Research map / Image generation / Fast diffusion sampling
LinCa: Accelerating Diffusion Models via Learnable Decomposed Feature Caching
Jinshan Liu, Haoran Qin, Xiaobing Tu, Jiacheng Liu, Jiahui Hu, zhengan yan, Yukun Xie, Kerui Shen, Jinkui Ren, Yuqi Lin, Xiantao Zhang, Linfeng Zhang
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 LinCa learned decomposition instead of TaylorSeer at similar caching interval | FLUX.1-dev feature caching · DrawBench (200 prompts) | ImageReward: 1.0018 -> 1.0175 | Improved |
| Use LinCa instead of HyCa at higher acceleration | FLUX.1-dev feature caching · DrawBench (200 prompts) | ImageReward: 1.0096 -> 1.0228 | Improved |
| Use LinCa instead of TaylorSeer on Qwen-Image at N=3 | Qwen-Image feature caching · DrawBench (200 prompts) | ImageReward: 1.0674 -> 1.2329 | Improved |
| Use LinCa instead of TaylorSeer at high speedup on Qwen-Image | Qwen-Image feature caching · DrawBench (200 prompts) | ImageReward: 0.7318 -> 1.0524 | Improved |
| Use LinCa instead of TaylorSeer on HunyuanVideo | HunyuanVideo feature caching · VBench | VBench score (%): 79.93 -> 80.16 | Improved |
| Use LinCa instead of TaylorSeer on Qwen-Image-Edit at N=7 | Qwen-Image-Edit feature caching · GEdit-CN Overall Score | Overall Score: 6.92 -> 7.45 | Improved |
| Use LinCa instead of TaylorSeer at N=9-10 on Qwen-Image-Edit | Qwen-Image-Edit feature caching · GEdit-EN Overall Score | Overall Score: 6.31 -> 7.40 | Improved |
| Stack LinCa on model-distilled FLUX.1-lite-8B | FLUX.1-lite-8B · DrawBench (200 prompts) | ImageReward: 0.8936 -> 0.9070 | Improved |
| Stack LinCa on step-distilled FLUX.1-schnell | FLUX.1-schnell · DrawBench (200 prompts) | ImageReward: 0.9692 -> 0.9843 | Improved |
| Stack LinCa on INT8-quantized FLUX.1-dev | FLUX.1-dev-int8 · DrawBench (200 prompts) | ImageReward: 0.9744 -> 1.0036 | Improved |
| Use invertible network instead of non-invertible learnable MLP for decomposition (ablation) | LinCa · FLUX.1-dev and Qwen-Image | generation quality | Improved |
| Use learnable decomposition instead of non-learnable strategies (ablation) | LinCa · FLUX.1-dev and Qwen-Image | generation quality | Improved |
Closest papers by meaning
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.90)
- HiCache: A Plug-in Scaled-Hermite Upgrade for Taylor-Style Cache-then-Forecast Diffusion Acceleration (ICLR 2026 · similarity 0.90)
- DeepCache: Accelerating Diffusion Models for Free (CVPR 2024 · similarity 0.89)
- Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model (CVPR 2025 · similarity 0.88)
- MagCache: Fast Video Generation with Magnitude-Aware Cache (NeurIPS 2025 · similarity 0.87)
- Accelerating Diffusion Transformers with Token-wise Feature Caching (ICLR 2025 · 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.
