acceptodds
Withdrawn from ICLR 2027

Coupling Trajectory Rectification and Perceptual Supervision in Rectified Flow

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Rectified Flow straightens the trajectories of a flow model so that they can be integrated accurately with few steps, but it supervises the velocity along each path rather than the resulting image. Another family of few-step methods instead supervises that image directly against a teacher, often under a perceptual distance, and reaches high quality at small budgets, but usually trains each model for a fixed number of steps. We propose RefineFlow, which unites both forms of supervision in a single prediction. This prediction, the endpoint estimated by the model, is matched to the teacher's image under a perceptual distance and simultaneously defines the rectification target. Supervision on the output thereby shapes the velocity field along the whole trajectory and straightens it, within a reflow stage and without extra network evaluation or teacher integration. On text-to-image backbones, RefineFlow improves few-step generation quality and prompt alignment over the corresponding reflow baselines with far fewer trainable parameters, without training a separate model for each number of steps.

Primary area:
generative models
Keywords:
Rectified Flow, Reflow, Flow Matching, Few-Step Generation, Perceptual Supervision, Text-to-Image Generation

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_WvxCOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

RefineFlow aims to improve few-step generation with Rectified Flow. Given an interpolated state from a fixed noise–image pair, the model predicts a velocity and constructs an endpoint estimate through one Euler extrapolation. This endpoint is used both for perceptual supervision and, with stop-gradient, to construct the velocity regression target. The paper evaluates SD v1.5, SDXL, CIFAR-10, and ImageNet, and includes a Diff2Flow integration and distance-function ablations. The text-to-image experiments use LoRA and E-LatentLPIPS, with one SD v1.5 checkpoint supporting multiple sampling budgets.

Critical strengths
  1. Matched comparisons provide useful evidence of effectiveness. Appendix C / Table 9 controls the backbone, pair-generation pipeline, LoRA configuration, and training hyperparameters. SDXL four-step FID improves from 24.30 to 21.18, and CLIP improves from 0.318 to 0.333. Table 5 retains Diff2Flow’s pair construction and replaces only the training objective, improving two-/four-step FID from 32.31 / 25.29 to 28.74 / 24.10. These results support the practical value of the combined objective.

  2. The method is straightforward to integrate. Training reuses fixed pairs and requires one velocity-network evaluation per iteration without online multi-step teacher integration. The same checkpoint supports the different sampling budgets in Table 1. This is useful in practice, although the perceptual network still adds training overhead.

Critical weaknesses

My concern is the central mechanistic interpretation and its supporting evidence. The velocity regression term is exactly gradient-equivalent to a time-weighted conventional velocity MSE, while the experiments do not sufficiently establish the incremental value of the remaining loss combination over prior work. Addressing this requires substantially reframing the main contribution and adding key controls, with no guarantee that a sufficient contribution remains. I therefore view this as more than a local clarification or a small experimental addition.

  1. The central shared-endpoint construction is gradient-equivalent to a time-weighted conventional velocity loss. This is my main concern about the methodological contribution. Let u=x1−x0u=x_1-x_0 and v=vθ(xt,t)v=v_\theta(x_t,t), with xt=(1−t)x0+tx1x_t=(1-t)x_0+tx_1. The endpoint estimate satisfies

    x^1(t)−x1=(1−t)(v−u).\hat x_1^{(t)}-x_1=(1-t)(v-u).

    A squared L2L_2 endpoint loss is therefore exactly a (1−t)2(1-t)^2-weighted velocity MSE. The actual perceptual distance involves nonlinear features and does not reduce to this scalar weighting; these cases should be distinguished.

    For the velocity regression term, u^t=sg⁡[x^1(t)]−x0\hat u_t=\operatorname{sg}[\hat x_1^{(t)}]-x_0, giving v−u^t=t(v−u)v-\hat u_t=t(v-u) numerically. With stop-gradient, the actual parameter gradient is

    =\mathbb E[2tJ_\theta^\top(v-u)] =\nabla_\theta\mathbb E[t\|v-u\|_2^2],$$ where $J_\theta=\partial v/\partial\theta$. The gradient corresponds to weight $t$, although the forward loss value is $t^2\|v-u\|_2^2$. Thus, Algorithm 1 has the same parameter gradient as a $t$-weighted velocity MSE plus endpoint perceptual supervision. The shared-endpoint construction introduces no additional parameter-update mechanism beyond this equivalent objective. My concern therefore goes beyond a missing explanation: acknowledging the equivalence requires substantially reframing the main contribution and establishing the value of the resulting loss combination. Additional generation results cannot remove this algebraic equivalence, although the equivalence itself does not invalidate the empirical gains.
  2. The ablations do not fully disentangle the contributions of time weighting and endpoint perceptual supervision, and a matched comparison with the closest objective is missing. Tables 6–7 provide evidence about the choice of distance function: for example, Table 7 improves CIFAR-10 one-step FID from 8.10 to 6.00 when replacing the squared distance with the perceptual distance. Table 9 supports the full objective over ordinary reflow. However, these results do not fully disentangle time weighting from endpoint perceptual supervision. The relevant controls are ordinary velocity MSE, tt-weighted MSE alone, ordinary MSE plus the same perceptual loss, and the full method under matched conditions.

  3. The paper lacks a comparison and analysis of other relevant, important methods.

    Lee et al., Improving the Training of Rectified Flows (RF++), Section 4.2, already combines LPIPS on an Euler-predicted endpoint with a time-weighted Pseudo-Huber velocity loss during reflow. This is not identical to RefineFlow, but it is close enough to require a direct discussion and a matched comparison. The manuscript already cites this work; the issue is the comparison and the incremental contribution. Without this comparison and the component controls above, the contribution remaining after the mechanistic interpretation is corrected is insufficiently supported.

    Li and He, Back to Basics: Let Denoising Generative Models Denoise (JiT), Section 3.2 / Table 1, is also relevant to the interpretation: it distinguishes network prediction space from loss space and discusses time reweighting when converting squared losses. Clarifying this distinction would make the contribution more precise.

Questions for authors
  1. Does the implementation follow Algorithm 1 exactly, using the same current velocity prediction to construct the endpoint and detached target? If so, do you agree with the gradient equivalence above, and what is the additional contribution under this formulation? If the implementation differs, which operation breaks the equivalence?

  2. With identical initialization, training pairs, time-sampling distribution, LoRA configuration, training budget, and evaluation settings, how do ordinary MSE, tt-weighted MSE alone, ordinary MSE plus the same perceptual loss, the full method, and an RF++-style objective compare? A tt-weighted MSE plus the same perceptual loss should be equivalent to the full method under the derivation above; it can check the implementation but should not be presented as an independent method demonstrating an improvement.

Reviewer_P6SQOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes RefineFlow: within a reflow stage, a single Euler projection yields an endpoint estimate x^1(t)=xt+(1−t)vθ\hat{x}_1^{(t)} = x_t + (1-t)v\theta, which is aligned to the paired endpoint x1x_1 under the E-LatentLPIPS perceptual distance and simultaneously replaces the fixed displacement as the velocity-regression target. The method is validated with LoRA fine-tuning on SD v1.5 and SDXL.

Critical strengths

None

Critical weaknesses

W1. The paper's core claim is that a single predicted endpoint serves as both the perceptual target and the rectification target, so that output supervision shapes the velocity along the trajectory. But plugging Eq. 2 and Eq. 4 into Eq. 6 gives v − û_t = t[v − (x₁ − x₀)], i.e. the flow term is just the standard reflow loss scaled by t². Its minimizer is unchanged, and the stop-gradient blocks exactly the coupling the paper advertises. What is left is reweighted reflow plus an off-the-shelf latent perceptual loss (E-LatentLPIPS, from Kang et al. 2024). The obvious ablation — a decoupled variant with the fixed target x₁ − x₀ plus the perceptual term — is never run, so there is no evidence the central mechanism does anything. This undercuts the title, Figure 1, and the first contribution.

W2. The headline SD v1.5 claim (beating 2-RF from the first reflow stage) compares against numbers quoted from the original paper, with different pair data and training budget. The only controlled, in-house comparison is SDXL in Appendix C, where the gain is much smaller (24.30 → 21.18). Combined with the narrow setup — two legacy UNet backbones, FID/CLIP/GenEval2 only, single runs with no error bars — I cannot attribute the reported gains to the proposed objective.

Questions for authors

1.In Table 4, ProReflow has the best FID but its GenEval2 entry is "–", and GenEval2 is the one metric where RefineFlow leads. Can you fill it in? Also, SDXL-Turbo is evaluated at 512 while the rest are at 1024 — is its FID comparable in that column?

2.Both backbones are latent diffusion models — SD v1.5 and SDXL are trained with an ε-prediction diffusion objective, and the rectified-flow property is obtained post hoc by reflow (your pairs are generated with 25-step DDIM on the pretrained base model). Models pretrained with flow matching at scale — SD3/3.5, FLUX.1, Qwen-Image — are never evaluated.

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

This paper proposes RefineFlow, which combines perceptual endpoint supervision with trajectory rectification for few-step generative modeling. It uses a single Euler projection to estimate the endpoint, which serves both as the perceptual matching target and the velocity regression target, without additional network evaluations or teacher integration. Experiments on SD v1.5, SDXL, CIFAR-10, and ImageNet demonstrate improved few-step generation quality and prompt alignment over reflow baselines, with parameter-efficient training via LoRA.

Critical strengths
  1. Novel Methodology. RefineFlow combines trajectory rectification and perceptual supervision through a shared predicted endpoint, improving both trajectory straightness and image quality within a unified training objective.

  2. Low Training Overhead. The method requires no additional network forward passes or repeated teacher integration during training, making it more computationally efficient than methods that rely on multi-step teacher rollouts.

  3. Strong Few-Step Generation Performance. RefineFlow significantly improves image quality at low sampling budgets. On Stable Diffusion v1.5, compared with the Rectified Flow baseline, it reduces FID from 47.91 to 31.92 for one-step sampling and from 31.35 to 26.55 for two-step sampling.

Critical weaknesses
  1. Approximate Endpoint Estimation. RefineFlow estimates the final output using a single Euler projection, assuming that the current velocity remains constant over the remaining time interval. This approximation may be inaccurate for highly nonlinear trajectories, potentially limiting the quality of the supervision signal compared with multi-step teacher integration.

  2. Limited Exploration of Training Strategies. The method uses fixed Reflow pairs throughout training without regenerating or updating them. Although this design reduces computational overhead, the paper does not investigate whether dynamically updating the pairs could further improve image generation quality.

  3. Incomplete Evaluation of Broader Applicability. The quantitative evaluation on SDXL focuses primarily on four-step sampling, and the paper does not investigate whether RefineFlow can be combined effectively with local trajectory rectification methods. Therefore, its performance across different sampling budgets and its potential benefits when combined with more advanced rectification strategies remain insufficiently explored.

Questions for authors

1. Endpoint Estimation: How sensitive is RefineFlow's performance to the accuracy of the single-step Euler endpoint estimation? Have the authors investigated whether higher-order approximations could further improve generation quality?

2. Loss Weighting: Why are the flow loss and perceptual loss assigned equal weights? Have the authors conducted sensitivity analyses to evaluate the impact of different loss weights on generation quality and trajectory straightness?

3. Evaluation on SDXL: Since the SDXL experiments focus primarily on four-step sampling, can the authors provide results for other sampling budgets, such as one, two, and eight steps, to further demonstrate the flexibility of RefineFlow?

4. Fixed Reflow Pairs: How does using fixed reflow pairs throughout training affect the final generation quality? Could periodically regenerating these pairs provide additional improvements despite the increased computational cost?

Reviewer_x3RgOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes RefineFlow, a training objective for Rectified Flow models that combines trajectory rectification with perceptual supervision through a shared endpoint prediction. The endpoint is estimated by a single Euler projection and used both for perceptual matching against the teacher output and, with stop-gradient, as the velocity regression target. The method uses fixed offline training pairs and supports different sampling budgets with a single checkpoint.

Critical strengths

(1) The idea is simple and intuitive. Supervising the predicted endpoint provides a natural way to incorporate perceptual information into velocity-field training. The method fits into an existing Rectified Flow training pipeline without additional generative-model evaluations or online teacher integration, making it practically appealing. (2) The gains are particularly clear at small sampling budgets. On SD v1.5, one-step FID improves from 47.91 to 31.92, while performance at larger sampling budgets remains competitive. The ability to use the same checkpoint across different step counts adds practical flexibility. (3) The matched SDXL comparison keeps the backbone, training pairs, LoRA configuration, and hyperparameters fixed, showing a four-step FID improvement from 24.30 to 21.18. Improvements when applying the objective within Diff2Flow provide additional evidence that its benefits are not restricted to one pair-construction pipeline.

Critical weaknesses

(1) The connection to closely related objectives could be clearer. Improving the Training of Rectified Flows (Lee et al., NeurIPS 2024) already combines velocity regression with perceptual supervision of an endpoint reconstructed from the velocity prediction. A more explicit discussion of this connection, particularly the distinction introduced by the shared-endpoint formulation, would help position the contribution. (2) Some empirical conclusions require qualification. The strongest improvements occur at one or two sampling steps, while gains at larger budgets are smaller and FID does not consistently surpass stronger competing methods.

Questions for authors

(1) Could the authors briefly discuss the relationship to the LPIPS-Huber objective of Lee et al. (2024), clarifying the main differences in the objective and the role of the shared endpoint? (2) Could the authors clarify which baseline results were reproduced under the stated evaluation protocol and which were taken from prior publications? A brief discussion of the additional perceptual-loss overhead would also help contextualize the efficiency claims.

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

Related papers

Discussion (0)

Sign in to comment.