Research map / Video / Video generation
LatSearch: Latent Reward-Guided Search for Faster Inference-Time Scaling in Video Diffusion
Zengqun Zhao, Ziquan Liu, Yu Cao, Shaogang Gong, Zhensong Zhang, Song Jifei, Jiankang Deng, ioannis Patras
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply LatSearch latent reward-guided search | Wan2.1-1.3B · VBench-2.0 average | score: 51.90 -> 53.84 | Improved |
| LatSearch vs EvoSearch | LatSearch · VBench-2.0 | score and time: 55.25 vs 55.01 average; 164.41 s vs 783.76 s | Mixed |
| initial-noise optimisation (FreeInit, FreqPrior) | Wan2.1-1.3B · VBench-2.0 average | score: 51.90 -> 49.82 / 50.37 | Hurt |
| increase search budget N from 4 to 8 (ablation) | LatSearch · VBench-2.0 average | score: 52.81 -> 54.13 | Improved |
| add RGRP resampling instead of beam-style selection (ablation) | LatSearch · VBench-2.0 average | score: 52.35 -> 53.84 | Improved |
| add preference loss to latent reward model (ablation) | LatSearch · VBench-2.0 average | score: 53.42 -> 53.84 | Improved |
| cosine similarity credit assignment instead of uniform/exponential/L2 (ablation) | LatSearch · VBench-2.0 average | score: 52.31 -> 53.84 | Improved |
| use scoring schedule {10,15,20} instead of five points (ablation) | LatSearch · VBench-2.0 average | score: 52.34 -> 53.84 | Improved |
| use scoring schedule {10,15,20} instead of {10,15} (ablation) | LatSearch · VBench-2.0 average | score: 52.17 -> 53.84 | Improved |
| vary temperature tau 0.5, 1.0, 2.0 (ablation) | LatSearch · VBench-2.0 average | score: 53.56 / 53.84 / 53.33 | No clear effect |
| apply LatSearch to CogVideoX-2B | CogVideoX-2B · VBench-2.0 average | score: 45.13 -> 48.14 | Improved |
| apply LatSearch to Wan2.1-14B | Wan2.1-14B · VBench-2.0 average | score: 52.58 -> 53.61 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- FreqPrior — FreqPrior: Improving Video Diffusion Models with Frequency Filtering Gaussian Noise (ICLR 2025) · better on 1 of 2 shared results
- CogVideoX-2B — CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer (ICLR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search (NeurIPS 2025 · similarity 0.91)
- FreeNoise: Tuning-Free Longer Video Diffusion via Noise Rescheduling (ICLR 2024 · similarity 0.89)
- AdaDiff: Adaptive Step Selection for Fast Diffusion Models (AAAI 2025 · similarity 0.88)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.88)
- InstructVideo: Instructing Video Diffusion Models with Human Feedback (CVPR 2024 · similarity 0.88)
- Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
