推論系統
DriveCache Allocates Denoising Cache Based on Planned Vehicle Trajectories, Maintaining Higher Fidelity at 1.97× Speed on Wan2.2 A14B
DriveCache uses vehicle translation and rotation known before generation to schedule diffusion-model feature reuse across different driving scenarios. On Wan2.2 A14B, it delivers 1.276 dB higher PSNR than SeaCache at a nearly identical acceleration level, though the code has not yet been released.
Diffusion-based driving world models must repeatedly run large backbones across multiple denoising steps. Existing caching methods generally determine when to reuse computation using either fixed intervals or feature changes observed during generation. DriveCache, submitted on August 17, instead reads the ego trajectory already provided by the planning system and uses cumulative translation distance and rotation angle to estimate each scene’s tolerance for caching error. Stationary or straight-line scenes can reuse more features, while turning scenarios are handled more conservatively.
The system first runs complete denoising trajectories on two anchors—low-motion and moving-turning scenarios—then uses a Jacobian-vector product to estimate the combined perturbation that consecutive reuse introduces into the final latent. Dynamic programming selects the number and positions of cache operations within an error budget. During inference, if input drift exceeds the calibrated range, the controller cancels that reuse operation, refreshes the features, and replans the remaining steps. It does not modify model weights or the number of sampling steps, but it must respect Wan2.2 A14B’s two-expert boundary: features cannot be reused across the high-noise and low-noise experts.
The authors tested Wan2.2 5B, A14B, and Epona on a single H20 GPU with batch size 1. The 5B configuration achieved a 1.84× speedup and 34.744 dB PSNR; A14B reached a 1.97× speedup and 32.778 dB, outperforming SeaCache by 1.884 dB and 1.276 dB, respectively, at comparable speedup levels. Epona achieved a 1.89× speedup. Calibrating A14B takes about nine minutes, while controller overhead accounts for less than 0.7% of end-to-end latency.
The engineering value lies in incorporating prior signals from the control system into the generator’s scheduling rather than relying solely on generic image-caching rules. However, the evaluation remains limited to three author-selected generation configurations and a single-GPU environment. PSNR, LPIPS, and trajectory metrics also cannot substitute for closed-loop driving safety tests. The paper currently only promises that the code will be released later, so external teams cannot yet fully reproduce the results.