Test-Time Scaling for Video Diffusion Models via Diagnosis-Guided Candidate Recycling

Hangzhou He, Lunhao Duan, Shanshan Zhao♯, Kaiwen Li, Qing-Guo Chen, Weihua Luo, Yanye Lu♯

Peking University, Alibaba Group

SIGGRAPH ASIA 2026 & TOG | Paper

Introduction

This is the implementation of GEARS(Guided Editing for Adaptive Recycling Search) from paper "Test-Time Scaling for Video Diffusion Models via Diagnosis-Guided Candidate Recycling". GEARS is a test-time scaling framework that improves text-to-video generation by turning low-scoring yet structurally plausible candidates into editable priors instead of discarding them after evaluation. GEARS consists of two components:

1. Stage-Aware Scheduler determines what to repair, when to repair it, and which candidates should be preserved, recycled, or discarded.
2. Candidate Recycler diagnoses recoverable failures from keyframes and multi-dimensional reward feedback, derives candidatespecific repair prompts, and repairs the corresponding candidates through manifold-aware latent SDEdit.

Quick Start

The code was tested with Python 3.10.20, PyTorch 2.7.0 with CUDA 12.8, Transformers 5.3.0, Diffusers 0.37.0 and Flash-Attn 2.7.3. We have also modified the VideoAlign codebase to to match with the transformers version.

Download the following resources:

1. A Wan2.1 checkpoint, such as Wan2.1-T2V-1.3B or Wan2.1-T2V-14B.
2. The VideoReward checkpoint.
3. The standard Vbench prompts arranged as one text file per category. Each non-empty line must contain one prompt.

Enter API key, endpoint, and generation configs in run_wan_gears_vbench.sh then run the script.

Acknowledgements

We thank the authors of Wan2.1, VideoReward, EvoSearch, and Vbench for their open-source contributions.

Citation

If you find GEARS useful for your research, please consider citing:

@article{he2026gears,
  title     = {Test-Time Scaling for Video Diffusion Models via Diagnosis-Guided Candidate Recycling},
  author    = {He, Hangzhou and Duan, Lunhao and Zhao, Shanshan and Li, Kaiwen and Chen, Qing-Guo and Luo, Weihua and Lu, Yanye},
  journal   = {ACM Transactions on Graphics},
  year      = {2025},
  note      = {SIGGRAPH Asia 2026},
}