LinCa predictors (ECCV 2026)

Learnable decomposed feature-caching predictors from LinCa: Accelerating Diffusion Models via Learnable Decomposed Feature Caching, accepted to ECCV 2026.

Code: github.com/QHR69/LinCa

Model repo: huggingface.co/QHRQQQ/LinCa

LinCa decomposes a cached DiT feature into sub-components with different continuity, predicts each with a matching order, then reconstructs it through a strictly invertible mapping. Under 0.2% extra parameters, up to 7.08× fewer FLOPs with near-lossless quality.

All speedups below are FLOPs speedups relative to the original 50-step model.

Model	Setting	FLOPs speedup	Quality
FLUX.1-dev	N=6	4.52×	ImageReward 1.0228 (orig. 0.9930)
Qwen-Image	N=6	4.90×	ImageReward 1.2163 (orig. 1.2532)
Qwen-Image-Edit	N=7	5.52×	GEdit-EN OS 7.56 (orig. 7.54)
HunyuanVideo	N=6	5.50×	VBench 80.16 (orig. 80.66)

Files

These files are predictor weights only. Download FLUX / Qwen backbones from their official sources.

File	Backbone
qwen-image_checkpoint.pt	Qwen-Image
qwen-image-edit_checkpoint.pt	Qwen-Image-Edit
flux/best_predictor_stage{0,1,2}.pt	FLUX.1-dev (3 stages, splits 17/17/16)

Use the script defaults in the GitHub repo when loading these checkpoints.

We do not release HunyuanVideo or Wan2.1 predictors.

Citation

@inproceedings{liu2026linca,
  title={LinCa: Accelerating Diffusion Models via Learnable Decomposed Feature Caching},
  author={Liu, Jinshan and Qin, Haoran and Tu, Xiaobing and Liu, Jiacheng and Hu, Jiahui and Yan, Zhengan and Xie, Yukun and Shen, Kerui and Ren, Jinkui and Lin, Yuqi and Zhang, Xiantao and Zhang, Linfeng},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}

Contact: qinhaoran68@gmail.com