ResilPhase: Plug-and-Play Phase Mapping and Noise-Resilient Macro-Trajectory Extrapolation for Diffusion Acceleration

2026-06-25 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

ResilPhase is a new plug-and-play framework designed to accelerate diffusion model inference, specifically addressing the high latency of powerful Diffusion Transformers (DiTs). Existing "cache-then-forecast" methods suffer significant quality degradation at high acceleration ratios due to discrete extrapolation on misaligned and numerically unstable representations, leading to accumulated spatial errors and noisy derivative amplification. ResilPhase reformulates accelerated inference as stable macro-trajectory extrapolation within ordinary differential equation (ODE) space. It aligns forecasting with the model's Global Drift (GD) to resolve feature inconsistency and memory overhead. Furthermore, it incorporates a derivative-free barycentric Lagrange extrapolator to circumvent derivative instability and approximation errors, alongside a bounded Phase Mapping that regularizes the extrapolation domain to suppress oscillatory error growth. Experiments on FLUX.1-dev and HunyuanVideo demonstrate high fidelity even under aggressive acceleration ratios.

Key takeaway

For Machine Learning Engineers optimizing diffusion model inference, ResilPhase offers a critical advancement. If you are struggling with quality degradation at high acceleration ratios using existing "cache-then-forecast" schemes, you should evaluate ResilPhase. Its approach of stable macro-trajectory extrapolation in ODE space, combined with derivative-free methods and bounded phase mapping, significantly improves fidelity. This enables you to achieve superior performance on models like FLUX.1-dev and HunyuanVideo, reducing latency without sacrificing output quality.

Key insights

ResilPhase accelerates diffusion models by reformulating inference as stable ODE-space macro-trajectory extrapolation, using derivative-free methods and phase mapping.

Principles

• Discrete extrapolation causes quality degradation in diffusion acceleration.
• Align forecasting with Global Drift for consistency.
• Derivative-free extrapolation bypasses noise instability.

Method

Reformulate inference as stable macro-trajectory extrapolation in ODE space, align forecasting with Global Drift, use a derivative-free barycentric Lagrange extrapolator, and apply bounded Phase Mapping.

In practice

• Apply ResilPhase to accelerate DiTs.
• Improve fidelity in video diffusion models.
• Reduce inference latency for generative AI.

Topics

• Diffusion Models
• Diffusion Acceleration
• Diffusion Transformers
• ODE Extrapolation
• Lagrange Extrapolator
• Video Generation

Code references

• yuanzhi-zhu/DiffPIR
• hustvl/DiffusionDrive

Best for: Research Scientist, AI Engineer, Computer Vision Engineer, AI Scientist, Machine Learning Engineer

Related on AIssential

• SURGE: Approximation-free Training Free Particle Filter for Diffusion Surrogate — SURGE offers an approximation-free particle filter for diffusion models, rigorously fusing observations with simulations via path-space resampling.
• Accelerating Disaggregated RL for Visual Generative LLMs with Diffusion-Based Parallelism and Trainer-Assisted Generation — DigenRL disaggregates RL for visual generative LLMs, using diffusion-based parallelism and trainer assistance for 1.56-2.10x throughput.
• Guided Trajectory Optimization with Sparse Scaling for Test-Time Diffusion — Reward-guided Trajectory Scaling (RTS) optimizes diffusion model noise paths via sparse, PCA-driven key-step selection and a coarse-to-fine search.
• Budget-Constrained Step-Level Diffusion Caching — Optimizing diffusion model cache policies for fixed compute budgets significantly improves generation quality.
• Simple Approximation and Derivative Free Inference-Time Scaling for Diffusion Models via Sequential Monte Carlo on Path Measures — URGE improves diffusion model sample quality via derivative-free pathwise importance reweighting and resampling.
• Wasserstein Convergence of ODE-Based Samplers in Decentralized Diffusion Model via Velocity Field Decomposition — The first W₂ convergence guarantee for decentralized diffusion models with ODE-based sampling is established, showing ℮(N⁻¹/²+ε) rate.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.