Research map / Image generation / Autoregressive and latent generation
When Local Rules Create Global Order: Self-Organized Representation Learning for Latent Diffusion Models
Junrong Lian, Weijian Deng, Pengxu Wei, Yaqin Chen, Qixiang Ye, Liang Lin
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 |
|---|---|---|---|
| add local attraction (noise-perturbed reconstruction) to reconstruction loss (ablation) | SORL · CelebA-HQ generation | FID: 80.02 -> 45.54 | Improved |
| add local repulsion alone to reconstruction loss (ablation) | SORL · CelebA-HQ generation | FID: 80.02 -> 98.82 | Hurt |
| add local repulsion alone to reconstruction loss (ablation) | SORL · CelebA-HQ generation | IS: 3.13 -> 4.28 | Improved |
| combine local attraction and repulsion (ablation) | SORL · CelebA-HQ generation | FID: 45.54 -> 23.55 | Improved |
| increase centroid count K from 1024 to 8192 (ablation) | SORL · CelebA-HQ generation | FID: 28.86 -> 23.55 | Improved |
| increase centroid count K from 8192 to 16384 (ablation) | SORL · CelebA-HQ generation | FID: 23.55 -> 24.67 | No clear effect |
| set pull weight lambda_pull to 1e-2 instead of 1 or 1e-4 (ablation) | SORL · CelebA-HQ generation | FID: 45.97 (1) / 35.54 (1e-4) -> 23.55 | Improved |
| use SORL latent space versus VAE for reconstruction | SORL · CelebA-HQ reconstruction | PSNR: 28.40 -> 31.51 | Improved |
| use SORL latent space versus VAE for unconditional generation | SORL · CelebA-HQ unconditional generation | FID: 13.02 -> 9.74 | Improved |
| use SORL autoencoder for latent diffusion versus SiT-XL baseline | SiT-XL/2 · ImageNet 100K steps | FID: 19.06 -> 15.12 | Improved |
| use SORL autoencoder versus iREPA on ImageNet | SiT-XL/2 · ImageNet 100K steps | Recall: 0.61 -> 0.58 | Mixed |
Closest papers by meaning
- What Matters for Diffusion-Friendly Latent Manifold? Prior-Aligned Autoencoders for Latent Diffusion (NeurIPS 2026 · similarity 0.89)
- How Diffusion Models Learn to Factorize and Compose (NeurIPS 2024 · similarity 0.88)
- Isometric Representation Learning for Disentangled Latent Space of Diffusion Models (ICML 2024 · similarity 0.87)
- CORAL: Disentangling Latent Representations in Long-Tailed Diffusion (NeurIPS 2025 · similarity 0.87)
- Compositional Discrete Latent Code for High Fidelity, Productive Diffusion Models (NeurIPS 2025 · similarity 0.87)
- Why DiT Models Underperform as Representation Learners without Long Skip Connections (NeurIPS 2026 · 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.
