AI Paper Atlas

Research map / Image generation / Autoregressive and latent generation

CMuon: Accelerating and Stabilizing Diffusion Transformer Training via Chunked Momentum Orthogonalization

Chuyan Chen, Peng Sun, Kun Yuan

ECCV 2026 · Image generation › Autoregressive and latent generation · approach: High-resolution diffusion synthesis · 3 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
chunk fused QKV, FFN and AdaLN matrices before orthogonalizationMuon · ImageNet 256 DiT-XL VA-VAE ep200FID: 1.29 -> 1.18Improved
chunk fused QKV, FFN and AdaLN matrices before orthogonalizationMuon · ImageNet 256 DiT-B VA-VAE ep200FID: 3.31 -> 3.03Improved
chunk fused matricesMuon · ImageNet 256 DiT-B VA-VAE ep400FID: 2.78 -> 2.57Improved
chunk fused matricesMuon · ImageNet 256 DiT-B SD-VAE ep200FID: 3.30 -> 3.06Improved
replace AdamW with CMuonAdamW · ImageNet 256 DiT-XLFID: 1.21 at 400 epochs -> 1.18 at 200 epochsImproved
use Muon instead of AdamWAdamW · ImageNet 256 DiT-B VA-VAE ep400FID: 2.78 -> 2.78No clear effect
chunk only FFN projections (ablation)Muon · DiT-B ep200FID: 3.31 -> 3.28No clear effect
chunk only QKV projections (ablation)Muon · DiT-B ep200FID: 3.31 -> 3.35Hurt
chunk only AdaLN projections (ablation)Muon · DiT-B ep200FID: 3.31 -> 3.23Improved
chunk FFN, QKV and AdaLN jointly rather than any single block (ablation)Muon · DiT-B ep200FID: 3.31 -> 3.02Improved
use MoonLight scaling instead of Vanilla scaling (ablation)CMuon · DiT-B ep200FID: 6.94 -> 3.03Improved
add sqrt(N_chunk) learning-rate rescale (ablation)CMuon · DiT-B ep40FID: 5.32 -> 3.78Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy