Exponential feature forecasting for training-free diffusion caching

Feature caches (TaylorSeer, HiCache) skip the network on most denoising steps and forecast the cached features from recent compute-step anchors. HiCache++ adds the exponential (DMD/Prony) basis and a training-free holdout selector, with the basis picked by model family.

pip install hicache-pp

Evidence

Synthetic feature-ODE forecast, H=8	2.2e-7 rel. L2 error for the exponential basis vs 6.5e0 (monomial) and 1.2e0 (rational)
Regime-switch detection (auto holdout)	120/120 windows detected; long-horizon error contained at 1.56 vs 9.40 forced-exponential
Hunyuan3D-2.1, interval 5	0.860 F-score@0.05 (DMD) vs 0.735 Hermite arm at 1.79x speedup; lead grows to +0.24 at i6
Hunyuan3D-2-mini, interval 5	0.794 = baseline, exactly lossless F-score@0.05
SAM3D, DMD interval 6	F1 = 1.000 geometry-lossless vs baseline at 1.56x speedup (CD 0.013)
TRELLIS v1 sparse-structure stage	0.829 F-score@0.05 at 2.76x, −0.010 vs vanilla (Hermite 0.825, −0.014)
DiT-XL/2 corrected polynomial ladder	2.27 paired-noise FID drift (TaylorSeer, 3.81x); corrected Hermite 3.54 / 6.46 / 10.74 at i4 / i6 / i8
Per-window eigendecomposition cache	483 ms/img for dmd_i4 (was 566; 3.17x → 3.71x), 2.5–3.1x cheaper per forecast on CPU

Media

Runtime decision path: a content-bound budget selects the backend, audits held-out error, then forecasts or evaluates.

FLUX A/B on Chroma1-HD: vanilla vs TaylorSeer vs DMD — FLUX-class A/B on Chroma1-HD (same seed): vanilla / TaylorSeer / DMD at 1.00x / 1.05x / 1.11x with parity images.

Install & verify

python -m hicache_pp.hermite     # Hermite basis + schedule + sign-convention regressions
python -m hicache_pp.dmd         # exponential basis + auto holdout modes + eigencache
python -m hicache_pp.tree        # tree-aware Hermite + exponential + Adaptive-CFG + eigencache
python tests/run_tests.py        # full CPU suite, passes 30 tests; also pytest-discoverable

Limits

What this release does not claim (from the release notes):

GPU-gated benchmark cells (FLUX/DMD head-to-heads need CUDA).

Family

Downstream adapters	TRELLIS v1: faster-trellis, faster-trellis-plus-plus · TRELLIS.2: fast-trellis2, hermit-trellis2, hermit-trellis2-plus-plus · Hunyuan3D-2: hunyuan2-plus, hunyuan2-plus-plus · Hunyuan3D-2.1: hunyuan2.1-plus, hunyuan2.1-plus-plus · SAM: sam3d-plus, sam3d-plus-plus, fastsam3d-plus, fastsam3d-plus-plus
ComfyUI nodes	ComfyUI-HiCache · ComfyUI-TRELLIS-HiCache · ComfyUI-TRELLIS2-HiCache
Builds on	TaylorSeer forecasting (cache-dit implementation) · HiCache Hermite basis
External	TaylorSeer arXiv:2503.06923 · HiCache arXiv:2508.16984 · reference impl