CachedSearch

CachedSearch

Training-free cached exploration for test-time search in video diffusion

Shreshth Saini · Neil Birkbeck · Yilin Wang · Balu Adsumilli · Alan C. Bovik

The University of Texas at Austin · Google

The University of Texas at Austin LIVE Lab Google YouTube

The sample you would have got, and the one search found

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.

One sample-1.43
CachedSearch+1.45
A polar bear wearing sunglasses floating on a pink inflatable flamingo in a swimming pool +2.89 gap · LTX-2 19B · 1920×1088 · same seed as full compute
One sample-0.28
CachedSearch+1.71
An octopus playing a grand piano on the seabed, bubbles rising through sunbeams +2.00 gap · LTX-2 19B · 1920×1088 · same seed as full compute
One sample-1.40
CachedSearch+1.54
A lion's face in extreme close-up at golden hour, mane lifting in the wind, eyes locked on the camera +2.94 gap · LTX-2 19B · 1920×1088 · same seed as full compute

The problem

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.

One sample-0.59
CachedSearch+0.81
A rally car sliding through a forest hairpin, gravel spraying into low sun +1.40 gap · LTX-2 19B · 1920×1088 · different seed from full compute

The observation

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.

Full compute
Cached draft
Full compute67s+1.61 verifier
Cached draft30s+1.74 verifier
Cost45%and scores +0.13 higher

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.

Best-of-8 at full compute 99 min CachedSearch: eight drafts, then one commit 57 min · 57%

Measured on Wan2.1-T2V 14B at 1280×720: eight full rollouts against eight drafts plus one commit.

Approach

A wrapper around generation, not a new model. Any published training-free cache works as the engine, any verifier works as the judge.

N cached drafts cache ON · ≈ half cost each 54% skipped score verifier ranks them argmax over drafts winner commit that seed cache OFF · 1 rollout deliver full-compute sample

Algorithm

Algorithm 1  CachedSearch
Input prompt c; seeds S = {s1,…,sN}; verifier V; sampler Φ; threshold τ
Output a full-compute sample x
  1. ▷ explore: one cached rollout per seed
  2. for i = 1 to N do
  3. i ← Φ(c, si ∣ cache = on, τ)cost Cc
  4. end for
  5. ▷ rank: only the ordering is consumed
  6. i* ← arg maxi∈[N] V(i, c)
  7. ▷ commit: regenerate the winner exactly
  8. x ← Φ(c, si* ∣ cache = off)cost Cf
  9. return x
Cost.  Cours = N·Cc + Cf  and  Cours / Cbest-of-N = Cc/Cf + 1/N
Exactness.  if arg maxi V(i, c) = arg maxi V(xi, c) then x = xbest-of-N bit for bit, because Φ is deterministic in the seed.
Requirement.  the drafts need only preserve the ordering induced by V, not approximate xi. Rank agreement, not reconstruction error, is the quantity that matters.

Quantitative results

From the paper: 50 prompts, eight seeds each, every candidate generated twice and scored by the same verifier.

What is measuredResultMeaning
Rank correlation, cached against full0.905the cached ordering is the full-compute ordering
Winner agreement72%the draft that wins is the sample that would have won
Search gain retained94.7%almost all the quality search buys, kept
Cost of full best-of-863%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.

Calibrated threshold per model

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.

ModelFamilyCalibrated thresholdGain retainedExploration speedup
Wan2.1-T2V-1.3BWan0.2088.3%2.41×
Wan2.1-T2V-14BWan0.1087.5%2.05×
Wan2.2-TI2V-5BWan0.1086.0%2.05×
CogVideoX-5BCogVideoX0.0585.9%1.78×
HunyuanVideoHunyuanVideo0.0585.1%1.77×
LTX-VideoLTX0.0279.6%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.

Recalibrated on the models in this gallery

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%.

ModelFamilyResolutionτ Gain retainedWinner agreementRank corr. Exploration speedup
LTX-2 19BLTX1920×10880.0285.0%59%0.741.64×
Wan2.1-T2V 14BWan1280×7200.1085.9%56%0.862.23×
HunyuanVideo 13BHunyuanVideo1280×7200.0594.8%65%0.871.77×
CogVideoX 5BCogVideoX720×4800.0286.9%61%0.831.16×
Wan2.1-T2V 1.3BWan832×4800.2090.0%67%0.872.28×

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.

Exploration engine

CachedSearch takes the cache as a component. Any published engine can fill the slot, trading fidelity against cost along one frontier.

EngineGain retainedCost of best-of-8
No caching (full best-of-8)100.0%621 s
PAB99.5%493 s
FasterCache CFG-Cache99.6%465 s
TeaCache93.2%365 s
EasyCache90.1%346 s

Search width 8; cost includes the full-compute commit.

More comparisons

Grouped by backbone, across five models from 1.3B to 19B and resolutions from 480p to 1080p.

LTX-2 19B 1920×1088 · 21 pairs

One sample-1.86
CachedSearch+1.41
A cat conducting an orchestra of mice with a matchstick baton in a candlelit hall +3.27 gap · LTX-2 19B · 1920×1088 · same seed as full compute
One sample-1.30
CachedSearch+1.56
A peacock unfurling its tail in slow motion, iridescent feathers catching low sun +2.86 gap · LTX-2 19B · 1920×1088 · same seed as full compute
One sample-1.58
CachedSearch+0.62
A hawk diving through a canyon, wings tucked, rock face rushing past +2.20 gap · LTX-2 19B · 1920×1088 · same seed as full compute
One sample-0.78
CachedSearch+1.14
Extreme slow motion close-up of a hummingbird hovering beside a scarlet flower, backlit wings, shallow depth of field +1.92 gap · LTX-2 19B · 1920×1088 · same seed as full compute
One sample-0.21
CachedSearch+1.71
A hot air balloon inflating at dawn on a frosty field, crew silhouetted against the burner flame +1.92 gap · LTX-2 19B · 1920×1088 · different seed from full compute
One sample-1.62
CachedSearch+0.23
A steam train thundering past a platform, leaves swirling in the draught +1.86 gap · LTX-2 19B · 1920×1088 · same seed as full compute
One sample-1.35
CachedSearch+0.35
An otter in a yellow lifejacket confidently riding a surfboard across turquoise tropical water, spray backlit by the sun +1.69 gap · LTX-2 19B · 1920×1088 · different seed from full compute
One sample-1.31
CachedSearch+0.27
Extreme slow motion of a swimmer breaking the surface, water sheeting off the face, droplets suspended, backlit +1.58 gap · LTX-2 19B · 1920×1088 · same seed as full compute
One sample-1.12
CachedSearch+0.38
Extreme slow motion of a violinist mid-stroke, rosin dust drifting through a raking stage light +1.50 gap · LTX-2 19B · 1920×1088 · different seed from full compute
One sample-1.64
CachedSearch-0.15
A sloth in a pinstripe suit riding an escalator in a glass office atrium +1.48 gap · LTX-2 19B · 1920×1088 · different seed from full compute
One sample+0.23
CachedSearch+1.67
A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage, wearing a black leather jacket and a long red dress, confident and casual +1.44 gap · LTX-2 19B · 1920×1088 · different seed from full compute
One sample-1.71
CachedSearch-0.28
A gorgeously rendered papercraft coral reef, folded paper fish and sea creatures drifting through layered paper water +1.42 gap · LTX-2 19B · 1920×1088 · same seed as full compute
One sample-0.60
CachedSearch+0.81
A hot rod burning out under stadium lights, tyre smoke billowing +1.42 gap · LTX-2 19B · 1920×1088 · different seed from full compute
One sample-1.95
CachedSearch-0.67
A humpback whale breaching the ocean surface at sunset, backlit spray, slow motion +1.28 gap · LTX-2 19B · 1920×1088 · different seed from full compute
One sample+0.41
CachedSearch+1.60
Extreme slow motion of a dancer spinning in a shaft of dusty light, dress and hair trailing, studio darkness +1.19 gap · LTX-2 19B · 1920×1088 · same seed as full compute
One sample-0.44
CachedSearch+0.69
A snail with a glowing mushroom shell crossing a mossy log at dusk +1.13 gap · LTX-2 19B · 1920×1088 · different seed from full compute
One sample-0.14
CachedSearch+0.93
Several giant woolly mammoths tread across a snowy meadow, long wool blowing in the wind, snow-capped mountains behind them +1.07 gap · LTX-2 19B · 1920×1088 · same seed as full compute
One sample-1.93
CachedSearch-0.89
A snowboarder carving down a steep face, powder exploding behind +1.04 gap · LTX-2 19B · 1920×1088 · same seed as full compute
One sample-0.25
CachedSearch+0.77
A red fox trotting through fresh powder snow in a pine forest at golden hour, shallow depth of field +1.02 gap · LTX-2 19B · 1920×1088 · same seed as full compute
One sample-0.56
CachedSearch+0.37
A cyclist sprinting up an alpine switchback, dust trailing from the wheels +0.93 gap · LTX-2 19B · 1920×1088 · different seed from full compute
One sample-0.63
CachedSearch-0.17
Slow motion water droplet striking a still pond, a perfect crown rising, morning light +0.46 gap · LTX-2 19B · 1920×1088 · different seed from full compute

Wan2.1-T2V 14B 1280×720 · 9 pairs

One sample+0.77
CachedSearch+1.73
A lighthouse beam sweeping through heavy fog over black rocks +0.95 gap · Wan2.1-T2V 14B · 1280×720 · same seed as full compute
One sample+0.07
CachedSearch+0.94
A blacksmith quenching a glowing blade, steam bursting upward +0.87 gap · Wan2.1-T2V 14B · 1280×720 · different seed from full compute
One sample+1.05
CachedSearch+1.85
A samurai in full lacquered armour standing still in falling snow, cherry petals drifting past, cinematic +0.80 gap · Wan2.1-T2V 14B · 1280×720 · different seed from full compute
One sample-0.22
CachedSearch+0.29
An astronaut floating outside a space station with Earth turning below, the blue planet mirrored in the visor +0.51 gap · Wan2.1-T2V 14B · 1280×720 · different seed from full compute
One sample+0.56
CachedSearch+1.68
A monk in maroon robes moves along a low altar in a dim temple, touching a taper to one butter lamp after another until a long row of small flames is burning. Smoke rises in thin threads and gathers under the painted beams above. Hundreds of unlit lamps stretch away behind him. Medium shot, eye level, slow lateral dolly, 50mm lens. Practical light only, from the lamps themselves, warm orange against deep brown shadow. Low key, warm tones, cinematic, high detail. +1.12 gap · Wan2.1-T2V 14B · 1280×720 · same seed as full compute
One sample+1.05
CachedSearch+1.89
A woman in a full kimono walking through a bamboo grove holding a paper umbrella +0.83 gap · Wan2.1-T2V 14B · 1280×720 · different seed from full compute
One sample-1.35
CachedSearch+1.28
A gondola gliding under a stone bridge in early morning mist +2.63 gap · Wan2.1-T2V 14B · 1280×720 · different seed from full compute
One sample-1.36
CachedSearch+0.96
A diver breaking through the surface into a cenote, blue light rising from far below +2.31 gap · Wan2.1-T2V 14B · 1280×720 · same seed as full compute
One sample-1.72
CachedSearch-0.04
A falcon striking a leather lure in mid air, feathers scattering, the falconer blurred behind +1.67 gap · Wan2.1-T2V 14B · 1280×720 · different seed from full compute

HunyuanVideo 13B 1280×720 · 18 pairs

One sample+0.16
CachedSearch+1.73
A red double-decker bus moving through rainy London at night, headlights streaking on wet asphalt +1.57 gap · HunyuanVideo 13B · 1280×720 · same seed as full compute
One sample-0.07
CachedSearch+1.26
Aerial view of a lone red kayak crossing a turquoise glacial lake, pale ice floes drifting past +1.33 gap · HunyuanVideo 13B · 1280×720 · same seed as full compute
One sample-1.19
CachedSearch+0.12
Close-up of a barista pouring latte art, a rosetta forming in the crema as steam rises +1.30 gap · HunyuanVideo 13B · 1280×720 · same seed as full compute
One sample+0.39
CachedSearch+1.53
Macro shot of a hummingbird hovering beside a bright pink flower, wings blurred, slow motion +1.14 gap · HunyuanVideo 13B · 1280×720 · same seed as full compute
One sample-1.26
CachedSearch-0.46
A humpback whale breaching in the open ocean, water sheeting off its body in slow motion +0.80 gap · HunyuanVideo 13B · 1280×720 · different seed from full compute
One sample+1.11
CachedSearch+1.71
A tiger walking through a misty bamboo forest at sunrise, stripes cutting through shafts of light +0.60 gap · HunyuanVideo 13B · 1280×720 · different seed from full compute
One sample+0.90
CachedSearch+1.45
A cheetah sprinting across the savanna in slow motion, dust rising behind its paws +0.54 gap · HunyuanVideo 13B · 1280×720 · different seed from full compute
One sample+1.32
CachedSearch+1.83
A ballerina turning a slow pirouette on a dark stage under a single overhead spotlight +0.51 gap · HunyuanVideo 13B · 1280×720 · same seed as full compute
One sample+1.47
CachedSearch+1.86
An astronaut walking across red desert dunes on Mars, fine dust swirling around the boots +0.39 gap · HunyuanVideo 13B · 1280×720 · same seed as full compute
One sample+1.71
CachedSearch+1.90
A wolf howling on a snowy ridge under green aurora, wide cinematic shot +0.19 gap · HunyuanVideo 13B · 1280×720 · same seed as full compute
One sample-0.37
CachedSearch+0.65
A jellyfish pulsing in dark water, its translucent bell glowing faint blue +1.02 gap · HunyuanVideo 13B · 1280×720 · same seed as full compute
One sample+0.45
CachedSearch+0.84
A bee landing on a lavender stem in macro, pollen visible on its legs +0.39 gap · HunyuanVideo 13B · 1280×720 · different seed from full compute
One sample+1.55
CachedSearch+1.89
A vintage red convertible on a coastal cliff road at sunset, the ocean glittering below +0.34 gap · HunyuanVideo 13B · 1280×720 · same seed as full compute
One sample+1.75
CachedSearch+1.94
A wolf standing on a rock outcrop at dawn, breath steaming, a mist filled valley below +0.20 gap · HunyuanVideo 13B · 1280×720 · different seed from full compute
One sample+0.33
CachedSearch+1.93
A street sweeper in a navy work jacket moves a bamboo broom across the stones of a narrow Kyoto lane in the early morning. Fallen cherry petals gather ahead of the broom in a soft pink line, and a few lift and drift as he sweeps. Wooden shopfronts with closed shutters line both sides of the lane. Medium wide shot, static camera at eye level. Soft overcast dawn light, cool and even, with no hard shadows. Muted colours, quiet atmosphere, realistic style. +1.59 gap · HunyuanVideo 13B · 1280×720 · same seed as full compute
One sample+0.69
CachedSearch+1.45
An ice climber in a red shell jacket swings an ice axe into the blue face of a frozen waterfall. Chips of ice burst away and fall past the camera. Her crampons are set into the wall below and her rope hangs slack beside her. Medium shot from slightly below and to the side, camera static. Cold diffuse daylight from above, the ice glowing internal blue where it is thickest. Blue and white palette, high contrast, realistic style. +0.76 gap · HunyuanVideo 13B · 1280×720 · different seed from full compute
One sample-0.11
CachedSearch+1.18
A firefighter silhouetted against orange flames, an arc of water crossing the frame +1.28 gap · HunyuanVideo 13B · 1280×720 · different seed from full compute
One sample-0.95
CachedSearch+0.85
A shepherd in a brown waxed coat and a black and white border collie move a flock of sheep along a stone walled lane. The dog runs low and wide along the flank, turning the leaders back into the lane. Dust and loose wool hang in the air behind the flock. Wide shot from the front, camera tracking slowly backward. Warm low side light from a setting sun, long shadows across the track. Golden tones, pastoral atmosphere, realistic style. +1.79 gap · HunyuanVideo 13B · 1280×720 · different seed from full compute

CogVideoX 5B 720×480 · 9 pairs

One sample-1.84
CachedSearch+0.43
A chameleon painting a canvas with its tail, colours shifting on its skin +2.26 gap · CogVideoX 5B · 720×480 · different seed from full compute
One sample-0.71
CachedSearch+1.42
A field of sunflowers turning as storm clouds sweep overhead +2.12 gap · CogVideoX 5B · 720×480 · same seed as full compute
One sample-1.57
CachedSearch-0.02
A goldfish swimming through a flooded city street, sunlight rippling on the walls +1.55 gap · CogVideoX 5B · 720×480 · different seed from full compute
One sample+0.78
CachedSearch+1.79
A lantern-lit wooden boat drifting through a flooded forest at dusk +1.01 gap · CogVideoX 5B · 720×480 · same seed as full compute
One sample+0.68
CachedSearch+1.56
A lantern festival at night, hundreds of paper lanterns rising into the dark sky +0.88 gap · CogVideoX 5B · 720×480 · same seed as full compute
One sample+1.15
CachedSearch+1.94
A hot air balloon rising past a waterfall in a jungle gorge, mist drifting +0.79 gap · CogVideoX 5B · 720×480 · same seed as full compute
One sample+1.07
CachedSearch+1.76
A snow globe with a tiny village inside, snow swirling as it settles +0.69 gap · CogVideoX 5B · 720×480 · different seed from full compute
One sample+1.43
CachedSearch+1.81
A red panda eating bamboo in a misty forest +0.38 gap · CogVideoX 5B · 720×480 · same seed as full compute
One sample+1.42
CachedSearch+1.72
A flamingo standing on one leg on a surfboard riding a small wave +0.30 gap · CogVideoX 5B · 720×480 · different seed from full compute

Wan2.1-T2V 1.3B 832×480 · 13 pairs

One sample-1.36
CachedSearch+1.91
An astronaut feeding ducks on a sunny afternoon, reflection from the water. +3.26 gap · Wan2.1-T2V 1.3B · 832×480 · same seed as full compute
One sample-1.58
CachedSearch+1.67
An astronaut planting a flag on a red rocky planet, dust drifting +3.25 gap · Wan2.1-T2V 1.3B · 832×480 · same seed as full compute
One sample-1.69
CachedSearch+0.99
Morning fog lifting off a mirror-calm alpine lake +2.68 gap · Wan2.1-T2V 1.3B · 832×480 · same seed as full compute
One sample-1.97
CachedSearch+0.33
An artist brush painting on a canvas close up. +2.30 gap · Wan2.1-T2V 1.3B · 832×480 · same seed as full compute
One sample-0.41
CachedSearch+1.89
A capybara soaking in a steaming hot spring while snow falls around it +2.30 gap · Wan2.1-T2V 1.3B · 832×480 · same seed as full compute
One sample-0.16
CachedSearch+1.87
A stop-motion clay dinosaur stomping through a paper city +2.03 gap · Wan2.1-T2V 1.3B · 832×480 · same seed as full compute
One sample-0.93
CachedSearch+1.06
A surfer riding inside a large barrel wave +1.99 gap · Wan2.1-T2V 1.3B · 832×480 · same seed as full compute
One sample-0.01
CachedSearch+1.94
A panda standing on a surfboard in the ocean in sunset. +1.95 gap · Wan2.1-T2V 1.3B · 832×480 · same seed as full compute
One sample-1.65
CachedSearch+0.25
A space shuttle launching into orbit, with flames and smoke billowing out from the engines +1.90 gap · Wan2.1-T2V 1.3B · 832×480 · different seed from full compute
One sample-0.23
CachedSearch+1.64
a motorcycle gliding through a snowy field +1.86 gap · Wan2.1-T2V 1.3B · 832×480 · different seed from full compute
One sample-0.37
CachedSearch+1.43
Fireworks exploding over a city skyline reflected in a river +1.80 gap · Wan2.1-T2V 1.3B · 832×480 · same seed as full compute
One sample-1.09
CachedSearch+0.71
A tropical beach at sunrise, with palm trees and crystal-clear water in the foreground +1.80 gap · Wan2.1-T2V 1.3B · 832×480 · same seed as full compute
One sample-2.06
CachedSearch+0.08
An old film projector flickering in a dark room +2.13 gap · Wan2.1-T2V 1.3B · 832×480 · same seed as full compute

Citation

@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}
}