ArcFlow僅5%參數加速FLUX/Qwen推理40倍

Post LinkedIn

🧠閱讀原文： 机器之心

#few-step-generation#lora-accelerationarcflowarcflowfluxqwenmicrosoft-asia-research

💡40x faster diffusion inference w/ 5% params—game-changer for real-time image gen (code out now)

⚡ 30 秒速覽

有什麼變化

透過沿教師模型曲線軌跡「漂移」實現少步生成

為什麼重要

大幅降低擴散模型推理延遲，實現生產環境即時AI影像生成，同時最小化運算成本與參數開銷。

下一步行動

Clone the ArcFlow GitHub repo and benchmark 40x speedup on your FLUX or Qwen diffusion models.

誰應關注：Researchers & Academics

關鍵要點

•透過沿教師模型曲線軌跡「漂移」實現少步生成
•經LoRA僅用5%參數，訓練速度比全參數微調快4倍
•在FLUX與Qwen-Image上實現40倍推理加速且無品質損失
•解決先前如Progressive Distillation的幾何失配問題
•開源程式碼：https://github.com/pnotp/ArcFlow

🧠 深度解析

背景與延伸：來自公開資料，非原文內容。引用 9 個來源。

🔑 增強重點摘要

•ArcFlow achieves 2-step (2 NFEs) text-to-image generation, reducing inference from multi-step teacher models to seconds while matching quality on benchmarks[1][2][3].
•Authors are from Fudan University (Zihan Yang, Shuyuan Tu, Licheng Zhang, Yu-Gang Jiang, Zuxuan Wu) and Microsoft Research Asia (Qi Dai), with paper published on arXiv as 2602.09014[3].
•Code, technical report, and basic model checkpoints released on GitHub on 2026-02-09, with upcoming ArcFlow-Qwen-20B and ArcFlow-FLUX-12B models planned[7].

🛠️ 技術深入

•ArcFlow parameterizes the velocity field as a mixture of continuous momentum processes to capture velocity evolution and form continuous non-linear trajectories within each denoising step[2][3].
•Uses analytical integration of non-linear trajectories to avoid numerical discretization errors, enabling high-precision approximation of teacher model paths[2][3].
•Implements trajectory distillation with lightweight LoRA adapters on teachers like Qwen-Image-20B (20B params) and FLUX.1-dev (12B params)[2][3][7].

🔮 前景展望基於引用來源的 AI 分析

ArcFlow LoRA adapters will integrate into existing diffusion pipelines without full retraining

LoRA enables lightweight adaptation to pre-trained models, allowing developers to add 2-step speedups to infrastructures like FLUX without overhauling systems[1][2].

2-NFE diffusion distillation will become standard for real-time image generation apps

40x speedup with preserved quality on large-scale models demonstrates viability for edge and production deployment beyond research[2][3].

⏳ 時間線

2026-02

ArcFlow paper published on arXiv (2602.09014)

2026-02-09

Code, technical report, and basic checkpoints released on GitHub

📎 來源 (9)

Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.

🧠閱讀原文： 机器之心

📰

AI 週報

閱讀本週精選 AI 大事摘要 →

👉相關動態

同主題

探索 #few-step-generation

同產品

更多關於 arcflow

同來源