Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ArcFlow: Unleashing 2-Step Text-to-Image Generation via High-Precision Non-Linear Flow Distillation

Zihan Yang, Shuyuan Tu, Licheng Zhang, Qi Dai, Yu-Gang Jiang, Zuxuan Wu

TL;DR

Diffusion-based image synthesis is hindered by high inference cost due to many denoising steps. ArcFlow introduces a non-linear flow distillation framework that parameterizes the velocity field as a mixture of continuous momentum processes, enabling exact integration via a closed-form analytic solver and aligning the student trajectory with the pre-trained teacher using velocity matching. By fine-tuning only lightweight LoRA adapters and an output head, ArcFlow achieves high-quality, diverse generation with as few as 2 NFEs, reporting up to 40× speedups on large backbones like Qwen-Image-20B and FLUX. The approach yields robust improvements across multiple benchmarks, faster convergence, and stable training without adversarial objectives, highlighting the value of respecting intrinsic flow dynamics for efficient generative inference. A notable limitation is under extreme 1-NFE settings, where gamma modeling becomes challenging, pointing to future work in richer parameterizations.

Abstract

Diffusion models have achieved remarkable generation quality, but they suffer from significant inference cost due to their reliance on multiple sequential denoising steps, motivating recent efforts to distill this inference process into a few-step regime. However, existing distillation methods typically approximate the teacher trajectory by using linear shortcuts, which makes it difficult to match its constantly changing tangent directions as velocities evolve across timesteps, thereby leading to quality degradation. To address this limitation, we propose ArcFlow, a few-step distillation framework that explicitly employs non-linear flow trajectories to approximate pre-trained teacher trajectories. Concretely, ArcFlow parameterizes the velocity field underlying the inference trajectory as a mixture of continuous momentum processes. This enables ArcFlow to capture velocity evolution and extrapolate coherent velocities to form a continuous non-linear trajectory within each denoising step. Importantly, this parameterization admits an analytical integration of this non-linear trajectory, which circumvents numerical discretization errors and results in high-precision approximation of the teacher trajectory. To train this parameterization into a few-step generator, we implement ArcFlow via trajectory distillation on pre-trained teacher models using lightweight adapters. This strategy ensures fast, stable convergence while preserving generative diversity and quality. Built on large-scale models (Qwen-Image-20B and FLUX.1-dev), ArcFlow only fine-tunes on less than 5% of original parameters and achieves a 40x speedup with 2 NFEs over the original multi-step teachers without significant quality degradation. Experiments on benchmarks show the effectiveness of ArcFlow both qualitatively and quantitatively.

ArcFlow: Unleashing 2-Step Text-to-Image Generation via High-Precision Non-Linear Flow Distillation

TL;DR

Diffusion-based image synthesis is hindered by high inference cost due to many denoising steps. ArcFlow introduces a non-linear flow distillation framework that parameterizes the velocity field as a mixture of continuous momentum processes, enabling exact integration via a closed-form analytic solver and aligning the student trajectory with the pre-trained teacher using velocity matching. By fine-tuning only lightweight LoRA adapters and an output head, ArcFlow achieves high-quality, diverse generation with as few as 2 NFEs, reporting up to 40× speedups on large backbones like Qwen-Image-20B and FLUX. The approach yields robust improvements across multiple benchmarks, faster convergence, and stable training without adversarial objectives, highlighting the value of respecting intrinsic flow dynamics for efficient generative inference. A notable limitation is under extreme 1-NFE settings, where gamma modeling becomes challenging, pointing to future work in richer parameterizations.

Abstract

Diffusion models have achieved remarkable generation quality, but they suffer from significant inference cost due to their reliance on multiple sequential denoising steps, motivating recent efforts to distill this inference process into a few-step regime. However, existing distillation methods typically approximate the teacher trajectory by using linear shortcuts, which makes it difficult to match its constantly changing tangent directions as velocities evolve across timesteps, thereby leading to quality degradation. To address this limitation, we propose ArcFlow, a few-step distillation framework that explicitly employs non-linear flow trajectories to approximate pre-trained teacher trajectories. Concretely, ArcFlow parameterizes the velocity field underlying the inference trajectory as a mixture of continuous momentum processes. This enables ArcFlow to capture velocity evolution and extrapolate coherent velocities to form a continuous non-linear trajectory within each denoising step. Importantly, this parameterization admits an analytical integration of this non-linear trajectory, which circumvents numerical discretization errors and results in high-precision approximation of the teacher trajectory. To train this parameterization into a few-step generator, we implement ArcFlow via trajectory distillation on pre-trained teacher models using lightweight adapters. This strategy ensures fast, stable convergence while preserving generative diversity and quality. Built on large-scale models (Qwen-Image-20B and FLUX.1-dev), ArcFlow only fine-tunes on less than 5% of original parameters and achieves a 40x speedup with 2 NFEs over the original multi-step teachers without significant quality degradation. Experiments on benchmarks show the effectiveness of ArcFlow both qualitatively and quantitatively.
Paper Structure (31 sections, 3 theorems, 25 equations, 14 figures, 8 tables, 1 algorithm)

This paper contains 31 sections, 3 theorems, 25 equations, 14 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Momentum Parameterization of Probability Flow
  5. Analytic ODE Solvers
  6. Flow Distillation with Analytic Solvers
  7. Experiments
  8. Implementation Details
  9. Comparison Study
  10. Ablation Study
  11. Conclusion
  12. Preliminaries
  13. Discussions on Qwen-Image-Lightning
  14. Additional Technical Details
  15. Mixed trajectory Integration
  16. ...and 16 more sections

Key Result

Theorem 1

Consider the velocity field predicted by ArcFlow at any sampled latent $\mathbf{y}$ and timestep $t$, parameterized as $\mathbf{v}_{\theta}(\mathbf{y}, t) = \sum_{k=1}^K \pi_k(\mathbf{y})\, \mathbf{v}_k(\mathbf{y})\, \gamma_k(\mathbf{y})^{1-t}$ according to Eq. eq:mixture_expectation. Let $\mathbf{u

Figures (14)

  • Figure 1: Comparisons between images generated by ArcFlow and other state-of-the-art distillation methods based on Qwen-Image-20B, demonstrating the power of ArcFlow for few-step high-fidelity generation while maintaining remarkable parameter efficiency.
  • Figure 2: Comparison of FID scores across training iterations for different methods. ArcFlow achieves superior convergence speed.
  • Figure 3: ArcFlow Framework. (a) The forward pipeline of ArcFlow. Given an input $x_t$, the condition $c$ and timestep $t$, a DiT backbone with three projection heads predicts the parameters $v$, $\omega$, $\gamma$ across $K$ dynamic modes, which respectively denote the mode-specific velocities, momentum factors, and the gating probabilities used to reconstruct the teacher velocity field. (b) A comparison of flow trajectories produced by the multi-step teacher model, the few-step linear student model, and our ArcFlow.
  • Figure 4: Qualitative comparisons with methods distilled on Qwen-Image-20B (2NFE). Every column contains two images which are generated from the same batch of initial noise. ArcFlow generates diverse samples that better align with teacher than competitors.
  • Figure 5: Qualitative comparisons with Qwen-Image-Lightning. Our ArcFlow exhibits visibly clearer details.
  • ...and 9 more figures

Theorems & Definitions (5)

  • Theorem 1
  • Definition 1: Chebyshev System
  • Lemma 1: Haar Condition cheney1966introduction
  • Proposition 1
  • proof