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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| chunk fused QKV, FFN and AdaLN matrices before orthogonalization | Muon · ImageNet 256 DiT-XL VA-VAE ep200 | FID: 1.29 -> 1.18 | Improved |
| chunk fused QKV, FFN and AdaLN matrices before orthogonalization | Muon · ImageNet 256 DiT-B VA-VAE ep200 | FID: 3.31 -> 3.03 | Improved |
| chunk fused matrices | Muon · ImageNet 256 DiT-B VA-VAE ep400 | FID: 2.78 -> 2.57 | Improved |
| chunk fused matrices | Muon · ImageNet 256 DiT-B SD-VAE ep200 | FID: 3.30 -> 3.06 | Improved |
| replace AdamW with CMuon | AdamW · ImageNet 256 DiT-XL | FID: 1.21 at 400 epochs -> 1.18 at 200 epochs | Improved |
| use Muon instead of AdamW | AdamW · ImageNet 256 DiT-B VA-VAE ep400 | FID: 2.78 -> 2.78 | No clear effect |
| chunk only FFN projections (ablation) | Muon · DiT-B ep200 | FID: 3.31 -> 3.28 | No clear effect |
| chunk only QKV projections (ablation) | Muon · DiT-B ep200 | FID: 3.31 -> 3.35 | Hurt |
| chunk only AdaLN projections (ablation) | Muon · DiT-B ep200 | FID: 3.31 -> 3.23 | Improved |
| chunk FFN, QKV and AdaLN jointly rather than any single block (ablation) | Muon · DiT-B ep200 | FID: 3.31 -> 3.02 | Improved |
| use MoonLight scaling instead of Vanilla scaling (ablation) | CMuon · DiT-B ep200 | FID: 6.94 -> 3.03 | Improved |
| add sqrt(N_chunk) learning-rate rescale (ablation) | CMuon · DiT-B ep40 | FID: 5.32 -> 3.78 | Improved |
Closest papers by meaning
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 2025 · similarity 0.90)
- DC-DiT: Adaptive Compute and Elastic Inference for Visual Generation via Dynamic Chunking (NeurIPS 2026 · similarity 0.88)
- DiC: Rethinking Conv3x3 Designs in Diffusion Models (CVPR 2025 · similarity 0.88)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.88)
- Stretching Each Dollar: Diffusion Training from Scratch on a Micro-Budget (CVPR 2025 · similarity 0.87)
- Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers (ICML 2024 · similarity 0.87)
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.
