- 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
-
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.
-
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.
-
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 and , with . The endpoint estimate satisfies
A squared endpoint loss is therefore exactly a -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, , giving 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. -
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, -weighted MSE alone, ordinary MSE plus the same perceptual loss, and the full method under matched conditions.
-
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
-
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?
-
With identical initialization, training pairs, time-sampling distribution, LoRA configuration, training budget, and evaluation settings, how do ordinary MSE, -weighted MSE alone, ordinary MSE plus the same perceptual loss, the full method, and an RF++-style objective compare? A -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.