This repository contains the baseline model used in LLaVA-Scissor. This model is an enhanced version of LLaVA-OneVision model with SIGLIP vision encoder and Qwen2.5-0.5B-Instruct large language model and is finetuned with Oryx data.

Quick Start

Here we provide a script for LLaVA-Scissor full token inference (without token compression).

python

from operator import attrgetter
from llava.model.builder import load_pretrained_model
from llava.mm_utils import get_model_name_from_path, process_images, tokenizer_image_token
from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, IGNORE_INDEX
from llava.conversation import conv_templates, SeparatorStyle

import torch
import cv2
import numpy as np
from PIL import Image
import requests
import copy
import warnings
from decord import VideoReader, cpu

warnings.filterwarnings("ignore")
# Load the OneVision model
pretrained = "model_zoo/BBBBCHAN/LLaVA-Scissor-baseline-0.5B"
model_name = "llava_qwen"
device = "cuda"
device_map = "auto"
tokenizer, model, image_processor, max_length = load_pretrained_model(pretrained, None, model_name, device_map=device_map, attn_implementation="sdpa")

model.eval()


# Function to extract frames from video
def load_video(video_path, max_frames_num):
    if type(video_path) == str:
        vr = VideoReader(video_path, ctx=cpu(0))
    else:
        vr = VideoReader(video_path[0], ctx=cpu(0))
    total_frame_num = len(vr)
    uniform_sampled_frames = np.linspace(0, total_frame_num - 1, max_frames_num, dtype=int)
    frame_idx = uniform_sampled_frames.tolist()
    spare_frames = vr.get_batch(frame_idx).asnumpy()
    return spare_frames  # (frames, height, width, channels)


# Load and process video
video_path = "Your/path/to/the/video"
video_frames = load_video(video_path, 16)
print(video_frames.shape)
image_tensors = []
frames = image_processor.preprocess(video_frames, return_tensors="pt")["pixel_values"].half().cuda()
image_tensors.append(frames)

# Prepare conversation input
conv_template = "qwen_2"
question = f"{DEFAULT_IMAGE_TOKEN}
Describe this video."
conv = copy.deepcopy(conv_templates[conv_template])
conv.append_message(conv.roles[0], question)
conv.append_message(conv.roles[1], None)
prompt_question = conv.get_prompt()

input_ids = tokenizer_image_token(prompt_question, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(device)
image_sizes = [frame.size for frame in video_frames]

# Generate response
cont = model.generate(
    input_ids,
    images=image_tensors,
    image_sizes=image_sizes,
    do_sample=False,
    temperature=0,
    max_new_tokens=4096,
    modalities=["video"],
)
text_outputs = tokenizer.batch_decode(cont, skip_special_tokens=True)
print(text_outputs[0])

Citation

If you find our repo useful for your research, please consider citing our paper:

bibtex

@article{sun2025llava,
  title={LLaVA-Scissor: Token Compression with Semantic Connected Components for Video LLMs},
  author={Sun, Boyuan and Zhao, Jiaxing and Wei, Xihan and Hou, Qibin},
  journal={arXiv preprint arXiv:2506.21862},
  year={2025}
}