Research map / Image generation / Autoregressive and latent generation
Spanning Tree Autoregressive Visual Generation
Sangkyu Lee, Changho Lee, Janghoon Han, Hosung Song, Tackgeun You, Hwasup Lim, Stanley Jungkyu Choi, Honglak Lee, Youngjae Yu
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 spanning tree BFS orders in training instead of random permutation throughout (ablation) | STAR-B · ImageNet-1k generation | FID: 3.57 -> 2.24 | Improved |
| use spanning tree orders at inference for a permutation-trained model (ablation) | STAR-B · ImageNet-1k generation | FID: 3.57 -> 3.58 | No clear effect |
| anneal permutation to spanning tree orders at end of training (ablation) | STAR-B · ImageNet-1k generation | FID: 3.58 -> 2.31 | Improved |
| train with spanning tree from the start rather than annealing from permutation (ablation) | STAR-B · ImageNet-1k generation | FID: 2.31 -> 2.24 | Improved |
| use spanning tree orders instead of raster-scan for training and inference (ablation) | STAR-B · ImageNet-1k inpainting | FID: 3.91 -> 2.39 | Improved |
| use spanning tree orders instead of raster-scan for training and inference (ablation) | STAR-B · ImageNet-1k generation | FID: 2.04 -> 2.24 | Hurt |
| use BFS instead of DFS for spanning tree traversal (ablation) | STAR · rejection sampling over masks 0.1-0.9 | Trials at mask ratio 0.1: 60.45 -> 5.59 | Improved |
| use BFS with farthest root instead of random root (ablation) | STAR · rejection sampling over masks 0.1-0.9 | Trials at mask ratio 0.1: 5.59 -> 4.41 | Improved |
| replace DiT inpainting by latent blending with STAR native prefix inpainting | STAR-L · ImageNet-1k inpainting (L-size) | FID: 4.58 -> 2.20 | Improved |
| use STAR instead of RandAR | STAR-L · ImageNet-1k inpainting (L-size) | FID: 2.66 -> 2.20 | Improved |
| use STAR instead of RAR at L-size | STAR-L · ImageNet-1k generation (L-size) | FID: 1.70 -> 1.98 | Hurt |
| use STAR instead of RAR at inpainting | STAR-L · ImageNet-1k inpainting (L-size) | FID: 3.56 -> 2.20 | Improved |
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.
- RandAR-XL — RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (CVPR 2025) · better on 4 of 4 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
- EditAR: Unified Conditional Generation with Autoregressive Models (CVPR 2025 · similarity 0.89)
- RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (CVPR 2025 · similarity 0.89)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.89)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.88)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.88)
- Denoising Autoregressive Transformers for Scalable Text-to-Image Generation (ICLR 2025 · 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.
