Research map / Image generation / Autoregressive and latent generation
SPRINT: Sparse-Dense Residual Fusion for Efficient Diffusion Transformers
Dogyun Park, Moayed Haji-Ali, Yanyu Li, Willi Menapace, Sergey Tulyakov, Hyunwoo J. Kim, Aliaksandr Siarohin, Anil Kag
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 SPRINT sparse-dense residual fusion with 75% token dropping | Improved SiT-XL/2 · ImageNet 256 (SD-VAE, 400K iters, w/o CFG) | FID: 12.8 -> 9.30 | Improved |
| add SPRINT | Improved SiT-XL/2 · ImageNet 256 (SD-VAE, 1M iters, w/ CFG) | FID: 2.36 -> 2.29 | Improved |
| add Tread or MDTv2 token dropping at 75% drop ratio | Improved SiT-XL/2 · ImageNet 256 (SD-VAE, 1M iters, w/ CFG) | FID: 2.36 -> 2.82 (Tread), 4.68 (MDTv2) | Hurt |
| add MicroDiT token dropping | Improved SiT-XL/2 · ImageNet 256 (SD-VAE, 1M iters, w/ CFG) | FID: 2.36 -> 2.53 | Hurt |
| add SPRINT to REPA | SiT-XL/2 REPA · ImageNet 256 (SD-VAE, 400K iters, w/o CFG) | FID: 10.0 -> 8.68 | Improved |
| add SPRINT to U-ViT | U-ViT-XL/2 · ImageNet 256 (SD-VAE, 400K iters, w/o CFG) | FID: 12.1 -> 9.20 | Improved |
| replace CFG with Path-Drop Guidance | SPRINT · ImageNet 256 system-level | FID: 1.96 -> 1.62 | Improved |
| add dense shallow residual path (ablation) | SPRINT (SiT-B/2) · ImageNet 256 ablation | FID: 85.1 -> 27.5 | Improved |
| add structured group-wise token sampling instead of uniform random (ablation) | SPRINT (SiT-B/2) · ImageNet 256 ablation | FID: 30.1 -> 27.5 | Improved |
| raise token drop ratio from 0 to 75% (ablation) | SPRINT (SiT-B/2) · ImageNet 256 ablation | FID: 54.1 -> 27.5 | Improved |
| raise drop ratio from 75% to 87.5% (ablation) | SPRINT (SiT-B/2) · ImageNet 256 ablation | FID: 27.5 -> 50.2 | Hurt |
| shift layers from middle block to encoder/decoder (3-6-3, 5-2-5) (ablation) | SPRINT (SiT-B/2) · ImageNet 256 ablation | FID: 27.3 -> 29.1 (3-6-3), 49.2 (5-2-5) | Hurt |
Closest papers by meaning
- SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · similarity 0.91)
- TinyFusion: Diffusion Transformers Learned Shallow (CVPR 2025 · similarity 0.90)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.90)
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 2025 · similarity 0.90)
- Stretching Each Dollar: Diffusion Training from Scratch on a Micro-Budget (CVPR 2025 · similarity 0.90)
- ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation (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.
