Work

DTC: Dependency-Aware Cache Execution and Deferred Trajectory Correction

Year: 2026
Type: preprint
Abstract: Learned feature caching reduces diffusion model inference latency by replacing full denoising forwards with low-cost predictions. We propose a joint approach combining Dependency-Aware Cache Execution (DACE) and Deferred Trajectory Correction (DTC). When cache decomposition, prediction, and output projection are token-independent, DACE computes features only for target tokens and manages history liveness by future read positions, reducing cache-path computation and storage. DTC reuses scheduled full steps to measure endpoint counterfactual defects before old history refresh, approximates the error integral using signed sampler step sizes and cache age, and applies correction after normal state updates. A unified dependency plan makes execution pruning compatible with feedback-required history; both layers preserve model weights and the scheduled full DiT call budget. On Qwen-Image-Edit background-replacement tasks, the full method reduces GPU-resident N7 latency by 12.1% and mean LPIPS distance by 18.9% relative to LinCa; DTC adds 2.1% overhead to DACE while further reducing LPIPS by 18.6%.
Institution: IFlyTek (China)
Language: English
Open Access status: green
OpenAlex ID: W7233906259
Cites: 32
Cited by:
Related to: 10