Cross-Resolution Distribution Matching for Diffusion Distillation

2026-03-06

AI Summary

Existing diffusion distillation methods suffer from a significant degradation in cross-resolution generation quality when reducing inference steps, primarily due to the mismatch between low- and high-resolution data distributions. To address this issue, this work proposes a Resolution-aware Matching Distillation (RMD) framework that systematically mitigates cross-resolution distribution discrepancies. RMD partitions timesteps according to the logSNR schedule, introduces a logSNR mapping to compensate for resolution-induced shifts, and aligns distributions along the resolution trajectory. Coupled with a predicted noise re-injection mechanism, RMD substantially enhances training stability and generation fidelity. The method achieves up to 33.4× and 25.6× inference speedups on SDXL and Wan2.1-14B, respectively, while preserving high visual quality.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Distributed Machine Learning & Federated Learning
• Natural Language Processing: Generation

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Diffusion distillation is central to accelerating image and video generation, yet existing methods are fundamentally limited by the denoising process, where step reduction has largely saturated. Partial timestep low-resolution generation can further accelerate inference, but it suffers noticeable quality degradation due to cross-resolution distribution gaps. We propose Cross-Resolution Distribution Matching Distillation (RMD), a novel distillation framework that bridges cross-resolution distribution gaps for high-fidelity, few-step multi-resolution cascaded inference. Specifically, RMD divides the timestep intervals for each resolution using logarithmic signal-to-noise ratio (logSNR) curves, and introduces logSNR-based mapping to compensate for resolution-induced shifts. Distribution matching is conducted along resolution trajectories to reduce the gap between low-resolution generator distributions and the teacher's high-resolution distribution. In addition, a predicted-noise re-injection mechanism is incorporated during upsampling to stabilize training and improve synthesis quality. Quantitative and qualitative results show that RMD preserves high-fidelity generation while accelerating inference across various backbones. Notably, RMD achieves up to 33.4X speedup on SDXL and 25.6X on Wan2.1-14B, while preserving high visual fidelity.

Problem

• diffusion distillation
• cross-resolution distribution gap
• low-resolution generation
• inference acceleration
• distribution mismatch

Innovation

• Cross-Resolution Distribution Matching
• Diffusion Distillation
• logSNR-based Mapping
• Noise Re-injection
• Multi-resolution Cascaded Inference