Video-Reason/VBVR-Wan2.1-diffsynth
629
1---2base_model:3- Wan-AI/Wan2.1-I2V-14B-720P-Diffusers4library_name: diffusers5license: apache-2.06pipeline_tag: image-to-video7datasets:8- Video-Reason/VBVR-Dataset9---10# VBVR: A Very Big Video Reasoning Suite11 12<a href="https://video-reason.com" target="_blank">13 <img alt="Project Page" src="https://img.shields.io/badge/Project%20-%20Homepage-4285F4" height="20" />14</a>15<a href="https://github.com/Video-Reason/VBVR-EvalKit" target="_blank">16 <img alt="Code" src="https://img.shields.io/badge/Evaluation_code-VBVR_Bench-100000?style=flat-square&logo=github&logoColor=white" height="20" />17</a>18<a href="https://github.com/Video-Reason/VBVR-Wan2.2" target="_blank">19 <img alt="Code" src="https://img.shields.io/badge/Training_code-VBVR_Wan2.2-100000?style=flat-square&logo=github&logoColor=white" height="20" />20</a>21<a href="https://github.com/Video-Reason/VBVR-DataFactory" target="_blank">22 <img alt="Code" src="https://img.shields.io/badge/Data_code-VBVR_DataFactory-100000?style=flat-square&logo=github&logoColor=white" height="20" />23</a>24<a href="https://huggingface.co/papers/2602.20159" target="_blank">25 <img alt="arXiv" src="https://img.shields.io/badge/arXiv-VBVR-red?logo=arxiv" height="20" />26</a>27<a href="https://huggingface.co/datasets/Video-Reason/VBVR-Dataset" target="_blank">28 <img alt="Dataset" src="https://img.shields.io/badge/%F0%9F%A4%97%20_VBVR_Dataset-Data-ffc107?color=ffc107&logoColor=white" height="20" />29</a>30<a href="https://huggingface.co/datasets/Video-Reason/VBVR-Bench-Data" target="_blank">31 <img alt="Bench Data" src="https://img.shields.io/badge/%F0%9F%A4%97%20_VBVR_Bench-Data-ffc107?color=ffc107&logoColor=white" height="20" />32</a>33<a href="https://huggingface.co/spaces/Video-Reason/VBVR-Bench-Leaderboard" target="_blank">34 <img alt="Leaderboard" src="https://img.shields.io/badge/%F0%9F%A4%97%20_VBVR_Bench-Leaderboard-ffc107?color=ffc107&logoColor=white" height="20" />35</a>36 37## Overview38Video reasoning grounds intelligence in spatiotemporally consistent visual environments that go beyond what text can naturally capture, 39enabling intuitive reasoning over motion, interaction, and causality. Rapid progress in video models has focused primarily on visual quality. 40Systematically studying video reasoning and its scaling behavior suffers from a lack of video reasoning (training) data. 41 42To address this gap, we introduce the Very Big Video Reasoning (VBVR) Dataset, an unprecedentedly large-scale resource spanning 200 curated reasoning tasks 43and over one million video clips—approximately three orders of magnitude larger than existing datasets. We further present VBVR-Bench, 44a verifiable evaluation framework that moves beyond model-based judging by incorporating rule-based, human-aligned scorers, 45enabling reproducible and interpretable diagnosis of video reasoning capabilities. 46 47Leveraging the VBVR suite, we conduct one of the first large-scale scaling studies of video reasoning and observe early signs of emergent generalization 48to unseen reasoning tasks. **Together, VBVR lays a foundation for the next stage of research in generalizable video reasoning.**49 50The model was presented in the paper [A Very Big Video Reasoning Suite](https://huggingface.co/papers/2602.20159).51 52## Models Zoo53 54| Model | Base Architecture | Other Remarks |55|-------|-------------------|---------------|56| [VBVR-Wan2.1](https://huggingface.co/Video-Reason/VBVR-Wan2.1) | Wan2.1-I2V-14B-720P | Diffusers format |57| [VBVR-Wan2.2](https://huggingface.co/Video-Reason/VBVR-Wan2.2) | Wan2.2-I2V-A14B | Diffusers format |58| [**VBVR-Wan2.1-diffsynth**](https://huggingface.co/Video-Reason/VBVR-Wan2.1-diffsynth) | Wan2.1-I2V-14B-720P | DiffSynth LoRA format |59| [VBVR-Wan2.2-diffsynth](https://huggingface.co/Video-Reason/VBVR-Wan2.2-diffsynth) | Wan2.2-I2V-A14B | DiffSynth LoRA format |60| [VBVR-LTX2.3-diffsynth](https://huggingface.co/Video-Reason/VBVR-LTX2.3-diffsynth) | LTX-Video-2.3 | DiffSynth LoRA format |61 62## Release Information63VBVR-Wan2.1 is trained from Wan2.1-I2V-14B-720P without architectural modifications, as the goal of VBVR is to *investigate data scaling behavior* and provide *strong baseline models* for the video reasoning research community. Leveraging the VBVR-Dataset, which constitutes one of the largest video reasoning datasets to date, the VBVR model family achieved highest scores on VBVR-Bench.64 65In this release, we present 66[**VBVR-Wan2.1**](https://huggingface.co/Video-Reason/VBVR-Wan2.1) (Diffusers format),67[**VBVR-Wan2.1-diffsynth**](https://huggingface.co/Video-Reason/VBVR-Wan2.1-diffsynth) (DiffSynth LoRA format), and68[**VBVR-LTX2.3-diffsynth**](https://huggingface.co/Video-Reason/VBVR-LTX2.3-diffsynth) (DiffSynth LoRA format; Diffusers does not yet support LTX-Video-2.3, so only the DiffSynth LoRA format is released for this model).69 70<table>71 <tr>72 <th>Model</th>73 <th>Overall</th>74 <th>ID</th>75 <th>ID-Abst.</th>76 <th>ID-Know.</th>77 <th>ID-Perc.</th>78 <th>ID-Spat.</th>79 <th>ID-Trans.</th>80 <th>OOD</th>81 <th>OOD-Abst.</th>82 <th>OOD-Know.</th>83 <th>OOD-Perc.</th>84 <th>OOD-Spat.</th>85 <th>OOD-Trans.</th>86 </tr>87 <tbody>88 <tr>89 <td><strong>Human</strong></td>90 <td>0.974</td><td>0.960</td><td>0.919</td><td>0.956</td><td>1.00</td><td>0.95</td><td>1.00</td>91 <td>0.988</td><td>1.00</td><td>1.00</td><td>0.990</td><td>1.00</td><td>0.970</td>92 </tr>93 <tr style="background:#F2F0EF;font-weight:700;text-align:center;">94 <td colspan="14"><em>Open-source Models</em></td>95 </tr>96 <tr>97 <td>CogVideoX1.5-5B-I2V</td>98 <td>0.273</td><td>0.283</td><td>0.241</td><td>0.328</td><td>0.257</td><td>0.328</td><td>0.305</td>99 <td>0.262</td><td><u>0.281</u></td><td>0.235</td><td>0.250</td><td><strong>0.254</strong></td><td>0.282</td>100 </tr>101 <tr>102 <td>HunyuanVideo-I2V</td>103 <td>0.273</td><td>0.280</td><td>0.207</td><td>0.357</td><td>0.293</td><td>0.280</td><td><u>0.316</u></td>104 <td>0.265</td><td>0.175</td><td><strong>0.369</strong></td><td>0.290</td><td><u>0.253</u></td><td>0.250</td>105 </tr>106 <tr>107 <td><strong>Wan2.2-I2V-A14B</strong></td>108 <td><strong>0.371</strong></td><td><strong>0.412</strong></td><td><strong>0.430</strong></td>109 <td><strong>0.382</strong></td><td><strong>0.415</strong></td><td><strong>0.404</strong></td>110 <td><strong>0.419</strong></td><td><strong>0.329</strong></td>111 <td><strong>0.405</strong></td><td>0.308</td><td><strong>0.343</strong></td>112 <td>0.236</td><td><u>0.307</u></td>113 </tr>114 <tr>115 <td><u>LTX-2</u></td>116 <td><u>0.313</u></td><td><u>0.329</u></td><td><u>0.316</u></td>117 <td><u>0.362</u></td><td><u>0.326</u></td><td><u>0.340</u></td>118 <td>0.306</td><td><u>0.297</u></td>119 <td>0.244</td><td><u>0.337</u></td><td><u>0.317</u></td>120 <td>0.231</td><td><strong>0.311</strong></td>121 </tr>122 <tr style="background:#F2F0EF;font-weight:700;text-align:center;">123 <td colspan="14"><em>Proprietary Models</em></td>124 </tr>125 <tr>126 <td><u>Seedance 2.0</u></td>127 <td><u>0.544</u></td><td><strong>0.570</strong></td><td>0.593</td><td><u>0.498</u></td><td><strong>0.618</strong></td><td><u>0.514</u></td><td><strong>0.602</strong></td>128 <td><u>0.517</u></td><td><strong>0.643</strong></td><td>0.398</td><td><u>0.492</u></td><td>0.427</td><td><strong>0.556</strong></td>129 </tr>130 <tr>131 <td>Runway Gen-4 Turbo</td>132 <td>0.403</td><td>0.392</td><td>0.396</td><td>0.409</td><td>0.429</td><td>0.341</td><td>0.363</td>133 <td>0.414</td><td>0.515</td><td><u>0.429</u></td><td>0.419</td><td>0.327</td><td>0.373</td>134 </tr>135 <tr>136 <td><strong>Sora 2</strong></td>137 <td><strong>0.546</strong></td><td><u>0.569</u></td><td><u>0.602</u></td>138 <td>0.477</td><td><u>0.581</u></td><td><strong>0.572</strong></td>139 <td><u>0.597</u></td><td><strong>0.523</strong></td>140 <td><u>0.546</u></td><td><strong>0.472</strong></td><td><strong>0.525</strong></td>141 <td><strong>0.462</strong></td><td><u>0.546</u></td>142 </tr>143 <tr>144 <td>Kling 2.6</td>145 <td>0.369</td><td>0.408</td><td>0.465</td><td>0.323</td><td>0.375</td><td>0.347</td><td>0.519</td>146 <td>0.330</td><td>0.528</td><td>0.135</td><td>0.272</td><td>0.356</td><td>0.359</td>147 </tr>148 <tr>149 <td>Veo 3.1</td>150 <td>0.480</td><td>0.531</td><td><strong>0.611</strong></td>151 <td><strong>0.503</strong></td><td>0.520</td><td>0.444</td>152 <td>0.510</td><td>0.429</td>153 <td><u>0.577</u></td><td>0.277</td><td>0.420</td>154 <td><u>0.441</u></td><td>0.404</td>155 </tr>156 <tr style="background:#F2F0EF;font-weight:700;text-align:center;">157 <td colspan="14"><em>Data Scaling Strong Baseline</em></td>158 </tr>159 <tr>160 <td><strong>VBVR-LTX2.3</strong></td>161 <td>0.516</td><td>0.580</td><td>0.608</td><td>0.631</td><td>0.529</td><td>0.454</td><td>0.680</td>162 <td>0.453</td><td>0.608</td><td>0.577</td><td><u>0.409</u></td><td>0.414</td><td><u>0.388</u></td>163 </tr>164 <tr>165 <td><strong>VBVR-Wan2.1</strong></td>166 <td><u>0.592</u></td><td><u>0.724</u></td><td><u>0.705</u></td><td><u>0.710</u></td><td><u>0.727</u></td><td><u>0.719</u></td><td><u>0.784</u></td>167 <td><u>0.461</u></td><td><u>0.674</u></td><td><strong>0.592</strong></td><td>0.387</td><td><u>0.461</u></td><td>0.387</td>168 </tr>169 <tr>170 <td><strong>VBVR-Wan2.2</strong></td>171 <td><strong>0.685</strong></td><td><strong>0.760</strong></td><td><strong>0.724</strong></td>172 <td><strong>0.750</strong></td><td><strong>0.782</strong></td><td><strong>0.745</strong></td>173 <td><strong>0.833</strong></td><td><strong>0.610</strong></td>174 <td><strong>0.768</strong></td><td><u>0.572</u></td><td><strong>0.547</strong></td>175 <td><strong>0.618</strong></td><td><strong>0.615</strong></td>176 </tr>177 </tbody>178</table>179 180## QuickStart181 182### Inference183For running inference, please refer to the [**official guide**](https://github.com/Video-Reason/VBVR-Wan2.2?tab=readme-ov-file#wan21-inference) in the VBVR-Wan2.2 GitHub repository.184This repository contains the latest instructions, configurations, and examples for performing inference with the VBVR family models.185 186## Citation187 188```bibtex189@article{vbvr2026,190 title = {A Very Big Video Reasoning Suite},191 author = {Wang, Maijunxian and Wang, Ruisi and Lin, Juyi and Ji, Ran and192 Wiedemer, Thadd{\"a}us and Gao, Qingying and Luo, Dezhi and193 Qian, Yaoyao and Huang, Lianyu and Hong, Zelong and Ge, Jiahui and194 Ma, Qianli and He, Hang and Zhou, Yifan and Guo, Lingzi and195 Mei, Lantao and Li, Jiachen and Xing, Hanwen and Zhao, Tianqi and196 Yu, Fengyuan and Xiao, Weihang and Jiao, Yizheng and197 Hou, Jianheng and Zhang, Danyang and Xu, Pengcheng and198 Zhong, Boyang and Zhao, Zehong and Fang, Gaoyun and Kitaoka, John and199 Xu, Yile and Xu, Hua bureau and Blacutt, Kenton and Nguyen, Tin and200 Song, Siyuan and Sun, Haoran and Wen, Shaoyue and He, Linyang and201 Wang, Runming and Wang, Yanzhi and Yang, Mengyue and Ma, Ziqiao and202 Milli{\`e}re, Rapha{\"e}l and Shi, Freda and Vasconcelos, Nuno and203 Khashabi, Daniel and Yuille, Alan and Du, Yilun and Liu, Ziming and204 Lin, Dahua and Liu, Ziwei and Kumar, Vikash and Li, Yijiang and205 Yang, Lei and Cai, Zhongang and Deng, Hokin},206 journal = {arXiv preprint arXiv:2602.20159},207 year = {2026},208 url = {https://arxiv.org/abs/2602.20159}209}210```