ACID: Adaptive Caching for vIDeo generation

Abstract: Video diffusion models produce high-quality generations but remain slow at inference due to their sequential denoising procedure. Caching-based acceleration methods address this by reusing intermediate model outputs: leading dynamic approaches such as TeaCache, EasyCache, and DiCache accumulate a drift signal and skip expensive model evaluations when accumulated drift stays below a fixed threshold τ. This threshold controls an apparent tradeoff - raising it yields faster generation at the cost of visual quality, while lowering it preserves quality but sacrifices speed. We show this tradeoff is not fundamental; it is an artifact of holding τ constant throughout denoising. We identify the existence of critical steps - timesteps where the drift signal changes rapidly - and show that applying a low threshold selectively at these steps while caching aggressively elsewhere recovers most of the quality of conservative caching at substantially higher inference speeds. Building on this insight, we propose ACID, a lightweight, training-free wrapper that monitors the rate of change of each method's existing drift signal to dynamically switch between a low and a high threshold. ACID is signal-agnostic and modular: it requires no retraining and plugs directly into existing dynamic caching methods without modifying their core mechanisms. Evaluated across three caching methods (TeaCache, EasyCache, DiCache) and three open-source video diffusion models (HunyuanVideo, Wan 2.1, CogVideoX), ACID consistently expands the Pareto frontier of visual quality versus inference speed beyond what any fixed threshold achieves. In particular, on TeaCache and HunyuanVideo, ACID achieves up to 2.16x speedup over the no-caching baseline, and up to 38% additional speedup over the conservative fixed-threshold baseline with negligible (<0.3 dB PSNR, <0.01 SSIM, <0.01 LPIPS) quality degradation.

• The paper introduces ACID, a training-free wrapper that switches between low and high caching thresholds based on changes in each method’s drift signal, preserving conservative reuse during quality-critical steps and enabling aggressive caching during stable phases.
• Across TeaCache, EasyCache, and DiCache on three video diffusion models, ACID consistently expands the speed–quality Pareto frontier, reaching up to 2.16× speedup with near-conservative fidelity and up to 38% additional speed over a conservative fixed-threshold baseline.
• ACID requires only three hand-tuned parameters per method–model configuration and improves caching without changing the underlying reconstruction logic, but automated threshold selection, broader benchmarks, and human perceptual studies remain open challenges.

Dynamic caching methods such as TeaCache, EasyCache, and DiCache accelerate video diffusion inference by accumulating a drift signal and skipping full model evaluations whenever the accumulated drift stays below a user-specified threshold $\tau$. The paper "ACID: Adaptive Caching for vIDeo generation" (2607.12358) argues that the quality–latency tradeoff induced by this threshold is not fundamental but an artifact of holding $\tau$ constant across the entire denoising trajectory. The authors propose ACID, a training-free, signal-agnostic wrapper that switches between a low and a high threshold based on the local rate of change of the underlying caching signal, and demonstrate that it consistently expands the Pareto frontier of visual quality versus speed across three caching methods and three open-source video diffusion models.

Background and motivation

Video diffusion transformers (DiT backbones) generate content through sequential denoising, with inference times of 3 to 23 minutes for roughly 5-second videos on a single A100 GPU. Caching-based acceleration reuses intermediate outputs to reduce this cost. TeaCache estimates drift via the L1 distance between consecutive Timestep Embedding Modulated Noisy Inputs (TEMNI); EasyCache uses a transformation rate $k$ (output change over input change) under an assumed locally linear input–output regime; DiCache runs only the first few transformer blocks as a shallow probe and measures feature change. Despite differing signals and output-reconstruction mechanisms, all three share the same algorithmic skeleton: accumulate drift, compare against a fixed $\tau$, run the full model when exceeded, and reset the accumulator.

The fixed threshold forces a single global caching aggressiveness. A high $\tau$ yields aggressive caching and degraded quality; a low $\tau$ preserves quality at reduced speedup. The central claim of the paper is that this tradeoff is structural rather than inherent: because denoising dynamics vary substantially across timesteps, no constant $\tau$ can be simultaneously conservative where quality is critical and aggressive where the process is stable.

Analysis: critical steps and signal structure

The paper's empirical analysis rests on two observations about caching signals. First, signals such as $\Delta$TEMNI are relatively consistent across diverse prompts, so their structure is a property of the model rather than of individual generations. Second, this structure is model-dependent: HunyuanVideo and Wan 2.1 exhibit a characteristic U-shaped drift profile, whereas CogVideoX shows a distinctly different distribution. Any adaptive scheme must therefore be robust across both pattern types, which motivates evaluation on all three models.

A controlled experiment using TeaCache compared low and high fixed thresholds against dual-threshold configurations with varying transition points. Optimal transition points were found to coincide with the timestep at which the second derivative of the caching signal stabilizes—i.e., where the signal shifts from a variable regime to a stable one. This identifies critical steps: regions where the drift signal changes rapidly and aggressive caching causes disproportionate quality loss. Importantly, the same gains from adaptive thresholding appear for EasyCache and DiCache despite their entirely different signals, supporting the claim that the limitation is structural to any fixed-threshold method rather than an artifact of one particular metric.

Method

ACID maintains two thresholds, $\tau_{\text{low}}$ and $\tau_{\text{high}}$, and selects between them by monitoring the first difference of the base method's signal $\tau$0. When $\tau$1—a stable phase—the high threshold applies, permitting aggressive caching; when the difference escapes this band—a dynamic, quality-critical phase—the low threshold applies, restricting cache reuse. The accumulated-drift logic, warmup behavior, and output-reconstruction mechanism of the underlying method are left untouched, making ACID a lightweight modular wrapper requiring no retraining. The design has three hyperparameters per configuration ($\tau$2, $\tau$3, $\tau$4), which the paper treats as tunable rather than learned; no automatic selection procedure is proposed.

Experimental results

Evaluation covers TeaCache (on HunyuanVideo, Wan 2.1, CogVideoX 1.5-5B), EasyCache and DiCache (on HunyuanVideo and Wan 2.1, restricted to officially supported configurations), using 33 VBench prompts spanning all 16 VBench dimensions with a fixed seed, on A100 80GB GPUs. Fidelity is measured as PSNR, SSIM, and LPIPS against the uncached baseline.

Representative results:

Method	Model	Mode	Speedup	PSNR	SSIM	LPIPS
TeaCache	HunyuanVideo	Fixed low / Adaptive / Fixed high	1.57× / 2.16× / 3.20×	24.69 / 24.41 / 19.98	0.819 / 0.813 / 0.693	0.187 / 0.196 / 0.353
TeaCache	Wan 2.1	Fixed low / Adaptive / Fixed high	1.42× / 1.96× / 2.86×	24.93 / 22.10 / 17.10	0.845 / 0.757 / 0.538	0.083 / 0.150 / 0.335
TeaCache	CogVideoX	Fixed low / Adaptive / Fixed high	1.41× / 1.87× / 2.19×	37.32 / 32.02 / 16.35	0.958 / 0.930 / 0.580	0.029 / 0.058 / 0.463
EasyCache	HunyuanVideo	Fixed low / Adaptive / Fixed high	2.06× / 2.39× / 2.55×	33.51 / 30.62 / 28.99	0.926 / 0.894 / 0.851	0.058 / 0.094 / 0.153
DiCache	HunyuanVideo	Fixed low / Adaptive / Fixed high	1.92× / 2.04× / 3.02×	30.96 / 29.42 / 20.15	0.904 / 0.888 / 0.720	0.088 / 0.105 / 0.328
DiCache	Wan 2.1	Fixed low / Adaptive / Fixed high	2.27× / 2.73× / 3.97×	13.41 / 13.29 / 11.79	0.356 / 0.355 / 0.282	0.508 / 0.541 / 0.644

The headline result is on TeaCache with HunyuanVideo: ACID achieves up to $\tau$5 speedup over the no-caching baseline and up to $\tau$6 additional speedup over the conservative fixed-threshold baseline with negligible degradation ($\tau$7 dB PSNR, $\tau$8 SSIM, $\tau$9 LPIPS). Across all seven method–model combinations, adaptive configurations dominate or extend the fixed-threshold Pareto frontier, typically achieving near-low-threshold fidelity at speedups substantially closer to the high-threshold regime. The appendix tables show this holds along the entire frontier—for example, DiCache on HunyuanVideo reaches $k$0 speedup adaptively while matching the quality of the $k$1 fixed configuration. One caveat worth noting: on DiCache with Wan 2.1, absolute fidelity metrics are poor even for the uncached-relative comparisons (PSNR around 13 dB), indicating that DiCache itself degrades quality heavily on that model; ACID improves the tradeoff but does not remedy the base method's weakness there.

Limitations and open questions

Several limitations are acknowledged or evident. The stability bound $k$2 and the two thresholds must be hand-tuned per method–model pair; the paper provides no principled or automated selection rule, and the reported gains depend on choosing well-performing $k$3 pairs. The transition-point analysis is empirical and grounded in the second derivative of specific signals; whether the stabilization criterion generalizes to future caching signals or image diffusion models is untested. Evaluation uses 33 prompts with a fixed seed, which is modest relative to full benchmark suites, and human perceptual evaluation is absent—fidelity is measured only against uncached baseline outputs. Finally, EasyCache and DiCache could not be evaluated on CogVideoX due to missing official implementations, leaving coverage of that model limited to TeaCache.

Conclusion

ACID demonstrates that the quality–latency tradeoff in dynamic caching for video diffusion models stems from fixed thresholds rather than any fundamental limit. By detecting critical steps through the rate of change of each method's existing drift signal and switching between conservative and aggressive thresholds accordingly, it improves the Pareto frontier uniformly across three caching methods and three models, with up to $k$4 end-to-end speedup at near-conservative quality. The main open question left by the work is whether the threshold-switching criterion can be calibrated automatically, removing the per-configuration hyperparameter tuning that currently accompanies its deployment.