Skip to content

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

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

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

Authors: Zhirong Shen, Rui Huang, Jiacheng Liu, Chang Zou, Peiliang Cai, Shikang Zheng, Zhengyi Shi, Liang Feng, Linfeng Zhang Date: 2026-04-29 Paper ID: openalex:2604.26365

Summary

This paper addresses the inference latency bottleneck in Diffusion Transformers by replacing traditional hand-crafted feature caching formulas with a data-driven approach called Learnable Linear Predictor (L2P). L2P learns per-timestep weights to reconstruct current features from historical trajectories, enabling significantly more aggressive skipping strategies than static methods. Experiments across FLUX.1-dev and Qwen-Image models demonstrate substantial latency improvements and FLOPs reduction without sacrificing visual quality. The framework is highly efficient, requiring only approximately 20 seconds of training on a single GPU.

Key Contributions

  • Introduced L2P (Learnable Linear Predictor), a training-free acceleration framework for Diffusion Transformers (DiTs) using per-timestep learnable weights.
  • Demonstrated 4.55x FLOPs reduction and 4.15x latency speedup on FLUX.1-dev while maintaining superior visual fidelity compared to fixed-formula caching methods.
  • Showcased robustness of learned predictors at aggressive acceleration factors (up to 7.18x) on Qwen-Image models where baseline methods suffer from quality degradation.

Open Questions & Future Work

Key Concepts

  • learnable-linear-predictor: A data-driven feature caching framework for diffusion models that replaces fixed forecasting formulas with learnable per-timestep weights.

Archivist Review

I approved the Learnable Linear Predictor concept as a reusable, architecture-agnostic mechanism for accelerating diffusion model inference. I also approved a refined open question regarding the transferability of adaptive acceleration strategies, as current data-driven approaches require specific calibration that limits their universal deployment. I maintained strict selectivity by rejecting redundant or overly narrow descriptors.

Approved Concepts

  • Learnable Linear Predictor (L2P): It provides a systematic, data-driven alternative to rigid, heuristic-based feature caching in diffusion models, enabling more aggressive skip-sampling without quality degradation.

Approved Open Questions

  • Adaptive Diffusion Acceleration Transferability: Solving this would enable fully ‘plug-and-play’ acceleration, eliminating the overhead of task-specific training or calibration for efficient inference.

Metadata & Links

url
https://arxiv.org/abs/2604.26365
paper_id
2604.26365
paper_source
openalex
domain
computer-vision
tags
diffusion-modelsinference-accelerationtransformerscachingefficient-sampling
architectures
datasets
concept_slugs
learnable-linear-predictor
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-02T05:50:03Z