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 GEARS (Guided Editing for Adaptive Recycling Search), 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 Stage-Aware Scheduler determines what to repair, when to repair it, and which candidates should be preserved, recycled, or discarded. The 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.
Figures & tables
Figure 1. Left: Test-time scaling paradigms along the denoising axis ( t=1→t=0 ). Video-T1 ( Liu et al., 2025a ) samples independent trajectories and selects the best video at the end; DLBS ( Oshima et al., 2025 ) seeks a better diffusion path by keeping K elites for each step as beam search; EvoSearch ( He et al., 2025a ) adds mid-denoising mutation around elites but still clusters candidates locally; GEARS further recycles low-scoring candidates via diagnostic latent editing (green dashed arrows), rerouting them back into denoising at two scheduled stages to reach regions noise-space search alone cannot. Right: Qualitative comparison. Baselines suffer from subject-background inconsistency or stagnant motion, while GEARS produces a coherent koala playing piano throughout.
Figure 2. Left : Joint and marginal distributions of Base and GEARS candidates in the VQ–MQ space. Contours visualize the joint density, while the top and right curves show the VQ and MQ marginal profiles, respectively. GEARS redirects recoverable candidates toward higher joint quality. Right : High-noise stages establish coarse motion and layout, whereas low-noise stages refine appearance and texture.
Figure 3. Overview of GEARS. Along the denoising trajectory from pure noise to clean latent, GEARS inserts multiple generation-to-editing checkpoints {s0,s1,…,sk} . At each checkpoint, candidates are first fully denoised, decoded, and evaluated as clean videos; noisy latents are never scored. The Scheduler determines the current repair focus and routes the evaluated candidates into preserved elites, recoverable candidates for the Recycler, and discarded samples. The Recycler diagnoses and repairs the recoverable candidates through prompt-guided latent editing, after which they are returned to the search pool via re-noising. High-noise checkpoints prioritize motion correction, while low-noise checkpoints focus on visual refinement.
VBench
VideoAlign
Method
Total Score
Quality Score
Semantic Score
Background Consistency
Imaging Quality
Motion Smoothness
Subject Consistency
Visual Quality
Motion Quality
Wan2.1-T2V-14B
0.8369
0.8559
0.7611
0.9809
0.6943
0.9830
0.9752
1.3807
0.9499
HunyuanVideo 13B
0.8324
0.8509
0.7582
0.9776
0.6756
0.9899
0.9737
1.4240
1.0920
Wan2.1-T2V-1.3B
0.8192
0.8471
0.7076
0.9655
0.6844
0.9674
0.9687
1.3737
0.9552
+ Video-T1
0.8294
0.8490
0.7508
0.9656
0.6832
0.9644
0.9684
1.4464
1.1222
+ DLBS
0.8224
0.8483
0.7188
0.9485
0.7039
0.9619
0.9627
1.4405
1.0834
Table 1. Quantitative Comparison Results. (Results of ImagerySearch are from ( Wu et al., 2026 ) )
Figure 4. Stage-aware candidate repair by the Scheduler and Recycler. Top: At Stage 1 (high noise), the candidate under-realizes the requested walking motion. The Recycler restores coherent subject motion while retaining the rabbit and the fantasy scene. The rightmost insets summarize background-compensated optical flow across the three frames under a shared scale; warmer colors indicate larger residual motion, and arrows show its direction. Bottom: At Stage 2 (low noise), the candidate preserves the beach layout and wave dynamics but misses the requested cyberpunk appearance. The Recycler introduces neon structures while maintaining the coastal composition, as highlighted by the enlarged regions. Together, the two cases illustrate how the Scheduler assigns motion and appearance deficiencies to their appropriate repair stages.
CogVideoX-2B
Wan2.1-VACE-1.3B
Method
Quality
Semantic
Quality
Semantic
Base
0.8097
0.7639
0.8480
0.6573
+ Video-T1
0.8070
0.7679
0.8454
0.7415
+ DLBS
0.8079
0.7716
0.8382
0.6985
+ EvoSearch
0.8006
0.7639
0.8703
0.6512
+ GEARS ( Ours )
0.8166
0.7973
0.8527
0.7616
Table 2. Generalization across Backbones .
Figure 5. Scaling behavior on Wan2.1-T2V-1.3B and Wan2.1-VACE-1.3B.
Compared Method
Δ Score
Prompt Wins
Human W/T/L
Wan2.1-T2V-1.3B
+ 0.018
97/110
75/11/14
+ Video-T1
+ 0.007
79/110
68/14/18
+ EvoSearch
+ 0.019
85/110
85/3/12
Wan2.1-T2V-14B
+ 0.004
71/110
68/3/29
Table 4. Prompt-level stability and blinded human evaluation of Wan2.1-T2V-1.3B + GEARS. Prompt-level results average five independent seeds for each of the 110 prompts; human results report Win/Tie/Loss.
Method
PFLOPs
Time (s) 33 frames
Wan2.1-T2V-14B
50.53
200.37±3.04
Wan2.1-T2V-1.3B + GEARS
52.77
189.66±4.91
Table 5. End-to-end inference cost. PFLOPs are measured with a local Qwen3.5-9B. GEARS wall-clock time includes MLLM API latency.
Figure 6. Discarded-to-Improved Trajectory in GEARS. A low-ranked but semantically valid candidate, which would be discarded by conventional generate-and-select TTS, is recycled through the Scheduler and Recycler. GEARS first repairs its motion deficiency at a high-noise checkpoint and later enhances visual details at a low-noise checkpoint. The repaired candidate is then returned to the final candidate pool and selected as the best video.
Figure 7. Visualized examples on Wan2.1-VACE-1.3B.
Checkpoint Setting
Focus
Total Score
{s0=1}
No adapt
0.8192
{s0=1,s1=0.6}
MQ only
0.8221
{s0=1,s1=0.3}
VQ only
0.8239
{s0=1,s1=0.6,s2=0.3}
α=1
0.8375
{s0=1,s1=0.6,s2=0.45,s3=0.3}
α=1
0.8304
{s0=1,s1=0.6,s2=0.45,s3=0.3}
α=2
0.8392
Table 6. Ablation of Stage-Aware Scheduler Settings. α is the transition index that partitions the checkpoints into high- and low-noise stages.
MLLM Diagnosis
MA-SDE
VBench
Enabled
Model
Total
Sub. Con.
Mot. Smo.
×
-
×
0.8192
0.9687
0.9674
✓
Qwen3.5-9B
×
0.8281
0.9782
0.9689
✓
Qwen3.5-Plus
×
0.8325
0.9807
0.9692
✓
Qwen3.5-9B
✓
0.8357
0.9812
0.9715
✓
Qwen3.5-Plus
✓
0.8375
0.9815
0.9727
Table 7. Ablation of Candidate Recycler Components.