Research map / Image generation / Fast diffusion sampling
Rethinking Token Reduction for Diffusion Models via Output-Similarity-Awareness
Hangyeol Lee, Hyojeong Lee, Joo-Young Kim
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 output-similarity-aware matching (DiTo) instead of input-similarity token reduction (best baseline ToMA) | FLUX.1-dev, ratio 0.50 · ImageNet-1k class prompts | PSNR vs vanilla: 20.49 -> 22.04 dB | Improved |
| DiTo vs baselines on FID at 0.50 reduction ratio | FLUX.1-dev · ImageNet-1k class prompts | FID: 43.71-90.00 (baselines) -> 33.47 | Improved |
| DiTo vs baselines on SD3 | SD3 · ImageNet-1k class prompts | FID, PSNR: PSNR 21.23 (SiTo) -> 23.16 at 0.50 | Improved |
| DiTo at 0.25 ratio on FID relative to the vanilla model | FLUX.1-dev · ImageNet-1k class prompts | FID: 31.62 -> 31.95 | No clear effect |
| DiTo reduces latency by reusing matching results | FLUX.1-dev, ratio 0.50 · ImageNet-1k class prompts | latency: 26.64 -> 21.69 s (-18.6%) | Improved |
| Reuse matching results across Reduction steps without frequency penalty (ablation) | DiTo · qualitative | blocking artifacts | Hurt |
| Add output-based matching instead of input-based matching (ablation) | token reduction · 500 random samples | recovery error | 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.
- ToMA — ToMA: Token Merge with Attention for Diffusion Models (ICML 2025) · 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
- ToMA: Token Merge with Attention for Diffusion Models (ICML 2025 · similarity 0.91)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.90)
- SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · similarity 0.90)
- Attend to Not Attended: Structure-then-Detail Token Merging for Post-training DiT Acceleration (CVPR 2025 · similarity 0.89)
- Which Tokens to Merge? Diffusion Dynamics for Efficient Image Generation (NeurIPS 2026 · similarity 0.88)
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 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.
