Research map / Image generation / Autoregressive and latent generation
Scalable Adaptive Computation for Iterative Generation
Allan Jabri, David J. Fleet, Ting Chen
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 |
|---|---|---|---|
| latent self-conditioning (training rate 0.9 vs 0) (ablation) | RIN · ImageNet 64x64 and 128x128 ablation | FID/IS: clear correlation between self-conditioning rate and sample quality | Improved |
| multiple read-write routing rounds (more stacked blocks) vs single read-write (ablation) | RIN · ImageNet ablation over model size | FID/IS: saturates later with significantly better sample quality | Improved |
| sigmoid noise schedule (temperature 0.9) instead of cosine for training (ablation) | RIN · ImageNet larger resolutions | training stability | Improved |
| noise schedule choice at sampling time (sigmoid vs cosine) (ablation) | RIN · ImageNet sampling | FID | No clear effect |
| large tokenization (1x1 patches, 4096 interface tokens) vs 8x8 patches (ablation) | RIN · ImageNet ablation | FID/IS | Mixed |
| input scaling | RIN at 256x256 · ImageNet 256x256 | FID: 4.51 -> 3.42 | Improved |
| RIN architecture vs U-Net diffusion baselines with fewer FLOPs | RIN · ImageNet 64x64 to 256x256 | FID and GFLOPs: e.g. 1.23 FID at 106 GFLOPs vs ADM 210 GFLOPs | Improved |
| RIN vs CF-guidance at 128x128 without guidance | RIN · ImageNet 128x128 | FID: 2.75 vs 2.43 | Hurt |
| RIN without guidance vs ADM with guidance at 256x256 | RIN · ImageNet 256x256 | FID: RIN+inp.scale 3.42 vs ADM+guid 4.59 (better); RIN 4.51 vs ADM+guid 4.59 (roughly tied) | Mixed |
| RIN video model vs Video Diffusion with reconstruction guidance | RIN · Kinetics-600 16x64x64 | FVD: 10.8 vs 16.6 | Improved |
| sampling steps 400 vs 1000 for RIN video (ablation) | RIN · Kinetics-600 16x64x64 | FVD: 11.5 -> 10.8 | Improved |
Closest papers by meaning
- ToMA: Token Merge with Attention for Diffusion Models (ICML 2025 · similarity 0.86)
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.86)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.86)
- ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation (NeurIPS 2025 · similarity 0.85)
- Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation (NeurIPS 2025 · similarity 0.85)
- RepLDM: Reprogramming Pretrained Latent Diffusion Models for High-Quality, High-Efficiency, High-Resolution Image Generation (NeurIPS 2025 · similarity 0.85)
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.
