Source-linked AI summary
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
TL;DR
Video test-time scaling often discards low-scoring candidates even when they retain useful motion, layout, or appearance structure. GEARS diagnoses recoverable deficiencies, edits candidates at stage-appropriate denoising points, and recycles them into the search pool; under matched budgets, it consistently improves over existing noise-space TTS methods, with a 1.3B model reaching a total score comparable to a 14B counterpart.
Problem
Existing video TTS methods mainly search over noise or denoising trajectories and discard low-scoring candidates, wasting useful structure and inference computation.
Method
GEARS uses a Stage-Aware Scheduler and Candidate Recycler to diagnose recoverable failures, edit re-noised latent states with candidate-specific conditions, and recycle repaired candidates.
Results
GEARS consistently outperforms existing noise-space TTS methods under matched compute budgets; on Wan2.1-T2V-1.3B, it improves the base model from 0.8192 to 0.8375, comparable to Wan2.1-T2V-14B at 0.8369.
Takeaways & Limitations
Diagnosis-guided recycling shifts video test-time scaling from passive noise selection toward active structured refinement of already-generated candidates.
Takeaways & Limitations
The reported global ranking score uses a default unweighted sum without tuning on VBench or human preferences.
Abstract
from arXiv · showhide
Recent video diffusion models have achieved remarkable generation quality, but high-fidelity results still largely depend on closed-source systems or costly large-scale infrastructure. Test-time scaling (TTS) offers a training-free way to improve lightweight generators by spending additional inference compute, yet existing methods mostly remain within a noise-search paradigm: they sample, select, or perturb denoising trajectories and discard low-scoring candidates after expensive generation. This generate-and-discard process wastes not only computation but also the partial motion, layout, or appearance structure already encoded in recoverable samples. We present \textbf{GEARS} (\textbf{G}uided \textbf{E}diting for \textbf{A}daptive \textbf{R}ecycling \textbf{S}earch), a training-free framework that introduces {diagnosis-guided candidate recycling} into video TTS by turning such candidates into editable priors through a generation-evaluation-editing loop. GEARS consists of two collaborative components. The \textbf{Stage-Aware Scheduler} determines what to repair, when to repair it, and which candidates should be preserved, recycled, or discarded. The \textbf{Candidate Recycler} diagnoses recoverable failures from keyframes and multi-dimensional reward feedback, derives candidate-specific repair prompts, and repairs the corresponding candidates through manifold-aware latent SDEdit. The repaired candidates are recycled into the search pool, creating refinement paths beyond standard noise perturbation while preserving useful structure. Under matched NFE budgets, GEARS consistently outperforms existing video TTS methods on VBench, bringing a 1.3B model to a total score comparable to a 14B counterpart, and ablations verify the necessity of adaptive scheduling, diagnosis-conditioned editing, and manifold-aware re-denoising. Code is available on GitHub.
1 Introduction
GEARS reframes video test-time scaling as diagnosis-guided candidate recycling: it repairs recoverable low-scoring candidates at controllable denoising stages and returns them to the search pool. This complements noise exploration with structured refinement while preserving useful candidate structure.
- Lightweight open-source video models retain a substantial quality gap relative to larger counterparts, motivating inference-time methods that add computation without changing model weights.
- Low-scoring candidates may remain plausible in some dimensions while failing in a recoverable aspect, so discarding them wastes encoded structure and prior generation compute.
- GEARS diagnoses dimension-specific failures, edits re-noised latent states at controllable denoising stages, and recycles corrected candidates into the search pool.
- The Stage-Aware Scheduler preserves elites, discards severely misaligned samples, and routes semantically valid but dimensionally deficient candidates to the Recycler.
- The Candidate Recycler uses visual anchors and multi-dimensional rewards to generate repair instructions, then applies manifold-aware SDE during re-denoising to reduce off-manifold artifacts and temporal jitter.
- Under matched compute budgets, GEARS consistently outperforms existing noise-space TTS methods and opens refinement paths beyond local noise perturbations.
2 Related Work
Related work develops training-free video editing and test-time scaling, but these lines largely remain separate. GEARS connects them by embedding diagnosis-conditioned latent editing into the inference-time search loop instead of discarding costly low-ranked candidates.
- Training-free video editing uses diffusion or flow-matching models to modify existing samples through noise inversion, latent manipulation, noise splicing, or context-aware refinement.
- Existing editing methods are typically standalone or post-processing procedures, leaving their correction capability disconnected from test-time scaling.
- Video TTS is challenging because candidates are expensive to generate and evaluate and quality depends on coupled factors including fidelity, motion, consistency, and text alignment.
- Consequently, generate-and-discard search can waste useful structure or motion patterns already encoded in low-ranked candidates.
- Verifier-focused work studies complementary metrics and stronger reward models, while search-focused work explores prompt rewriting, Best-of-N sampling, and Tree-of-Frames strategies.
- GEARS departs from prevailing video TTS by recycling recoverable low-scoring candidates through diagnosis-conditioned latent editing, converting spent compute into structured diversity.
3 Methodology
GEARS organizes video test-time scaling as repeated generation-to-editing checkpoints, where candidates are fully evaluated, selectively preserved or repaired, and recycled through stage-aware latent re-denoising. Its Scheduler assigns repair priorities across noise levels, while the Recycler uses visual and reward feedback to generate candidate-specific prompts and restore recoverable deficiencies.
- Overview: GEARS evaluates fully denoised, decoded videos at multiple checkpoints along the noise trajectory rather than scoring noisy latents.Each checkpoint defines an outer-loop recycling stage from a selected noise level to the clean latent.
- Stage-Aware Scheduler: High-noise checkpoints prioritize motion and coarse structure, whereas low-noise checkpoints prioritize appearance, texture, and local visual fidelity.An empirically chosen transition index α partitions the checkpoints into high-noise and low-noise regimes.
- Stage-Aware Scheduler: At each checkpoint, the Scheduler filters poor text alignment, selects weak-but-promising candidates for editing, and preserves the remaining valid candidates as elites.Candidates are ranked by the stage-dependent focus after semantic filtering, with the bottom q_edit fraction sent to the Recycler.
- Candidate Recycler: The Recycler uses keyframes, reward scores, and the stage focus to produce enhanced prompts that preserve original semantics while emphasizing the diagnosed deficiency.An MLLM generates candidate-specific prompts from four uniformly sampled keyframes and multi-dimensional feedback.
- Candidate Recycler: Repaired candidates are re-noised to the active checkpoint and re-denoised with a manifold-aware SDE sampler before replacing the original candidate in the survivor pool.The repaired candidates and preserved elites are merged for subsequent checkpoints, while elites retain their clean latents, prompts, and reward scores.
4 Experiments
GEARS is evaluated across multiple video backbones, durations, compute budgets, and independent evaluation settings. It consistently improves quality, scales effectively with additional compute, and recycles structurally useful candidates rather than discarding them.
- Generalization across Different Backbones: GEARS consistently improves semantic scores and maintains stable quality gains on CogVideoX-2B and Wan2.1-VACE-1.3B.These results indicate applicability across different backbones.
- Scaling Behavior at Test-Time: At small budgets such as NFEs=400, GEARS underperforms Video-T1, but its performance gains become more efficient as the compute budget increases.Recycling and editing become increasingly effective once sufficient candidate diversity is available.
- Longer-Video Generation: On approximately 5-second Wan2.1-T2V-1.3B videos, GEARS reaches a VBench Total of 0.8371 and outperforms the base model, Video-T1, and EvoSearch under comparable budgets.The main experiments use approximately 2-second videos, while this evaluation tests a longer duration.
- Scaling to a Larger Backbone: On Wan2.1-T2V-14B, GEARS improves the base model from 0.8369 to 0.8425, outperforming Video-T1 at 0.8410 and EvoSearch at 0.8388.Candidate recycling remains effective on a larger backbone.
- Prompt- and Seed-Level Stability: GEARS improves normalized prompt-level scores over the 1.3B base model, Video-T1, and EvoSearch by 0.018, 0.007, and 0.019, respectively, winning on 97, 79, and 85 of 110 prompts.The evaluation averages five independent seeds per prompt.
Wan2.1-VACE-1.3B
The ablations evaluate stage-aware scheduling and Candidate Recycler components for video generation. Results show that combining adaptive checkpoint scheduling with diagnosis and manifold-aware re-denoising produces the strongest reported scores.
- Stage-Aware Scheduler Settings: Combining high-noise motion and low-noise visual checkpoints raises the score to 0.8375, compared with 0.8192 for the base model.Single editing checkpoints achieve only 0.8221 for motion-focused editing and 0.8239 for visual-focused editing.
- Editing Ratio: VBench Total scores of 0.8359, 0.8375, and 0.8384 result from editing ratios of 0.25, 0.50, and 0.75 within the 1000-NFE baseline budget.The scores remain stable across settings, while allocating more compute to recycling provides a modest improvement.
- Candidate Recycler Components: MLLM diagnosis increases the base score from 0.8192 to 0.8281 with Qwen3.5-9B and 0.8325 with Qwen3.5-Plus.Reward- and frame-conditioned diagnosis turns weak-but-promising candidates into useful editing targets.
- Candidate Recycler Components: Replacing standard SDE with MA-SDE further improves scores to 0.8357 and 0.8375, with the combined Recycler achieving the best performance.The improvement is associated with better subject consistency and motion smoothness, reducing subject drift and temporal jitter after editing.
5 Conclusion, Limitations and Future Work
GEARS shifts video test-time scaling from passive noise-space exploration to active recycling of plausible candidates with recoverable deficiencies. Its effectiveness is strongest when candidates preserve useful structure, while severe text misalignment, entangled failures, and empirically chosen schedules remain limitations.
- Conclusion: GEARS recycles semantically plausible low-scoring candidates by repairing recoverable deficiencies at denoising stages where the attributes are most controllable.This converts discarded candidates into structured refinement paths and shifts inference-time search toward active candidate recycling.
- Conclusion: GEARS consistently improves generation quality compared to baseline methods.
- Limitations: Severe text misalignment or multiple entangled failures can remain difficult to repair reliably.GEARS is most effective when low-scoring candidates remain semantically plausible and preserve recoverable structure.
- Future Work: The current checkpoint schedule and transition boundary are empirically chosen and may not be optimal for every backbone, prompt, or budget.