Research map / Image generation / Fast diffusion sampling

Is Noise Conditioning Necessary for Denoising Generative Models?

ICML 2025 · Image generation › Fast diffusion sampling · approach: Denoising diffusion probabilistic models · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Systematically tests removing noise conditioning from diffusion, flow-matching, and consistency models, finds most degrade gracefully (flow models sometimes improve), derives a computable error bound, and proposes uEDM, a noise-unconditional EDM variant reaching FID 2.23 on CIFAR-10.

Problem

Noise conditioning (feeding the noise level to the denoiser) is widely regarded as indispensable for denoising generative models; whether it is actually necessary is unclear.

Approach

Motivated by blind image denoising, the authors reformulate denoising generative models so the network sees only the noisy input, derive an effective regression target and a sampling error bound computable from schedules and data alone, and show most models degrade gracefully without the noise level input; they also design uEDM by setting c_out(t)=1.

Key results

• Flow-based FM (1-RF) improves FID on CIFAR-10 without noise conditioning (Euler 3.01 to 2.61).
• uEDM without noise conditioning reaches FID 2.23 on CIFAR-10 versus 1.99 for EDM with conditioning (Heun, 35 NFE).
• DDIM with the deterministic ODE sampler fails catastrophically without noise conditioning (FID 3.99 to 40.90), while most other variants degrade modestly.

Measured effects reported in the paper

Surprising findings

• Some flow-matching variants achieve better FID without noise conditioning than with it.
• Removing noise conditioning from DDIM with an ODE sampler causes catastrophic failure, though stochastic samplers degrade gracefully.

Limitations

• Error bound relies on assumptions (single data point approximation, Lipschitz-type conditions) that are only approximately satisfied on real data, and the bound magnitudes are very loose.
• Hyperparameters were tuned only on noise-conditional models and reused for noise-unconditional ones.
• Experiments are mostly on small-resolution image datasets and the CFG study is limited to SiT-B/2.

Experimental setup

Datasets: CIFAR-10 ImageNet 32x32 FFHQ 64x64 AFHQ-v2 ImageNet 256x256

Benchmarks: CIFAR-10 ImageNet 32x32 FFHQ 64x64 ImageNet 256x256

Metrics: FID NFE

Baselines: iDDPM DDIM ADM EDM FM (1-RF) iCT ECM SiT-B/2

Builds on: EDM DDPM DDIM Rectified Flow iCT

Closest papers by meaning

• Denoising Task Difficulty-based Curriculum for Training Diffusion Models (ICLR 2025 · similarity 0.89)
• Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.89)
• NoiseCtrl: A Sampling-Algorithm-Agnostic Conditional Generation Method for Diffusion Models (CVPR 2025 · similarity 0.89)
• Compressed Image Generation with Denoising Diffusion Codebook Models (ICML 2025 · similarity 0.89)
• Does Generation Require Memorization? Creative Diffusion Models using Ambient Diffusion (ICML 2025 · similarity 0.88)
• Assessing the quality of denoising diffusion models in Wasserstein distance: noisy score and optimal bounds (NeurIPS 2025 · similarity 0.88)