <= pos_e, (num_frames, s, e)

    base_idx = torch.nonzero(data['input_ids'][0] == model.config.vision_start_token_id).item()
    pos_s, pos_e = pos_s * window + base_idx + 1, pos_e * window + base_idx + 2

    input_ids = data['input_ids'][0].tolist()
    input_ids.insert(pos_s, model.config.seg_s_token_id)
    input_ids.insert(pos_e, model.config.seg_e_token_id)
    data['input_ids'] = torch.LongTensor([input_ids])
    data['attention_mask'] = torch.ones_like(data['input_ids'])
    # ===========================================

    data = data.to(device)

    # switch adapter to *verifier*
    model.base_model.disable_adapter_layers()
    model.base_model.enable_adapter_layers()
    model.set_adapter('verifier')

    # run inference
    with torch.inference_mode():
        logits = model(**data).logits[0, -1].softmax(dim=-1)

    # NOTE: magic numbers here
    # In Qwen2-VL vocab: 9454 -> Yes, 2753 -> No
    score = (logits[9454] - logits[2753]).sigmoid().item()
    probs.append(score)

# sort predictions by verifier's confidence
ranks = torch.Tensor(probs).argsort(descending=True).tolist()

pred = [pred[idx] for idx in ranks]
conf = [conf[idx] for idx in ranks]

print(f'Verifier Re-ranked Timestamps: {pred}')

# ==================== Answerer ====================

# select the best candidate moment
s, e = parse_span(pred[0], duration, 32)

messages = [{
    'role':
    'user',
    'content': [{
        'type': 'video',
        'video': video_path,
        'video_start': s,
        'video_end': e,
        'min_pixels': 128 * 28 * 28,
        'max_pixels': 256 * 28 * 28,
        'max_frames': 32,
        'fps': 2.0
    }, {
        'type': 'text',
        'text': question
    }]
}]

text = processor.apply_chat_template(messages, add_generation_prompt=True)
images, videos = process_vision_info(messages)
data = processor(text=[text], images=images, videos=videos, return_tensors='pt').to(device)

# remove all adapters as *answerer* is the base model itself
with model.disable_adapter():
    output_ids = model.generate(**data, do_sample=False, temperature=None, top_p=None, top_k=None, max_new_tokens=256)

# decode output ids
output_ids = output_ids[0, data.input_ids.size(1):-1]
response = processor.decode(output_ids, clean_up_tokenization_spaces=False)

print(f'Answerer Response: {response}')

📖 Citation

Please kindly cite our paper if you find this project helpful.

@article{liu2025videomind,
  title={VideoMind: A Chain-of-LoRA Agent for Long Video Reasoning},
  author={Liu, Ye and Lin, Kevin Qinghong and Chen, Chang Wen and Shou, Mike Zheng},
  journal={arXiv preprint arXiv:2503.13444},
  year={2025}
}