Pith. sign in

REVIEW 3 major objections 4 minor 50 references

Lossy caching preserves candidate rankings in video test-time search, so search can explore cheaply and commit only the winner at full compute.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 03:24 UTC pith:I6KOPGEB

load-bearing objection Good empirical study of whether cached rollouts preserve verifier rankings for video test-time search, but the 'verifier-agnostic' framing overreaches and the headline numbers are tied to ImageReward. the 3 major comments →

arxiv 2607.23159 v2 pith:I6KOPGEB submitted 2026-07-25 cs.AI cs.CVcs.MM

CachedSearch: Training-Free Cached Exploration for Test-Time Search in Video Diffusion

Shreshth Saini , Neil Birkbeck , Yilin Wang , Balu Adsumilli , Alan C. Bovik This is my paper
classification cs.AI cs.CVcs.MM
keywords test-time searchvideo diffusiontraining-free cachingranking preservationbest-of-N searchverifier-guided generationexplore-cheap commit-fullseed determinism
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that training-free caching, which makes each video-diffusion rollout about twice as fast, does not corrupt the candidate rankings that test-time search depends on. This fact is turned into a concrete method: explore all candidates with aggressive caching, then regenerate only the winning seed at full compute. If correct, the main cost barrier to test-time video search—2 to 10 times the cost of a single sample—drops by roughly half while retaining most of the quality gain, and the benefit grows as the search width increases. The claim is supported by a seed-matched protocol: cached and full rollouts from the same seeds are both scored by a reward verifier, yielding high per-prompt rank correlation and errors that concentrate where candidates are nearly tied, making wrong picks cheap. The paper argues the result transfers across model families after recalibrating one caching threshold and stacks multiplicatively with pruning-based search.

Core claim

The central discovery is that within-prompt verifier rankings of video-diffusion candidates survive aggressive training-free caching, and therefore search can be split into cheap exploration and full-compute delivery. On Wan2.1-T2V-1.3B with a roughly two-fold faster adaptive caching wrapper, ImageReward scores of seed-matched cached and full rollouts have median per-prompt Spearman rank correlation 0.905 and 72% top-1 agreement on the VBench suite; VBench-2.0 replicates this with 0.881. Recomputing the cached winner at full compute retains 90–94% of the full-search reward gain, and the residual errors concentrate on near-tied candidates, making corruption self-limiting. CachedSearch operati

What carries the argument

The load-bearing mechanism is the adaptive transformation-vector cache wrapper. Instead of caching transformer outputs, it caches the transformation vector Δ = vθ(x_ref) − x_ref; a skipped call returns x + Δ, and an accumulated relative-input-drift counter decides when to recompute, with warmup and cooldown windows and independent state per classifier-free-guidance branch. This perturbs seed-matched trajectories rather than sampling a different distribution, which the paper argues is exactly what search needs. The second piece is the explore-cheap/commit-full protocol: cached rollouts serve as drafts for ranking; the winning seed is regenerated at full compute, and seed determinism makes the

Load-bearing premise

The headline numbers assume ImageReward's frame-level scores are a faithful proxy for the value users care about; if a deployment's verifier does not share ImageReward's biases—especially its tolerance of motion dampening—the measured 94.7% gain capture and the ranking-preservation conclusion could change materially.

What would settle it

Run the same seed-matched protocol with a verifier that penalizes reduced motion, such as a motion-aware reward trained on human preference or direct human ratings of cached versus full winners, and measure median per-prompt Spearman correlation and N=8 gain capture on the VBench suite; if median ρ falls below roughly 0.7 or capture below 85%, the claim that caching preserves search rankings would be refuted for that verifier.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Test-time video search can run at roughly half the cost while keeping close to 95% of the best-of-N quality gain, because only the winning seed is recomputed at full precision.
  • At a fixed budget, cached exploration searches about twice as wide, with a measured 38% reward gain over full-compute best-of-4.
  • The protocol is orthogonal to the verifier and the search algorithm, so it can be layered onto existing pruning-based search; measured stacking yields a 3.11x exploration speedup at 88.6% gain capture.
  • Ranking corruption is self-limiting: errors concentrate on near-tied prompts where a wrong pick costs little, and even the corrupted tail retains roughly 68% of attainable gain.
  • A 25-prompt pilot suffices to recalibrate the cache threshold for a new model family, giving a cheap transfer recipe across architectures.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The ranking-plus-regret audit used here could be applied to any lossy accelerator before composing it with search; the paper's own comparisons show truncated or downscaled drafts preserve far less ranking, suggesting trajectory-preserving approximations are the right class to use for exploration.
  • The finding that both ImageReward and VideoScore prefer motion-dampened cached twins implies learned verifiers may systematically reward reduced motion; deployments with motion-critical prompts should audit with optical flow and LPIPS rather than trusting reward-model scores alone.
  • Because gain capture rises with search width and approaches roughly 95%, the method is most valuable precisely where full-compute search is most expensive, making it a natural companion to wider or more aggressive search budgets.
  • The rejection of a two-sample adaptive threshold leaves open the possibility that a free signal from the cache's internal accumulated-drift indicator could outperform a fixed threshold; that is a testable extension the paper does not pursue.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper asks whether training-free caching of video-diffusion rollouts preserves verifier rankings well enough to support test-time best-of-N search, and proposes CachedSearch: explore all candidates under aggressive caching, score the drafts, then re-render only the winning seed at full compute. On Wan2.1-T2V-1.3B with an EasyCache-style adaptive wrapper at τ=0.10, seed-matched cached and full rollouts scored by ImageReward yield median per-prompt Spearman ρ=0.905 and 72% top-1 agreement on VBench; VBench-2.0 replicates the pattern. The method is reported to retain 94.7% of best-of-8 reward gain at 63.3% of the cost. The authors argue that ranking errors concentrate on near-tied candidates, making corruption self-limiting, and that the protocol transfers across six models and four families after τ recalibration. Appendix B develops a Gaussian-copula ranking-noise model, and Appendix D includes extensive robustness checks including a second verifier, batching controls, pruning composition, and temporal metrics.

Significance. If the headline result holds, this is a substantial and well-executed empirical contribution: it is the first direct study of whether caching corrupts candidate ranking in video test-time search, and the explore-cheap/commit-full template is practically valuable. The seed-matched paired-rollout protocol, the three benchmark suites, the six-model breadth, the bit-exact determinism controls, and the direct temporal auditing are real strengths. The claims are clearly separated in many places: the authors openly report the LTX boundary, the nominal nature of keep-draft capture, and the conservative behavior of the Appendix B model. However, the external-validity claims outrun the evidence. The measured ranking-preservation and capture numbers are verifier-specific, and the paper's own VideoScore audit substantially weakens the 'verifier-agnostic' claim. The generality claim across six models also needs qualification because one of the six models cannot be calibrated into the operating band. These are fixable by reframing and additional reporting rather than by new theory.

major comments (3)
  1. [§D.6, §5.5, Abstract] The abstract and Section 1 label CachedSearch 'verifier-agnostic,' and Section 5.5 summarizes Appendix D.6 as bounding 'caching-attributable cross-verifier loss at −2.5 points.' But D.6 reports that on the same rollouts, VideoScore-v1.1 gives median per-prompt ρ=0.762, top-1 agreement 54%, and gain capture 65.6%, versus 0.905, 72–73%, and ~90% under ImageReward. Those are materially different cached-vs-full ranking-fidelity numbers. The −2.5-point comparison is the full-compute VideoScore gain of the ImageReward-selected cached winner relative to the ImageReward-selected full winner; it does not measure CachedSearch's ranking preservation when VideoScore itself is the deployment verifier. Since VideoScore also prefers motion-dampened cached outputs (Appendix F), no tested verifier is independent of the artifact. The central claim is therefore established for ImageReward only. I recommend
  2. [§5.6, Table 11, Abstract] The abstract says the result 'holds from 1.3B–14B across six models and four families' and the text says ports require recalibrating a single parameter. Table 11 shows that LTX-Video-2B cannot reach 85% capture at any measured threshold: its best operating point is 79.6% at τ=0.02, and at fixed τ=0.10 its median ρ is 0.536 with 67.6% capture (Table 5). CogVideoX and Hunyuan also need τ*=0.05 and land near 85%, while the Wan family qualifies at higher τ. Thus the mechanism transfers, but the statement that a single recalibrated parameter places every family in the operating band is contradicted by the paper's own boundary case. This is load-bearing because Figure 1(c) and the abstract's cross-model generality are headline claims; they should be narrowed to 'the audit protocol transfers; five of six models can be calibrated into the band.'
  3. [Table 1, §4.1] The paper states that all uncertainty uses a prompt bootstrap with B=10^4, but Table 1, which carries the headline 94.7% capture and 63.3% cost, reports only point estimates. No confidence interval is given for the N=8 commit row, for the 38% iso-cost gain, or for the width extensions (95.7%, 95.2%). The paper itself notes that gate-grid and VBench-suite point estimates are statistically consistent despite numerical differences. Without intervals, differences such as N=8 94.7% versus N=16 95.7% cannot be evaluated as support for 'capture rises with width.' Since these are central empirical claims based on n=50 prompts and ratio-of-means estimators, I ask that intervals accompany all headline capture, cost, and gain numbers.
minor comments (4)
  1. [Table 6, §B.2] The caption calls the model 'parameter-free,' but the latent correlation r is calibrated from the measured Spearman ρ via Corollary 1. The agreement in Table 7 is therefore a calibration/consistency check, not an out-of-sample prediction. Please adjust the wording to avoid implying independence.
  2. [§D.6] The ImageReward top-1 agreement is quoted as 72% in Section 4.5 and 73% in D.6 for the validated subset. Please align the numbers or explain the subset difference.
  3. [Figure 8, §5.2] The frontier capture(%) = 104 − 19.1 ln(speedup) is fitted to six caching points and called a 'law' in places. Given small n and the paper's own caution about per-family frontiers, I suggest labeling this a descriptive fit rather than a universal law.
  4. [§5.4] The self-limiting argument uses full-compute score spread as the explanatory variable. It may be worth stating explicitly that spread is not known at exploration time without an extra scoring pass, which is presumably why the adaptive threshold in D.3 is tested and rejected.

Circularity Check

0 steps flagged

No load-bearing circularity: headline ranking and capture numbers are direct measurements; the only calibration loop is an explicitly labeled conservative appendix model.

full rationale

The paper's central claims are direct measurements. Median Spearman rho=0.905, 72% top-1 agreement, 94.7% gain capture, and 63.3% cost are computed from seed-matched cached/full rollout pairs under a specified verifier (ImageReward) on three suites and six models; these are empirical quantities, not derived from the method's own assumptions. CachedSearch itself is a protocol; its cost identities (Eqs. 6-7) use measured latencies, and recommit is guaranteed by seed determinism, not by a circular argument. The rank-noise model in Appendix B calibrates r from the same Spearman rho and then predicts capture=r under a Gaussian-copula assumption; if read as an independent prediction this would be nearly tautological, but the paper explicitly presents it as a conservative consistency check ('We keep the headline empirical and use the model only as a conservative envelope'), and the headline numbers are measured, so it is not load-bearing. The capture-versus-speedup fit is descriptive and not used to establish the method. The self-citations ([30], [31]) are contextual and do not justify the central premise. The abstract's 'verifier-agnostic' claim conflicts with the paper's own VideoScore audit (median rho=0.762, 65.6% capture in D.6) and with Appendix J's admission that VideoScore shares the motion-dampening preference; that is an external-validity/overclaim issue, not circularity. No definitional reduction, fitted-parameter-renamed-as-prediction, or load-bearing self-citation chain was found.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

No new physical or theoretical entities are introduced; CachedSearch is a protocol rather than a new force, particle, or conserved quantity. The central claim rests on measured constants (tau, gamma, rho), a Gaussian-copula model, seed determinism, and the validity of the chosen verifier as value.

free parameters (5)
  • cache threshold tau = 0.10 default; per-model tau* = 0.20 (Wan2.1-1.3B), 0.10 (Wan2.2-5B, Wan2.1-14B), 0.05 (CogVideoX-5B, Hunyuan-13B); no qu
    Controls how aggressively the wrapper skips transformer calls. The paper recalibrates it per model family to keep >=85% gain capture, so all headline numbers depend on this chosen operating point.
  • warmup/cooldown steps Kw=Kc=5 = 5
    Hand-picked always-compute windows that affect the skip schedule and thus speed/fidelity. Not swept, and not justified by a separate experiment.
  • verifier frame count K=8 = 8
    ImageReward is averaged over 8 uniformly spaced frames. Ranking values are relative to this evaluation choice, though the paper audits with VideoScore and VQAScore.
  • latent rank correlation r = 0.913 (median-calibrated), 0.833 (mean-calibrated)
    Appendix B maps measured Spearman rho to a Gaussian-copula correlation r and uses it in Propositions 1 and 5. It is estimated from the same seed-matched data and used to predict capture and iso-cost advantage, but not fitted to the headline capture.
  • frontier coefficients for capture-vs-speedup = capture = 104.2 - 19.1 ln(speedup), R^2 = 0.92 (Wan-1.3B); per-family variants
    This is a descriptive fit to six measured caching operating points. It is called a 'law' in places but is not needed for the main CachedSearch results.
axioms (6)
  • domain assumption Deterministic ODE solver: for fixed (c,s,tau), cached and full rollouts are deterministic, so recommit from a seed exactly reproduces the full-compute sample.
    Section 2 states the ODE solver is deterministic and the seed fully determines the sample; Section 3.2 relies on this for the exactness of commit. A bit-exact no-cache control is used as evidence, but determinism remains an implementation/hardware assumption.
  • domain assumption Assumption 1: candidate pairs (S_i, S_hat_i) are i.i.d. across seeds within a prompt.
    Appendix B.1 uses exchangeability to derive regret and capture formulas. Seeds are drawn independently, but prompt-level heterogeneity means this is an idealization.
  • ad hoc to paper Assumption 2: within a prompt, (S_i, S_hat_i) follows a bivariate Gaussian copula.
    Appendix B.1 introduces this as a ranking-noise model. The paper reports it is conservative in the observed direction, but it is not validated against the full score distribution.
  • domain assumption Assumption 3: the full verifier score S is taken as the value of a candidate.
    Appendix B.1 states verifier mis-specification is treated separately. All gain and capture numbers in the paper inherit this value assumption.
  • domain assumption Public checkpoints and the wrapper reproduce the measured latencies and scores on NVIDIA GH200 hardware.
    All experimental results depend on public model checkpoints and a single GPU environment; this cannot be independently verified from the manuscript alone.
  • domain assumption The Wan2.1-1.3B cost model measured at 480x832, 81 frames, T=50, batch size 1 is representative of the speedup ratio gamma ~ 0.51.
    Section 3.3 uses this cost model; Section D.5 shows batching does not help for Wan, but other models and hardware may have different gamma values.

pith-pipeline@v1.3.0-alltime-deepseek · 38693 in / 13603 out tokens · 133709 ms · 2026-08-01T03:24:51.925011+00:00 · methodology

0 comments
read the original abstract

Test-time search lets small video diffusion models rival larger ones, but costs 2-10x more. All candidates are fully denoised, although most are discarded. Training-free caching makes each rollout 2-3x faster at near-lossless quality. Composition is safe only if lossy caching preserves verifier rankings. We present the first study of whether caching corrupts candidate ranking in video test-time search. On Wan2.1-T2V-1.3B with an adaptive caching wrapper (~2x per-candidate speedup), ImageReward scores seed-matched cached and full rollouts. Median per-prompt Spearman rank correlation is 0.905, with 72% top-1 agreement on the VBench suite. VBench-2.0 replicates this result on a harder suite. Recomputing the cached winner at full compute retains 90-94% of the full-search gain. Errors cluster among near-tied candidates, making corruption self-limiting. This finding leads to CachedSearch. It explores every candidate with aggressive caching, then re-generates only the winner at full compute. At N=8, it captures 94.7% of best-of-N's gain at 63% of the cost. Capture rises with width. At matched budget, it searches twice as wide for 38% more gain. The result holds from 1.3B-14B across six models and four families: Wan, LTX, CogVideoX, and Hunyuan. Wan2.1-14B matches the 1.3B model's fidelity. Mid-trajectory pruning multiplies the exploration saving to 3.11x at 88.6% capture. Ports to other model families require recalibrating a single parameter, showing that fidelity tracks architecture rather than parameter count. CachedSearch is training-free, verifier-agnostic, and orthogonal to the search algorithm, making it a plug-in multiplier for test-time scaling.

Figures

Figures reproduced from arXiv: 2607.23159 by Alan C. Bovik, Balu Adsumilli, Neil Birkbeck, Shreshth Saini, Yilin Wang.

Figure 1
Figure 1. Figure 1: CachedSearch explores cheaply and commits at full compute. It keeps most of test￾time search’s quality at a fraction of the cost and generalizes across models. (a) Cached and full-compute rollouts of the same seeds rank candidates alike (Wan2.1-T2V-1.3B [34], τ=0.10; each dot a candidate, one prompt’s 8 highlighted): median per-prompt Spearman ρ = 0.905, recurring on three independent suites. (b) At N=8, c… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of CachedSearch. Cache every candidate, score the drafts, then re-generate the winning seed at full compute. the probability-flow ODE over a decreasing grid of T timesteps 1 = t1 > · · · > tT > tT +1 = 0 with an Euler-type update, xtk+1 = xtk + (tk+1 − tk) ˜vk, (1) followed by VAE decoding of xtT +1 into pixels. With classifier-free guidance (CFG) [10] at scale w, each step evaluates the transform… view at source ↗
Figure 3
Figure 3. Figure 3: Within-prompt candidate ranking survives caching: median Spearman ρ = 0.905. Cached vs. full scores for 50 prompts and 8 seeds at τ = 0.10. Color encodes within-prompt score spread; each dot is one measured rollout pair. 0 10 20 30 prompts τ = 0.05 med 0.90 p10 0.74 median p10 0 10 20 τ = 0.10 med 0.90 p10 0.61 0.0 0.5 1.0 per-prompt ρ 0 10 20 prompts τ = 0.20 med 0.86 p10 0.47 0.0 0.5 1.0 per-prompt ρ 0 2… view at source ↗
Figure 5
Figure 5. Figure 5: Above break-even, every caching engine beats full-compute best￾of-N. Delivered reward vs. end-to-end budget on the 50-prompt gate grid. Markers denote N = 2, 4, 8; step truncation is the matched￾cost alternative. 4.2 Candidate-ranking preservation On the VBench suite, median within-prompt Spearman correlation is ρ = 0.905, top-1 agreement is 72%, and p10 is 0.690. The gate pilot gives the same median, with… view at source ↗
Figure 6
Figure 6. Figure 6: Regret stays far below random pick￾ing at every caching level. Per-prompt regret CDFs at N = 8 against pooled random baselines. Dotted curves show the VBench [12] suite; zero re￾gret occurs on 64% of gate prompts and 72% offi￾cial. 0.2 0.4 0.6 0.8 1.0 1.2 candidate score spread (full) 0.0 0.2 0.4 0.6 0.8 1.0 ran king preservation ρ corr(spread, ρ) = +0.25 ρ < 0.7: 9/50 prompts regret 0 0.3 0.6 [PITH_FULL_… view at source ↗
Figure 8
Figure 8. Figure 8: Caching engines share one capture￾vs-speedup frontier. Filled markers show PAB [44], CFG-Cache [25], TeaCache [23], and Easy￾Cache [48]; blue points sweep τ and × marks the independent default rerun. The dashed fit covers six caching points (R2 = 0.92), each with n = 50 prompts. 300 400 500 600 wall-clock cost per prompt (s), N = 8 88 90 92 94 96 98 100 102 104 delivered gain (% of best-of-8's) full best-o… view at source ↗
Figure 10
Figure 10. Figure 10: One frontier fits reuse choices within a backbone, but architecture shifts the frontier. Gain capture vs. candidate speedup; up-left is better. Panel A varies reuse axes on Wan2.1-1.3B [34]; Panel B sweeps τ across six models. Each point has 44–50 prompts. At fixed τ = 0.10, [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: CachedSearch dominates best-of￾N beyond the N∗ ≈ 2 break-even. Reward gain vs. wall-clock; up-left is better. Curves compare full, keep, and commit for N = 2, 4, 8 on 50 prompts; the gold diamond adds pruning. 1.4 1.6 1.8 2.0 2.2 2.4 exploration speedup (×) 88 89 90 91 92 93 94 search-gain capture (%) τ = 0.05 τ = 0.10 fixed τ τ = 0.20 adaptive (probe K = 2) [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Commit matches best-of-8 quality at 61% of its transformer FLOPs. Mean ImageRe￾ward [39], VideoScore [9], and same-seed LPIPS [43] on the held-out VBench [12] subset; bars are 95% prompt-bootstrap intervals [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Commit tracks best-of-8 across the full metric suite. Four strategies on the held-out VBench [12] subset, with raw metric scales and 95% prompt-bootstrap intervals. Dotted lines mark best￾of-8; the last column reports wall-clock and transformer FLOPs. verifier, matching the +0.037 bias in Section 5.5. Yet keep has LPIPS [43] 0.114, aesthetic −0.012 , and imaging −1.0 . Its temporal signature is mild at th… view at source ↗
Figure 15
Figure 15. Figure 15: Schedule length, not pixel count, controls redundancy. Skip fraction (left) and gain capture (right) vs. denoising steps at τ = 0.10 and n = 50. Dashed lines are log-linear trends with bootstrap bands; open diamonds show the 720p contrast. threshold (Section 5.1) and architecture (Section 5.6). Finer schedules take smaller steps, so more updates fall below the same drift threshold. From 25 to 100 steps, s… view at source ↗
Figure 16
Figure 16. Figure 16: Delivered capture rises toward 95% with search width while top-1 falls. Width scaling at τ = 0.10 on 50 prompts; open marks are seed-subset simulations through N = 16, and N = 32 is measured directly. The gray band is the rank-noise prediction. below cached-only exploration (278 s), i.e. the measured 1.96× (caching) and 1.59× (pruning) factors multiply to the observed 3.1×; delivering the recommitted winn… view at source ↗
Figure 17
Figure 17. Figure 17: Fidelity tracks architecture family, while absolute savings track model cost. Median ρ (A) and seconds saved per candidate (B) at τ = 0.10, with 8 seeds and 50 prompts per model [PITH_FULL_IMAGE:figures/full_fig_p028_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Search yields visible gains, and cached exploration finds the same winner cheaper. Selected VBench [12] prompts compare single, full best-of-8, and commit at τ = 0.10. Chips give measured ImageReward [39] and seed; red boxes mark discriminating regions. Protocol, seeds, verifier. Grids: 50 prompts × seeds 0–7 × {full, cached}, i.e. 800 scored rollouts per model at τ=0.10, plus 400 cached rollouts per addi… view at source ↗
Figure 19
Figure 19. Figure 19: Extended qualitative comparisons confirm the same search behavior. Additional VBench [12] prompts under the [PITH_FULL_IMAGE:figures/full_fig_p034_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: All four delivery strategies compared under measured budgets. Selected VBench [12] prompts form rows; columns are single, best-of-8, keep, and commit. Headers give cost relative to full best-of-8; corner chips give measured delivered ImageReward [39]. I.3 Cross-model strips [PITH_FULL_IMAGE:figures/full_fig_p034_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Calibrated backbones track their full-compute twins; over-driven LTX [7] visibly drifts. One shared prompt and seed across six models at τ = 0.10, with full and cached rows under each model-card recipe. Chips report ImageReward [39] and reused denoising steps. J Limitations We use ImageReward [39] as the verifier. A better temporally aligned reward model would be preferable, but no publicly available one … view at source ↗
Figure 22
Figure 22. Figure 22: Cached twins preserve subject, layout, and palette; differences concentrate in detail and motion phase. Four same-seed winner pairs at τ = 0.10, with two frames per video. Chips report measured ImageReward [39] and pairwise LPIPS [43]. Full mean |flow| 7.04 Cached mean |flow| 4.79 τc = 0.20: −32% motion mean |flow| map Full mean |flow| 2.78 Cached mean |flow| 2.45 τc = 0.20: −12% motion ``An eagle soaring… view at source ↗
Figure 23
Figure 23. Figure 23: Flow maps expose motion dampening that learned judges miss. Two same-seed, high￾motion winner pairs at τ = 0.20: full compute above, cached below. Each block shares the color scale for temporal-mean Farneback flow magnitude. 36 [PITH_FULL_IMAGE:figures/full_fig_p036_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Static content is preserved under keep-draft; motion is dampened (−8% mean flow at this τ , Section 5.5). Keep (top rows) vs. commit (bottom rows) winner videos at τ = 0.20, four evenly spaced frames; adversarial selection: pairs chosen by largest mean frame difference among the 50 winner pairs. 37 [PITH_FULL_IMAGE:figures/full_fig_p037_24.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 30 linked inside Pith

  1. [1]

    Recache: Learning budget-aware caching schedules for diffusion models via reinforce

    Mishan Aliev, Eva Neudachina, Ilya Bykov, Aleksandr Oganov, Kirill Struminsky, Aibek Alanov, and Denis Rakitin. Recache: Learning budget-aware caching schedules for diffusion models via reinforce. arXiv preprint arXiv:2606.06060 , 2026

    Pith/arXiv arXiv 2026

  2. [2]

    Flowcast: Trajectory forecasting for scalable zero-cost speculative flow matching

    Divya Jyoti Bajpai, Shubham Agarwal, Apoorv Saxena, Kuldeep Kulkarni, Subrata Mitra, and Man- jesh Kumar Hanawal. Flowcast: Trajectory forecasting for scalable zero-cost speculative flow matching. arXiv preprint arXiv:2602.01329 , 2026. ICLR 2026

    arXiv 2026

  3. [3]

    Bwcache: Accelerating video diffusion transformers through block-wise caching

    Hanshuai Cui, Zhiqing Tang, Zhifei Xu, Zhi Yao, Wenyi Zeng, and Weijia Jia. Bwcache: Accelerating video diffusion transformers through block-wise caching. arXiv preprint arXiv:2509.13789 , 2025

    arXiv 2025

  4. [4]

    Accelerated diffusion models via speculative sampling

    Valentin De Bortoli, Alexandre Galashov, Arthur Gretton, and Arnaud Doucet. Accelerated diffusion models via speculative sampling. arXiv preprint arXiv:2501.05370 , 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Lemica: Lexicographic minimax path caching for efficient diffusion-based video generation

    Huanlin Gao, Ping Chen, Fuyuan Shi, Chao Tan, Zhaoxiang Liu, Fang Zhao, Kai Wang, and Shiguo Lian. Lemica: Lexicographic minimax path caching for efficient diffusion-based video generation. arXiv preprint arXiv:2511.00090 , 2025. NeurIPS 2025 Spotlight

    arXiv 2025

  6. [6]

    Blade: Block-sparse attention meets step distillation for efficient video generation

    Youping Gu, Xiaolong Li, Yuhao Hu, Minqi Chen, and Bohan Zhuang. Blade: Block-sparse attention meets step distillation for efficient video generation. arXiv preprint arXiv:2508.10774 , 2025. 12 Technical report

    arXiv 2025

  7. [7]

    Ltx-video: Realtime video latent diffusion

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103 , 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Scaling image and video generation via test-time evolutionary search

    Haoran He, Jiajun Liang, Xintao Wang, Pengfei Wan, Di Zhang, Kun Gai, and Ling Pan. Scaling image and video generation via test-time evolutionary search. arXiv preprint arXiv:2505.17618 , 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Videoscore: Building automatic met- rics to simulate fine-grained human feedback for video generation

    Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chan- dra, Ziyan Jiang, Aaran Arulraj, Kai Wang, Quy Duc Do, Yuansheng Ni, Bohan Lyu, Yaswanth Narsu- palli, Rongqi Fan, Zhiheng Lyu, Yuchen Lin, and Wenhu Chen. Videoscore: Building automatic met- rics to simulate fine-grained human feedback for video generation. arX...

    Pith/arXiv arXiv 2024

  10. [10]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 , 2022

    Pith/arXiv arXiv 2022

  11. [11]

    Speculative decoding for autoregressive video generation

    Yuezhou Hu and Jintao Zhang. Speculative decoding for autoregressive video generation. arXiv preprint arXiv:2604.17397 , 2026

    Pith/arXiv arXiv 2026

  12. [12]

    Vbench: Comprehensive benchmark suite for video genera- tive models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Vbench: Comprehensive benchmark suite for video genera- tive models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2024...

    Pith/arXiv arXiv 2024

  13. [13]

    Forcing-kv: Hybrid kv cache compression for efficient autoregressive video diffusion models

    Yicheng Ji, Zhizhou Zhong, Jun Zhang, Qin Yang, XiTai Jin, Ying Qin, Wenhan Luo, Shuiyang Mao, Wei Liu, and Huan Li. Forcing-kv: Hybrid kv cache compression for efficient autoregressive video diffusion models. arXiv preprint arXiv:2605.09681 , 2026

    Pith/arXiv arXiv 2026

  14. [14]

    Temporal backtracking search for test-time generative video reasoning

    Sejoon Jun, Zheng Ding, Huangyuan Su, Weirui Ye, and Yilun Du. Temporal backtracking search for test-time generative video reasoning. arXiv preprint arXiv:2606.13861 , 2026

    arXiv 2026

  15. [15]

    Hunyuanvideo: A systematic framework for large video generative models

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603 , 2024

    Pith/arXiv arXiv 2024

  16. [16]

    William H. Kruskal. Ordinal measures of association. Journal of the American Statistical Association , 53(284):814–861, 1958

    1958

  17. [17]

    Early failure detec- tion and intervention in video diffusion models

    Byung-Ki Kwon, Sohwi Lim, Hyeon-Woo Nam, Ye-Bin Moon, and Tae-Hyun Oh. Early failure detec- tion and intervention in video diffusion models. arXiv preprint arXiv:2603.14320 , 2026

    arXiv 2026

  18. [18]

    Sol video inference engine: Agent-native full-stack acceleration framework for efficient video generation

    Yitong Li, Junsong Chen, Haopeng Li, Haozhe Liu, Jincheng Yu, Ligeng Zhu, Ping Luo, Song Han, and Enze Xie. Sol video inference engine: Agent-native full-stack acceleration framework for efficient video generation. arXiv preprint arXiv:2606.23743 , 2026

    Pith/arXiv arXiv 2026

  19. [19]

    Evaluating text-to-visual generation with image-to-text generation

    Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision (ECCV) , 2024

    2024

  20. [20]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow match- ing for generative modeling. In International Conference on Learning Representations (ICLR) , 2023

    2023

  21. [21]

    Fastcache: Fast caching for diffusion transformer through learnable linear approximation

    Dong Liu, Yanxuan Yu, Jiayi Zhang, Yifan Li, Ben Lengerich, and Ying Nian Wu. Fastcache: Fast caching for diffusion transformer through learnable linear approximation. arXiv preprint arXiv:2505.20353, 2025

    arXiv 2025

  22. [22]

    Video-t1: Test-time scaling for video generation

    Fangfu Liu, Hanyang Wang, Yimo Cai, Kaiyan Zhang, Xiaohang Zhan, and Yueqi Duan. Video-t1: Test-time scaling for video generation. arXiv preprint arXiv:2503.18942 , 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Timestep embedding tells: It’s time to cache for video diffusion model

    Feng Liu, Shiwei Zhang, Xiaofeng Wang, Yujie Wei, Haonan Qiu, Yuzhong Zhao, Yingya Zhang, Qixiang Ye, and Fang Wan. Timestep embedding tells: It’s time to cache for video diffusion model. arXiv preprint arXiv:2411.19108 , 2024. CVPR 2025. 13 Technical report

    Pith/arXiv arXiv 2024

  24. [24]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In International Conference on Learning Representations (ICLR) , 2023

    2023

  25. [25]

    Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, and Kwan-Yee K. Wong. Fastercache: Training-free video diffusion model acceleration with high quality. arXiv preprint arXiv:2410.19355, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    Navicache: Test-time self-calibration caching for video generation

    Zheqi Lv, Zhibo Zhu, Jinke Wang, Qi Tian, Shengyu Zhang, Zhengyu Chen, Chengxi Zang, Zhou Zhao, and Fei Wu. Navicache: Test-time self-calibration caching for video generation. arXiv preprint arXiv:2606.26795, 2026. ICML 2026

    Pith/arXiv arXiv 2026

  27. [27]

    Inference-time scaling for diffusion models beyond scaling denoising steps

    Nanye Ma, Shangyuan Tong, Haolin Jia, Hexiang Hu, Yu-Chuan Su, Mingda Zhang, Xuan Yang, Yandong Li, Tommi Jaakkola, Xuhui Jia, and Saining Xie. Inference-time scaling for diffusion models beyond scaling denoising steps. arXiv preprint arXiv:2501.09732 , 2025

    Pith/arXiv arXiv 2025

  28. [28]

    Latte: Latent diffusion transformer for video generation

    Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048 , 2024

    Pith/arXiv arXiv 2024

  29. [29]

    T-stitch: Accelerating sampling in pre-trained diffusion models with trajectory stitching

    Zizheng Pan, Bohan Zhuang, De-An Huang, Weili Nie, Zhiding Yu, Chaowei Xiao, Jianfei Cai, and Anima Anandkumar. T-stitch: Accelerating sampling in pre-trained diffusion models with trajectory stitching. arXiv preprint arXiv:2402.14167 , 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Shreshth Saini, Shashank Gupta, and Alan C. Bovik. Rectified-cfg++ for flow based models. In Advances in Neural Information Processing Systems (NeurIPS) , 2025

    2025

  31. [31]

    Shreshth Saini, Ru-Ling Liao, Yan Ye, and Alan C. Bovik. LGDM: Latent guidance in diffusion models for perceptual evaluations. In International Conference on Machine Learning (ICML) , 2025

    2025

  32. [32]

    Hybridstitch: Pixel and timestep level model stitching for diffusion acceleration

    Desen Sun, Jason Hon, Jintao Zhang, and Sihang Liu. Hybridstitch: Pixel and timestep level model stitching for diffusion acceleration. arXiv preprint arXiv:2603.07815 , 2026

    arXiv 2026

  33. [33]

    Stream-t1: Test-time scaling for streaming video generation

    Yijing Tu, Shaojin Wu, Mengqi Huang, Wenchuan Wang, Yuxin Wang, Chunxiao Liu, and Zhendong Mao. Stream-t1: Test-time scaling for streaming video generation. arXiv preprint arXiv:2605.04461 , 2026

    Pith/arXiv arXiv 2026

  34. [35]

    Wan2.2: Open and advanced large-scale video generative models

    Wan Team. Wan2.2: Open and advanced large-scale video generative models. https://github.com/ Wan-Video/Wan2.2, 2025. Open-weights release including Wan2.2-TI2V-5B; technical report shared with Wan2.1 (arXiv:2503.20314)

    Pith/arXiv arXiv 2025

  35. [36]

    Bovik, Hamid R

    Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing , 13(4):600–612,

  36. [37]

    Williams

    Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3–4):229–256, 1992. doi: 10.1007/BF00992696

    doi:10.1007/bf00992696 1992

  37. [38]

    Quantcache: Adaptive importance-guided quantization with hierarchical latent and layer caching for video genera- tion

    Junyi Wu, Zhiteng Li, Zheng Hui, Yulun Zhang, Linghe Kong, and Xiaokang Yang. Quantcache: Adaptive importance-guided quantization with hierarchical latent and layer caching for video genera- tion. arXiv preprint arXiv:2503.06545 , 2025

    Pith/arXiv arXiv 2025

  38. [39]

    Imagereward: Learning and evaluating human preferences for text-to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems (NeurIPS) , 2023. arXiv:2304.05977

    Pith/arXiv arXiv 2023

  39. [40]

    Vision- reward: Fine-grained multi-dimensional human preference learning for image and video generation

    Jiazheng Xu, Yu Huang, Jiale Cheng, Yuanming Yang, Jiajun Xu, Yuan Wang, Wenbo Duan, Shen Yang, Qunlin Jin, Shurun Li, Jiayan Teng, Zhuoyi Yang, Wendi Zheng, Xiao Liu, Dan Zhang, Ming Ding, Xiaohan Zhang, Xiaotao Gu, Shiyu Huang, Minlie Huang, Jie Tang, and Yuxiao Dong. Vision- reward: Fine-grained multi-dimensional human preference learning for image and...

    Pith/arXiv arXiv 2024

  40. [41]

    Cogvideox: Text-to-video diffusion models with an expert transformer

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072 , 2024

    Pith/arXiv arXiv 2024

  41. [42]

    Gonzalez, Jianfei Chen, and Jun Zhu

    Jintao Zhang, Kaiwen Zheng, Kai Jiang, Haoxu Wang, Ion Stoica, Joseph E. Gonzalez, Jianfei Chen, and Jun Zhu. Turbodiffusion: Accelerating video diffusion models by 100-200 times. arXiv preprint arXiv:2512.16093, 2025

    arXiv 2025

  42. [43]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 586–595, 2018. arXiv:1801.03924

    Pith/arXiv arXiv 2018

  43. [44]

    Real-time video generation with pyramid attention broadcast

    Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast. arXiv preprint arXiv:2408.12588 , 2024. ICLR 2025

    Pith/arXiv arXiv 2024

  44. [45]

    Latsearch: Latent reward-guided search for faster inference-time scaling in video diffusion

    Zengqun Zhao, Ziquan Liu, Yu Cao, Shaogang Gong, Zhensong Zhang, Jifei Song, Jiankang Deng, and Ioannis Patras. Latsearch: Latent reward-guided search for faster inference-time scaling in video diffusion. arXiv preprint arXiv:2603.14526 , 2026. ECCV 2026

    Pith/arXiv arXiv 2026

  45. [46]

    Vbench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness

    Dian Zheng, Ziqi Huang, Hongbo Liu, Kai Zou, Yinan He, Fan Zhang, Lulu Gu, Yuanhan Zhang, Jing- wen He, Wei-Shi Zheng, Yu Qiao, and Ziwei Liu. Vbench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness. arXiv preprint arXiv:2503.21755 , 2025

    Pith/arXiv arXiv 2025

  46. [47]

    Open-sora: Democratizing efficient video production for all

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404, 2024

    Pith/arXiv arXiv 2024

  47. [48]

    Less is enough: Training-free video diffusion acceler- ation via runtime-adaptive caching

    Xin Zhou, Dingkang Liang, Kaijin Chen, Tianrui Feng, Xiwu Chen, Hongkai Lin, Yikang Ding, Feiyang Tan, Hengshuang Zhao, and Xiang Bai. Less is enough: Training-free video diffusion acceler- ation via runtime-adaptive caching. arXiv preprint arXiv:2507.02860 , 2025. 15 Technical report Appendix contents A Extended related work . . . . . . . . . . . . . . ....

    Pith/arXiv arXiv 2025

  48. [50]

    model sd

    gates reuse at DiT-block granularity through relative- L1 similarity. FastCache learns a linear approxi- mation of cached features, and NaviCache self-calibrates its schedule [ 21, 26]. LeMiCa [ 5] casts scheduling as a shortest-path problem and uses lexicographic minimax optimization to bound globally accumulated error. ReCache [ 1] instead learns budget...

  49. [51]

    Redundancy and fidelity do not track parameter count

    and LTX-Video-2B [ 7] also lose ranking fidelity at fixed τ (capture 79.9% / 67.6%), while matching or exceeding Wan speedups. Redundancy and fidelity do not track parameter count. We therefore measure r for each model. A cheap pilot needs one prompt set and two rollout modes, exactly the calibration in Appendix B. Even CogVideoX’s uncalibrated r retains ...

  50. [2004]

    doi: 10.1109/TIP.2003.819861

    arXiv 2003