Both are full-compute rollouts of the same prompt on the same model at the same settings. The only difference is which seed was chosen.
Ask a video model for the same thing twice and you get two different videos. Change the seed and the subject moves, the light changes, and sometimes the thing you asked for simply is not there. Anyone who has shipped generated video knows the routine: run it a few times, keep the one that worked.
That routine has a name. Best-of-N search runs the prompt N times and keeps whichever a verifier scores highest. It works. It also costs N times as much, and N−1 of those samples go straight in the bin.
On the right is one prompt at two seeds on the same model. Nothing separates them but the number the sampler started from. One of them is what you would have received.
Caching is normally measured by how close a cached sample is to the real thing. For search that is the wrong test. The draft never ships. All it has to do is come out in the right order, and that is a far weaker requirement than looking identical.
One seed on Wan2.1-T2V 1.3B, generated twice: at full compute on the left, under caching on the right. The draft costs 45% as much and the verifier prefers it. Fidelity to the full rollout is not what search needs, and not what it gets.
The sampler is deterministic in the seed, so regenerating the winner with caching off gives back a real full-compute sample. When the cached ranking and the full-compute ranking agree on which seed wins, what ships is bit-identical to what full best-of-N would have handed you. The caching changed the choice, not the output.
Across the 104 pairs on this page the drafts picked the same winner on 64 of them.
Measured on Wan2.1-T2V 14B at 1280×720: eight full rollouts against eight drafts plus one commit.
A wrapper around generation, not a new model. Any published training-free cache works as the engine, any verifier works as the judge.
Every pair is two full-compute rollouts of one prompt on the same model, with the verifier score each received.
From the paper: 50 prompts, eight seeds each, every candidate generated twice and scored by the same verifier.
| What is measured | Result | Meaning |
|---|---|---|
| Rank correlation, cached against full | 0.905 | the cached ordering is the full-compute ordering |
| Winner agreement | 72% | the draft that wins is the sample that would have won |
| Search gain retained | 94.7% | almost all the quality search buys, kept |
| Cost of full best-of-8 | 63% | the same search, for under two thirds of the compute |
Wan2.1-1.3B with EasyCache at τ = 0.20, the paper's headline configuration.
The exploration threshold is the one model-specific number. It tracks architecture family rather than parameter count, so it is calibrated once per family and reused.
| Model | Family | Calibrated threshold | Gain retained | Exploration speedup |
|---|---|---|---|---|
| Wan2.1-T2V-1.3B | Wan | 0.20 | 88.3% | 2.41× |
| Wan2.1-T2V-14B | Wan | 0.10 | 87.5% | 2.05× |
| Wan2.2-TI2V-5B | Wan | 0.10 | 86.0% | 2.05× |
| CogVideoX-5B | CogVideoX | 0.05 | 85.9% | 1.78× |
| HunyuanVideo | HunyuanVideo | 0.05 | 85.1% | 1.77× |
| LTX-Video | LTX | 0.02 | 79.6% | 1.71× |
The threshold is τ, the accumulated-change tolerance the cache reuses features under. Gain retained is the fraction of what full best-of-8 search buys. Never copy a threshold across families: driving LTX-2 at the Wan value halved winner agreement on identical rollouts.
The gallery runs newer and larger checkpoints than the paper's grid, on prompts of its own, so every backbone shown here was recalibrated from its own rollouts rather than inheriting a threshold. The rule is the paper's: take the largest τ whose gain retained still clears 85%.
| Model | Family | Resolution | τ | Gain retained | Winner agreement | Rank corr. | Exploration speedup |
|---|---|---|---|---|---|---|---|
| LTX-2 19B | LTX | 1920×1088 | 0.02 | 85.0% | 59% | 0.74 | 1.64× |
| Wan2.1-T2V 14B | Wan | 1280×720 | 0.10 | 85.9% | 56% | 0.86 | 2.23× |
| HunyuanVideo 13B | HunyuanVideo | 1280×720 | 0.05 | 94.8% | 65% | 0.87 | 1.77× |
| CogVideoX 5B | CogVideoX | 720×480 | 0.02 | 86.9% | 61% | 0.83 | 1.16× |
| Wan2.1-T2V 1.3B | Wan | 832×480 | 0.20 | 90.0% | 67% | 0.87 | 2.28× |
Measured on 222 prompts × 8 seeds × 2 arms, the same rollouts the galleries below are drawn from. Gain retained is the share of the best-of-8 improvement over an average draw that survives ranking drafts instead of finished samples; speedup is measured wall-clock per draft. Copying a threshold across families is the one thing that breaks. LTX-2 driven at the Wan value collapses to 50.5% retained on identical rollouts, which is why it runs at 0.02. CogVideoX is the other case: the paper's 0.05 retains only 71.4% on these prompts and 0.03 only 81.2%, so it runs at 0.02 and buys a smaller, honest saving. Every comparison below uses its model's calibrated threshold.
CachedSearch takes the cache as a component. Any published engine can fill the slot, trading fidelity against cost along one frontier.
| Engine | Gain retained | Cost of best-of-8 |
|---|---|---|
| No caching (full best-of-8) | 100.0% | 621 s |
| PAB | 99.5% | 493 s |
| FasterCache CFG-Cache | 99.6% | 465 s |
| TeaCache | 93.2% | 365 s |
| EasyCache | 90.1% | 346 s |
Search width 8; cost includes the full-compute commit.
Grouped by backbone, across five models from 1.3B to 19B and resolutions from 480p to 1080p.
@article{saini2026cachedsearch,
title = {CachedSearch: Training-Free Cached Exploration for Test-Time Search in Video Diffusion},
author = {Saini, Shreshth and Birkbeck, Neil and Wang, Yilin and Adsumilli, Balu and Bovik, Alan C.},
year = {2026},
eprint = {2607.23159},
archivePrefix = {arXiv},
primaryClass = {cs.AI}
}