Research map / Video / Video generation
Importance-Based Token Merging for Efficient Image and Video Generation
Haoyu Wu, Jingyi Xu, Hieu Le, Dimitris Samaras
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 |
|---|---|---|---|
| importance-based destination selection instead of random (ToMeSD) | Stable Diffusion 2 token merging · MS-COCO 30K, r=0.75 | FID: 20.89 -> 17.75 | Improved |
| importance-based merging at low merging ratio | Stable Diffusion 2 token merging · MS-COCO 30K, r=0.3 | FID: 12.20 -> 12.20 | No clear effect |
| importance-based merging on diffusion transformer | PixArt-alpha · MS-COCO, r=0.7 | FID: 95.46 -> 50.76 | Improved |
| importance-based merging on multi-view diffusion | Zero123++ · GSO, r=0.75 | PSNR: 13.12 -> 14.58 | Improved |
| importance-based merging on multi-view diffusion | Zero123++ · GSO, r=0.75 | SSIM: 0.789 -> 0.784 | No clear effect |
| importance-based merging on video diffusion | AnimateDiff v3 · VBench, r=0.75 | VBench total: 78.59 -> 79.12 | Improved |
| merge tokens in temporal layers of video model | AnimateDiff v3 · VBench | generated dynamics | Hurt |
| choose destination tokens by top-k importance instead of random from pool (ablation) | importance-based merging · MS-COCO SD2, r=0.3 | FID: 12.20 -> 12.56 | Hurt |
| draw independent tokens only from the important pool (ablation) | importance-based merging · MS-COCO SD2, r=0.7 | FID: 16.43 -> 16.22 | Improved |
| use CFG instead of cross-attention maps as importance signal (ablation) | importance-based merging · MS-COCO SD2, r=0.75 | FID: 18.17 -> 17.75 | Improved |
| add important token pool (p from 0 to 0.4) (ablation) | importance-based merging · MS-COCO SD2, r=0.3 | FID: 12.87 -> 12.19 | Improved |
| use importance-based instead of random merging | Stable Diffusion 2 · inference cost, r=0.7 | latency and TFLOPs: 5.7 -> 5.8 s latency; 3.36 TFLOPs | No clear effect |
Closest papers by meaning
- Which Tokens to Merge? Diffusion Dynamics for Efficient Image Generation (NeurIPS 2026 · similarity 0.92)
- Multimodal Promptable Token Merging for Diffusion Models (AAAI 2025 · similarity 0.89)
- ToMA: Token Merge with Attention for Diffusion Models (ICML 2025 · similarity 0.88)
- Attend to Not Attended: Structure-then-Detail Token Merging for Post-training DiT Acceleration (CVPR 2025 · similarity 0.88)
- VidToMe: Video Token Merging for Zero-Shot Video Editing (CVPR 2024 · similarity 0.87)
- Video Token Merging for Long Video Understanding (NeurIPS 2024 · similarity 0.87)
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.
