Skip to main content
Back to timeline
researchsquare.comSource publication:

Diffusion cache skips the full recompute: DACE+DTC cuts Qwen-Image-Edit latency 12.1% and LPIPS 18.9%

Synopsis

Diffusion inference uses learned feature caching to skip full denoising forwards, but cache prediction errors are written into the sampling state and accumulate. DACE runs cache decomposition and prediction only for target tokens and manages history by future read positions; DTC measures an endpoint counterfactual at scheduled full steps, estimates the interval defect integral with signed step sizes, and applies correction after the normal update. On Qwen-Image-Edit background replacement under the N7 schedule, the pair lowers GPU-resident latency by 12.1% and mean LPIPS distance by 18.9% relative to LinCa, while DTC adds 2.1% overhead to DACE and further lowers LPIPS by 18.6%.

Source-provided article image: DTC: Dependency-Aware Cache Execution and Deferred Trajectory Correction
Figure 1 · researchsquare.com · Page 4

Interpretation

When accelerating diffusion inference with caching, features can be computed only for the target image tokens that are finally consumed, skipping cache decomposition, prediction and output projection for condition tokens. Earlier cache methods still ran decomposition and prediction over all tokens at cache steps, although condition-token cache features were not used in final consumption. On Qwen-Image-Edit background replacement under the fixed N7 schedule, DACE alone cut latency from LinCa's 30.270 s to 26.061 s, a 13.91% reduction, and peak memory from 61.219 GiB to 59.116 GiB.

State displacement caused by cache prediction can be corrected retroactively at later full steps without changing model weights or adding full DiT calls. Previously the sampler wrote cache prediction error directly into the latent state, and refreshing history at full steps updated later predictions while the accumulated displacement stayed in the sampling state. Adding DTC on top of DACE under N7 lowered LPIPS from 0.1888 to 0.1537, an 18.60% reduction, at a cost of 0.545 s, about 2.09% extra time; the paired LPIPS difference versus LinCa was −0.0358 with a 95% interval of [−0.0580, −0.0136].

Execution pruning and endpoint feedback can share one dependency plan, keeping pruned history alive until the feedback read completes. Releasing history purely by output dependency would remove the old-history read DTC needs before refresh at a full step, so the two layers could not hold together. Endpoint reads are counted in the full step's read set; N7 keeps 18 CFG-branch full forwards and raises predictor inverse transforms from 82 to 96, with final latents elementwise identical to the zero-gain control.

Perspective

The demonstrated setting is Qwen-Image-Edit background-replacement editing at 1024×1024 resolution with 50-step Euler sampling and CFG strength 4, using LinCa's pretrained predictor with both sets of weights fixed and 18 CFG-branch full calls under the N7 schedule; in external comparisons text conditions are pre-encoded and the encoder moved to CPU, while the module ablation keeps the full pipeline GPU-resident, and the two timing protocols are reported separately.

The confirmation cohort has only eight images, and per-image results show LPIPS improving on six and worsening on two while PSNR improves on seven and worsens on one; image 8 loses 3.937 dB and affects the mean. The predeclared test that N9+DTC would reduce both LPIPS and runtime relative to N7 DACE did not pass, with an LPIPS difference of +0.00764 and interval [−0.02460, 0.04531], four images improving and four regressing. Under N7 the last five cache steps and under N9 the last two receive no correction.