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

cs.CVAug 18, 2026

LinCa: Accelerating Diffusion Models via Learnable Decomposed Feature Caching

Authors: Jinshan Liu, Haoran Qin, Xiaobing Tu, Jiacheng Liu, Jiahui Hu, Zhengan Yan, Yukun Xie, Kerui Shen, +4 more

Organizations: Shanghai Jiao Tong University, China · Xi’an Jiaotong University, China · Shandong University, China · Terminal Intelligent Computing Division, Alibaba Cloud, China · South China University of Technology, China · Jilin University, China

Abstract

Diffusion models have achieved remarkable success in image and video generation, yet the high computational cost of iterative sampling remains a critical bottleneck for practical deployment. Feature caching has emerged as a promising acceleration paradigm by reusing or predicting intermediate features across timesteps. However, existing training-free methods apply uniform prediction strategies that cannot adapt to the heterogeneous feature dynamics, causing significant quality degradation under high acceleration ratios. We propose LinCa, a feature caching framework based on learnable invertible networks. LinCa decomposes cached features into sub-components with distinct continuity properties via a lightweight invertible network and applies differentiated prediction orders matched to each component. The strict invertibility guarantees lossless reconstruction back to the original feature space, forming a unified Decompose-Predict-Reconstruct pipeline. By training separate predictors for different models and timestep segments, LinCa adapts to heterogeneous feature dynamics. Experiments on FLUX, Qwen-Image, and HunyuanVideo demonstrate that LinCa, with less than 0.2% additional parameters, significantly outperforms existing methods and maintains near-lossless quality at 5-7x speedup. Code: https://github.com/QHR69/LinCa

Figures & tables

Figure 1: Images sampled by Qwen-Image with LinCa at 6.95 $\times$ acceleration.

Figure 2: Analysis of dynamics mismatch. (a): Denoise trajectories visualized via Principal Component Analysis (PCA) within FLUX.1-dev and Qwen-Image. Significant continuity differences exist across different models and denoising stages. (b)-(c): Denoise trajectories of specific dimensions visualized via PCA within FLUX.1-dev and Qwen-Image. Even within the same model and stage, significant continuity differences still exist across different feature dimensions. (b) shows dimensions with low continuity and abrupt mutations, while (c) shows dimensions with high continuity and stability.

Figure 3: Overview of the LinCa framework. (a) Feature Caching via LinCa: Feature caching caches features during computation steps and skips computation by utilizing historical cached features during prediction steps, where LinCa provides an end-to-end “Decompose-Predict-Reconstruct” pipeline. (b) LinCa Pipeline: Cached features are decomposed into sub-components with distinct continuity, predicted by differentiated order polynomial extrapolation, and losslessly reconstructed back to the original feature space. (c) Invertible Projection: Each invertible block consists of an invertible

1\times 1

convolution and an additive coupling layer with lightweight MLPs

F

and

G

. The strict invertibility of each sub-layer guarantees lossless reconstruction.

Method	Acceleration				ImageReward $\uparrow$	CLIP Score $\uparrow$
Method	Latency(s) $\downarrow$	Speed $\uparrow$	FLOPs(T) $\downarrow$	Speed $\uparrow$	ImageReward $\uparrow$	CLIP Score $\uparrow$
Original: 50 steps	23.10	1.00 $\times$	3719.50	1.00 $\times$	0.9930 ( $+$ 0.0%)	32.61 ( $+$ 0.0%)
$60\%$ steps	14.87	1.55 $\times$	2231.70	1.67 $\times$	0.9693 ( $-$ 2.4%)	32.50 ( $-$ 0.3%)
$\Delta$ $-$ DiT ( $\mathcal{N}=2$ ) [ 7 ]	15.96	1.45 $\times$	2480.01	1.50 $\times$	0.9471 ( $-$ 4.6%)	32.46 ( $-$ 0.5%)
$\Delta$ $-$ DiT ( $\mathcal{N}=3$ ) [ 7 ]	11.63	1.98 $\times$	1686.76	2.21 $\times$	0.8750 ( $-$ 11.9%)	32.29 ( $-$ 1.0%)
FORA $(\mathcal{N}=3)$ [ 41 ]	9.08	2.54 $\times$	1320.07	2.82 $\times$	0.9802 ( $-$ 1.3%)	32.45 ( $-$ 0.5%)

Table 1: Quantitative comparison of text-to-image generation for FLUX.1-dev. Best results are highlighted in bold , and second-best are underlined .

Figure 4: On FLUX.1-dev, LinCa delivers higher speedup with better text consistency, fine-grained details, and spatial relationships.

Method	Acceleration				Quality Metrics		Perceptual Metrics
Method	Latency(s) $\downarrow$	Speed $\uparrow$	FLOPs(T) $\downarrow$	Speed $\uparrow$	ImageReward $\uparrow$	CLIP $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Original: 50 steps	126.60	1.00 $\times$	12917.56	1.00 $\times$	1.2532 ( $+$ 0.0%)	35.52	$\infty$	1.00	0.00
$50\%$ steps	63.62	1.99 $\times$	6458.78	2.00 $\times$	1.2023 ( $-$ 4.1%)	35.24	30.55	0.75	0.27
$20\%$ steps	25.89	4.89 $\times$	2583.51	5.00 $\times$	0.9223 ( $-$ 26.4%)	34.94	28.60	0.60	0.52
TaylorSeer ( $\mathcal{N}$ =3) [ 25 ]	62.36	2.03 $\times$	4646.60	2.78 $\times$	1.0674 ( $-$ 14.8%)	34.78	28.14	0.53	0.65
HyCa ( $\mathcal{N}$ =3) [ 55 ]	59.72	2.12 $\times$	4646.60	2.78 $\times$	1.2316 ( $-$ 1.7%)	35.01	30.21	0.80	0.25

Table 2: Quantitative comparison of text-to-image generation for Qwen-Image. Best results are highlighted in bold , and second-best are underlined .

Figure 5: On Qwen-Image, LinCa retains superior image quality and details as the acceleration ratio increases, while Taylorseer degrades with blurring and detail loss.

Method	Efficient	Acceleration				VBench $\uparrow$
Method	Attention	Latency(s) $\downarrow$	Speed $\uparrow$	FLOPs(T) $\downarrow$	Speed $\uparrow$	Score(%)
Original: 50 steps	✔	185.00	1.00 $\times$	29773.0	1.00 $\times$	80.66 ( $+$ 0.0%)
$22\%$ steps	✔	40.66	4.55 $\times$	6550.1	4.55 $\times$	78.74 ( $-$ 2.4%)
FORA $(\mathcal{N}=5)$ [ 41 ]	✔	43.84	4.22 $\times$	5960.4	5.00 $\times$	78.83 ( $-$ 2.3%)
ToCa $(\mathcal{N}=5,R=90\%)$ [ 60 ]	✘	49.07	3.77 $\times$	7006.2	4.25 $\times$	78.86 ( $-$ 2.2%)
DuCa $(\mathcal{N}=5,R=90\%)$ [ 61 ]	✔	40.39	4.58 $\times$	6483.2	4.48 $\times$	78.72 ( $-$ 2.4%)

Table 3: Quantitative comparison of text-to-video generation for HunyuanVideo. Best results are highlighted in bold , and second-best are underlined .

Figure 6: On HunyuanVideo, LinCa maintains high-quality generation under higher acceleration ratio, while other methods suffer from lossing background scene, mistaken spatial relationships, and missing motional details.

Method	Acceleration				GEdit-CN (Full)			GEdit-EN (Full)
Method	Latency(s) $\downarrow$	Speed $\uparrow$	FLOPs(T) $\downarrow$	Speed $\uparrow$	SC $\uparrow$	PQ $\uparrow$	OS $\uparrow$	SC $\uparrow$	PQ $\uparrow$	OS $\uparrow$
Original: 50 steps	284.51	1.00 $\times$	28190.88	1.00 $\times$	7.68	7.51	7.41	7.82	7.54	7.54
50% steps	143.29	1.99 $\times$	14095.44	2.00 $\times$	7.70	7.53	7.44	7.77	7.52	7.47
20% steps	58.45	4.87 $\times$	5638.18	5.00 $\times$	7.65	7.42	7.35	7.73	7.46	7.44
FORA ( $\mathcal{N}$ =5) [ 41 ]	63.15	4.51 $\times$	5643.13	5.00 $\times$	7.60	7.31	7.25	7.62	7.34	7.28
DuCa ( $\mathcal{N}$ =7, $\mathcal{R}$ =95%) [ 61 ]	69.54	4.09 $\times$	5699.89	4.95 $\times$	7.73	7.44	7.44	7.80	7.40	7.45

Table 4: Quantitative comparison of image editing for Qwen-Image-Edit. Best results are highlighted in bold , and second-best results are underlined .

Figure 7: On Qwen-Image-Edit, LinCa delivers better prompt comprehension and generates high-fidelity images while maintaining consistency in non-edited regions.

Method	Acceleration				ImageReward $\uparrow$	CLIP Score $\uparrow$
Method	Latency(s) $\downarrow$	Speed $\uparrow$	FLOPs(T) $\downarrow$	Speed $\uparrow$	ImageReward $\uparrow$
FLUX.1-lite-8B: 28 steps	8.21	1.00 $\times$	1291.49	1.00 $\times$	0.8936 (+0.0%)	32.12 (+0.0%)
LinCa ( $\mathcal{N}$ =3): 28 steps	3.34	2.46 $\times$	556.21	2.32 $\times$	0.9070 (+1.5%)	32.36 (+0.7%)
FLUX.1-schnell: 4 steps	2.48	1.00 $\times$	283.56	1.00 $\times$	0.9692 (+0.0%)	32.54 (+0.0%)
LinCa ( $\mathcal{N}$ =3): 4 steps	1.38	1.80 $\times$	142.16	1.99 $\times$	0.9843 (+1.6%)	32.67 (+0.4%)
FLUX.1-dev-int8: 50 steps	12.55	1.00 $\times$	1888.07	1.00 $\times$	0.9744 (+0.0%)	32.55 (+0.0%)

Table 5: Quantitative comparison of text-to-image generation for FLUX.1-lite-8B, FLUX.1-schnell and FLUX.1-dev-int8.

Figure 8: Ablation results of LinCa. (a)-(b) : Comparison of decomposition network architectures on FLUX.1-dev and Qwen-Image. The learnable invertible network consistently outperforms both MLP and untrained alternatives. (c) Comparison of prediction order strategies. Differentiated multi-order prediction maintains superior quality over single-order prediction.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9: Prediction error between predicted and ground-truth features. LinCa obtains lower MSE by decomposing features into order-specific subspaces and applying matched prediction orders.

Figure 10: Hyperparameter sensitivity analyses of LinCa. (a) : Impact of timestep segment number

S

S=3

yields consistently strong performance, with further increases bringing only marginal gains. (b) : Impact of loss coefficient

\lambda

\lambda=1

achieves the best performance across most intervals. (c)-(e) : Architecture of the decomposition network.

L=2

invertible blocks with hidden dimension

h=128

achieves the best performance across all intervals.

Method	Predictor	Speed $\uparrow$	ImageReward $\uparrow$	CLIP Score $\uparrow$
TaylorSeer	Taylor	4.16 $\times$	0.9793	32.63
LinCa	Hermite	5.51 $\times$	1.0162	32.72
LinCa	Taylor	5.51 $\times$	1.0139	32.58
LinCa	Chebyshev	5.51 $\times$	1.0023	32.60
LinCa	Lagrange	5.51 $\times$	1.0144	32.60
LinCa	Laguerre	5.51 $\times$	0.9264	32.50

Appendix

Table 6: Comparison of high-order predictors under similar acceleration ratios.

Method	Steps	Speed $\uparrow$	ImageReward $\uparrow$
FLUX.1-dev (Original)	50	$1.00\times$	0.99
Comparison with distilled models
FLUX.1-lite-8B	28	$1.79\times$	0.89
LinCa $(N=4)$	14	$\textbf{3.57}\times$	1.02
Comparison with training-based methods
LESA $(N=5)$	12	$4.17\times$	1.00

Appendix

Table 7: More results on FLUX.1-dev.

Method	Steps	Speed $\uparrow$	ImageReward $\uparrow$
Qwen-Image (Original)	50	$1.00\times$	1.25
Comparison with distilled models
Qwen-Image-Distill-Full	15	$3.33\times$	1.02
Qwen-Image-Distill-LoRA	15	$3.33\times$	0.95
LinCa $(N=4)$	14	$\textbf{3.57}\times$	1.19
Comparison with training-based methods

Appendix

Table 8: More results on Qwen-Image.

Method	Steps	Speed $\uparrow$	ImageReward $\uparrow$	GenEval $\uparrow$
Qwen-Image-Lightning	8	$1.00\times$	1.28	0.84
LinCa $(N=3)$	4	$2.00\times$	1.26	0.85

Appendix

Table 9: Compatibility with Qwen-Image-Lightning.

Method	Steps	Speed $\uparrow$	VBench Score(%) $\uparrow$
Wan2.1-1.3B (Original)	50	$1.00\times$	83.72
TaylorSeer $(N=5)$	12	$4.17\times$	81.07
LinCa $(N=6)$	10	$\textbf{5.00}\times$	82.56

Appendix

Table 10: Text-to-video generation result on Wan2.1.

Explore similar work

LearniBridge: Learnable Calibration of Feature Caching for Diffusion Models Acceleration
Jun 25, 2026Xuyue Huang, Zhe Chen, Wang Shen +1Video Diffusion Models Diffusion Transformer
Shenzhen Ubiquitous Data Enabling Key Laboratory, Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Central Media Technology Institute, Huawei Technologies Co., Ltd., Shenzhen, China
Beyond Fixed Formulas: Data-Driven Linear Predictor for Efficient Diffusion Models
Apr 29, 2026Zhirong Shen, Rui Huang, Jiacheng Liu +6Diffusion Transformer Diffusion Model Caching
Shanghai Jiao Tong University · University of Electronic Science and Technology of China · Shandong University +2
LESA: Learnable Stage-Aware Predictors for Diffusion Model Acceleration
Feb 24, 2026Peiliang Cai, Jiacheng Liu, Haowen Xu +3Diffusion Model Inference Acceleration Diffusion Models
Shanghai Jiao Tong University

Jun 25, 2026J/K move · Enter open · S save

LearniBridge: Learnable Calibration of Feature Caching for Diffusion Models Acceleration

Xuyue Huang, Zhe Chen, Wang Shen, Xiao-Ping Zhang

Shenzhen Ubiquitous Data Enabling Key Laboratory, Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Central Media Technology Institute, Huawei Technologies Co., Ltd., Shenzhen, China

Video Diffusion Models Diffusion Transformer Diffusion-Based Image Generation Diffusion Model Caching Diffusion Model Inference Acceleration

Open paper AskPDF arXiv

Diffusion Transformers (DiTs) have driven substantial progress in image and video generation but suffer from prohibitive computational costs. Feature caching accelerates inference by reusing intermediate representations. Existing methods rely on historical features for implementation simplicity, yet suffer from severe error accumulation at high acceleration ratios. To address this limitation, we investigate the nature of the requisite feature correction. We demonstrate that the optimal calibration update is characterized by a shared low-rank subspace across diverse prompts. Guided by this structural insight, we propose LearniBridge, a learnable calibration mechanism for feature caching that bridges multiple timesteps through lightweight LoRA updates. This mechanism enables effective calibration requiring only 3-5 training samples. Extensive experiments on image and video generation show that LearniBridge achieves up to

5.87\times

5.75\times

, and

4.10\times

acceleration on FLUX, HunyuanVideo, and WAN2.1, respectively. On WAN2.1, it improves VBench by 1.28% over the previous SOTA at

4.10\times

acceleration. Our code is available at https://github.com/Iiiiiiirene/LearniBridge.

Abstract

Explore similar work

LearniBridge: Learnable Calibration of Feature Caching for Diffusion Models Acceleration

Beyond Fixed Formulas: Data-Driven Linear Predictor for Efficient Diffusion Models

LESA: Learnable Stage-Aware Predictors for Diffusion Model Acceleration