Research map / Image generation / Fast diffusion sampling
SiT: Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformers
Nanye Ma, Mark Goldstein, Michael Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, Saining Xie
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 velocity-based SiT recipe (continuous time, velocity prediction, Linear interpolant, SDE sampler) instead of DiT recipe | DiT backbone, all sizes · ImageNet 256x256 (400K steps) | FID-50K: 43.5 -> 33.0 (B), 19.5 -> 17.2 (XL) | Improved |
| Switch training from discrete to continuous time (ablation) | SiT-B · ImageNet 256x256 | FID-50K | Improved |
| Predict velocity or use weighted score loss instead of plain score loss (ablation) | SBDM-VP SiT-B · ImageNet 256x256 | FID-50K | Improved |
| Replace SBDM-VP interpolant with GVP or Linear interpolant (ablation) | SiT-B velocity model · ImageNet 256x256 | FID-50K | Improved |
| Sample with SDE instead of ODE (ablation) | SiT-B · ImageNet 256x256, 400K | FID-50K: 34.84 -> 33.02 | Improved |
| Sample with SDE instead of ODE at low NFE (ablation) | SiT · ImageNet 256x256 | FID vs NFE | Mixed |
| Use KL-optimal diffusion coefficient w^KL for GVP velocity model (ablation) | SiT-B GVP velocity · ImageNet 256x256 | FID-50K: 33.4 (sigma_t) -> 32.9 (w^KL) | Improved |
| Use cost-regularized w^{KL,eta} for Linear interpolant (ablation) | SiT-B Linear velocity · ImageNet 256x256 | FID-50K: 33.6 (w^KL) -> 33.0 | Improved |
| Use regularized w^{KL,eta} for SBDM-VP (ablation) | SiT-B SBDM-VP velocity · ImageNet 256x256 | FID-50K: 37.8 (w^KL) -> 41.1 | Hurt |
| Learn score instead of velocity (ablation) | SiT-B Linear interpolant · ImageNet 256x256 | FID-50K: 33.6 -> 41.0 (w^KL) | Hurt |
| Apply classifier-free guidance (cfg=1.5) to SiT-XL at 7M steps | SiT-XL · ImageNet 256x256 | FID-50K: 8.3 -> 2.06 | Improved |
| Increase SiT model size (S to XL) (ablation) | SiT · ImageNet 256x256, 400K | FID-50K: 57.6 -> 17.2 | Improved |
Closest papers by meaning
- Stochastic Interpolants with Data-Dependent Couplings (ICML 2024 · similarity 0.90)
- Stochastic Interpolants via Conditional Dependent Coupling (NeurIPS 2026 · similarity 0.90)
- Multitask Learning with Stochastic Interpolants (NeurIPS 2025 · similarity 0.90)
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 2025 · similarity 0.89)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.89)
- A solvable model of learning generative diffusion: theory and insights (NeurIPS 2025 · similarity 0.89)
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.
