acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Quality Tuning for Rectified Flow Models from a Rotational Perspective

Abstract

Quality tuning (QT) is critical for diffusion and rectified-flow models. It adapts pretrained generators to curated high-quality data and supports later post-training. However, the pretraining-QT data gap may require more updates, while over-tuning on limited data can cause overfitting and lower generation quality. Standard QT also uses a fixed noise shift, so noise exposure remains unchanged during training. We revisit QT from a rotational perspective, motivated by two observations. First, earlier direction correction improves reconstruction, even though it affects more steps. Second, at similar noise levels, angular errors vary across examples. Based on these observations, we introduce Rotational Quality Tuning (RoQT). At the macro level, a noise curriculum anneals a strong high-noise shift through an intermediate shift toward the reference shift. At the micro level, angular weighting emphasizes difficult examples within each noise region. We further decompose the flow-matching mean-squared error into magnitude and rotational components, and leverage the rotational component to derive an explicit, scale-calibrated direction correction. RoQT improves convergence efficiency and generation quality while mitigating the performance degradation caused by over-tuning on limited data. Experiments on SD3.5-Medium and Z-Image demonstrate that RoQT consistently outperforms standard QT on GenEval and Qwen-Image-Bench under the same training data and matched inference settings. We will release the code and data on GitHub.

Primary area:
generative models
Keywords:
Diffusion models, fine-tuning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_haiJOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Rotational Quality Tuning (RoQT), a framework designed to improve the quality tuning of pretrained rectified-flow image generation models. The method is motivated by two observations: correcting velocity directions at higher noise levels can improve reconstruction quality, and samples at similar noise levels exhibit different angular prediction errors. Based on these observations, RoQT introduces three components: (1) a macro-level noise curriculum that progressively shifts training from high-noise regions toward the reference timestep distribution, (2) a micro-level angular weighting scheme that emphasizes samples with larger relative directional errors, and (3) an additional direction correction term derived from the rotational component of the flow-matching MSE. The authors evaluate RoQT on SD3.5-Medium and Z-Image using approximately 50K curated image-text pairs. The method improves overall GenEval and Qwen-Image-Bench scores compared with standard supervised fine-tuning (SFT), and the authors provide component ablations and training dynamics to support its effectiveness. While the proposed framework is computationally straightforward and demonstrates encouraging empirical improvements, I have several concerns regarding its methodological novelty, the justification of its core design choices, and the completeness of the experimental evaluation.

Critical strengths
  1. The paper addresses the challenge of efficiently adapting pretrained rectified-flow models to limited high-quality datasets while mitigating performance degradation during prolonged fine-tuning. This is a relevant problem for modern generative-model training pipelines.
  2. RoQT does not require modifications to the backbone architecture, flow-matching trajectory, or inference procedure. The method can be integrated into existing LoRA fine-tuning pipelines with a single trainable-model forward pass.
  3. The decomposition of the flow-matching MSE into magnitude and rotational components provides an intuitive mathematical interpretation of the additional direction correction. The scale-calibrated angular objective also makes the intended directional gradient explicit.
  4. The experiments demonstrate improvements in overall GenEval and Qwen-Image-Bench scores on both SD3.5-Medium and Z-Image. The component and curriculum ablations provide preliminary evidence supporting the effectiveness of the combined framework.
Critical weaknesses
  1. The three components of RoQT largely resemble established training strategies, including timestep curriculum learning, difficulty-based sample reweighting, and angular regularization. In particular, the proposed direction correction is derived directly from the standard flow-matching MSE decomposition and essentially increases the relative strength of directional gradients already present in the original objective. The paper does not sufficiently demonstrate why this particular formulation is preferable to simpler alternatives, such as timestep-weighted flow matching, direct cosine-based regularization, or conventional hard-example weighting. Without comparisons against these alternatives, it is difficult to determine whether the improvement comes from the proposed rotational perspective or from more general loss-reweighting effects.

  2. The direction-correction experiment in Figure 1 is used to motivate emphasizing high-noise supervision. However, applying corrections from higher noise levels also affects more ODE integration steps. Consequently, the observed reconstruction improvements cannot be attributed solely to the greater importance of high-noise directional errors. Furthermore, the experiment relies on oracle reference velocities computed using ground-truth clean latents. Equation (5) establishes a relationship between timestep and one-step reconstruction error, but this does not directly establish the optimal supervision allocation along a complete generative trajectory. Additionally, Figure 2(a) shows elevated angular errors near both noise endpoints, which does not immediately explain why training should initially concentrate so heavily on high-noise samples.

  3. The micro-level weighting scheme assumes that larger prediction-target angular errors correspond to samples requiring stronger supervision. However, in conditional flow matching, the optimal velocity predictor represents a conditional expectation, and the target velocity may remain ambiguous given the noisy input and conditioning information. As a result, a larger angular error does not necessarily indicate a correctable model deficiency. This issue may be particularly relevant near the noise endpoints, where the available information for predicting an individual sample's target velocity can be limited. The paper does not investigate whether angular difficulty weighting primarily emphasizes learnable prediction errors or potentially amplifies supervision on intrinsically ambiguous samples.

  4. The experimental evaluation lacks widely adopted perceptual preference metrics such as PickScore, HPSv2, and ImageReward, relying primarily on GenEval and Qwen-Image-Bench. More importantly, despite approximately 17.2% of the curated fine-tuning dataset belonging to the Text category, the paper does not report dedicated OCR or text-rendering accuracy metrics. This omission is particularly concerning because the proposed curriculum strongly emphasizes high-noise supervision, potentially affecting the balance between global visual aesthetics and fine-grained textual fidelity. Improvements in aesthetics do not necessarily imply improvements in text-rendering accuracy, and the current evaluation cannot rule out potential trade-offs. Moreover, the improvements are not uniform across quality dimensions. On SD3.5-Medium, the Qwen-Image-Bench Quality score slightly decreases from 47.86 to 47.83, despite an improvement in Aesthetics. On Z-Image, Real-world Fidelity decreases from 43.00 to 42.13. The overall Qwen-Image-Bench improvement on SD3.5-Medium is also relatively modest, increasing from 43.53 to 44.04. Given that the primary objective is quality tuning, broader perceptual preference evaluation, dedicated OCR benchmarks, and human preference comparisons would provide substantially stronger evidence for the claimed improvements.

Questions for authors
  1. Can the authors provide a controlled experiment separating the effect of high-noise direction correction from the effect of correcting more ODE integration steps? How does this observation directly justify the proposed noise curriculum?

  2. How does RoQT compare against simpler alternatives under matched training budgets, including timestep-weighted flow matching, direct cosine regularization, and conventional difficulty-based loss weighting?

  3. Can the authors include PickScore, HPSv2, ImageReward, and dedicated OCR/text-rendering evaluations? In particular, does the improvement in aesthetics introduce any trade-off in text-rendering accuracy or other fine-grained generation capabilities?

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

The paper addresses slow adaptation and potential quality degradation from prolonged fine-tuning of rectified-flow models on limited high-quality data. The authors observe that starting velocity-direction correction at a higher noise level yields better reconstruction in the illustrated example, and that directional errors vary across samples at the same noise level. Motivated by these observations, they propose Rotational Quality Tuning (RoQT), combining a noise curriculum that transitions from a strong high-noise bias toward the reference sampling distribution, angular-difficulty weighting within noise regions, and an additional direction loss derived from the decomposition of velocity MSE. On SD3.5-M, RoQT improves GenEval from 0.6814 to 0.6943 and Qwen-Image-Bench from 43.53 to 44.04. On Z-Image, the corresponding scores improve from 0.7344 to 0.7562 and from 42.48 to 44.16. The training curves show that RoQT maintains better evaluation performance than standard SFT in later training stages.

Critical strengths
  1. A practically relevant problem. The paper examines supervision allocation and over-tuning during fine-tuning on high-quality data, with a clear approach to improving both noise-level allocation and velocity-direction learning.
  2. A lightweight, easy-to-adopt method. RoQT changes only sampling and loss computation, without modifying the flow path, velocity target, or model architecture. It requires only one model forward pass and can be readily integrated into existing fine-tuning pipelines.
  3. Overall gains on two backbones. On both SD3.5-M and Z-Image, overall scores on both benchmarks exceed those of standard SFT. The divergence between metrics in the training curves also highlights that monitoring a single metric may conceal quality degradation associated with over-tuning.
Critical weaknesses
  1. The technical novelty is limited. The noise curriculum, angular-difficulty weighting, and additional direction loss correspond to relatively common ideas in noise scheduling, hard-example weighting, and directional alignment. The method primarily combines these strategies for quality tuning, with limited novelty in the individual components.
  2. The ablations do not sufficiently isolate each component's contribution. In Table 3, the noise-curriculum-only variant achieves a GenEval score of 0.6711, below standard SFT's 0.6814. Retaining angular weighting and the direction loss while fixing the shift at the reference value also fails to outperform standard SFT on GenEval. This does not invalidate the full method, but the interactions between components and their individual benefits require clearer analysis.
  3. Evidence for robustness and perceptual-quality improvements remains limited. Some overall gains are small, such as the 0.51-point improvement on Qwen-Image-Bench for SD3.5-M, and results across multiple random seeds or confidence intervals are not reported. Evaluation mainly relies on compositional alignment and vision-language-model scores, leaving human-preference and aesthetic benefits insufficiently validated.
  4. Implementation and hyperparameter details are insufficient. The precise noise-curriculum schedule, online estimation of regional mean angular errors, some weighting hyperparameters, and inference settings are not adequately specified. More complete details and sensitivity analysis would help assess reproducibility and the method's range of applicability.
Questions for authors

See "Weakness".

1 comment
AuthorsOpenReview ↗

We thank the reviewer for highlighting the relevance of quality tuning, RoQT's single-pass design, and its empirical gains.

Q1. Technical Novelty and Coordinated Formulation

RoQT's contribution is a lightweight, single-pass framework that coordinates noise allocation (Macro), regional angular-difficulty reweighting (Micro), and scale-calibrated direction correction. The comparisons in Table R2 of our General Response show that generic loss reweighting and standard cosine regularization do not match its performance, supporting the practical value of this coordination.

Q2. Component Contributions and Complementary Gains

Table R1 in our General Response provides the component ablations. Adding Micro to Macro increases GenEval from 0.6711 to 0.6822 and improves all three preference metrics. Removing Macro lowers performance relative to full RoQT, which achieves the highest scores across all five metrics. These comparisons support complementary gains from the components.

Q3. Perceptual Preference, Human Evaluation, and Robustness

We conducted a double-blind human preference evaluation on SD3.5-M using 500 prompts randomly sampled from Qwen-Image-Bench and GenEval. Image pairs used matched training steps; each pair was assessed by at least 5 of 10 raters under a forced-choice protocol (no ties). Across all pairwise evaluations, RoQT achieved 89.0% wins over Base and 71.4% over standard SFT, supporting the preference gains in Table R1. Generation-seed stability is reported in Table R6 (response to Vuzr, Q6).

Q4. Implementation Details and Reproducibility

We will include the following settings in the final version:

  1. Macro schedule: Over 3,000 training steps, the shift changes from 15→915 \to 9 at step 500 and 9→39 \to 3 at step 1,500.
  2. Regional mean: aˉb(t)\bar{a}_{b(t)} is computed online across GPU ranks per microbatch for t≥0.7t \ge 0.7, 0.3≤t<0.70.3 \le t < 0.7, and t<0.3t < 0.3.
  3. Weighting: η=0.5\eta=0.5, ϵ=10−5\epsilon=10^{-5}, and weights clipped to [0.5,1.5][0.5, 1.5].
  4. Inference: Official evaluation settings use 40 steps with CFG 4.5 for SD3.5-Medium and CFG 4.0 for Z-Image (seed 42).
Reviewer_VuzrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes roqt for rectified flow models. It has three parts: 1. a noise curriculum from high noise to low noise, 2. an angular-based sample weighting method, and 3. an extra direction loss. Experiments on sd3.5-m and z-Image show improvements on geneval and qwen-image-bench.

Critical strengths
  1. The method is simple and easy to apply to existing rectified flow models. It does not need extra model forward passes or change the original flow path.
  2. The results on two models show some consistent improvements over sft especially on z-Image.
Critical weaknesses
  1. The main gaol is to improve quality tuning by focusing more on high-noise steps and velocity directions. However, the experiments do not clearly support this idea. In Table 3(a), the Macro noise curriculum alone only achieves 0.6711 on GenEval, which is even lower than standard SFT (0.6814 in Table 1). The improvement of the full method may mainly come from the other loss terms, rather than the noise curriculum. Also, the noise-level analysis in Appendix B is not completed. This raises a serious concern about the main motivation and effectiveness of the proposed method.
  2. Lacks comparisons eg. MSE-based hard sample weighting, a standard cosine direction loss, and timestep-dependent loss weighting. It is unclear how much the proposed methods improves over these simple baselines.
Questions for authors
  1. In Figure 1, correcting the velocity at high noise gives better reconstruction. However, this correction also affects more sampling steps. It is unclear whether the gain comes from the high-noise region or simply from more corrected steps.
  2. Figure 2 shows large angular errors at both high and low noise levels. Why does the proposed curriculum mainly focus on high noise?
  3. Equation 5 only describes one-step reconstruction error. It does not directly explain the final generation quality after multiple sampling steps.
  4. The improvements on SD3.5-M are relatively small (0.6814 to 0.6943 on GenEval and 43.53 to 44.04 on Qwen-Image-Bench). Suggest to run multiple random seeds or error bars.
2 comments
AuthorsOpenReview ↗

We thank the reviewer for recognizing RoQT's simplicity, computational efficiency, and consistent empirical gains across backbones. We address your specific questions below.

Q1. Role of the Macro Noise Curriculum and Perceptual Metrics

GenEval measures compositional accuracy, while quality tuning also targets perceptual quality and human preference. In Table R1 of our General Response, Macro only improves PickScore (22.81) and ImageReward (0.8710) over Base and SFTover Base and SFT, while Micro only maintains a GenEval score close to standard SFT (0.6812 vs. 0.6814) Removing the Macro schedule (RoQT w/o Macro) drops overall performance relative to full RoQT,demonstrating that the Macro curriculum plays an essential, complementary role within the full framework.

Q2. Appendix B Noise-Curriculum Diagnostics

Table R4 tracks validation FM loss and angular error on SD3.5-M at 25%, 50%, and 100% of training. Both decrease across high-, mid-, and low-noise bands, shows that validation FM loss and angular error decline across all three noise bands as training progresses from 25% to 100%.

Table R4. Noise-curriculum validation diagnostics for RoQT on SD3.5-M.

ProgressNoise BandFM Loss (LFM\mathcal{L}_{\text{FM}}) ↓\downarrowAngular Error (1−cos⁡φt1-\cos\varphi_t) ↓\downarrow
25% (Step 750)High (t≥0.7t \ge 0.7)0.2150.076
Mid (0.3≤t<0.70.3 \le t < 0.7)0.1880.046
Low (t<0.3t < 0.3)0.2520.190
50% (Step 1500)High (t≥0.7t \ge 0.7)0.1980.058
Mid (0.3≤t<0.70.3 \le t < 0.7)0.1710.035
Low (t<0.3t < 0.3)0.2220.158
100% (Step 3000)High (t≥0.7t \ge 0.7)0.1850.045
Mid (0.3≤t<0.70.3 \le t < 0.7)0.1620.029
Low (t<0.3t < 0.3)0.1950.131

Q3. Comparisons with Simple Baselines

Please refer to Table R2 in our General Response for full comparative results. RoQT consistently outperforms standard SFT, offline hard-example weighting, and direct cosine regularization across Qwen-Image-Bench, PickScore, HPSv2, and ImageReward on Z-Image.

Q4. Correction Timing versus Number of Corrected Steps

Please refer to Table R3 in our General Response. The 50 ODE steps denote the full sampling trajectory. To control for correction count, we added runs with exactly one correction at each selected noise level. A single high-noise correction performs best and even exceeds 10 consecutive low-noise corrections (19.66 vs. 17.72 dB). Thus, the advantage cannot be explained solely by the larger number of corrections in the original Figure 1 experiment. These results support the importance of correction timing beyond the number of corrected updates.

Q5. High-Noise Emphasis vs. Endpoint Angular Errors

Figure 2(a) shows elevated angular errors near both endpoints. Eq. (5), ∥x0−x^0∥2=t2∥vt−u∥2\Vert{}x_0 - \hat{x}_0\Vert{}^2 = t^2 \Vert{}v_t - u\Vert{}^2, gives a local rationale for emphasizing high noise: for a fixed velocity error, the squared error of the instantaneous clean-sample estimate is weighted by t2t^2. The single-correction control in Table R3 provides complementary evidence at the final trajectory endpoint, where high-noise correction performs best.

We also tested an alternating bimodal high/low-noise curriculum (15→9→315 \to 9 \to 3 shifting down to 0.10.1) on SD3.5-M. Its scores (Table R5) fall below the Base results in Table R1, possibly reflecting a mismatch between the bimodal training distribution and the continuous inference schedule.

Table R5. High/low-noise curriculum comparison on SD3.5-M.

MethodGenEval ↑Qwen-Image-Bench ↑
RoQT with bimodal high/low shifts0.643039.10
RoQT (Ours)0.694344.04

Q6. Multi-Seed Robustness

Table R6 reports mean ± standard deviation across three generation seeds (42, 43, 44) on SD3.5-M. Both Base and RoQT show small variations in GenEval and Qwen-Image-Bench, supporting inference stability across these runs.

Table R6. Multi-seed evaluation performance comparison on SD3.5-M (mean ± std across seeds 42, 43, 44).

MethodGenEval ↑Qwen-Image-Bench ↑
Base0.6688 ± 0.002043.19 ± 0.06
RoQT (Ours)0.6943 ± 0.003044.04 ± 0.05

Dear Reviewer Vuzr,

Thank you again for your detailed comments. We have provided additional evidence on high-noise steps and velocity directions of Figure 1, appendix B Noise-Curriculum Diagnostics, more comparisons with simple baselines, additional semantic metric comparisons and improvements on SD3.5. When you have a chance, could you let us know whether our responses address your concerns? We would be happy to clarify any remaining points.

Best regards, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Clarification of RoQT’s Methodological Contribution

Thank you for revisiting our paper and acknowledging that the additional evidence addresses some concerns. We understand that methodological novelty remains your primary concern. We would like to clarify how the direction-error analysis informs RoQT’s formulation and how the additional experiments support this design.

1. The ablations suggest an interaction between noise allocation and the training objective.

The GenEval results in Table R1 can be reorganized as follows:

Training objectiveWithout MacroWith MacroMacro effect
Standard SFT objective0.68140.6711−0.0103
Micro weighting + direction correction0.68110.6943+0.0132

In the reported configurations, adding Macro lowers GenEval under standard SFT but raises it when paired with Micro weighting and direction correction. This reversal suggests that the benefit of noise allocation depends on the accompanying training objective, supporting the joint design of supervision allocation and directional optimization.

2. The controlled trajectory diagnostic provides complementary evidence on correction timing.

Table R3 compares correction locations using 50 ODE steps for every trajectory. Each intervention applies exactly one correction; the uncorrected trajectory serves as the baseline.

Correction locationCorrected updatesFinal PSNR (dB) ↑
No correction014.26
Low noise117.33
Mid noise118.58
High noise119.66

The high-noise intervention achieves the highest final PSNR among the tested locations, supporting the importance of correction timing at a fixed correction count in this oracle reconstruction setting. We use this diagnostic as empirical motivation for supervision allocation across noise levels; the training ablations in Table R1 evaluate the resulting combined method.

3. RoQT connects the analysis to a concrete joint formulation.

The rotational decomposition of flow-matching MSE motivates the scale-calibrated direction term. The controlled diagnostic and training ablations provide empirical support for coupling directional optimization with progressive noise allocation and regional angular weighting. The matched comparisons in Table R2 further evaluate this formulation against generic weighting and uncalibrated cosine alternatives.

We acknowledge the established precedents for the individual ingredients. Our contribution is this specific formulation and the empirical evidence for designing noise allocation and directional optimization jointly in quality tuning. We will make this contribution more explicit in the revised manuscript and would appreciate your reconsideration in light of this clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.