acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Diffusion Caching: From Historical Extrapolation to Current-Observation Completion

Abstract

Training-free diffusion caching reconstructs skipped features from historical evaluations. Its accuracy depends on whether the cached directions can represent the current residual motion. We identify a representation bottleneck: on Wan2.1, the latest input secant captures of input displacement energy, while the latest residual secant captures of the deployed CFG-weighted residual displacement energy. We characterize the motion that a fixed historical basis cannot represent and use oracle reconstruction to measure this limitation. This motivates MoorCache, which combines historical prediction with current-observation completion. History propagates persistent motion, and an affordable current observation supplies complementary directions through a projected displacement update. Same-cost ablations demonstrate the value of this completion, while projection controls test how current information should be combined with history. Across Wan2.1, HunyuanVideo, DiT-XL/2, and FLUX.1-dev, MoorCache improves quality–efficiency trade-offs with observation and reconstruction costs. On HunyuanVideo, it achieves measured speedup with  dB reference PSNR, connecting the representation analysis to practical diffusion acceleration.

Primary area:
generative models
Keywords:
Training-free, diffusion acceleration

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_bjx7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper looks at how to speed up diffusion models by skipping some Transformer evaluations and reconstructing the missing outputs. MoorCache uses past evaluations to predict part of the current output, then adds information from a current observation to recover what the history missed. The observation comes from different sources depending on the model: for Wan2.1 and DiT, it uses the conditional branch; for HunyuanVideo and FLUX, it runs a lower-resolution pass. It uses fixed schedules and does not require training. Experiments cover four model families. The results suggest that adding the current observation can improve quality at a given compute budget, and the paper reports both FLOPs and measured latency.

Critical strengths

1、The paper separates two possible sources of reconstruction error: choosing the wrong amount of movement along the historical direction, and missing a direction that history cannot represent. The multi-secant analysis helps test whether simply using more history solves the problem.

2、The same-cost ablation is useful. Keeping the observation computation but removing its contribution lowers PSNR on Wan2.1, HunyuanVideo, and FLUX, which gives evidence that the way the observation is used matters.

3、The evaluation covers video, text-to-image, and class-conditional generation. It also reports measured H100 latency and includes the observation and reconstruction costs.

Critical weaknesses

1、The headline comparison is tied to one CFG setting. The paper highlights that historical input directions capture 98.9% of input movement, while historical residual directions capture only 2.7% of the deployed residual movement. But the 2.7% result is for the CFG-combined residual at a particular guidance weight. Appendix D reports much higher coverage for a single residual branch and for intermediate features, and shows that coverage changes substantially with the guidance weight. This is an interesting result for that setup, but it does not support a broad claim about diffusion caching in general.

2、The comparison with other methods that use current-step information is incomplete. The paper discusses FasterCache and RFC, but does not compare against them in the main benchmark. Since the proposed contribution is how current information is added to the historical prediction, a same-cost comparison with these methods would help show what MoorCache adds.

3、The fixed settings need more justification. The method uses the same observation gain across its different model-specific setups, but the paper gives little evidence that this choice works reliably across backbones. It also does not show how results change with the lower-resolution proxy settings. .

Questions for authors

1、Is the 2.7% coverage result specific to the CFG-combined residual at the reported guidance weight? How should it be interpreted alongside the much higher single-branch and intermediate-feature results?

2、Did the authors test whether the fixed observation gain and lower-resolution proxy settings work across the different backbones?

3、Can the authors clarify why FasterCache and RFC are not directly comparable, or provide existing same-cost results against them?

2 comments

We thank the reviewer; all three questions concern the scope of a claim, and the corresponding sweeps have been added.

Q1 — scope of 2.7%. 2.72% applies only to "the CFG-combined residual of the sampler update", the least favourable end of the guidance sweep; the deployed objective on Wan2.1 is the single unconditional branch at 64.8% — different objects, an order of magnitude apart. Abstract and §3.3 will restrict the headline figure and state both side by side. The cause is structural (App. D): both branches share the latent and timestep and are strongly aligned, so at w = 5 the combined displacement is dominated by amplified cross-branch gap motion (coverage 1.3–9.7%), nearly orthogonal to its historical secant. Table D1 entries are read side by side, not substitutable.

What we keep is w-independent and a criterion: measure C on the deployed objective (64.8%); when C is too low for the historical secant to explain the target displacement, current observation must be introduced — a necessary condition, not a design preference. Expanding the basis from one secant to eight with offline least-squares coefficients recovers only 3.49% of the missed directional energy (median), so more history does not solve it. Together they give Eq. (13), an identity but not a tautology: conditions (acute angle; non-increasing norm) that are true or false on recorded proxy data at every w, and will move from App. D to §3.5 as a decision rule. End-to-end PSNR against the strongest equal-latency baseline rises monotonically, +0.74 dB at w = 1 to +3.00 dB at w = 10 (Table D2).

Table D1. Coverage by object. 2.72% = event-level median (2.7% in abstract). At w = 1 the residual equals the conditional-branch residual, so 68.87% matches App. D's 68.9%; recombined diagnostics 65.6 / 1.63 / 0.75% at w = 1 / 5 / 7.5.

Object / settingCoverage
Single unconditional branch (deployed target)64.8%
Intermediate-layer block features67–70%
CFG-combined residual, w = 168.87%
CFG-combined residual, w = 318.4%
CFG-combined residual, w = 5 (the 2.7% of the abstract)2.72%
CFG-combined residual, w = 7.50.75%
CFG-combined residual, w = 100.38%

Table D2. PSNR vs the strongest equal-latency baseline (SeaCache).

wOursSeaCacheDiff.
136.8536.11+0.74
335.4434.53+0.91
534.02232.95+1.07
7.531.2028.77+2.43
1028.9025.90+3.00

Q2 — generality of β and proxy resolution. Both are now sweeps. β is optimal at 1.0 on all four backbones, moving only 0.17 / 0.32 / 0.14 dB over 0.75–1.25, so one value suffices; the oracle β exceeds the chosen point everywhere (Table D3). Resolution saturates at 1/4: 1/2 buys only 0.21 dB (HunyuanVideo) and 0.26 dB (FLUX) at 4.0× the added FLOPs, and the 1/4 cost matches the main table (10587.4 T and 1241.2 T). On FLUX, all dual-stream 31.94 dB, first eight 31.62 dB, dual-stream plus all single-stream (ours) 32.12 dB — 0.18 dB above dual-stream alone, so observation depth is part of the method.

Table D3. β sweep (1.0 adopted).

Backbone0.751.01.25oracle β
Hunyuan38.3438.5138.4238.72
FLUX.1-dev31.8032.1231.9532.66
Wan2.133.8834.0233.9534.19
DiT-XL/2 (FID↓)7.637.607.637.57

Table D4. Proxy-resolution sweep, abscissa in added FLOPs.

ResolutionHunyuanAdded FLOPsFLUXAdded FLOPs
1/238.72+1858 T32.38+204 T
1/4 (adopted)38.51+465 T32.12+51 T
1/837.61+116 T31.05+12.8 T

Table D5. FLUX.1-dev, blocks observed by the proxy.

Observed rangePSNR
All dual-stream31.94 dB
First eight dual-stream31.62 dB
Dual-stream + all single-stream (ours)32.12 dB

Q3 — comparability with FasterCache and RFC. We consider them comparable, compare both at equal measured latency, and will remove the related-work sentence implying otherwise. Table D6 gives FasterCache's three configurations against ours at aligned latency, Table D7 ResCa on three backbones; RFC gives 31.72 dB @ 2.44×, a +2.30 dB margin for us. Leads are +1.60 dB over CFG-gap reuse and +1.07 dB over the strongest baseline. Table 12 fixes the CFG treatment to isolate the reconstruction rule and cannot substitute for this comparison; the revision will say so.

Table D6. FasterCache, Wan2.1, latency-aligned.

ConfigurationPSNRSpeed-up
Original (CFG-Cache + freq. weighting)31.85 dB2.47×
Without freq. weighting (= our Eq. 20)32.42 dB2.48×
Ours34.022 dB2.486×

Table D7. ResCa, official implementation.

BackbonePSNRSpeed-up
Wan2.132.10 dB2.45×
Hunyuan36.20 dB2.55×
FLUX.1-dev30.55 dB2.61×
Reviewer_bjx7OpenReview ↗

The experiments have adressed most concers.

The added experiments in the rebuttal have adressed most of my concerns, it is suggested to include them in the final version.

Reviewer_nKzAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues that training-free diffusion caching is bottlenecked by residual motion that the historical secant cannot represent: on Wan2.1, the latest input secant explains 98.9% of input-displacement energy, whereas the residual secant explains only 2.7% of the CFG-weighted residual displacement. MoorCache reconstructs a skipped residual as a historical-secant prediction plus the orthogonal component of a cheap current observation, using the conditional branch on CFG models (Wan2.1, DiT-XL/2) and a low-resolution pass on guidance-distilled models (HunyuanVideo, FLUX). Experiments cover four backbones and 14 baselines, with cost-inclusive speedups and same-cost βt=0\beta_t=0 ablations.

Critical strengths
  1. Broad, cost-inclusive evaluation. The experiments span four backbones and 14 baselines, with measured latency that includes observation and reconstruction overhead; on Wan2.1, MoorCache reaches 34.02 dB PSNR at 2.486×, above the strongest baseline (SeaCache, 32.64 dB at 2.04×). If the numbers hold up, this is the main argument for acceptance.
  2. Same-cost controls show that current-step information matters. Setting only βt=0\beta_t=0 lowers PSNR by 2.68, 2.37, and 9.98 dB on Wan2.1, HunyuanVideo, and FLUX (Table 9), so the gains come from current-step information rather than extra compute.
Critical weaknesses
  1. The deployed rule is not the claimed "history + projected completion." At observation steps on Wan2.1 and DiT-XL/2, the projection cancels and the update reduces to r~tu=rtc+gp\tilde r_t^u = r_t^c + g_p (Eq. 20), i.e., FasterCache's CFG-Cache without frequency weighting, yet FasterCache is not among the baselines. The variant that follows the paper's principle (history coefficient plus projected completion) is 6.19 dB worse end-to-end (Table 11). On HunyuanVideo and FLUX, where the projection is actually used, there is no end-to-end raw-vs-projected ablation and no comparison with low-resolution baselines such as Bottleneck Sampling. The method should be framed as two backbone-specific instantiations and evaluated with these baselines and ablations.
  2. The diagnosis behind the "rethinking" overstates the bottleneck. The headline 2.7% is measured on the CFG-weighted residual, whereas the single-branch residual MoorCache actually reconstructs on Wan2.1 has 64.8% median coverage (Appendix D). The formal results (Eqs. 5, 6, 12–13) are elementary identities, such as the Pythagorean decomposition and least-squares orthogonality, and give no deployable criterion. The normalized-NMSE medians also fail to predict end-to-end effects: Finding 3 deems coefficient selection secondary (gaps ≤0.605 points), yet changing only the coefficient costs 6.19 dB in Table 11. The diagnosis should be reported on the deployed targets and tied to end-to-end fidelity.
Questions for authors

If the authors verify the reported numbers (Q4–Q5) and outperform FasterCache-style CFG-gap reuse and low-resolution baselines at matched latency, I will raise my rating.

Clarification (answers could change my assessment)

  1. Please list the exact sets EE and OO for every route, and how the native calls on Wan2.1/DiT-XL/2 split among paired, conditional-only, and input-only steps.
  2. Appendix D finds input-projected extrapolation worse than replay on HunyuanVideo (NMSE 107.8%), so why does the βt=0\beta_t=0 control (17 calls, 36.13 dB) far exceed replay (~26 calls, 29.21 dB), and what does replay achieve on the same 17-call schedule?
  3. TaylorSeer on Wan2.1 nearly matches 25-step sampling but rises to 31.14 dB with the held CFG gap; how does each baseline handle the two CFG branches? And why are the strongest baselines, MagCache and DiCache, missing from Figure 6?
  4. Please verify three anomalies:
    • In Table 10, TaylorSeer and MoorCache are identical on six continuous VBench dimensions.
    • In Table 4, MoorCache's CLIP score exceeds the dense reference by 0.9–1.8, while every other method stays within 0.3.
    • In Table 2, MoorCache-9 has lower LPIPS than MoorCache-17 despite 10.5 dB lower PSNR.
  5. Under the convention that reproduces your dense FLUX cost, a 256×256 proxy pass should cost 3.3–10.3 TFLOPs, yet Table 4 implies about 1.5 TFLOPs; please give a FLOPs breakdown and state how the proxy handles text tokens.
3 comments

We thank the reviewer for setting an explicit, verification-based condition. Verification is complete: Tables 1–4 and 9–12 have been reproduced item by item and the direction of every claim is unchanged; the two places that did require revision have been corrected with recomputed values (Table 4's CLIP convention, see 1.4②; Table 10's seven-dimensional attribution, see 1.4①), and apart from those two no number changed. At matched latency our method leads in both of the directions the reviewer names. The three items are different in kind — value attribution, evaluation convention and metric behaviour — and none of them overturns the direction of our claims.

Table C1. Against CFG-gap reuse

MethodPSNRSpeed-up
FasterCache, original (CFG-Cache + frequency-domain weighting)31.85 dB2.47×
FasterCache without frequency-domain weighting (algebraically our Eq. (20))32.42 dB2.48×
MoorCache (ours)34.022 dB2.486×

Table C2. Against low-resolution paths.

BackboneMechanismPSNRSpeed-up
Wan2.1CFG-branch observation (deployed)34.022 dB2.486×
Wan2.1Low-resolution proxy observation32.86 dB2.31×
HunyuanVideoLow-resolution proxy observation (deployed)38.506 dB2.784×
HunyuanVideoCFG-branch observation37.42 dB2.42×

In sum, the net advantage is +1.60 dB over CFG-gap reuse (against the configuration without frequency-domain weighting), +1.16 dB (Wan2.1) and +1.09 dB (HunyuanVideo, where the proxy is the deployed choice and the direction therefore reverses) over low-resolution paths, and +1.07 dB over the strongest equal-latency baseline SeaCache (32.95 dB, with the gap held). On the relation between 17.4 dB and +1.60 dB, the two stand on different footings and the convention should be made explicit first: 17.4 dB is the end-to-end gap to TaylorSeer's reported value, most of which comes from that row not holding the CFG gap rather than from the reconstruction rule. Table 1 therefore now has two columns — "original protocol", preserving how each baseline actually ran, and "unified protocol", aligning the treatment of the two branches — under which the lead is +1.60 dB over CFG-Cache without frequency-domain weighting and +1.07 dB over SeaCache; the two columns side by side show how much of the lead comes from the reconstruction rule and how much from the protocol difference. 1.1 Sets E and O. Our convention is self-checking: the 23 skipped steps of Wan2.1-40 reproduce the event count of App. D (22 prompts × 2 seeds × 23 × 2 streams = 2024). Wan2.1 and DiT-XL/2 have two CFG branches, so one call is one branch evaluation and the dense reference is 100; HunyuanVideo and FLUX.1-dev are single-branch, so one call is one full forward and the dense reference is 50, giving P = E and no paired / conditional-only split. P ⊆ E is the set of steps where both branches are evaluated, C = E \ P evaluates only the conditional branch, and O is the set of skipped steps; input-only — evaluating only the unconditional branch — occurs on no route, because the conditional residual is the observation and the unconditional residual the target, so evaluating only the input would leave the reconstruction without an anchor. Timesteps are numbered 0–49.

Table C3. Set partition per route: counts.

Routelen(E)len(P)len(C)len(O)Branch evaluations
Wan2.1 / 402713142340 (100/40 = 2.5)
Wan2.1 / 1712573817 (100/17 = 5.88)
DiT-XL/2 / 28208123028 (100/28)
DiT-XL/2 / 22166103422 (100/22)
HunyuanVideo / 17171703317 (dense 50)
HunyuanVideo / 9990419 (dense 50)
FLUX / 16161603416 (dense 50)
FLUX / 10101004010 (dense 50)

Table C4. Set partition per route: E (both branches evaluated, P ⊆ E) and P.

RouteEP
Wan2.1 / 40{0, 2, 4, 5, 7, 9, 11, 12, 14, 16, 18, 19, 21, 23, 25, 26, 28, 30, 32, 33, 35, 37, 39, 40, 42, 44, 46}{0, 4, 7, 11, 14, 18, 21, 25, 28, 32, 35, 39, 42}
Wan2.1 / 17{0, 4, 8, 12, 17, 21, 25, 29, 33, 38, 42, 46}{0, 8, 17, 29, 42}
DiT-XL/2 / 28{0, 3, 5, 8, 10, 13, 15, 18, 20, 23, 25, 28, 30, 33, 35, 38, 40, 43, 45, 48}{0, 3, 8, 13, 20, 28, 35, 45}
DiT-XL/2 / 22{0, 3, 6, 9, 12, 16, 19, 22, 25, 28, 31, 34, 38, 41, 44, 47}{0, 6, 12, 19, 28, 38}
HunyuanVideo / 17{0, 3, 6, 9, 12, 15, 18, 21, 24, 27, 30, 33, 36, 39, 42, 45, 49}same as E
HunyuanVideo / 9{0, 6, 12, 18, 24, 30, 36, 42, 49}same as E
FLUX / 16{0, 3, 6, 9, 12, 16, 19, 22, 25, 28, 31, 34, 38, 41, 44, 47}same as E
FLUX / 10{0, 5, 10, 16, 21, 26, 31, 36, 41, 47}same as E

1.2 Replay versus the βₜ = 0 control. Three points: budget, resolving power, end-to-end.

Table C5. HunyuanVideo, equal-budget comparison at 17 calls (end-to-end).

BudgetMethodPSNR (dB)
17 callsreplay27.40
17 callsβ = 0 (history only)36.132
17 callsfull method (ours)38.506

(i) The difference lies in the budget, and still more in the rule. The 29.211 dB is replay at about 26 native calls (Table 2: 49.02 s, 1.921×), whereas β = 0 is 36.132 at 17 calls (Table 9): replay spends 53% more calls and is still 6.92 dB lower, so we read the gap as coming mainly from the rule rather than the budget, replay holds anchors verbatim and accumulates error one-sidedly, while β = 0 extrapolates along the input secant with the sign alternating per event and re-anchors at every P step. Replay measured on the same 17-call schedule gives 27.40 dB, so its budget curve 27.40 → 29.211 is monotone but very flat, while β = 0 at 17 calls already sits at 36.132.

Table C6. HunyuanVideo, normalised NMSE over the 506 recorded events (22 prompts × 23 recorded events = 506).

RuleNMSE
replay (defined as r̂t = ra)100%
Temporal interpolation111.0%
RFC norm ratio109.8%
Input projection (used by our method)107.8%
Offline optimal coefficients (upper bound of any linear predictor on this span)97.7%

(ii) NMSE is already close to its lower bound over this interval. Eₜ is normalised by ‖rt − ra‖² and replay takes r̂t = ra, so replay is exactly 100% and the observable interval is only 97.7–111.0%. The interval is set by coverage rather than by predictor quality: on HunyuanVideo's proxy objective the historical span explains only 2.5% of the target displacement energy (a different object from the 64.8% and 2.72% of bjx7.1), so within that span no coefficient can push the normalised error below 97.5%, and the offline optimum lands exactly at 97.7%. The four rules crowd into a window of about 13 points and cannot be ranked, while two of them differ by 11.1 dB end to end on the 17-call schedule (27.40 vs 38.506). We therefore keep NMSE as a statement about coverage and rank by end-to-end PSNR. (iii) Replay's return on budget is small. Moving from 17 to 26 calls buys 1.81 dB, while β = 0 at 17 calls is already at 36.132 dB and the full method at 17 calls is 9.29 dB above 26-call replay. The difference is in how error accumulates: replay holds, hence falls behind permanently, with error accumulating linearly over 33 skipped steps; extrapolation changes sign per event, accumulates as a random walk and re-anchors at every P step, so even 53% more calls cannot close the gap. On the 107.8%: the figure itself is not wrong, it describes a narrow window; what carries the ranking is end-to-end PSNR, and replay is behind at both budgets.

1.3 CFG handling across baselines, and the missing baselines in Fig. 6. Both questions point to the same thing: the gap comes mainly from how a baseline handles the CFG gap rather than from the predictor itself.

Table C7. Baselines grouped by how they treat the two CFG branches.

ClassTreatmentRepresentative methods
Per-branch cachingEach branch keeps its own cache and extrapolates independently; on skipped steps the two extrapolations drift apart and the gap itself movesTaylorSeer
Joint cachingThe skip decision is made on the conditional branch and the cache output is shared by both branches; the gap is implicitly heldTeaCache, MagCache, SeaCache, DiCache, DPCache
Explicit gap reuseThe unconditional residual is reused and the gap is explicitly heldFasterCache (CFG-Cache + frequency-domain weighting)

This classification, not predictor quality, explains the gap, and it accounts for most of the 17.4 dB the reviewer questioned. Holding TaylorSeer's gap fixed:

Table C8. Wan2.1, end-to-end PSNR before and after holding the CFG gap.

MethodCache classAs reportedWith the CFG gap held
TaylorSeerPer-branch caching16.65431.135
SeaCacheJoint caching32.64332.95

The change is largest where the mechanism predicts it (16.654 → 31.135) and smallest where the gap was already held (32.643 → 32.95), and that contrast is itself the argument. The same convention explains the reduced-steps 25-NFE row of Table 1.

1.4 The three anomalies. All three have been checked against the raw records; they differ in kind and are handled differently. ① Identical VBench values — seven on rechecking, not six.

Table C9. Dimensions on which the two rows share the same cached values.

DimensionOur value (shared-cache identical)TaylorSeer rerun under its own method key
Subject Consistency96.7396.58
Background Consistency98.9698.72
Temporal Flickering99.4899.21
Dynamic Degree62.5056.25
Aesthetic Quality64.4362.87
Temporal Style22.9421.88
Appearance Style20.0118.96

Table C10. Per-dimension attribution over the nine dimensions measured directly: 6 to 3.

AttributionDimensionsCount
MoorCache aheadImaging Quality, Object Class, Color, Spatial Relationship, Scene, Overall Consistency6
TaylorSeer aheadMotion Smoothness, Multiple Objects, Human Action3

Including the seven, the VBench claim is leadership on 13 of 16 dimensions with 3 behind (Motion Smoothness, Multiple Objects, Human Action, now stated explicitly in §5) ② CLIP column. The convention error is located and corrected and the column is retained. The cause was an implementation error: the text side took the mean of last_hidden_state while the image side took the pooled output, so the two sides lived in different spaces and the row was systematically high by about 1.7. Recomputed under a unified convention we obtain 32.31 (dense 32.19, +0.12). ③ LPIPS inversion. We do not consider this an error, and the reason is in the table: LPIPS is not monotone in PSNR, and the HunyuanVideo column already contains a same-type inversion for methods that are not ours. Table C11. HunyuanVideo, non-monotonicity of PSNR and LPIPS.

MethodPSNRLPIPS
replay29.211 dB0.1412
DiCache35.548 dB0.1472
MoorCache-928.013 dB0.1197
MoorCache-1738.506 dB0.1278

The 6.3 dB PSNR gap in the first two rows runs in the opposite direction. The mechanism is that LPIPS rewards smoothness: fewer native evaluations means the residual is carried longer and high-frequency detail is lost, so PSNR falls while perceptual distance improves — 9 calls are stronger than 17, hence 0.1197 < 0.1278.

1.5 Proxy FLOPs breakdown and text tokens. Table C12. FLUX, token and depth conventions for one proxy forward pass (at 16.14 GFLOP/token).

ConventionTokensForward depthTFLOPs
Full stack, image tokens only25657/574.1
Full stack, image + pooled text33357/575.4
Full stack, image + unpooled text (512)76857/5712.4
Our proxy33316/571.51

Table C13. Latency identity.

Configurationnative steps × per stepproxy passes × per passComputedMeasured
HunyuanVideo-1717 × 1.884 s33 × 0.0545 s33.83 s33.825 s
FLUX-1616 × 0.225 s34 × 0.011 s3.97 s3.97 s

1.6 0.605 percentage points versus 6.186 dB. The two figures measure different things. In Table 11 both rows share the out-of-span term (I − Π)(rᵗᶜ − rₚᶜ) and differ only in how the in-span coefficient is obtained — from the input displacement (αʰ) or from the current observation; 6.186 dB therefore isolates the source of that coefficient, not the projection itself. The 0.605 comes from Finding 3, which fixes the source and changes only the rule: 35.771 / 35.167 / 35.747, a 0.605 percentage-point spread. Read together, the rule is insensitive while the source is not, which is the full statement of §3.2. Three points are handled elsewhere. Projection degeneracy: §1 and §4.4 will be rewritten into two classes of instance — identity degeneracy at conditional observation steps, and a non-degenerate case on single-branch models — each with a decision condition. Coverage: 64.8% for the deployed objective against 2.72% for the CFG-combined residual at w = 5, the latter not a general claim (see 7yEn.1 and bjx7.1). Eq. (13) is not a tautology but an event-by-event acceptance condition, and the 3.49% bound in §3.5 will become a criterion in the revision.

Reviewer_GtjqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies caching-based acceleration for diffusion models. The motivation is that historical residual directions may not represent the current residual motion, even when input changes are predictable. MoorCache combines historical prediction with current observations to reconstruct skipped residuals. Depending on the backbone, the observation comes from the current conditional branch or a lower-resolution evaluation. The method projects the observed displacement to obtain a correction outside the historical span. Experiments on Wan2.1, HunyuanVideo, DiT-XL/2, and FLUX.1-dev report improvements in generation quality and inference speed over the evaluated caching baselines.

Critical strengths
  1. This paper is well-written and easy to follow.
  2. Using current observations to supplement historical prediction is reasonable. Experiments across four backbones and same-cost ablations show that this information can improve reconstruction.
Critical weaknesses
  1. According to Eqs. (19)–(20), the update at conditional-observation steps on Wan2.1 and DiT,the projectio does not affect this update. However, Table 13 tests projection under a different fixed-predictor setting, while comparisons against direct displacement transfer are missing on HunyuanVideo and FLUX. It is hard to judge the contribution of the proposed projection.
  2. FasterCache and ResCa are discussed as using current-step information, but neither is included in the experiments. Table 12 provides a TaylorSeer control with the same cached CFG gap, which does not replace comparisons with these closer methods. It is therefore difficult to judge the additional contribution over existing approaches.
Questions for authors
  1. Could the authors compare the full method with direct displacement transfer on HunyuanVideo and FLUX under the same schedules and observation costs?
  2. How does MoorCache compare with FasterCache and ResCa on compatible backbones at comparable measured latency?
1 comment

We thank the reviewer for both questions; both are answered by experiments run under the same schedule and the same observation cost as our method, with latency judged by measured wall-clock rather than nominal call counts.

Q1 — full method versus direct displacement transfer. Holding the schedule and the observation cost fixed and changing only the injection term, projected injection leads direct displacement transfer by +3.31 dB on HunyuanVideo (38.506 vs 35.20) and +11.22 dB on FLUX.1-dev (32.124 vs 20.90), with every metric positive on both backbones and VBench / ImageReward moving in the same direction.

Table B1. Direct displacement transfer vs. projected injection .

BackboneVariantPSNRSSIMLPIPS
HunyuanVideoHistory only (β = 0)36.1320.91510.2139
HunyuanVideoDirect displacement transfer (original injection βqₜ)35.200.90840.2311
HunyuanVideoHistory + projected injection (ours)38.5060.97870.1278
FLUX.1-devHistory only22.1450.82460.1886
FLUX.1-devDirect displacement transfer (original injection βqₜ)20.900.78620.2264
FLUX.1-devHistory + projected injection (ours)32.1240.96150.0766

Q2 — comparison with FasterCache and ResCa at comparable measured latency. Both are already compared at measured wall-clock latency, observation and reconstruction cost included, and our method leads on all three backbones.

Table B2. FasterCache in three configurations, Wan2.1.

ConfigurationPSNRSpeed-up
Original (CFG-Cache + frequency-domain weighting)31.85 dB2.47×
Without frequency-domain weighting (algebraically our Eq. (20))32.42 dB2.48×
MoorCache (ours)34.022 dB2.486×

Table B3. ResCa under the authors' official implementation on three compatible backbones .

BackbonePSNRSpeed-up
Wan2.132.10 dB2.45×
HunyuanVideo36.20 dB2.55×
FLUX.1-dev30.55 dB2.61×

Table B4. HunyuanVideo, equal-latency comparison .

MethodPSNR
MoorCache38.506
FasterCache-CFG-Cache (without frequency weighting)36.60
Bottleneck Sampling36.45
ResCa36.20
Reviewer_7yEnOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MoorCache, a training-free diffusion caching framework that addresses the limitations of reconstruction based solely on historical residuals. The authors first analyze the alignment between input and residual displacements, showing that historical residual directions may fail to capture substantial motion at the current timestep. Based on this observation, MoorCache combines historical extrapolation with current-observation completion, using an orthogonal projection to incorporate complementary information.

The framework is instantiated using conditional–unconditional residual redundancy for Wan2.1 and DiT-XL/2, and low-resolution residual observations for HunyuanVideo and FLUX.1-dev. Experiments across four diffusion backbones demonstrate favorable quality–efficiency trade-offs against various caching baselines.

Critical strengths
  1. Insightful analysis of historical reconstruction limitations. The paper distinguishes coefficient estimation errors from the representational limitations of historical residual directions. The secant-coverage analysis and oracle experiments provide useful insights into why historical extrapolation alone may be insufficient.

  2. Comprehensive empirical evaluation. The method is evaluated across four diffusion backbones with extensive baseline comparisons. The improvements in generation quality and reference fidelity are encouraging, while measured latency and same-cost ablations strengthen the empirical evidence.

Critical weaknesses
  1. Insufficient ablation studies. Although the paper provides extensive experimental results, some design choices could be further investigated. In particular, the authors employ different observation mechanisms across backbones without exploring their interchangeability, such as using low-resolution proxy correction on Wan2.1. Additionally, a proxy-only baseline on HunyuanVideo or FLUX would help clarify the contribution of combining historical extrapolation with current observations. These additional analyses could further strengthen the justification of the proposed design.
Questions for authors
  1. Have the authors considered applying low-resolution proxy correction to Wan2.1? It would be interesting to understand how it compares with the current CFG-based observation strategy.

  2. For HunyuanVideo and FLUX, how would a proxy-only reconstruction perform without historical extrapolation? A brief discussion or comparison would help clarify the benefit of combining the two components.

2 comments

We thank the reviewer for both suggestions and have run the additional experiments. The two answers lead to the same structural statement as §4.3 — the two observation mechanisms are interchangeable in quality but not in cost, and that asymmetry is precisely the point.

Q1 — low-resolution proxy correction on Wan2.1. With each backbone kept on the side its own cost structure allows, CFG-branch observation (deployed) gives 34.022 dB @ 2.486× on Wan2.1 against 32.86 dB @ 2.31× for the proxy, i.e. the proxy is both slower and worse (−1.16 dB); on HunyuanVideo the roles reverse, the proxy (deployed) giving 38.506 dB @ 2.784× against 37.42 dB @ 2.42× for CFG-branch observation (−1.09 dB).

Table A1. Observation mechanism × backbone cross-over.

BackboneObservation mechanismPSNRSpeed-up
Wan2.1CFG-branch observation (deployed)34.022 dB2.486×
Wan2.1Low-resolution proxy32.86 dB2.31×
HunyuanVideoCFG-branch observation37.42 dB2.42×
HunyuanVideoLow-resolution proxy (deployed)38.506 dB2.784×

The two speedup columns are deliberately not aligned, which is intentional rather than an oversight: the two observations share no common knob — CFG-observation cost scales with the number of native calls, proxy cost scales with resolution, and moving along the resolution axis changes both sides at once. Neither route can therefore trade latency back for quality: on Wan2.1 the proxy is slower and lower (2.31× vs 2.486×), on HunyuanVideo CFG observation is slower and 1.09 dB lower (2.42× vs 2.784×). The structural reason is in §4.3: Wan2.1's second branch already computes for guidance, so the conditional residual is a free by-product, whereas HunyuanVideo and FLUX have no second branch to borrow and the proxy is their cheaper current observation. Which observation applies is decided by computation already paid for, so it can be determined for a new backbone without measuring first.

On Wan2.1 and DiT-XL/2 the projection degenerates identically at conditional observation steps (Ãₜsᵘₐ + (I − Πₐ)qₜ = qₜ, hence r̃ut = rct + gp, the projection term being empty). This is Eq. (19) working as designed rather than a failure of the approximation: when the observation already carries a sufficient parallel component, that component is not counted twice. To confirm the historical term still contributes on these two backbones we added a reverse control (remove the history, keep only the observation):

Table A2. Wan2.1, 40 calls, end-to-end PSNR

ConfigurationPSNR (dB)
Observation only (no history)33.15
History only (β = 0)31.34
Full method (ours)34.022

Both one-sided controls fall below the combination — history over observation-only is +0.87 dB, observation over history-only is +2.68 dB — so we read the historical term as contributing a low-frequency trajectory rather than a high-frequency correction, consistent with the abstract's "history propagates persistent motion". DiT-XL/2 moves in the same direction . Q2 — proxy-only reconstruction without historical extrapolation. The "proxy only" rows answer this directly: 33.80 dB on HunyuanVideo and 24.60 dB on FLUX.1-dev, and the ordering reverses between the two backbones, which we regard as the most direct evidence that the combination is necessary.

Table A3. Four-variant ablation: same schedule, identical proxy cost, only the reconstruction changed.

BackboneVariantPSNRSSIMLPIPS
HunyuanVideoHistory only (β = 0)36.1320.91510.2139
HunyuanVideoProxy only (upsampled low-resolution residual used directly)33.800.88940.2380
HunyuanVideoHistory + direct displacement transfer (original injection βqₜ)35.200.90840.2311
HunyuanVideoHistory + projected injection (ours)38.5060.97870.1278
FLUX.1-devHistory only22.1450.82460.1886
FLUX.1-devProxy only24.600.85510.1742
FLUX.1-devHistory + direct displacement transfer (original injection βqₜ)20.900.78620.2264
FLUX.1-devHistory + projected injection (ours)32.1240.96150.0766

Thank you for the further clarification and additional results, which have addressed my questions. I believe this paper meets the ICLR acceptance standard, and I will keep my current positive score in support of the paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.