acceptodds
Withdrawn from ICLR 2027

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

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

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 approximate velocity field 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 mitigates linear 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 of original parameters and achieves a 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.

Primary area:
generative models
Keywords:
image generation, diffusion model, few-step distillation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_pqtKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on accelerating text-to-image generation through few-step diffusion distillation. The authors observe that linear shortcuts poorly approximate curved teacher trajectories, which can reduce image quality and diversity in few-step generation. They therefore propose ArcFlow, which models nonlinear trajectories using continuous momentum mixtures, integrates them analytically, and learns them through lightweight adapters. To validate the method, they conduct several experiments, including benchmark comparisons on Qwen-Image and FLUX, trajectory analyses, convergence comparisons, and component ablations. The results show that ArcFlow enables two-step generation, preserves diversity, and achieves faster convergence.

Critical strengths
  1. This paper proposes ArcFlow, which constructs a closed-form nonlinear flow trajectory to approximate the teacher trajectory.

  2. To simplify training, the authors propose an analytic trajectory solver for ArcFlow.

  3. The paper conducts a range of experiments to demonstrate ArcFlow's advantages across different backbones.

Critical weaknesses
  1. Section 1 claims that ArcFlow is "the first distillation framework to explicitly construct a closed-form non-linear flow trajectory to approximate the teacher trajectory." However, "Hierarchical Koopman Diffusion: Fast Generation with Interpretable Diffusion Trajectory" (NeurIPS 2025) already evolves states in closed form, zt=eA(t−s)zsz_t = e^{A(t-s)} z_s, within a learned latent space. These trajectories are not straight lines. HKD is trained on PF-ODE trajectories from a pretrained EDM teacher (Alg. 1; Eq. (6)). ArcFlow's interval trajectory is likewise a superposition of exponential modes, yet the paper does not discuss this work. The priority claim therefore needs qualification.

  2. The Abstract claims generation "without significant quality degradation" at 2 NFEs. However, Table 1 reports that the FLUX text score drops from 0.556 to 0.368. This substantial decline does not support such a broad preservation claim.

  3. The claimed 40× inference speedup lacks a measured comparison with the teacher. The main text reports no latency results, and NFE counts alone do not establish the claimed wall-clock speedup.

  4. The 2-NFE baselines in Tables 1 and 2 are not fully specified. Section 4.2 notes that official pi-Flow checkpoints support only 4 NFEs, while the official Qwen-Image-Lightning repository releases only 4- and 8-step LoRAs. The paper does not state whether these baselines were retrained at 2 NFEs or simply run below their distilled step counts, nor does it explain how their training budgets compare with ArcFlow's. The reported 2-NFE advantages may therefore partly reflect a mismatch in step count or training budget.

  5. The core claim attributes quality degradation to linear shortcuts that cannot follow the teacher's changing tangents. However, a few-step sampler uses only interval endpoints, and a linear step with the correct average velocity can reach any endpoint. ArcFlow's own update, xte=xts−Φx_{t_e} = x_{t_s} - \Phi, is likewise a single displacement. Section 4.3 compares velocity changes at adjacent teacher timesteps and intermediate latent states with Qwen-Image-Lightning, which a constant-velocity student cannot match by construction. The linear control in Section 4.3 is trained under matched settings, i.e., to match instantaneous teacher velocities (Eq. (6)), which a constant velocity cannot satisfy. Its performance gap is nonetheless attributed to the "limited expressiveness" of linear modes. No linear student trained on the teacher's interval displacement under the same budget is reported. Thus, the benefit of nonlinear trajectories over a properly trained linear step remains untested.

Questions for authors
  1. Could the authors distinguish ArcFlow's momentum-based velocity parameterization from HKD's closed-form latent evolution and trajectory supervision? Please narrow the first-framework claim accordingly.

  2. Could the authors qualify the quality-preservation claim and report teacher-student comparisons with uncertainty, especially for text rendering? Please distinguish preserved capabilities from degraded ones.

  3. Could the authors report teacher and student latency under identical hardware, precision, batch size, and guidance settings? Please clarify how the 40× factor was obtained.

  4. Could the authors specify the checkpoint, native step count, and training budget of each 2-NFE baseline? Please compare pi-Flow and Qwen-Image-Lightning at their native step counts, or after retraining them at 2 NFEs under a matched budget.

  5. Could the authors compare ArcFlow with a γ≡1\gamma \equiv 1 student trained to regress the teacher's interval displacement using the same teacher-supervision budget per update? Please report results at the final training budget.

Reviewer_BxAdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ArcFlow, which predicts a mixture of time-evolving momentum modes from the beginning of each denoising interval. These modes define a nonlinear velocity curve with a closed-form integral. The student is trained through multi-timestep teacher velocity supervision on mixed teacher–student rollout states, using LoRA adapters and extended output heads. Experiments demonstrate the effectiveness of the proposed method.

Critical strengths
  • The method provides a coherent parameterization and inference mechanism. The same predicted modes determine both intermediate velocities and interval displacement. Analytical integration avoids additional backbone evaluations within each interval.

  • Comprehensive experiments cover multiple models like Qwen-Image and FLUX and benchmarks. Against Qwen-Image-Lightning, human judgments prefer ArcFlow in 52.15% of cases versus 33% for the baseline.

Critical weaknesses
  • The novelty of the supervision scheme is overstated. pi-Flow already predicts a shared policy once and supervises it at multiple teacher-query states along detached rollouts. Consequently, multi-timestep supervision of shared interval parameters is not itself the distinguishing contribution. ArcFlow’s distinction is its exponential momentum parameterization and directly integrable interval transition.

  • "Few-step Linear Flow Trajectory (Previous Student)" in Figure 2 seems misleading. Do you mean "Previous Student" as consistency models/meanflow-like students? If so, these methods learn the same teacher trajectory as ArcFlow. Euler parametrization does not mean linear flow. ArcFlow only differs in parametrization, rather than the way of learning flow trajectory.

Questions for authors

Please address the novelty concerns in weakness part.

Reviewer_FkDnOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

ArcFlow distills large flow-matching text-to-image models (Qwen-Image-20B, FLUX.1-dev) into 2-step generators. At the start of each interval, the student predicts a mixture of K “momentum modes”, each a velocity vector with an exponential time factor. This defines a non-linear velocity curve with a closed-form integral. Training matches the student’s analytic velocity to the teacher’s at intermediate states of the rollout, using LoRA adapters and output heads. ArcFlow reports state-of-the-art 2-NFE results and faster convergence than pi-Flow and TwinFlow.

Critical strengths
  • Strong 2-NFE results on large models with lightweight training. ArcFlow is competitive with or better than prior 2-NFE students on two 12–20B backbones while training under 5% of parameters, and it preserves diversity better than adversarial baselines (OneIG Diversity 0.182 vs. 0.098 for Qwen-Image-Lightning).
Critical weaknesses
  • Incremental over pi-Flow [1]. The framework is pi-Flow’s imitation distillation. The only change is the policy family: an exponential-mode mixture instead of pi-Flow’s Gaussian-mixture policy. Moreover, ArcFlow’s velocity is time-only within an interval, much like pi-Flow’s DX policy, which the paper does not discuss.
  • The gains over pi-Flow are not convincingly shown. At pi-Flow’s established 4-NFE setting, the two methods are essentially tied (Tables 1–2). The 2-NFE pi-Flow baseline appears to come from the authors’ short 3,000-step retraining (FID 20.07, Tables 2 and 8), while ArcFlow’s headline model seems to be trained longer. “Faster convergence” over 3,000 steps also does not show a better final model.

Reference:

[1] Hansheng Chen, Kai Zhang, Hao Tan, Leonidas Guibas, Gordon Wetzstein, Sai Bi. pi-Flow: Policy-Based Few-Step Generation via Imitation Distillation. ICLR 2026

Questions for authors
  • How does ArcFlow compare with pi-Flow’s DX policy, which is also time-only within an interval? What does the exponential basis add beyond it?
  • Was the 2-NFE pi-Flow baseline in Tables 1–2 trained for the same number of steps as the reported ArcFlow model? What are the results when both are trained to convergence?
  • Since ArcFlow’s velocity cannot react to deviations of the state within an interval, does it accumulate more error than pi-Flow’s state-dependent GM policy, e.g., under perturbations or at larger intervals?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.