[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-2236":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":16,"subscribersCount":16,"size":16,"stars1d":15,"stars7d":17,"stars30d":18,"stars90d":16,"forks30d":16,"starsTrendScore":13,"compositeScore":19,"rankGlobal":10,"rankLanguage":10,"license":20,"archived":21,"fork":21,"defaultBranch":22,"hasWiki":21,"hasPages":21,"topics":23,"createdAt":10,"pushedAt":10,"updatedAt":24,"readmeContent":25,"aiSummary":26,"trendingCount":16,"starSnapshotCount":16,"syncStatus":17,"lastSyncTime":27,"discoverSource":28},2236,"DCW","AMAP-ML\u002FDCW","AMAP-ML","[CVPR 2026] Elucidating the SNR-t Bias of Diffusion Probabilistic Models","",null,"Python",122,3,114,1,0,2,7,45.51,"Other",false,"main",[],"2026-06-12 04:00:13","# [CVPR 2026] Elucidating the SNR-t Bias of Diffusion Probabilistic Models\n\nDiffusion Probabilistic Models have demonstrated remarkable performance across a wide range of generative tasks. However, we have observed that these models often suffer from a Signal-to-Noise Ratio–timestep (SNR-t) bias. This bias refers to the misalignment between the SNR of the denoising sample and its corresponding timestep during the inference phase. Specifically, during training, the SNR of a sample is strictly coupled with its timestep. However, this correspondence is disrupted during inference, leading to error accumulation and impairing the generation quality. We provide comprehensive empirical evidence and theoretical analysis to substantiate this phenomenon and propose a simple yet effective differential correction method to mitigate the SNR-t bias. Recognizing that diffusion models typically reconstruct low-frequency components before focusing on high-frequency details during the reverse denoising process, we decompose samples into various frequency components and apply differential correction to each component individually. Extensive experiments show that our approach significantly improves the generation quality of various diffusion models (IDDPM, ADM, DDIM, A-DPM, EA-DPM, EDM, PFGM++, and FLUX) on datasets of various resolutions with negligible computational overhead.\n\n## Overview\n![framework](assets\u002Fframework.png)\n\nThis is the codebase for our paper **Elucidating the SNR-t Bias of Diffusion Probabilistic Models (CVPR 2026)**.\nThe repository is heavily based on [EDM](https:\u002F\u002Fgithub.com\u002FNVlabs\u002Fedm). For environment setup, datasets preparation, pre-trained models loading, and fid calculation, please refer to the official EDM code [repository](https:\u002F\u002Fgithub.com\u002FNVlabs\u002Fedm). \n\n## Requirements\nAfter setting up the environment required for the EDM base model, you also need to install the packages related to wavelet transform.\n```python\npip install pytorch_wavelets \npip install PyWavelets\n```\n## Introduction\nNow, we integrate the Differential Correction in Wavelet domain (DCW) into the EDM base model.\n\n(1) Import the required packages.\n```python\nfrom pytorch_wavelets import DWTForward, DWTInverse\n```\n\n(2) Define DWT and IDWT. \n```python\ndwt = DWTForward(J=1, mode='zero', wave='haar').cuda()\niwt = DWTInverse(mode='zero', wave='haar').cuda()\n```\nMoreover, it is possible to employ different wavelet bases, for instance, db4 and sym8.\n\n(3) Define various correction functions.\n```python\ndef dcw_pix(x,y,scaler):\n\t......\n\t\t\ndef dcw_low(x, y, scaler):\n    x = x.to(torch.float32)\n    y = y.to(torch.float32)\n    xl, xh = dwt(x)\n    yl, yh = dwt(y)\n    xl = xl + scaler * (xl - yl)\n    x_new = iwt((xl, xh))\n    return x_new\n\t\t\t\ndef dcw_high(x, y, scaler):\n\t......\n\t    \ndef dcw_double(x, y, low_scaler, high_scaler):\n\t......\n```\n(4) Apply the correction function.\n\nLow-frequency correction, as shown in Eqs.18 and 20 of our paper.\n```python\nx_next = dcw_low(x_next, denoised, (t_steps[i] \u002F sigma_max) * corr_scaler)\n```\n## FLUX-DCW\nHere, we provide the core code for FLUX-DCW. More specifically, we only need to modify the step function in FlowMatchEulerDiscreteScheduler. Simply perform the following operations in sequence to run the code successfully:\n\n(1) First, we assume that you are already able to directly use [FLUX](https:\u002F\u002Fhuggingface.co\u002Fblack-forest-labs\u002FFLUX.1-dev) for image generation. \n\n(2) Then, please install the wavelet transform-related libraries via the commands: pip install pytorch_wavelets and pip install PyWavelets. \n\n(3) Finally, you directly replace your FlowMatchEulerDiscreteScheduler.py with the one we provide, and pass in the parameter scaler by yourself to run it successfully.\n\n## Contact\nIf you have any questions, feel free to contact the first author. (Email: `mengyu23@sjtu.edu.cn`)\n\n## Citation\n```\n@article{yu2026eluci,\n  title={Elucidating the SNR-t Bias of Diffusion Probabilistic Models},\n  author={Meng Yu and Lei Sun and Jianhao Zeng and Xiangxiang Chu and Kun Zhan},\n  journal={arXiv preprint arXiv:2604.16044},\n  year={2026}\n}\n```\n","该项目旨在解决扩散概率模型在生成任务中遇到的信噪比-时间步（SNR-t）偏差问题。通过详细的实证分析与理论研究，项目揭示了训练和推理阶段信噪比与时序之间的不匹配现象，并提出了一种基于小波域差分校正的方法来减轻这种偏差。该方法通过对样本进行频率分解，并对各频率成分分别应用校正策略，显著提升了多种扩散模型（如IDDPM、ADM等）在不同分辨率数据集上的生成质量，且额外计算开销极小。适用于需要高质量图像生成的应用场景，特别是在对细节还原度有较高要求的情况下。","2026-06-11 02:49:01","CREATED_QUERY"]