Research map / Image generation / Fast diffusion sampling
Guiding Token-Sparse Diffusion Models
Felix Krause, Stefan Andreas Baumann, Johannes Schusterbauer, Olga Grebenkova, Ming Gui, Vincent Tao Hu, Björn Ommer
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 |
|---|---|---|---|
| replace CFG with Sparse Guidance | SiT-XL/2 + routing · ImageNet-256 (400 epochs) | FID: 2.57 -> 1.58 | Improved |
| use SG_FLOPS config instead of CFG | SiT-XL/2 + routing · ImageNet-256 (400 epochs) | GFLOPs: 228.84 -> 97.67 | Improved |
| replace CFG with SG on a masking-trained model | MaskDiT-style SiT-XL/2 · ImageNet-256 (masking, 160 epochs) | FID: 5.82 -> 5.73 | Improved |
| replace CFG with SG on a routing-trained model | TREAD-style SiT-XL/2 · ImageNet-256 (routing, 160 epochs) | FID: 2.95 -> 2.07 | Improved |
| replace CFG with SG | TREAD-style SiT-XL/2 · ImageNet-256 (routing, 160 epochs) | Inception Score: 233.3 -> 223.4 | Hurt |
| replace Sparse Guidance with CFG plus dense finetuning | sparsely trained SiT-XL/2 · ImageNet-256 | FID | Improved |
| combine SG with AutoGuidance | SiT-XL/2 + routing · ImageNet-256 (400 epochs) | FID: 2.95 -> 1.58 | Improved |
| add SG instead of ICG | SiT-XL/2 + routing · ImageNet-256 (400 epochs) | FID: 2.81 -> 1.58 | Improved |
| raise inference-time sparsity without guidance (omega = 1.0) (ablation) | token-sparse diffusion model · ImageNet-256 | image quality | Hurt |
| increase guidance scale omega together with sparsity (ablation) | SG · ImageNet-256 (FID@5k) | FID | No clear effect |
| use routing instead of masking as train-time sparsity under SG (ablation) | SG · ImageNet-256 | FID: 5.73 -> 2.07 | Improved |
| replace CFG with SG | TR-DIT-2.5B · HPSv3 | overall score: 9.21 -> 9.87 | 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.
- CogView4 — CogView: Mastering Text-to-Image Generation via Transformers (NeurIPS 2021) · better on 1 of 1 shared result
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
- Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.92)
- Plug-and-Play Diffusion Distillation (CVPR 2024 · similarity 0.91)
- REG: Rectified Gradient Guidance for Conditional Diffusion Models (ICML 2025 · similarity 0.91)
- Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models (NeurIPS 2024 · similarity 0.90)
- Feedback Guidance of Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.90)
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.
