Research map / Image generation / Fast diffusion sampling
WorldCache: Accelerating World Models for Free via Heterogeneous Token Caching
Weilun Feng, Guoxin Fan, Haotong Qin, Mingqiang Wu, Yuqi Li, Xiangqi Li, Zhulin An, Libo Huang, Dingrui Wang, Longlong Liao, Michele Magno, Yongjun Xu and 1 more
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 uniform caching with heterogeneous token caching (WorldCache) | diffusion world models · Voyager-13B image-to-world | PSNR vs no-cache: 21.76 (EasyCache) -> 23.49 | Improved |
| apply WorldCache to Voyager-13B | HunyuanVoyager-13B · WorldScore | WorldScore: 46.40 -> 45.43 | No clear effect |
| apply WorldCache to Voyager-13B | HunyuanVoyager-13B · end-to-end latency | speedup: 1054s -> 289s, 3.65x | Improved |
| use layer-wise caching baselines (ToCa, DuCa, TaylorSeer, HiCache) | Voyager-13B · memory | memory overhead: > 100GB | Hurt |
| replace uniform reuse, linear or damped prediction by curvature-based grouping (CHTP) (ablation) | WorldCache · Voyager/Aether | PSNR, SSIM, LPIPS | Improved |
| use random token grouping instead of curvature grouping (ablation) | WorldCache predictor · token prediction ablation | fidelity | Hurt |
| vary grouping percentiles over nine {ps,pc} pairs (ablation) | WorldCache · Voyager | PSNR: 22.77 to 23.52 | No clear effect |
| use curvature-only triggering for skipping (ablation) | adaptive skipping · skipping strategy ablation | fidelity | Hurt |
| use CAS instead of global difference or norm guided or fixed-interval skipping (ablation) | WorldCache · skipping strategy ablation | PSNR, SSIM, LPIPS | Improved |
| increase CAS threshold (ablation) | CAS on Aether · adaptive skipping threshold | fidelity vs latency | Mixed |
| use CAS instead of TeaCache skipping (ablation) | WorldCache · Aether | quality and speed | 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.
- HERO — HERO: Human-Feedback Efficient Reinforcement Learning for Online Diffusion Model Finetuning (ICLR 2025) · 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
- CacheQuant: Comprehensively Accelerated Diffusion Models (CVPR 2025 · similarity 0.87)
- DeepCache: Accelerating Diffusion Models for Free (CVPR 2024 · similarity 0.87)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.86)
- ManifoldCache: Training-Free Diffusion Acceleration via Constraint Manifold Caching (NeurIPS 2026 · similarity 0.86)
- Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (NeurIPS 2026 · similarity 0.86)
- Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching (NeurIPS 2024 · similarity 0.86)
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.
