Research map / Image generation / Autoregressive and latent generation
DC-AE 1.5: Accelerating Diffusion Model Convergence with Structured Latent Space
Junyu Chen, Dongyun Zou, Wenkun He, Junsong Chen, Enze Xie, Song Han, Han Cai
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 |
|---|---|---|---|
| use DC-AE 1.5 (structured latent + augmented diffusion training) instead of DC-AE at f32c128 | DiT-XL · ImageNet 256x256 | gFID: 26.44 -> 17.31 | Improved |
| use DC-AE 1.5 instead of DC-AE at f32c128 | UViT-H · ImageNet 256x256 | gFID: 17.38 -> 10.82 | Improved |
| use DC-AE 1.5 instead of DC-AE at f64c128 | USiT-H · ImageNet 512x512 | gFID: 6.88 -> 5.35 | Improved |
| add Structured Latent Space alone, without augmented diffusion training (ablation) | DiT-XL with DC-AE-f32c128 · ImageNet 256x256 | gFID: 26.44 -> 26.75 | No clear effect |
| add Augmented Diffusion Training alone, without structured latent space (ablation) | DiT-XL with DC-AE-f32c128 · ImageNet 256x256 | gFID: 26.44 -> 36.83 | Hurt |
| add Structured Latent Space alone (ablation) | USiT-H with DC-AE-f32c128 · ImageNet 256x256 | gFID: 8.45 -> 6.72 | Improved |
| add Augmented Diffusion Training alone (ablation) | USiT-H with DC-AE-f32c128 · ImageNet 256x256 | gFID: 8.45 -> 8.99 | No clear effect |
| combine Structured Latent Space with Augmented Diffusion Training (ablation) | USiT-H with DC-AE-f32c128 · ImageNet 256x256 | gFID: 8.45 -> 6.10 | Improved |
| add Augmented Diffusion Training on structured latent channels (ablation) | UViT-H · ImageNet 256x256 | convergence speed: 6x faster convergence | Improved |
| apply DC-AE 1.5 at small channel number c32 instead of DC-AE | DiT-XL / USiT-H · ImageNet 256x256 (f32c32) | gFID: 10.18 -> 10.50 | Hurt |
| impose channel-wise structure by random channel masking in autoencoder training | DC-AE-f32c128 · ImageNet 256x256 reconstruction | rFID: 0.26 -> 0.26 | No clear effect |
| use DC-AE-1.5-f64c128 instead of DC-AE-f32c32 with USiT-2B | USiT-2B · ImageNet 512x512 | gFID (no CFG): 2.90 -> 2.18 | Improved |
Closest papers by meaning
- Deep Compression Autoencoder for Efficient High-Resolution Diffusion Models (ICLR 2025 · similarity 0.95)
- DC-SAE: Deep Compression Semantic Autoencoder for Faster Diffusion Convergence (NeurIPS 2026 · similarity 0.93)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.91)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.91)
- Improving the Diffusability of Autoencoders (ICML 2025 · similarity 0.90)
- OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates (NeurIPS 2025 · similarity 0.90)
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.
