Research map / Image generation / Fast diffusion sampling
DiffusionBlocks: Block-wise Neural Network Training via Diffusion Interpretation
Makoto Shing, Masanori Koyama, Takuya Akiba
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 equi-probability instead of uniform partitioning, layers [4,4,4] (ablation) | DiffusionBlocks · CIFAR-10 | FID: 43.53 -> 38.03 | Improved |
| use equi-probability instead of uniform partitioning, layers [3,6,3] (ablation) | DiffusionBlocks · CIFAR-10 | FID: 43.59 -> 41.64 | Improved |
| use equi-probability instead of uniform partitioning, layers [6,4,2] (ablation) | DiffusionBlocks · CIFAR-10 | FID: 47.49 -> 45.42 | Improved |
| use equi-probability instead of uniform partitioning, layers [2,4,6] (ablation) | DiffusionBlocks · CIFAR-10 | FID: 42.37 -> 40.40 | Improved |
| use B=2 blocks instead of end-to-end (B=1) (ablation) | DiffusionBlocks DiT-L/2 · ImageNet | FID: 12.09 -> 9.90 | Improved |
| use B=3 blocks instead of end-to-end (B=1) (ablation) | DiffusionBlocks DiT-L/2 · ImageNet | FID: 12.09 -> 11.11 | Improved |
| use B=6 blocks instead of end-to-end (B=1) (ablation) | DiffusionBlocks DiT-L/2 · ImageNet | FID: 12.09 -> 14.43 | Hurt |
| use B=6 blocks instead of B=1 (ablation) | DiffusionBlocks DiT-S/2 · CIFAR-10 | FID: 39.83 -> 53.32 | Hurt |
| use B=2 blocks instead of B=1 (ablation) | DiffusionBlocks DiT-S/2 · CIFAR-10 | FID: 39.83 -> 35.47 | Improved |
| use B=4 blocks instead of B=2 (ablation) | DiffusionBlocks AR · LM1B | MAUVE: 0.61 -> 0.67 | Improved |
| use B=6 blocks instead of B=4 (ablation) | DiffusionBlocks AR · LM1B | MAUVE: 0.67 -> 0.62 | Hurt |
| apply DiffusionBlocks instead of Forward-Forward | ViT · CIFAR-100 | accuracy: 7.85 -> 59.30 | Improved |
Closest papers by meaning
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.89)
- BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers (CVPR 2025 · similarity 0.88)
- Rethinking Cross-Layer Information Routing in Diffusion Transformer (NeurIPS 2026 · similarity 0.88)
- Neural Diffusion Models (ICML 2024 · similarity 0.88)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.88)
- MAGE: All-[MASK] Block Already Knows Where to Look in Block Diffusion LLM (NeurIPS 2026 · similarity 0.88)
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.
