Source-linked AI summary
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
Few-step distillation must reduce diffusion sampling cost without losing fidelity, but linear shortcuts poorly match teachers’ changing tangent directions. ArcFlow uses analytically integrated non-linear momentum trajectories and lightweight adapters, achieving high-quality 2-NFE generation with less than 5% fine-tuned parameters and a 40× speedup.
Problem
Diffusion models require many sequential denoising steps, while linear distillation shortcuts struggle to match teacher trajectories whose tangent directions vary across timesteps.
Method
ArcFlow parameterizes velocity as a mixture of continuous momentum processes, analytically integrates the resulting non-linear trajectory, and trains it through trajectory distillation with lightweight adapters.
Results
ArcFlow achieves superior generation quality with fewer trainable parameters, a 40× inference speedup over the teacher, and at most 4× faster training convergence than prior methods.
Takeaways & Limitations
Respecting non-linear flow dynamics enables high-precision teacher alignment, faster convergence, and efficient few-step generative inference without invasive full-parameter training.
Takeaways & Limitations
ArcFlow, like prior image-generation methods, could be misused to generate misleading or harmful visual content and therefore requires responsible deployment measures.
Abstract
from arXiv · showhide
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.
1 Introduction
ArcFlow targets the quality–efficiency challenge in few-step image generation by replacing linear trajectory shortcuts with analytically integrated non-linear flow trajectories. On Qwen-Image-20B, it achieves high-fidelity 2-NFE generation, faster convergence, and parameter-efficient adaptation.
- Diffusion and flow-matching models typically require 40 to 100 denoising steps, limiting their practicality for real-time applications.
- Existing distillation methods approximate teacher trajectories with linear shortcuts despite changing tangent directions, causing geometric mismatch.
- ArcFlow parameterizes velocity as a weighted mixture of continuous momentum processes to construct non-linear trajectories that follow teacher tangent variation.
- With 2 NFEs, ArcFlow produces high-fidelity images comparable to Qwen-Image-20B while surpassing pi-Flow and TwinFlow in 2-step quality.
- An analytic trajectory solver enables high-precision teacher alignment and supports parameter-efficient distillation without full-model training.
- ArcFlow achieves a 40× inference speedup, at most 4× faster training convergence, and fine-tunes less than 5% of the original parameters.
2 Related Work
Related work accelerates diffusion sampling through trajectory simplification, adversarial or distribution-matching objectives, and velocity approximations. These approaches face challenges involving discretization error, unstable training, overhead, or limited precision at 2 steps.
- ArcFlow’s Figure 2 comparison reports superior convergence speed across training iterations using FID scores.
- Text-to-image generation: Diffusion and flow-matching models integrate probability-flow ODEs with iterative solvers requiring 40 to 100 function evaluations.
- Few-step Image Generation: Trajectory-simplification methods such as Progressive Distillation and Rectified Flow reduce NFEs by iteratively straightening the flow.
- Few-step Image Generation: VSD, DMD, and TwinFlow improve visual sharpness but rely on adversarial objectives associated with unstable training, mode collapse, and high memory overhead.
- Few-step Image Generation: Gaussian-mixture approaches approximate velocity evolution, but their probabilistic approximations lack precision at lower NFEs, particularly 2 steps.
3 Method
ArcFlow models the teacher’s changing velocity field with non-linear momentum trajectories and analytically integrates them for few-step distillation. Its mixed-mode parameterization and trajectory-alignment training target accurate teacher matching with efficient adaptation.
- Existing few-step distillation methods use linear shortcuts for teacher trajectories whose tangent directions vary across timesteps, causing geometric mismatch.
- 3.1 Momentum Parameterization of Probability Flow: ArcFlow parameterizes the velocity field as a weighted mixture of continuous momentum processes across multiple evolution modes.Each mode has a gating probability, basic velocity, and momentum factor.
- 3.1 Momentum Parameterization of Probability Flow: When K≥N, the momentum-based parameterization admits a configuration that fits sampled ground-truth velocities at N distinct timesteps.The theorem supports high-precision non-linear approximation of the teacher velocity field.
- 3.2 Analytic ODE Solvers: The momentum mixture is equivalent to a linear combination of exponential time factors, enabling closed-form integration across arbitrary timestep intervals.The analytic transition operator directly computes latent displacement, avoiding numerical integration within each sampling step.
- 3.2 Analytic ODE Solvers: The analytic solver remains numerically stable as γ→1, smoothly reducing its coefficient to the linear interval length t_s−t_e.This bridges the non-linear regime γ≠1 and the linear-flow regime γ=1.
- 3.3 Flow Distillation with Analytic Solvers: ArcFlow distills the teacher by matching instantaneous velocities over sampled timesteps, with mixed latent integration exposing the student to the full interval trajectory.Matching only n=2∼4 timesteps is reported as sufficient because the momentum parameterization captures inherent non-linearity.
- 3.3 Flow Distillation with Analytic Solvers: The method supports parameter-efficient training using lightweight LoRA adapters and an output head instead of full-model fine-tuning.
4 Experiments
ArcFlow is evaluated on Qwen-Image-20B and FLUX.1-dev across multiple benchmarks, qualitative comparisons, convergence tests, and ablations. It generally preserves quality and diversity while using parameter-efficient non-linear trajectory modeling.
- Quantitative Results: ArcFlow is evaluated on Qwen-Image-20B and FLUX.1-dev using 2 NFEs across Geneval, DPG-Bench, and OneIG-Bench.The evaluation also includes Align5000 and qualitative comparisons against distilled few-step methods.
- Quantitative Results: ArcFlow consistently outperforms or remains competitive with state-of-the-art few-step models across three benchmarks.It achieves the lowest FID and pFID across both backbones according to the reported comparison.
- Quantitative Results: +85.7% Diversity improvement is reported for ArcFlow on OneIG-Bench.The comparison states that ArcFlow preserves diversity relative to adversarial-based methods that suffer from mode collapse.
- Qualitative Results: ArcFlow preserves higher visual quality and generation diversity than competing linear distillation methods under matched initial noise.Competitors are reported to show mode collapse, degraded aesthetics, blurred textures, or structural artifacts.
- Qualitative Results: ArcFlow produces finer and more coherent details than Qwen-Image-Lightning in zoomed-in qualitative comparisons.The reported discrepancy is attributed to Lightning potentially trading fine-grained visual fidelity for semantic alignment.
- Ablation Study: Adaptive momentum improves teacher–student alignment, while removing the explicit momentum factor produces inferior FID scores.The momentum ablation uses the Align5000 prompt set with 1,500 training steps.
- Ablation Study: Neither restricted mixture configuration, (K, 1) or (1, K), is competitive with the default (K, K) configuration.The study interprets this as evidence that velocity and momentum mixtures should be decoupled.
- Ablation Study: Increasing the mixture size K generally improves performance, but the improvement from K=16 to K=32 is marginal.The reported diminishing returns are considered alongside increased parameter cost and complexity.
5 Conclusion
ArcFlow distills diffusion teachers with non-linear trajectories built from continuous momentum mixtures and analytically integrated to align student and teacher flows. The reported conclusion is improved quality, diversity, convergence, and parameter efficiency relative to linear baselines.
- Conclusion: ArcFlow explicitly models teacher dynamics with non-linear trajectories by parameterizing velocity as a mixture of continuous momentum processes.Its analytic solver enables high-precision trajectory integration for flow distillation.
- Conclusion: ArcFlow avoids unstable adversarial objectives and invasive full-parameter training, leading to faster convergence and more efficient distillation.The method uses analytical trajectories and parameter-efficient adaptation.
- Conclusion: Extensive experiments report superior generation quality with fewer trainable parameters than linear baselines.The conclusion frames respecting underlying flow dynamics as important for efficient generative inference.
Impact Statement
ArcFlow targets efficient, high-quality image generation by distilling iterative probability-flow dynamics into few-step inference. Its design uses non-linear trajectory modeling, analytic integration, and parameter-efficient adaptation, while responsible deployment remains necessary because generative image systems can produce harmful content.
- Responsible deployment remains important because image-generation methods may be misused to create misleading or harmful visual content.
- 40–100 NFEs make iterative probability-flow ODE sampling a computational bottleneck for real-time image generation.
- Flow-matching distillation trains a student velocity field to emulate a teacher across timesteps, reducing the number of required function evaluations.
- ArcFlow represents velocity evolution with continuous momentum processes and analytically integrates the resulting non-linear student trajectory over arbitrary intervals.
- The analytic transition avoids numerical solvers for student dynamics, while logarithmic momentum-factor parameterization and a reduced learning rate improve training stability.The momentum projection layer uses a learning rate of 0.1× that of other trainable layers.
D Training Details and Hyperparameters setting.
ArcFlow training freezes teacher backbones while adapting selected LoRA modules and output heads. Its analytic transition is continuous at the linear case and is implemented with stability-oriented momentum parameterization and initialization choices.
- Training setup: Teacher backbones are frozen while LoRA adapters and output heads predict velocities, momentum factors, and gating probabilities.
- Training setup: Rank-256 LoRA adapters are inserted into selected transformer, MLP, timestep-embedding, and projection modules for Qwen-Image-20B and FLUX.1-dev.
- Training setup: Training uses BF16 mixed precision on 96 H100 GPUs, with additional configurations reported in Table 6.
- Analytic transition: The analytic transition integrates the continuous student dynamics exactly over arbitrary intervals without numerical solvers.
- Analytic transition: At γ=1, the analytic expression is continuous and recovers the linear dynamic mode.
- Numerical stability: For numerical stability, the implementation uses the linear-case expression when |ln γ| < 10^-6.
E.2 Proof of Theorem 1
Theorem 1 establishes that ArcFlow’s momentum parameterization can represent arbitrary sampled velocity trajectories when the number of modes is at least the number of distinct sampled timesteps. The proof reduces this claim to invertibility of an exponential basis matrix via Chebyshev-system theory.
- Theorem statement: With K=N momentum modes, ArcFlow can theoretically approximate any ground-truth trajectory at N discrete timesteps.
- Problem reformulation: The representation problem is reformulated using composite weights, decoupling the D-dimensional fitting problem into identical scalar linear systems.
- Invertibility: Exact solvability for arbitrary sampled velocities is equivalent to invertibility of the basis matrix M.
- Chebyshev systems: A Chebyshev System is characterized by every non-trivial linear combination having at most N−1 distinct zeros.
- Conclusion: The Haar condition then guarantees matrix invertibility, yielding a parameter configuration that exactly aligns with arbitrary trajectory dynamics on the data manifold.
- Chebyshev systems: The exponential basis is shown to satisfy the Chebyshev zero-count property through induction and Rolle’s Theorem.
F.1 Ablations on Mixed Trajectory Integration
Ablations evaluate mixed trajectory integration on Qwen-Image-20B and FLUX.1-dev using teacher-alignment FID. The study compares ArcFlow models trained with and without this integration strategy.
- Mixed trajectory integration is ablated on Qwen-Image-20B and FLUX.1-dev after 3,000 training iterations with batch size 16.
- The ablation uses teacher-alignment FID on Align5000 to compare models with and without mixed trajectory integration.
- Figure 8 visualizes convergence across different student methods based on Qwen-Image-20B.
F.2 Convergence Visualization
ArcFlow shows fast, stable convergence in visualization experiments and maintains low-latency 2-NFE inference, while 1-NFE inference produces blurry results.
- Convergence: ArcFlow exhibits coherent global structure after 0.5K training iterations, with most stochastic artifacts and irregular noise suppressed.At this stage, the images mainly show mild over-smoothing rather than structural corruption.
- Convergence: By 3K iterations, ArcFlow produces images with no obvious visual defects under human inspection, demonstrating fast convergence and strong training stability.
- Baseline comparison: Compared with ArcFlow, pi-Flow retains residual noise artifacts throughout training, while TwinFlow converges more slowly and has inferior early and intermediate visual quality.The passage attributes TwinFlow’s behavior to its linear parameterization conflicting with the teacher’s complex trajectory.
- Inference efficiency: At 2 NFEs and 1024 × 1024 resolution, Table 8 compares the average inference time of student models run five times on the same prompt.The supplied passages state that ArcFlow-FLUX and ArcFlow-Qwen fall between the longest and shortest inference-time extremes, with negligible added floating-point operations.
- Failure case: At 1 NFE, ArcFlow produces blurry results, representing a severe quality degradation in the single-step setting.
G Limitations and Future Work
ArcFlow’s main reported limitation is severe generation-quality degradation when inference is forced into a single NFE.
- At 1 NFE, ArcFlow fails to produce meaningful results because modeling the momentum factor γ becomes highly sensitive and difficult.The authors propose deeper or more expressive layers for modeling γ and plan evaluation across models with diverse parameter scales.
H Additional Qualitative Results
The paper provides additional qualitative comparisons for students distilled from Qwen-Image-20B and FLUX.1-dev, alongside ArcFlow visualizations at 2 NFEs.
- Visualization protocol: The additional ArcFlow examples use randomly sampled prompts and are shown without manual selection or filtering.
- Qwen-Image-20B: Figure 10 presents additional qualitative comparisons among student models distilled from Qwen-Image-20B using the same batch of initial noise in each column.
- FLUX.1-dev: Figure 11 presents qualitative comparisons among student models distilled from FLUX.1-dev.
- ArcFlow-Qwen: Figures 12 and 13 visualize ArcFlow-Qwen outputs generated at 2 NFEs, with each image at 1024 × 1024 resolution.
- ArcFlow-FLUX: Figure 14 visualizes ArcFlow-FLUX outputs generated at 2 NFEs, with each image at 1024 × 1024 resolution.