Research map / Image generation / Fast diffusion sampling
LayoutDiffusion: Improving Graphic Layout Generation by Discrete Diffusion Probabilistic Models
Junyi Zhang, Jiaqi Guo, Shizhao Sun, Jian-Guang Lou, Dongmei Zhang
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 block-wise transition matrix with piece-wise linear schedule instead of generic D3PM absorbing/uniform (ablation) | LayoutDiffusion · RICO Un-Gen | FID: 4.985 -> 2.490 | Improved |
| add discretized Gaussian transition for coordinate tokens instead of uniform (ablation) | LayoutDiffusion · RICO Un-Gen | FID: 3.174 -> 2.490 | Improved |
| add absorbing transition for type tokens instead of uniform (ablation) | LayoutDiffusion · RICO Un-Gen | mIoU: 0.593 -> 0.620 | Improved |
| use piece-wise linear type schedule (type noise only late) instead of linear (ablation) | LayoutDiffusion · RICO Un-Gen | mIoU: 0.580 -> 0.620 | Improved |
| use beta_t=g/(T-t+eps)^h coordinate schedule instead of linear (ablation) | LayoutDiffusion · RICO Un-Gen | Align.: 0.202 -> 0.069 | Improved |
| reduce diffusion steps from 1000+ to 200 (ablation) | LayoutDiffusion · RICO Un-Gen | FID: 2.295 -> 2.490 | No clear effect |
| reduce diffusion steps to 100 (ablation) | LayoutDiffusion · RICO Un-Gen | FID: 2.490 -> 3.875 | Hurt |
| replace LayoutFormer++ with LayoutDiffusion without retraining for refinement | Refinement · RICO | FID: 3.666 -> 0.549 | Improved |
| replace LayoutFormer++ with LayoutDiffusion for unconditional generation | Un-Gen · PubLayNet | FID: 47.082 -> 8.625 | Improved |
| use LayoutDiffusion in place of Diffusion-LM | Un-Gen · RICO / PubLayNet | mIoU: 0.662 -> 0.620 (RICO) | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LayoutFormer++ — LayoutFormer++: Conditional Graphic Layout Generation via Constraint Serialization and Decoding Space Restriction (CVPR 2023) · better on 5 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Unifying Layout Generation With a Decoupled Diffusion Model (CVPR 2023 · similarity 0.93)
- LayoutDM: Discrete Diffusion Model for Controllable Layout Generation (CVPR 2023 · similarity 0.92)
- Towards Aligned Layout Generation via Diffusion Model with Aesthetic Constraints (ICLR 2024 · similarity 0.91)
- Visual Layout Composer: Image-Vector Dual Diffusion Model for Design Layout Generation (CVPR 2024 · similarity 0.88)
- NoiseCollage: A Layout-Aware Text-to-Image Diffusion Model Based on Noise Cropping and Merging (CVPR 2024 · similarity 0.87)
- DesignDiffusion: High-Quality Text-to-Design Image Generation with Diffusion Models (CVPR 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.
