OmniCache: Diffusion Transformer Acceleration

• OmniCache is a training-free acceleration technique that reuses intermediate computations in diffusion Transformers using trajectory-oriented cache scheduling.
• The method employs curvature-based scheduling and noise correction to insert cache reuse at low-curvature timesteps while mitigating errors during denoising.
• Experimental results demonstrate up to 2.50× speedup in image and video generation, highlighting its advantages over conventional local reuse methods.

OmniCache is a training-free acceleration method for diffusion Transformer (DiT) models that reuses intermediate computations during inference by taking a trajectory-oriented global perspective on the denoising process rather than relying only on local inter-step similarity. Introduced by Huanpeng Chu, Wei Wu, Guanyu Feng, and Yutao Zhang, it is designed for image and video generation settings in which reverse diffusion requires many denoising steps and each step executes a costly Transformer denoiser; its central claim is that cache reuse should be distributed across the full sampling trajectory and corrected for cache-induced noise, rather than concentrated mainly in later steps where adjacent features look most similar (Chu et al., 22 Aug 2025).

1. Problem formulation and conceptual premise

OmniCache addresses inference acceleration for diffusion Transformer models, especially video generation, without retraining the model. In the paper’s framing, DiT inference is expensive for two coupled reasons: diffusion models require many denoising or sampling steps, and each step invokes a deep Transformer stack with heavy attention and MLP computation over large latent tensors. The method therefore targets repeated per-step computation rather than model weights or training dynamics (Chu et al., 22 Aug 2025).

The proposal is explicitly positioned against prior cache-based diffusion acceleration methods that exploit local inter-step similarity and therefore tend to prioritize reusing later sampling steps. OmniCache argues that this criterion is incomplete. Later timesteps may indeed be more similar to one another, but they also have less denoising capacity left to repair directional errors introduced by cache reuse. Early reuse can inject larger raw error, yet subsequent denoising steps can still absorb and correct it. This motivates the paper’s “trajectory-oriented global perspective”: the relevant question is not merely where adjacent steps are most similar, but where along the entire denoising trajectory reuse can be inserted with minimal harm to the global sampling path (Chu et al., 22 Aug 2025).

A recurring misconception is to read OmniCache as a generic cache manager. In the paper, however, “cache” refers specifically to reusing intermediate module outputs inside a DiT during sampling. It is therefore an inference-time scheduling and correction method for diffusion trajectories, not a request cache, not a KV-cache optimizer, and not a general storage system.

2. Trajectory analysis and reuse scheduling

OmniCache has a calibration stage and an inference stage. The calibration stage collects hidden states $x_t$ along sampling trajectories, reconstructs a low-dimensional approximation of the trajectory, computes timestep-wise curvature, estimates cache-induced noise correlation coefficients $\gamma_t$, and produces a reuse schedule $\mathcal{S}$. The inference stage then applies cache reuse only at selected timesteps in $\mathcal{S}$ (Chu et al., 22 Aug 2025).

The paper studies trajectory geometry through a 3D approximation of the high-dimensional sampling path. Let the sampled path be $\{x_{t_i}\}$. It first defines the endpoint direction

$u = \frac{x_{t_N}-x_{t_0}}{\|x_{t_N}-x_{t_0}\|},$

then projects onto the orthogonal complement of $u$, performs PCA, and extracts principal directions $w_1$ and $w_2$. Each point is approximated in the resulting 3D subspace by

$$x_{t_i}^{\mathrm{proj}} = (x_{t_i}\cdot u)u + (x_{t_i}\cdot w_1)w_1 + (x_{t_i}\cdot w_2)w_2.$$

The empirical finding reported in the paper is that DiT sampling trajectories exhibit a regular, almost content-independent “boomerang” shape. This supports the use of global trajectory structure rather than only local feature similarity as the basis for scheduling reuse (Chu et al., 22 Aug 2025).

The schedule construction is curvature-based. OmniCache places reuse at low-curvature timesteps, on the premise that low curvature means the sampling direction is changing slowly and is therefore less sensitive to perturbation from reused activations. High-curvature regions are treated as critical directional transitions and are preserved as exact computation steps. The paper repeatedly states this rule qualitatively, but it does not provide an explicit curvature formula or an explicit thresholding rule for selecting $\gamma_t$0 in the text summarized here. It does, however, impose one hard constraint: no three consecutive timesteps may use cache reuse. That constraint is tied to the noise-correlation mechanism used later for correction (Chu et al., 22 Aug 2025).

For experiments, two schedules are defined. OmniCache-slow reuses cache on 50% of timesteps, and OmniCache-fast reuses cache on 60% of timesteps. These are the main operating points reported across image and video benchmarks.

3. Cache reuse mechanism and noise compensation

OmniCache performs cache reuse at each DiT block. The paper first illustrates the idea with attention outputs and then states that actual sampling replaces both cached attention outputs and MLP outputs. The method is therefore block- or module-output reuse inside Transformer layers, rather than reuse of top-level latents or LLM-style attention KV states (Chu et al., 22 Aug 2025).

If cache reuse is applied, the denoiser prediction becomes $\gamma_t$1 rather than the exact $\gamma_t$2. The cache-induced error is defined as

$\gamma_t$3

This term perturbs the denoising direction, and repeated perturbations can accumulate across steps. OmniCache’s second major idea is that this perturbation is temporally correlated and can therefore be estimated from nearby steps rather than ignored (Chu et al., 22 Aug 2025).

The paper models neighboring cache-induced noises by a first-order temporal approximation:

$\gamma_t$4

where $\gamma_t$5 is a noise correlation coefficient estimated during calibration. For consecutive reuse steps, the estimate is extended multiplicatively:

$\gamma_t$6

The conceptual link to the trajectory view is that the derivative of cache-induced noise is argued to be related to the second derivative of the sampling trajectory, hence to curvature. Low-curvature steps are therefore the ones for which the correlation-based approximation is expected to be most accurate (Chu et al., 22 Aug 2025).

OmniCache does not subtract the raw estimated noise directly. Instead, it applies stage-dependent filtering before correction. The paper states that early timesteps emphasize global structure and low-frequency information, whereas later timesteps emphasize fine detail and high-frequency information. Accordingly, early-stage correction uses low-pass filtering, and late-stage correction uses high-pass filtering. The correction is applied to the reused denoiser output as

$\gamma_t$7

The exact filter implementation is not specified in the summarized text, only the low-pass/high-pass distinction by diffusion stage (Chu et al., 22 Aug 2025).

This yields a two-part mechanism: curvature-based scheduling determines where reuse is safe, and correlation-based filtered correction limits the resulting directional error. A plausible implication is that OmniCache’s novelty lies less in the raw act of reusing activations than in combining global schedule design with explicit compensation for reuse-induced noise.

4. Experimental performance and ablations

The evaluation spans class-conditional image generation, text-to-video generation, and image-to-video generation. The reported models are DiT-XL/2, OpenSora, Latte, and CogVideoX-5b-I2V-distill. The reported hardware is NVIDIA 4090 GPUs. Datasets and benchmarks include ImageNet for image generation, VBench for text-to-video, and a 100-video image-to-video evaluation for CogVideoX-5b-I2V-distill (Chu et al., 22 Aug 2025).

The main reported results are summarized below.

Model / task	Baseline	OmniCache result
OpenSora text-to-video	VBench 79.22	slow: 78.83 at $\gamma_t$8; fast: 78.50 at $\gamma_t$9
Latte text-to-video	VBench 77.40	slow: 77.24 at $\mathcal{S}$0; fast: 77.09 at $\mathcal{S}$1
CogVideoX-5b-I2V-distill	Aesthetic 0.59, Q-Align 0.79	OmniCache: 0.621, 0.792 at $\mathcal{S}$2
DiT-XL/2 image generation	FID 2.31, sFID 4.98	OmniCache-fast: FID 2.69, sFID 5.86 at $\mathcal{S}$3

On OpenSora, baseline inference is reported as 25.72 s for 30 steps, while OmniCache-slow is reported as 12.80 s total. The same latency breakdown attributes 12.645 s to real inference steps, 0.15 s to cache reuse steps, $\mathcal{S}$4 s to high/low-pass filtering, and $\mathcal{S}$5 s to noise correction. This is the paper’s clearest evidence that the correction machinery is negligible relative to a full DiT forward pass (Chu et al., 22 Aug 2025).

The ablation on CogVideoX-5b-I2V-distill isolates the role of the correction pipeline at fixed $\mathcal{S}$6 acceleration. Baseline scores are Aesthetic Quality 0.59 and Q-Align 0.79. Cache reuse only gives 0.58 and 0.778. Cache reuse plus noise correction gives 0.593 and 0.788. Cache reuse plus noise correction plus filtering gives 0.621 and 0.792. This establishes that naive reuse degrades quality, dynamic correction recovers most of the loss, and stage-dependent filtering provides an additional gain (Chu et al., 22 Aug 2025).

The empirical picture is not that OmniCache dominates every prior method on every setting. On DiT-XL/2, for example, ToCa is reported at $\mathcal{S}$7 speed with FID 2.68, while OmniCache-fast is reported at $\mathcal{S}$8 speed with FID 2.69. The paper’s strongest advantage is instead on video generation and especially on lower-step or distilled regimes, where late-step similarity-based reuse is argued to be less reliable.

5. Position within the cache literature

Within diffusion-model acceleration, OmniCache is most directly related to training-free cache reuse methods such as AdaCache, TGATE, TeaCache, ToCa, FORA, $\mathcal{S}$9-DiT, and PAB, all of which the paper treats as baselines or immediate precursors. Its distinguishing claim is that those methods mostly exploit local inter-step similarity, whereas OmniCache derives reuse placement from global sampling-trajectory analysis and augments it with dynamic noise filtering (Chu et al., 22 Aug 2025).

The broader cache literature uses the word “cache” for substantially different computational objects. SemantiCache, for example, is a training-free KV-cache compression method for long-context autoregressive inference that preserves semantic structure through delimiter-based chunking, clustering over key similarity, and proportional attention after merging; its object of compression is the historical attention cache $\mathcal{S}$0, not intermediate DiT block outputs (Wu et al., 15 Mar 2026). OmniMem is an explicit full-range memory retrieval framework for chunk-based autoregressive video generation that performs sparse retrieval over a historical KV cache and addresses local bias and union explosion; it manages a growing history of chunk-level KV blocks rather than a diffusion sampling trajectory (Zhao et al., 28 May 2026). InstCache, by contrast, is a predictive exact prompt/response cache for LLM serving that stores full user instructions and corresponding answers in a hash table and bypasses inference on an exact hit (Zou et al., 2024).

Taken together, these systems suggest that contemporary “cache” research spans at least three distinct reuse granularities: historical attention states, request/response pairs, and intermediate neural activations. OmniCache belongs to the third category. It is therefore best understood not as a variant of KV-cache optimization, but as a trajectory-aware activation-reuse method specialized to diffusion Transformers.

6. Limitations, scope, and interpretation

The paper states one practical limitation explicitly: cache reuse cannot be applied for three consecutive steps, because the inter-step noise-estimation mechanism becomes unreliable under longer runs of skipped computation (Chu et al., 22 Aug 2025). Additional limitations are also clear from the reported methodology. A calibration stage is required; the exact curvature formula and the exact schedule-construction rule are not fully specified in the summarized text; and the method is tailored to DiT models rather than arbitrary diffusion backbones.

Its strongest empirical case is on video generation and on lower-step or distilled models such as CogVideoX-5b-I2V-distill, where the paper argues that existing cache methods can fail because there is less redundancy and fewer remaining steps to repair late-stage reuse errors. On high-step image generation, the relative advantage is smaller, and strong baselines can be competitive or slightly better on speed-quality tradeoffs (Chu et al., 22 Aug 2025).

The practical significance claimed by the paper is that OmniCache is training-free, does not require model retraining or distillation, has negligible correction overhead, and integrates with existing DiT inference as an inference-time scheduling and correction scheme. This suggests a specific interpretation of its contribution: OmniCache is less a generic cache framework than a statement about where reuse should occur in diffusion sampling. Its contribution is to turn the denoising trajectory itself into the organizing principle for reuse, and to treat cache-induced perturbation as an estimable, filterable noise source rather than as an unavoidable side effect.