Research map / Image generation / Autoregressive and latent generation
Decoupling Complexity from Scale in Latent Diffusion Model
Tianxiong Zhong, Xingye Tian, Xuebo Wang, Boyuan Jiang, Xin Tao, Pengfei Wan
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 entropy-guided content-adaptive token allocation across patches instead of uniform (ablation) | DCS-Tok · ImageNet val reconstruction (avg 2 levels per patch) | PSNR: 34.25 -> 34.99 | Improved |
| add L2 regularization with margin to latent denoising (ablation) | DCS-Tok · Tokenizer ablation, generation fragmentation | fragmentation / PSNR at 8 levels: 29.64 -> 32.61 PSNR at 8 levels; fragmentation removed | Improved |
| add latent denoising task alone (no L2 regularization) (ablation) | DCS-Tok · Tokenizer ablation, generation | generation fragmentation | No clear effect |
| add L2 regularization without margin (ablation) | DCS-Tok · Tokenizer ablation, reconstruction | PSNR at 5-8 levels: 33.15 -> 29.64 at 8 levels | Hurt |
| randomly drop tail token levels per patch during training (structured latents) (ablation) | DCS-Tok · qualitative reconstruction | hierarchy of latents | Improved |
| use linear level sampling probability instead of uniform (ablation) | DCS-Tok · reconstruction | contribution of later-level tokens | Improved |
| add more token levels in reconstruction (ablation) | DCS-Tok · Tokenizer ablation (no denoise, no L2) | PSNR: 24.22 -> 33.22 (1 -> 8 levels) | Improved |
| DCS-LDM with first-level tokens vs state-of-the-art LDMs on ImageNet | DCS-LDM · ImageNet 256x256 label-to-image (w/o CFG) | gFID: 2.17 -> 7.68 (LightningDiT vs DCS-LDM) | Hurt |
| DCS-LDM vs VFRTok on UCF101 video generation | DCS-LDM · UCF101 24fps 256x256 label-to-video (w/ CFG) | gFVD: 71.34 -> 86.87 | Hurt |
Closest papers by meaning
- Nested Diffusion Models Using Hierarchical Latent Priors (CVPR 2025 · similarity 0.87)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.87)
- Improving Progressive Generation with Decomposable Flow Matching (NeurIPS 2025 · similarity 0.87)
- Improving the Diffusability of Autoencoders (ICML 2025 · similarity 0.87)
- Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition (ICLR 2024 · similarity 0.87)
- TTS-VAR: A Test-Time Scaling Framework for Visual Auto-Regressive Generation (NeurIPS 2025 · 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.
