Elucidating the Design Space of Arbitrary-Noise-Based Diffusion Models

2025-07-24

AI Summary

Existing diffusion models enforce fixed Gaussian noise schedules, leading to image degradation artifacts, excessively long reverse trajectories, and high reconstruction complexity. To address this, we propose the Explicit Diffusion Adaptation (EDA) framework—the first approach to decouple noise design from the Gaussian constraint in diffusion modeling. We theoretically prove that arbitrary noise injection incurs no additional computational overhead. EDA preserves the architectural flexibility of EDM-style modules while introducing a five-step efficient sampling strategy. Evaluated on three distinct inverse problems—MRI bias field correction, CT metal artifact reduction, and natural image shadow removal—EDA achieves state-of-the-art performance: it surpasses most task-specific methods using only five sampling steps, significantly improving reconstruction accuracy, robustness to noise and corruption, and generalization across diverse degradation scenarios.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

EDM elucidates the unified design space of diffusion models, yet its fixed noise patterns restricted to pure Gaussian noise, limit advancements in image restoration. Our study indicates that forcibly injecting Gaussian noise corrupts the degraded images, overextends the image transformation distance, and increases restoration complexity. To address this problem, our proposed EDA Elucidates the Design space of Arbitrary-noise-based diffusion models. Theoretically, EDA expands the freedom of noise pattern while preserving the original module flexibility of EDM, with rigorous proof that increased noise complexity incurs no additional computational overhead during restoration. EDA is validated on three typical tasks: MRI bias field correction (global smooth noise), CT metal artifact reduction (global sharp noise), and natural image shadow removal (local boundary-aware noise). With only 5 sampling steps, EDA outperforms most task-specific methods and achieves state-of-the-art performance in bias field correction and shadow removal.

Problem

• Extends diffusion models beyond Gaussian noise limitations
• Reduces image corruption from forced Gaussian noise injection
• Enables efficient restoration with complex noise patterns

Innovation

• Expands noise pattern freedom in diffusion models
• Preserves module flexibility without computational overhead
• Validated on MRI, CT, and natural image tasks