yiyexy's picture
Model card: put badges on one line
7ce9a7e verified
|
Raw History Blame Contribute Delete
14.9 kB
---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
tags:
- multimodal
- video-understanding
- streaming
- llava-onevision
---
<div align="center">
# LLaVA-OneVision-2.5-4B-Instruct
<p align="center"><a href="https://github.com/EvolvingLMMs-Lab/LLaVA-OneVision-2.5"><img style="display:inline-block; margin:0 2px; vertical-align:middle;" alt="GitHub" src="https://img.shields.io/badge/GitHub-Code-181717?logo=github"></a> <img style="display:inline-block; margin:0 2px; vertical-align:middle;" alt="Project page: coming soon" src="https://img.shields.io/badge/Project%20Page-Coming%20soon-2a78d6"> <img style="display:inline-block; margin:0 2px; vertical-align:middle;" alt="Technical report: coming soon" src="https://img.shields.io/badge/Technical%20Report-Coming%20soon-b31b1b"> <img style="display:inline-block; margin:0 2px; vertical-align:middle;" alt="License" src="https://img.shields.io/badge/License-Apache--2.0-1baf7a"></p>
**English** | [中文](README_CN.md)
</div>
LLaVA-OneVision-2.5-4B-Instruct is a 4B vision-language model for single-image, multi-image and video input. Beyond the usual offline tasks, it is trained for **streaming** video: frames arrive one at a time, and at every step the model decides whether to stay silent or speak. The same weights and the same chat template serve both modes.
## Highlights
- **Offline and streaming in one model.** `apply_chat_template(..., streaming=True)` switches mode; eight protocol tokens carry the per-frame decision to stay silent, respond or caption. Nothing else about the input changes.
- **Codec-aligned video.** Besides uniform frame sampling, *codec mode* uses the encoder's bit-cost to pack the informative regions of a video into a handful of canvases.
- **9× spatial compression.** A patch-16 vision tower with 3×3 merging gives one visual token per 48×48 pixels, so long videos fit comfortably in the 262k context.
## Performance
### Video
<table style="width:100%; display:table;">
<thead><tr>
<th align="left" style="width:28%">Benchmark</th>
<th align="center" style="width:24%">OV2.5-4B (codec)</th>
<th align="center" style="width:24%">OV2.5-4B (frames)</th>
<th align="center" style="width:24%">Qwen3.5-4B</th>
</tr></thead>
<tbody>
<tr><td colspan="4" align="left"><b>General video understanding</b></td></tr>
<tr><td align="left">VideoMME (no subtitles)</td><td align="center"><b>71.9</b></td><td align="center">69.7</td><td align="center">71.7</td></tr>
<tr><td align="left">VideoMME (subtitles)</td><td align="center">74.7</td><td align="center">75.2</td><td align="center"><b>75.9</b></td></tr>
<tr><td align="left">MotionBench</td><td align="center"><b>64.2</b> <sup>TC128×313,600</sup></td><td align="center">62.4 <sup>128f×153,664</sup></td><td align="center">60.1 <sup>128f×153,664</sup></td></tr>
<tr><td align="left">TOMATO</td><td align="center">36.0 <sup>TC16×153,664</sup></td><td align="center"><b>39.2</b> <sup>16f×153,664</sup></td><td align="center">32.4 <sup>16f×153,664</sup></td></tr>
<tr><td align="left">MVBench</td><td align="center"><b>75.5</b> <sup>TC64×313,600</sup></td><td align="center">72.0 <sup>64f×153,664</sup></td><td align="center">64.0 <sup>64f×153,664</sup></td></tr>
<tr><td align="left">NextQA MC test</td><td align="center"><b>86.4</b> <sup>TC64×313,600</sup></td><td align="center">86.1 <sup>64f×153,664</sup></td><td align="center">80.7 <sup>64f×153,664</sup></td></tr>
<tr><td align="left">TempCompass MC</td><td align="center"><b>79.7</b> <sup>TC64×313,600</sup></td><td align="center">75.8 <sup>64f×153,664</sup></td><td align="center">73.4 <sup>64f×153,664</sup></td></tr>
<tr><td align="left">Video-MME-v2</td><td align="center"><b>17.2</b> <sup>TC64×313,600</sup></td><td align="center">16.4 <sup>64f×330,000</sup></td><td align="center">15.5 <sup>64f×330,000</sup></td></tr>
<tr><td align="left">EgoSchema subset</td><td align="center">79.8 <sup>TC64×313,600</sup></td><td align="center"><b>80.0</b> <sup>64f×153,664</sup></td><td align="center">68.2 <sup>64f×153,664</sup></td></tr>
<tr><td align="left">Video-Index</td><td align="center"><b>23.1</b> <sup>TC512×55,296</sup></td><td align="center">20.0 <sup>1fps, ≤512f×55,296</sup></td><td align="center">19.6 <sup>1fps, ≤512f×55,296</sup></td></tr>
<tr><td colspan="4" align="left"><b>Long video</b></td></tr>
<tr><td align="left">LVBench</td><td align="center">61.6 <sup>TC1024×156,800</sup></td><td align="center"><b>62.0</b> <sup>950f×200,000</sup></td><td align="center">56.5 <sup>2048f×120,000</sup></td></tr>
<tr><td align="left">MLVU dev</td><td align="center"><b>77.1</b> <sup>TC512×313,600</sup></td><td align="center">75.3 <sup>512f×72,900</sup></td><td align="center">73.8 <sup>512f×72,900</sup></td></tr>
<tr><td align="left">LongVideoBench val</td><td align="center"><b>71.1</b> <sup>TC768×235,200</sup></td><td align="center">69.0 <sup>1280f×80,000</sup></td><td align="center">62.7 <sup>700f×300,000</sup></td></tr>
<tr><td align="left">VideoEval-Pro</td><td align="center"><b>64.6</b> <sup>TC768×235,200</sup></td><td align="center">64.4 <sup>950f×200,000</sup></td><td align="center">62.8 <sup>2048f×100,000</sup></td></tr>
<tr><td colspan="4" align="left"><b>Temporal grounding (mIoU)</b></td></tr>
<tr><td align="left">TimeLens ActivityNet</td><td align="center"><b>59.1</b> <sup>TC128×313,600</sup></td><td align="center">58.8 <sup>2048f×409,600, fps=4</sup></td><td align="center">53.8 <sup>2048f×409,600, fps=4</sup></td></tr>
<tr><td align="left">TimeLens Charades</td><td align="center"><b>61.3</b> <sup>TC128×313,600</sup></td><td align="center">60.6 <sup>2048f×409,600, fps=4</sup></td><td align="center">48.1 <sup>2048f×409,600, fps=4</sup></td></tr>
<tr><td align="left">TimeLens QVHighlights</td><td align="center"><b>70.3</b> <sup>TC128×313,600</sup></td><td align="center">68.4 <sup>2048f×409,600, fps=4</sup></td><td align="center">67.5 <sup>2048f×409,600, fps=4</sup></td></tr>
</tbody>
</table>
### Streaming video
<table style="width:100%; display:table;">
<thead><tr>
<th align="left" style="width:40%">Benchmark</th>
<th align="center" style="width:20%">OV2.5-4B<br>(codec, 24 fps)</th>
<th align="center" style="width:20%">MOSS-VL-Realtime 11B<br>(1 fps)</th>
<th align="center" style="width:20%">JoyAI-VL-Interaction 8B<br>(1 fps)</th>
</tr></thead>
<tbody>
<tr><td colspan="4" align="left"><b>OVO-Bench</b></td></tr>
<tr><td align="left">Real-Time Visual Perception (RTVP)</td><td align="center"><b>81.0</b></td><td align="center">75.9</td><td align="center">76.4</td></tr>
<tr><td align="left">Backward Tracing (BT)</td><td align="center">68.8</td><td align="center"><b>72.6</b></td><td align="center">53.3</td></tr>
<tr><td align="left">Forward Active Responding (FAR)</td><td align="center">60.7</td><td align="center"><b>62.1</b></td><td align="center">54.6</td></tr>
<tr><td align="left">Overall</td><td align="center"><b>70.2</b></td><td align="center"><b>70.2</b></td><td align="center">61.4</td></tr>
<tr><td colspan="4" align="left"><b>StreamingBench</b></td></tr>
<tr><td align="left">Real-Time Visual Understanding (RTVU)</td><td align="center"><b>85.3</b></td><td align="center">82.9</td><td align="center">83.2</td></tr>
<tr><td align="left">Misleading Context Understanding (MCU)</td><td align="center">36.8</td><td align="center">--</td><td align="center"><b>37.2</b></td></tr>
<tr><td align="left">Anomaly Context Understanding (ACU)</td><td align="center"><b>67.6</b></td><td align="center">--</td><td align="center">51.6</td></tr>
<tr><td align="left">Sequential Question Answering (SQA)</td><td align="center"><b>54.8</b></td><td align="center">50.4</td><td align="center">53.6</td></tr>
<tr><td align="left">Proactive Output (PO)</td><td align="center"><b>63.6</b></td><td align="center">60.0</td><td align="center">39.6</td></tr>
<tr><td align="left">Contextual Understanding (CU)</td><td align="center">55.7</td><td align="center"><b>56.4</b></td><td align="center">45.5</td></tr>
<tr><td colspan="4" align="left"><b>Proactive and interactive</b></td></tr>
<tr><td align="left">OmniMMI: Proactive Alerting (PA)</td><td align="center"><b>68.5</b></td><td align="center">66.0</td><td align="center">13.5</td></tr>
<tr><td align="left">ProactiveVideoQA: PAUC (ω=0.5)</td><td align="center">33.1</td><td align="center"><b>47.2</b></td><td align="center">37.0</td></tr>
<tr><td align="left">InteractionBench: Total</td><td align="center">38.6</td><td align="center">--</td><td align="center">--</td></tr>
<tr><td align="left">StreamArena: Proactive Interaction (Pro)</td><td align="center"><b>25.5</b></td><td align="center">11.8</td><td align="center">9.4</td></tr>
</tbody>
</table>
### Image
<table style="width:100%; display:table;">
<thead><tr>
<th align="left" style="width:24%">General perception and reading</th>
<th align="center" style="width:13%">OV2.5-4B</th>
<th align="center" style="width:13%">Qwen3.5-4B</th>
<th align="left" style="width:24%">Grounding and spatial</th>
<th align="center" style="width:13%">OV2.5-4B</th>
<th align="center" style="width:13%">Qwen3.5-4B</th>
</tr></thead>
<tbody>
<tr><td align="left">MMStar</td><td align="center">63.9</td><td align="center"><b>66.9</b></td><td align="left">RefCOCO val (Acc@0.5)</td><td align="center"><b>92.6</b></td><td align="center">89.1</td></tr>
<tr><td align="left">MMBench-EN dev</td><td align="center">88.2</td><td align="center"><b>88.8</b></td><td align="left">RefCOCOg val (Acc@0.5)</td><td align="center"><b>88.9</b></td><td align="center">87.7</td></tr>
<tr><td align="left">RealWorldQA</td><td align="center">73.9</td><td align="center"><b>74.4</b></td><td align="left">RefSpatial-Bench</td><td align="center"><b>62.1</b></td><td align="center">29.1</td></tr>
<tr><td align="left">AI2D</td><td align="center">82.2</td><td align="center"><b>82.4</b></td><td align="left">CrossPoint</td><td align="center"><b>80.5</b></td><td align="center">33.7</td></tr>
<tr><td align="left">BLINK</td><td align="center"><b>67.1</b></td><td align="center">63.6</td><td align="left">CV-Bench 2D</td><td align="center"><b>83.3</b></td><td align="center">80.7</td></tr>
<tr><td align="left">V*-Bench</td><td align="center"><b>89.5</b></td><td align="center">81.7</td><td align="left">CV-Bench 3D</td><td align="center"><b>94.5</b></td><td align="center">92.0</td></tr>
<tr><td align="left">OCRBench</td><td align="center">79.8</td><td align="center"><b>85.2</b></td><td align="left">EmbSpatial</td><td align="center"><b>83.0</b></td><td align="center">74.7</td></tr>
<tr><td align="left">DocVQA val</td><td align="center">92.9</td><td align="center"><b>95.2</b></td><td align="left">MMSI-Bench</td><td align="center"><b>33.3</b></td><td align="center">31.4</td></tr>
</tbody>
</table>
<sub>Compared with Qwen3.5-4B, thinking disabled. **Bold** is the best score in each row; `--` means not available. The superscript on each video score is the setting that produced it: frames × pixels per frame, or for codec, canvases (TC) × pixels per canvas. VideoMME is the mean over the short, medium and long splits. Each side reports its best score over the settings tried, from single runs. The image table shows selected benchmarks; full results, settings and protocol will be given in the technical report.</sub>
## Model Overview
<table style="width:100%; display:table;">
<tbody>
<tr><td align="left" style="width:28%"><b>Type</b></td><td align="left">instruction-tuned vision-language model, LLaVA-OneVision-2.5 architecture</td></tr>
<tr><td align="left" style="width:28%"><b>Parameters</b></td><td align="left">5.25B total (4B language model plus vision tower and adapter)</td></tr>
<tr><td align="left" style="width:28%"><b>Language side</b></td><td align="left">32 layers, hidden size 2560, FFN 9216, 16 attention heads / 4 KV heads, linear attention with one full-attention layer every 4</td></tr>
<tr><td align="left" style="width:28%"><b>Vision side</b></td><td align="left">24 layers, hidden size 1024, <code>patch_size=16</code>, <code>spatial_merge_size=3</code>, attention within a frame</td></tr>
<tr><td align="left" style="width:28%"><b>Vocabulary</b></td><td align="left">248320, including the 8 streaming protocol tokens</td></tr>
<tr><td align="left" style="width:28%"><b>Context length</b></td><td align="left">262144</td></tr>
<tr><td align="left" style="width:28%"><b>Weights</b></td><td align="left">bfloat16, 3 safetensors shards, about 9.9 GB</td></tr>
</tbody>
</table>
## Quickstart
The modeling code ships with the weights, so `trust_remote_code=True` is required.
```bash
pip install "transformers>=5.7.0" accelerate torch torchvision
pip install flash-attn --no-build-isolation # optional, recommended
pip install decord # frame-sampling video path only
```
### Image
```python
import torch
from PIL import Image
from transformers import AutoProcessor, AutoModelForImageTextToText
MODEL = "lmms-lab-encoder/LLaVA-OneVision-2.5-4B-Instruct"
processor = AutoProcessor.from_pretrained(MODEL, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
MODEL, trust_remote_code=True, dtype=torch.bfloat16,
attn_implementation="flash_attention_2", # "sdpa" without flash-attn
device_map="auto",
).eval()
messages = [{"role": "user", "content": [
{"type": "image"},
{"type": "text", "text": "Describe this image in detail."},
]}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=text, images=[Image.open("example.jpg").convert("RGB")],
return_tensors="pt").to(model.device)
with torch.inference_mode():
out = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
```
### Video
```python
messages = [{"role": "user", "content": [
{"type": "video"},
{"type": "text", "text": "What is happening in this video?"},
]}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(
text=text, videos=["example.mp4"],
max_frames=64,
min_pixels=153664, max_pixels=153664, # set equal to fix the per-frame budget
return_tensors="pt",
).to(model.device)
```
For codec mode, install `codec-video-prep` and pass `video_backend="codec"`:
```bash
pip install -i https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple \
"codec-video-prep==0.3.0"
codec-video-prep-doctor
```
```python
inputs = processor(text=text, videos=["example.mp4"],
video_backend="codec", return_tensors="pt").to(model.device)
```
### Streaming
The checkpoint already supports streaming: pass `streaming=True` to `apply_chat_template`. The runtime library for real-time online inference is coming soon.
## License
Apache-2.0.