ResilPhase diffusion acceleration: macro-trajectory extrapolation

ResilPhase, a new noise-resilient acceleration framework by Qicheng Zhao, Yu Li, Qi Sun and Zheyu Yan, rethinks how diffusion inference is extrapolated to cut latency. The paper was submitted to arXiv on 25 Jun 2026 (arXiv:2606.26769) and is listed as accepted by ECCV 2026.

What is ResilPhase and how does it work?

ResilPhase reframes accelerated diffusion inference as stable macro-trajectory extrapolation in ODE space and aligns forecasting with the model's Global Drift (GD), the end-to-end state evolution. The framework removes feature inconsistency and memory overhead by predicting in ODE/GD space rather than on intermediate features. To avoid noisy higher-order derivatives, it uses a derivative-free barycentric Lagrange extrapolator and a bounded Phase Mapping that constrains the extrapolation domain and suppresses oscillatory error growth.

The paper names three core elements: Global Drift alignment, a derivative-free barycentric Lagrange extrapolator, and bounded Phase Mapping. Together these components form ResilPhase, which the authors describe as a "noise-resilient acceleration framework." The submission PDF on arXiv for this work is 41,484 KB in size and lists experiments on the FLUX.1-dev and HunyuanVideo datasets.

How does ResilPhase differ from "cache-then-forecast" and why prior methods fail?

ResilPhase targets the failure modes of recent "cache-then-forecast" schemes, which use derivative-based polynomials on intermediate features and suffer heavy quality degradation at high acceleration ratios. The authors identify the root cause as discrete extrapolation on representations that are misaligned with the continuous diffusion trajectory and numerically unstable. That misalignment produces accumulated spatial errors, noisy derivative amplification and high-order instability.

Instead of extrapolating feature-level representations, ResilPhase forecasts the macro-trajectory in ODE/GD space, eliminating feature inconsistency. The paper argues that even smooth macro-trajectories carry intrinsic noise in higher-order temporal derivatives, which is why the derivative-free barycentric Lagrange extrapolator is central: it bypasses derivative instability and approximation error that plague derivative-based polynomial forecasts.

What evidence do the authors provide?

The paper reports experimental validation on two named datasets, FLUX.1-dev and HunyuanVideo, and states that ResilPhase demonstrates state-of-the-art fidelity under aggressive acceleration ratios. The arXiv record (arXiv:2606.26769) and the submission metadata confirm the authorship and venue acceptance: submitted 25 Jun 2026 and accepted by ECCV 2026. The manuscript file available through arXiv is 41,484 KB.

Why it matters

ResilPhase tackles two practical obstacles for deploying diffusion-based models: inference latency and instability when extrapolating multiple steps. By moving forecasting into ODE/GD space and adopting a derivative-free extrapolator plus bounded Phase Mapping, the method directly addresses the numerical instabilities that cause quality collapse at high speedups. That approach could let practitioners push acceleration ratios higher without the sharp fidelity losses seen in derivative-based cache-then-forecast methods.

What to watch

Look for ECCV 2026 proceedings or the paper's final materials for quantitative benchmarks and implementation details; the arXiv entry lists experiments on FLUX.1-dev and HunyuanVideo but the submission itself provides the full methods and results. Also watch for released code or model artifacts tied to the ResilPhase implementation that confirm real-world speed and fidelity gains.

References: ResilPhase, Qicheng Zhao, Yu Li, Qi Sun, Zheyu Yan; arXiv:2606.26769, submitted 25 Jun 2026; accepted by ECCV 2026. Experiments cited on FLUX.1-dev and HunyuanVideo datasets.