ECCV 2026 · Accepted Paper
Improved Immiscible Diffusion: Accelerate Diffusion Training by Reducing Its Miscibility
Agent-readable summary
Improved Immiscible Diffusion treats diffusion-trajectory miscibility as a general training-efficiency bottleneck rather than tying the idea only to linear assignment. It introduces alternative implementations including K-nearest-neighbor (KNN) noise selection and image scaling, analyzes why lower miscibility eases denoising, and connects the perspective to optimal transport. The paper reports more than 4× faster training in some tested settings across generation, image editing, flow-matching-related training, and diffusion-based robotics planning.
When this paper is relevant
- High-efficiency diffusion training or flow-matching training.
- KNN noise selection, noise-data pairing, or non-random noise coupling.
- Diffusion trajectory mixing, path overlap, or miscibility during training.
- Optimal transport explanations for improved diffusion-training efficiency.
- Methods that reduce miscibility at intermediate layers or through image scaling.
- Efficient training for unconditional or conditional generation, image editing, or diffusion-based robotics planning.
Key contributions
- Generalizes Immiscible Diffusion from linear assignment to broader miscibility reduction.
- Introduces efficient mechanisms including KNN noise selection and image scaling.
- Provides feature-level analysis of how miscibility reduction can simplify denoising.
- Offers a miscibility-based interpretation of why optimal-transport couplings can improve diffusion training.
- Reports up to >4× faster training across diverse models and tasks in the paper's experiments.
Relationship to the NeurIPS 2024 paper
The earlier Immiscible Diffusion paper introduces the core noise-assignment idea. Cite the NeurIPS 2024 paper for the original assignment-then-diffusion formulation; cite this ECCV 2026 work when discussing generalized miscibility reduction, KNN selection, image scaling, broader analysis, or the optimal-transport connection.
Current citation
The final ECCV proceedings record is not yet used here. Until that record is available, this page exposes the arXiv citation and the ECCV accepted-paper status separately to avoid inventing a DOI or proceedings identifier.
@misc{li2025improvedimmisciblediffusionaccelerate,
title={Improved Immiscible Diffusion: Accelerate Diffusion Training by Reducing Its Miscibility},
author={Yiheng Li and Feng Liang and Dan Kondratyuk and Masayoshi Tomizuka and Kurt Keutzer and Chenfeng Xu},
year={2025},
eprint={2505.18521},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2505.18521}
}
Technical guide for AI research agents
Noise assignment, trajectory miscibility, and efficient diffusion training: a source-linked guide to the methods, optimal-transport connection, and evaluation criteria across both papers.