{"id":"e0d1acfd-ca3a-46b3-b226-1cfcdc55bdb4","arxiv_id":"2507.02860","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"EasyCache is a training-free, runtime-adaptive caching method that accelerates DiT video generation by reusing transformation vectors during stable denoising phases.","lead":"EasyCache speeds up video diffusion models by caching and reusing transformation vectors when the model's behavior is locally stable, without retraining or offline profiling. It reports 2.1 to 3.3x faster inference on OpenSora, Wan2.1, and HunyuanVideo while keeping outputs close to the original.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Core criterion in Eq.4–6 monitors relative changes in v, but the cached object is Δ=v−x; under the paper's own Euler update (Eq.7), Δ drifts with v·Δs, which E_t does not measure, so the stated stability mechanism does not actually control reuse error.","rationale":"The reader's weakest assumption is that the accumulated indicator E_t reliably detects when the transformation vector Δ_t is nearly constant. My stress test sharpens this into a precise internal inconsistency: the indicator is built from estimated relative changes of the model output v, but the cached object is Δ = v − x. Under the paper's own flow-matching Euler update (Eq.7), Δ changes by −vΔs even when v is perfectly constant, so E_t does not bound the error it is supposed to control. This is not merely a lack of empirical evidence for the stability assumption; it is a theoretical gap between the criterion and the cached quantity. The concern is load-bearing because the paper's 'key insight' and the design of Eq.6 both rest on this link. If the proposed concrete test shows that reuse steps have large true Δ error, then the method's success cannot be attributed to the stated stability mechanism, and the paper would need either a corrected analysis or an empirically validated criterion. If the test shows small true error at reuse steps, the mechanism may still be salvageable with a different explanation. Because the empirical results could remain valid even if the theory is flawed, a conditional acceptance is appropriate rather than outright rejection; the authors should add the proposed instrumentation or a correct theoretical bound before final acceptance. I agree with the reader's conditional verdict, though my reason is more specific than 'limited evidence'.","tokens_in":13559,"tokens_out":12914,"duration_ms":148119,"concrete_test":"Take 10 VBench prompts on HunyuanVideo (or Wan2.1-1.3B), run the original model for all 50 steps to obtain ground-truth v_t, and simulate EasyCache's decisions offline using its published (τ,R). At every step t where EasyCache would reuse Δ_i, compute the true relative cached-vector error e_t = ||Δ_t − Δ_i|| / ||Δ_i|| and record E_t. If a substantial fraction (e.g., >20%) of reuse steps have E_t < τ but e_t > 0.05, the indicator is not controlling the quantity it claims to control, and the stated central hypothesis is falsified; if instead all such steps have small e_t, the method is accidentally conservative in these samples and the explanation needs revision.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Sec.3.3 defines Δ_t := v_t − x_t and reuses a cached Δ_i while the accumulated indicator E_t = Σ ε_n (Eq.5) stays below τ, where ε_t ≈ k_i||x_t − x_{t-1}||/||v_{t-1}|| (Eq.4). This indicator estimates relative changes of the model output v, not of the cached vector Δ. Under the flow-matching Euler update that the paper itself uses in Eq.7, x_t = x_{t-1} + v_{t-1}Δs_t, so Δ_t − Δ_{t-1} = (v_t − v_{t-1}) − v_{t-1}Δs_t. Hence ||Δ_t − Δ_{t-1}|| ≥ Δs_t||v_{t-1}|| − ||v_t − v_{t-1}||. In the regime the paper calls 'stable' (small ||v_t − v_{t-1}|| relative to ||x_t − x_{t-1}||), the Δ drift is dominated by the input step, which E_t never accumulates. A constant-velocity field v_t = v_i would keep E_t = 0 while Δ_t moves linearly away from Δ_i; EasyCache would reuse indefinitely and produce a badly wrong trajectory. Thus the central logical link—'small output change implies Δ nearly constant'—is reversed for Euler sampling. Sec.3.4 infers 'stable k' implies 'reuse transformation vectors' without deriving any bound on ||Δ_t − Δ_i||. The empirical success, if real, is not explained by the stated mechanism; the criterion may work only because reuse happens in late stages where ||v_{t-1}|| is small, a property not encoded in E_t.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EasyCache, a training-free framework for accelerating DiT-based video diffusion models by reusing a cached transformation vector Δ_i = v_i − x_i at steps where a cumulative indicator E_t stays below a threshold τ. The indicator accumulates estimates of relative output change using a constant transformation rate k_i. Full model evaluations are triggered when E_t ≥ τ, during warm-up steps, and at the final step. Experiments on Open-Sora 1.2, Wan2.1-1.3B, HunyuanVideo, and FLUX.1-dev report speedups of 2.1–3.3× (up to 4.64× for text-to-image) with improved PSNR/SSIM/LPIPS over TeaCache and PAB, together with compatibility with the SVG attention acceleration. Ablations cover the threshold τ, warm-up steps R, the reuse criterion, and the update strategy for k.","tokens_in":13967,"tokens_out":10973,"duration_ms":107732,"significance":"If the underlying mechanism were established, this would be a practically valuable contribution: training-free, only two hyperparameters, model-agnostic across video and image DiTs, and accompanied by code. The empirical gains over strong baselines (e.g., PSNR 32.66 vs. 23.85 for TeaCache on HunyuanVideo at higher speedup) are substantial, and the ablations are informative. However, the paper's central theoretical claim—that the accumulated output-change indicator controls the reuse error of the cached transformation vector—is not supported by the submitted derivation, and the evaluation would benefit from matched-speedup comparisons. These issues are addressable in a revision but currently limit the scientific contribution.","major_comments":[{"comment":"The adaptive caching criterion does not control the actual reuse error ||Δ_t − Δ_i||. Under the Euler update in Eq. (7), Δ_t − Δ_{t−1} = (v_t − v_{t−1}) − v_{t−1} Δs_t; even when the output change v_t − v_{t−1} is negligible, the term v_{t−1} Δs_t causes Δ to drift at a rate of order Δs_t per step. The indicator E_t accumulates only an estimate of the relative output change, so it can remain near zero while the cached Δ drifts substantially (e.g., a constant-velocity field gives zero output change and unbounded Δ drift). The paper does not prove a bound on ||Δ_t − Δ_i|| in terms of E_t, and the claim in Sec. 3.3 that 'this stability, in turn, implies that the transformation vector Δ_t also remains nearly constant' is therefore not justified. Please either derive such a bound, modify the criterion to directly account for the step-size contribution (e.g., by accumulating Δs_t in E_t), or provide empirical evidence that E_t strongly correlates with the actual Δ drift across models and prompts.","section":"Sec. 3.3, Eqs. (4)-(6) and Sec. 3.4, Eq. (7)"},{"comment":"Under the approximation in Eq. (4) and the update in Eq. (7), ε_t ≈ k_i ||x_t − x_{t−1}|| / ||v_{t−1}|| = k_i Δs_t, since ||x_t − x_{t−1}|| = Δs_t ||v_{t−1}||. Consequently, the accumulated indicator E_t is approximately k_i times the elapsed continuous time since the last full computation, independent of the actual output dynamics. This makes the 'runtime-adaptive' criterion effectively a fixed time budget with a model-specific rate, which is a different mechanism from the local-stability detection described in the paper. The authors should acknowledge this interpretation, or change the indicator so that it does not cancel the velocity norm, otherwise the adaptive claim is overstated.","section":"Sec. 3.3, Eq. (4)"},{"comment":"The headline PSNR gains compare methods at different speedups (e.g., HunyuanVideo: TeaCache at 1.67× with PSNR 23.85 vs. EasyCache at 2.21× with PSNR 32.66). While a Pareto improvement is strong, the comparison would be more conclusive with matched-speedup runs or full trade-off curves for all three video models. The τ-ablation in Table 3 provides such a curve only for Wan2.1-1.3B and with a single seed per prompt; extending this analysis and reporting variance across seeds would substantially strengthen the quality claim.","section":"Table 1, Sec. 4.2"}],"minor_comments":[{"comment":"There is a grammatical error: 'a efficient' should be 'an efficient'.","section":"Abstract"},{"comment":"The figure does not state which model, sampling configuration, and prompt are used for the analysis; the caption should specify these details and define the averaging operation used for the L1 norm.","section":"Fig. 2"},{"comment":"The paper states that five samples per VBench prompt are generated, but it does not report the number of prompts or standard deviations for PSNR, SSIM, and LPIPS. Please include these, as the ablations in Sec. 4.4 use a single sample per prompt and are therefore not statistically robust.","section":"Sec. 4.1"},{"comment":"The definition of ε_t should make the percentage normalization explicit; the threshold values (e.g., τ = 5%) are given in percent, but the equation as written is a dimensionless ratio.","section":"Eq. (4)"},{"comment":"The text states that the VBench drop is 'negligible' and 'below 1%', but the table shows Open-Sora 79.40 to 78.74 (−0.66) and Wan2.1 81.30 to 80.49 (−0.81); please present the percentage changes explicitly and discuss whether these are within the benchmark's run-to-run variability.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely a good fit for a computer vision journal if the theoretical gap is addressed. My main concern is that the performance may be attributable to the time-budget behavior of the criterion rather than to the claimed stability detection; the authors should clarify this before acceptance. I also note that the comparison with TeaCache at higher speedups, while favorable, would be more persuasive with matched-speedup data."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The new thing is a runtime-adaptive cache trigger for DiT video diffusion, based on a per-step transformation rate k_t and an accumulated output-change indicator E_t. That replaces TeaCache's offline polynomial prior, and it's a genuinely different mechanism. The experiments are broad: three large video models, text-to-image, GPU architectures, compatibility with SVG, plus ablations on tau, R, and the k update rule. Code is released. On the numbers as reported, EasyCache beats TeaCache on speed and PSNR in most settings, and the gains are large enough to take seriously.\n\nThe biggest soft spot is conceptual and it's real. E_t monitors relative changes of the model output v_t, but the cached object is Delta_t = v_t - x_t. Under the paper's own Euler update, x_t = x_{t-1} + v_{t-1}*Delta_s, so Delta_t - Delta_{t-1} is approximately -v_{t-1}*Delta_s when v is nearly constant. A constant velocity field keeps E_t near zero while Delta moves linearly; reusing Delta_i then produces error proportional to distance traveled. So the criterion as stated does not control the reuse error. The stability of k_t does not imply stability of Delta_t. The paper asserts this in Sec.3.3 and gives an intuitive argument in Sec.3.4, but there's no bound. This doesn't automatically kill the empirical method -- it may work for other reasons, such as late-stage v norms being small -- but the explanation in the paper is not the right one, and a reviewer should demand either a bound or a corrected criterion.\n\nOther issues are less severe. The \"no tuning\" framing oversells: tau and R are per-model (2.5%/5%/10%, R 5-10), and the headline comparisons are at different speedups. No error bars. PSNR against the baseline favors conservative methods, and the VBench numbers are a bit mixed, though within a point.\n\nWho it's for: people working on training-free diffusion acceleration. It deserves a serious referee, but the revision needs to address the theory gap before I'd trust the stated mechanism. My own verdict: conditional; the empirical case is strong enough to publish if the theory is fixed.","headline":"A promising training-free caching method with broad empirical results, but the stated stability criterion does not actually control the reuse error as claimed.","tokens_in":14461,"tokens_out":5338,"would_cite":true,"duration_ms":60252,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"EasyCache claims that reusing a cached transformation vector during locally stable denoising phases accelerates DiT video generation 2.1x–3.3x with no training and no offline profiling.","keywords":["video generation","Diffusion Transformer","inference acceleration","feature caching","training-free acceleration","runtime-adaptive caching","denoising stability","latent caching"],"falsifier":"Record the actual per-step distances $\\|\\Delta_t - \\Delta_i\\|$ and the indicator $E_t$ across many prompts in a video DiT: if long stretches of later denoising show $E_t < \\tau$ while $\\|\\Delta_t - \\Delta_i\\|$ grows enough to produce visible frame-to-frame artifacts or a VBench drop beyond 1%, the stability assumption and its proxy are falsified.","tokens_in":13411,"feed_emoji":"🎬","tokens_out":7683,"duration_ms":75968,"temperature":0.7,"pith_summary":"EasyCache claims that through most of the denoising trajectory of a Diffusion Transformer video model, the per-step transformation — the vector difference between the model's output and its input latent — is locally almost constant, so previously computed transformations can be reused instead of running the full network. The paper proposes a runtime-adaptive criterion that accumulates an estimated relative output-change rate and reuses the cached transformation only while the accumulated deviation stays below a threshold $\\tau$. On OpenSora, Wan2.1, and HunyuanVideo this gives 2.12× to 2.54× speedups with no retraining and no offline profiling, rising to 3.33× when combined with an efficient-attention accelerator. If the claim is right, training-free caching becomes a plug-in technique that adapts per sample and per model, removing the dataset-specific tuning that limits prior caching schemes.","feed_headline":"Caching stable denoising steps makes video diffusion up to 3.3x faster","feed_subtitle":"A runtime rule reuses past transformation vectors during stable phases, keeping visual quality across three video models.","key_machinery":"The load-bearing object is the transformation vector $\\Delta_t := v_t - x_t$ together with the accumulated stability indicator $E_t = \\sum_{n=i+1}^{t} \\varepsilon_n$, where each local indicator $\\varepsilon_n$ approximates the output-change rate as $k_i\\|x_n - x_{n-1}\\| / \\|v_{n-1}\\|$ using the last fully computed rate $k_i$. When $E_t < \\tau$, Eq. 6 reuses the cached $\\Delta_i$; when the threshold is crossed, the full model runs, the cache is refreshed, and $E_t$ resets to zero. The warm-up count $R$ forces full computation during the early unstable phase, and the final step is always computed.","core_discovery":"The central discovery is that the relative transformation rate $k_t = \\|v_t - v_{t-1}\\| / \\|x_t - x_{t-1}\\|$ stabilizes quickly after a brief warm-up and stays nearly constant over most of the denoising process, so the transformation vector $\\Delta_t := v_t - x_t$ can be approximated by a recently computed $\\Delta_i$. EasyCache operationalizes this with the update rule $v_t = x_t + \\Delta_i$ whenever the accumulated stability indicator $E_t$ stays below the threshold $\\tau$, and a full forward pass otherwise, with the final step always fully computed. On HunyuanVideo this preserves PSNR 32.66 dB versus 23.85 dB for the previous dynamic-caching baseline at a higher speedup, and the same rule transfers to text-to-image generation.","pith_inferences":["A natural extension is to make $\\tau$ schedule-dependent, allowing more aggressive reuse in the final refinement steps where visual differences are subtle; the paper only reports a single fixed threshold per model.","The stability assumption is qualitative, so a quantitative test would measure the distribution of $\\|\\Delta_t - \\Delta_i\\|$ across many prompts and check whether $E_t$ is a calibrated predictor of that distance.","Because EasyCache targets the DiT block's input/output transformation, it should compose with token-pruning or quantization accelerators, potentially closing part of the gap to real-time generation that the paper names as its main limitation.","The same constant-transformation-rate idea could be tested on other flow-matching or score-based samplers, including image or audio DiTs, where the cheap $E_t$ proxy is computable from the same quantities."],"forward_implications":["On Open-Sora 1.2, Wan2.1-1.3B, and HunyuanVideo, EasyCache alone achieves 2.12×, 2.54×, and 2.21× speedups while improving PSNR, SSIM, and LPIPS over step-reduction, static-cache, PAB, and TeaCache at comparable settings.","On HunyuanVideo at 1280×720, EasyCache combined with SVG efficient attention reaches 3.33× total speedup with about a 1.1% PSNR drop, whereas the TeaCache+SVG combination loses 17.8% PSNR.","Because the criterion needs no offline profiling or dataset prior, a single $\\tau, R$ setting transfers across video models and to text-to-image FLUX, where EasyCache reaches 4.64× speedup with FID 23.2 versus 25.8 for the full model.","VBench scores stay within 1% of the unaccelerated model on the tested prompts, indicating the quality loss is largely imperceptible to human evaluators."],"supporting_citations":[{"why":"Introduces the Diffusion Transformer architecture whose per-step transformation dynamics EasyCache exploits.","marker":"[5]"},{"why":"HunyuanVideo is the most expensive testbed and the source of the headline 2.21× speedup and PSNR 32.66 result.","marker":"[6]"},{"why":"TeaCache is the prior dynamic-caching state of the art that EasyCache outperforms and whose offline-profiling reliance motivates the runtime-adaptive design.","marker":"[15]"},{"why":"SVG sparse attention is the orthogonal accelerator used to demonstrate 3.33× combined speedup.","marker":"[16]"},{"why":"Wan2.1 supplies a second video testbed and the ablations for threshold and warm-up settings.","marker":"[4]"},{"why":"Open-Sora 1.2 is the third video testbed used for the main comparison table.","marker":"[1]"},{"why":"Flow-matching ODE sampler formalism underpins the Sec. 3.4 explanation for why the transformation rate stabilizes.","marker":"[36]"},{"why":"VBench is the evaluation suite used for perceptual alignment scores across all video baselines.","marker":"[40]"}],"fun_headline_variants":["EasyCache accelerates video diffusion up to 3.3x without training","Runtime-adaptive caching speeds video generation 3.3x","Training-free caching outperforms previous dynamic caching","Video diffusion 3.3x faster via adaptive transformation caching","Less is enough: 3.3x video diffusion speedup without fine-tuning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that after a short warm-up the model's transformation vector stays nearly constant during denoising, and that the accumulated error estimate built from the last computed rate reliably detects when that constancy fails.","fun_headline_variants_meta":{"raw":{"variants":["EasyCache accelerates video diffusion up to 3.3x without training","Runtime-adaptive caching speeds video generation 3.3x","Training-free caching outperforms previous dynamic caching","Video diffusion 3.3x faster via adaptive transformation caching","Less is enough: 3.3x video diffusion speedup without fine-tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000586,"raw_usage":{"total_tokens":2753,"prompt_tokens":941,"completion_tokens":1812,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":1724}},"tokens_in":557,"tokens_out":1812,"duration_ms":16559,"temperature":1.0,"reasoning_tokens":1724,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:18:33.230258+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the actual per-step distances $\\|\\Delta_t - \\Delta_i\\|$ and the indicator $E_t$ across many prompts in a video DiT: if long stretches of later denoising show $E_t < \\tau$ while $\\|\\Delta_t - \\Delta_i\\|$ grows enough to produce visible frame-to-frame artifacts or a VBench drop beyond 1%, the stability assumption and its proxy are falsified.","supporting_citations":[{"cited_title":"Scalable diffusion models with transform- ers,","cited_arxiv_id":null,"evidence_quote":"Introduces the Diffusion Transformer architecture whose per-step transformation dynamics EasyCache exploits."},{"cited_title":"Timestep embedding tells: It’s time to cache for video diffusion model,","cited_arxiv_id":null,"evidence_quote":"TeaCache is the prior dynamic-caching state of the art that EasyCache outperforms and whose offline-profiling reliance motivates the runtime-adaptive design."},{"cited_title":"Sparse videogen: Accelerating video diffusion trans- formers with spatial-temporal sparsity,","cited_arxiv_id":null,"evidence_quote":"SVG sparse attention is the orthogonal accelerator used to demonstrate 3.33× combined speedup."},{"cited_title":"Flow matching for generative modeling,","cited_arxiv_id":null,"evidence_quote":"Flow-matching ODE sampler formalism underpins the Sec. 3.4 explanation for why the transformation rate stabilizes."},{"cited_title":"Vbench: Comprehensive benchmark suite for video generative models,","cited_arxiv_id":null,"evidence_quote":"VBench is the evaluation suite used for perceptual alignment scores across all video baselines."}],"review_version":1}