Test-Time Iterative Error Correction for Efficient Diffusion Models

TL;DR

This work tackles the challenge that efficiency techniques for diffusion models introduce approximation errors that can grow exponentially during denoising, hindering deployment on constrained devices. It proposes Iterative Error Correction (IEC), a test-time, plug-and-play refinement that iteratively corrects diffusion predictions using a contraction-mapping update, transforming exponential error growth into linear growth without retraining or architecture changes. The authors provide a theoretical analysis showing fixed-point convergence under appropriate \lambda and demonstrate that IEC bounds accumulated error and enables linear propagation across timesteps. Empirically, IEC yields consistent improvements across multiple diffusion architectures and efficiency techniques (quantization, DeepCache, CacheQuant) on datasets like CIFAR-10, LSUN, ImageNet, and MS-COCO, with configurable overhead. The approach offers a practical, generalizable strategy for boosting deployed diffusion models, enabling flexible accuracy-efficiency trade-offs in real-world applications.

Abstract

With the growing demand for high-quality image generation on resource-constrained devices, efficient diffusion models have received increasing attention. However, such models suffer from approximation errors introduced by efficiency techniques, which significantly degrade generation quality. Once deployed, these errors are difficult to correct, as modifying the model is typically infeasible in deployment environments. Through an analysis of error propagation across diffusion timesteps, we reveal that these approximation errors can accumulate exponentially, severely impairing output quality. Motivated by this insight, we propose Iterative Error Correction (IEC), a novel test-time method that mitigates inference-time errors by iteratively refining the model's output. IEC is theoretically proven to reduce error propagation from exponential to linear growth, without requiring any retraining or architectural changes. IEC can seamlessly integrate into the inference process of existing diffusion models, enabling a flexible trade-off between performance and efficiency. Extensive experiments show that IEC consistently improves generation quality across various datasets, efficiency techniques, and model architectures, establishing it as a practical and generalizable solution for test-time enhancement of efficient diffusion models.

Figures (11)

• Figure 1: Empirical results of (a) $\|A_t + B_t J_t\|$; (b) $\|\nabla G(x)\|$ under varying $\lambda$. Reported values are averaged over 100 sample generations using a DDIM pretrained on CIFAR-10.
• Figure 2: Error comparison across timesteps.
• Figure 3: Ablation study of applying IEC on quantization and Deepcache. The baseline does not use IEC, while "All Steps" applies IEC to all timesteps. The "$\pm$A" indicates that IEC is applied to both the first and last A timesteps.
• Figure 4: Qualitative comparison of Stable Diffusion on the COCO dataset: Baseline vs. CacheQuant (W8A8, N=5) with and without IEC.
• Figure 5: Qualitative comparison of Stable Diffusion on the COCO dataset: Baseline vs. CacheQuant (W4A8, N=10) with and without IEC.