# DCW — Wavelet-Domain Differential Correction Fixes a Fundamental SNR Bias in Diffusion Models

> DCW fixes a silent training-inference SNR mismatch in diffusion models, improving generation quality across FLUX, EDM, IDDPM, and six other architectures.

- Org: Alibaba AMAP
- Date: 2026-04-17
- Category: paper, algorithm
- Source: https://arxiv.org/abs/2604.16044
- Canonical page: https://ai-tldr.dev/releases/alibaba-amap-dcw-diffusion/

## In short

CVPR 2026 paper identifying a training-inference SNR mismatch in all major diffusion models, plus a wavelet-domain correction (DCW) that fixes it across FLUX, EDM, IDDPM, and six other architectures with negligible overhead.

## What it is

"Elucidating the SNR-t Bias of Diffusion Probabilistic Models" is a CVPR 2026 paper from Lanzhou University and Alibaba AMAP. It identifies a subtle but systematic flaw in how all major diffusion models are trained: during training, each noisy sample's Signal-to-Noise Ratio is tightly tied to its timestep. During inference (denoising), that coupling breaks down — the model accumulates errors because it expects a specific SNR at each step but does not get it.

## How it works

The fix is called DCW (Differential Correction in Wavelet domain). Because diffusion models reconstruct low-frequency details before high-frequency ones during denoising, DCW decomposes each sample into frequency bands using a wavelet transform and applies a separate timestep-dependent correction term to each band. This realigns the expected SNR to the actual SNR at each denoising step. The fix is architecture-agnostic — demonstrated across IDDPM, ADM, DDIM, A-DPM, EA-DPM, EDM, PFGM++, and FLUX — with negligible extra compute.

## Why it matters

FLUX is now one of the most widely used open-source image generation models. A training-aware fix that improves generation quality at no meaningful cost is directly useful to anyone fine-tuning or distilling FLUX-based models. The CVPR 2026 acceptance and 64 HuggingFace upvotes (top paper today) confirm the community sees this as a real, reproducible contribution.

## Who it's for

ML researchers and practitioners training or fine-tuning diffusion models, especially FLUX-based pipelines.

## Try it

git clone https://github.com/AMAP-ML/DCW

## Links

- [Paper](https://arxiv.org/abs/2604.16044)
- [Repo](https://github.com/AMAP-ML/DCW)
- [HF Papers](https://huggingface.co/papers/2604.16044)