Research map / Image generation / Autoregressive and latent generation
DynaST: Dynamic Sparse Transformer for Exemplar-Guided Image Generation
Songhua Liu, Jingwen Ye, Sucheng Ren, Xinchao Wang
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 |
|---|---|---|---|
| add dynamic attention pruning (ablation) | DynaST · DeepFashion (pose-guided generation) | FID: 9.32 -> 8.36 | Improved |
| add dynamic attention pruning (ablation) | DynaST · DeepFashion (pose-guided generation) | SWD: 21.8 -> 11.8 | Improved |
| add dynamic attention pruning (ablation) | DynaST · CelebA-HQ (edge-guided face synthesis) | FID: 15.3 -> 12.0 | Improved |
| add inner-scale sparse attention blocks (ablation) | DynaST · DeepFashion (pose-guided generation) | FID: 8.88 -> 8.36 | Improved |
| add inner-scale sparse attention blocks (ablation) | DynaST · CelebA-HQ (edge-guided face synthesis) | FID: 14.7 -> 12.0 | Improved |
| raise matching resolution from 32 to full 256 (ablation) | DynaST · DeepFashion (pose-guided generation) | FID: 8.55 -> 8.36 | Improved |
| raise matching resolution from 64 to full 256 (ablation) | DynaST · DeepFashion warped vs ground truth | L1: 0.063 -> 0.054 | Improved |
| raise matching resolution from 128 to full 256 (ablation) | DynaST · DeepFashion (pose-guided generation) | FID: 8.41 -> 8.36 | Improved |
| raise matching resolution from 32 to full 256 (ablation) | DynaST · DeepFashion warped vs ground truth | Time (x1e-2 s): 5.53 -> 9.63 | Hurt |
| use dynamic sparse attention instead of CoCosNet-v2 recurrent GRU matching | exemplar-guided generation · DeepFashion warped vs ground truth | Time (x1e-2 s): 21.7 -> 9.63 | Improved |
| use DynaST instead of CoCosNet-v2 | exemplar-guided generation · DeepFashion (pose-guided generation) | FID: 13.0 -> 8.36 | Improved |
Later papers that compared with this method
2 later accepted papers report DynaST as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- EBDM: Exemplar-guided Image Translation with Brownian-bridge Diffusion Models (ECCV 2024 · better on 0 of 2 shared results)
- Masked and Adaptive Transformer for Exemplar Based Image Translation (CVPR 2023 · better on 2 of 3 shared results)
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.
- CoCosNet-v2 — CoCosNet v2: Full-Resolution Correspondence Learning for Image Translation (CVPR 2021) · better on 8 of 8 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
- LaTtE-Flow: Layerwise Timestep-Expert Flow-based Transformer (ICML 2026 · similarity 0.86)
- Unleashing Diffusion Transformers for Visual Correspondence by Modulating Massive Activations (NeurIPS 2025 · similarity 0.85)
- SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · similarity 0.85)
- FiT: Flexible Vision Transformer for Diffusion Model (ICML 2024 · similarity 0.85)
- SparseFormer: Sparse Visual Recognition via Limited Latent Tokens (ICLR 2024 · similarity 0.85)
- Cross-view Masked Diffusion Transformers for Person Image Synthesis (ICML 2024 · 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.
