acceptodds
Under review as a conference paper at ICLR 2027

A Lightweight Velocity Correction Framework for Improving Instruction Following in Few-Step Text-to-Image Generation

Abstract

Few-step text-to-image generators produce high-quality images with few network forward passes, yet struggle with complex prompts involving multiple subjects, interactions, spatial relations, counting, and attribute binding. Can these errors be reduced without changing the generator's parameters or increasing its forward-pass count? We propose a lightweight correction framework for a frozen 6B-parameter generator. The generator predicts velocities that determine how its intermediate image representation is updated; a corrector with fewer than 20M parameters learns residual adjustments at selected sampling steps. Because earlier corrections change the states encountered at later steps, our On-Trajectory Flow-Bridge Supervision derives residual targets at the states actually reached during corrected generation. A single corrector is shared across selected steps and runs in parallel with the generator. Only the corrector is trained, requiring neither reward optimization nor re-distillation. To our knowledge, this is the first framework for lightweight velocity-residual adaptation of fully frozen few-step text-to-image generators that preserves their original sampling schedule and forward-pass count. On our target complex-scene distribution, the framework achieves a +17.60-percentage-point NetWin over frozen Z-Image-Turbo in pairwise instruction-following evaluation, with only 26.9 ms (0.84%) additional end-to-end latency per image and no increase in backbone evaluations.

Primary area:
generative models
Keywords:
text-to-image generation, few-step generative models, velocity residual correction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_pSchOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper freezes Z-Image-Turbo and trains a S3-DiT corrector to improve few step text-to-image generation. The corrector adds a residual to the backbone's velocity at selected steps. Its inputs are the current latent and previous step's cached velocity, so it can run alongside the backbone.

Critical strengths
  • Simple and deployable. The backbone stays untouched, the sampling schedule is unchanged, and overhead is small.
  • The state-target pairing ablation is informative. Corrected rollout training gives +17.6, Base-rollout gives -28.
  • Clear math. The eq6 split into reference velocity correction and trajectory recovery is clean, and easy to follow.
  • The step timing study (Fig 8) is useful and informative.
Critical weaknesses
  • How the 574 held out target domain prompts are produced is unclear. Please report i) how eval prompts were constructed, ii) similarity to training set, and how do you ensure out of distribution testing, iii) Fraction of evaluation prompts whose subjects/relation types appear in training as well, iv) results on split that holds out subjects and relation types completely.

  • The paper motivates the method with failures in spatial relations, counting, and multi-subject interactions (Abstract, Sec. 1), and the qualitative examples emphasize these. On T2I-CompBench++, however, the gains are concentrated in attribute binding (color, shape, texture), while 2D/3D spatial, numeracy, and non-spatial are essentially unchanged or slightly worse. 2D spatial is +0.001, 3D spatial is -0.005 and numeracy is +0.001. So these are mostly flat or negative. Can the authors break down target-domain NetWin by requirement type, to show which failure modes the corrector actually fixes?

  • The headline beats zero, but most ablation gaps are within noise: 17.6 vs 14.1 (current velocity), 14.5 ({4,5}), and 13.9 ({4})

  • D-OPSD is designed for exactly this setting, so it is surprising to see a -12.37 which suggests it does worse than doing nothing. Are the official code and hyper parameters used for this? If yes, can the authors explain why it is much worse than doing nothing ?

  • Also, DiffCon is implemented on all 8 steps, inspite of the authors finding later steps can be catastrophic wrt correction. How does diffcon do if it is also applied at the same steps as proposed method ?

  • The central motivation for keeping the generator frozen, that adapting a step-distilled generator degrades image quality under its original schedule (Sec. 1, Sec. 2), is not measured. Please report image-quality metrics (e.g., FID, aesthetic score, ImageReward) for all methods in Table 1, so that the claimed advantage over LoRA-based adaptation is supported directly.

  • Fig. 1 argues from the 50-step Z-Image vs. Turbo velocity gap, but the method never uses Z-Image. Using 50-step Z-Image outputs from the same ε as targets would directly test that motivation. Fig. 1c shows Z-Image at +19.5 vs. Turbo, so this is a natural ceiling.

  • Only one backbone is evaluated. The method should be verified on multiple backbones such as flux, Owen etc to confirm that the approach generalizes

Questions for authors

see weaknesses

8 comments

W1:

We thank the reviewer for the valuable suggestions. We clarify the prompt construction pipeline and the train–evaluation separation protocol as follows.

(i) Evaluation prompt construction.

Our prompt-construction pipeline shares the design principles of ConceptMix (Wu et al., NeurIPS 2024): sampling visual concepts, organizing their bindings into structured representations, generating natural-language prompts with an LLM, and validating the prompts for conflicts. Our implementation targets complex illustration scenes and includes explicit character, action, interaction, and scene specifications, together with consistency checking and atomic-requirement extraction. All compared methods are evaluated using the same resulting prompts and standardized formatting, providing a consistent basis for comparison.
Our prompts were generated through a four-stage, constraint-based pipeline:

  1. Scene composition: Structured scene specifications were constructed by combining predefined banks of characters, actions, interactions, environments, lighting, and composition attributes, with compatibility constraints and diversity controls.
  2. Language-based expansion: An LLM expanded each specification into a detailed Chinese description of a single static scene, covering visual content, subjects, spatial composition, and lighting.
  3. Consistency checking: Generated descriptions were checked for contradictions, incompatible actions or orientations, and temporal inconsistencies. Explicit, individually verifiable visual requirements were extracted, and invalid candidates were discarded.
  4. Prompt formatting: Validated descriptions and requirements were converted into standardized image-generation prompts, prioritizing core scene constraints and illustration-style specifications.

(ii–iii) Train–evaluation similarity and OOD guarantees.

Importantly, the training and evaluation sets were constructed using a type-level disjoint sampling protocol established before prompt generation. We first partitioned the predefined subject categories, character identities and aliases, and action/interaction relation types into non-overlapping training and evaluation pools.

All participating subjects—including primary characters, secondary characters, generic subjects, and background participants—were restricted to their respective pools. The same constraints also applied to auxiliary-character sampling and fallback action selection. The validation set used only training-side types.

Consequently, the evaluation set was designed to contain no overlap in predefined subject identities/types or action/interaction relation types with the training set, while retaining the same overall illustration domain and prompt-generation framework.

(iv) Results on fully held-out subjects and relations.

The original 574-prompt evaluation set already follows this joint subject-and-relation holdout protocol. Therefore, the reported target-domain results directly evaluate generalization to unseen subject and relation types, rather than requiring an additional split. Our Corrector achieves a NetWin of +17.25 percentage points over the frozen Z-Image-Turbo baseline on this set.

Due to data confidentiality restrictions, we cannot publicly release the underlying prompt collection. We plan to release the prompt-construction pipeline code, subject to the applicable disclosure restrictions, to facilitate reproducibility.

W2:

We sincerely thank the reviewer for highlighting this important concern. To identify which types of instruction-following errors our method actually improves, we conducted an additional requirement-level analysis of the 574 target-domain evaluation prompts.

Specifically, we reanalyzed the existing atomic-level requirement annotations and their evaluation outcomes, without regenerating images or rerunning the judge. For each requirement category within each prompt, we aggregated the corresponding atomic-level judgments into a single GSB outcome, following the original evaluation protocol that prioritizes core requirements over detailed requirements (core-then-detail). We then calculated category-specific NetWin scores and 95% confidence intervals using prompt-level bootstrap resampling.

Breakdown of Target-Domain NetWin by Requirement Type

Requirement TypePromptsWin / Tie / LossNetWin (pp)95% CI (pp)
Action / Pose / Orientation571160 / 315 / 96+11.21[+5.63, +16.43]
Position / Spatial Relations57467 / 477 / 30+6.45[+3.14, +9.76]
Character / Identity31423 / 283 / 8+4.78[+1.30, +8.23]
Counting / Numeracy20012 / 182 / 6+3.00[−1.01, +7.22]

Under the original evaluation protocol, action/pose/orientation and position/spatial relations exhibit the largest positive NetWin point estimates, with both confidence intervals excluding zero. Character-related requirements also show a positive improvement, while the gain in counting remains modest and statistically uncertain.

These findings clarify that our method does not uniformly improve all semantic attributes. Instead, its target-domain gains are more pronounced for action-related and spatial-relation requirements, consistent with our motivation to address complex structural instruction-following errors.

We also acknowledge that the limited improvements in spatial reasoning and numeracy on T2I-CompBench++ indicate that these benefits do not necessarily generalize uniformly across benchmarks. We will clarify the scope of these improvements and incorporate the requirement-level analysis into the revised paper.

W3:

We sincerely thank the reviewer for raising this important concern regarding the statistical reliability of our ablation results. To assess the impact of generation randomness, we repeated the evaluation using three different random seeds on the same 574 evaluation prompts and report the mean and standard deviation of GSB NetWin.

ConfigurationMean ± SD (pp)
Ours ({2,3,4})17.60 ± 0.45
Current velocity14.10 ± 0.51
Steps {4,5}14.50 ± 0.43
Step {4}13.90 ± 0.60

The results show that our full method maintains a higher mean NetWin than all ablated configurations, with improvements of 3.10–3.70 percentage points. These differences are substantially larger than the observed standard deviations (0.43–0.60 pp), providing additional evidence that the advantages of our selected configuration are consistent across different generation seeds, rather than being attributable solely to generation randomness.

W4:

We thank the reviewer for raising this important concern. To clarify the negative D-OPSD result, we provide details of our implementation, additional checkpoint evaluations, and a possible explanation supported by the original D-OPSD paper and related studies.

Implementation and checkpoint evaluation. We followed the authors' publicly released Z-Image-Turbo training script, using a learning rate of (1\times10^{-4}) and 2,000 optimizer steps, consistent with the released script. We additionally evaluated the checkpoints at steps 1,000 and 2,000. Both checkpoints yielded negative NetWin scores on our target-domain evaluation, with the final checkpoint achieving −12.37 percentage points. This indicates that the negative result is not solely attributable to the selection of the final checkpoint.

Possible explanation. Although D-OPSD is designed for adapting step-distilled models, its primary objective is to learn new concepts, styles, and domains while preserving few-step generation capability, rather than explicitly optimizing complex compositional instruction following. Its supervision aligns the text-conditioned student's velocity predictions with those of a multimodally conditioned teacher. Thus, its effectiveness depends on whether the teacher provides sufficiently informative supervision for the target task.

Importantly, the original D-OPSD paper (Jiang et al., NeurIPS 2026, Sec. 4, Fig. 7) explicitly identifies teacher capability as a critical limitation, showing that training may fail when the multimodally conditioned teacher cannot provide effective supervision. In contrast, our objective is to improve the model's own instruction-following capability, which may require supervisory signals beyond what the teacher itself can reliably provide.

Furthermore, its Table 2 reports that, under full fine-tuning for domain adaptation, D-OPSD improves FID and aesthetic quality on Z-Image-Turbo, while GenEval decreases from 0.7543 to 0.7170 and DPG decreases from 84.7645 to 84.1116. The authors discuss these changes as a trade-off associated with domain adaptation.

Related studies, including ClassDiffusion (Huang et al., ICLR 2025) and Preserve and Personalize (Kim et al., ICLR 2026), further demonstrate that fine-tuning for concept or subject fidelity can compromise compositional generation and prompt adherence. These findings suggest that improvements in adaptation-related objectives do not necessarily translate into better complex instruction following. Therefore, performance degradation on our task is consistent with the limitations and trade-offs documented in prior work.

W5:

We thank the reviewer for this insightful suggestion. To ensure a fairer comparison, we conducted an additional experiment restricting DiffCon's intervention steps to {2, 3, 4}, matching those of our proposed method.

MethodIntervention StepsGSB NetWin (pp)
DiffConAll 8 sampling steps+1.74
DiffCon{2, 3, 4}+4.70
Ours{2, 3, 4}+17.25

Restricting DiffCon to steps {2, 3, 4} increases its NetWin from +1.74 to +4.70 percentage points. This improvement suggests that early-step intervention also benefits DiffCon, supporting our motivation for restricting the correction steps.

Nevertheless, DiffCon's performance remains substantially below that of our method (+4.70 vs. +17.25 percentage points), even under the same intervention schedule. These results indicate that the performance advantage of our method cannot be explained solely by the selection of correction steps, suggesting additional benefits from our correction framework.

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

This paper proposes a lightweight velocity-residual corrector for improving instruction following in a frozen few-step text-to-image generator. An 18.6M-parameter network predicts corrections from the current latent, text conditioning, noise level, and cached previous-step backbone velocity, allowing concurrent execution with the backbone. Training recomputes residual targets on corrected sampler trajectories using a linear noise-to-target bridge, with offline edited images providing supervision. A shared corrector intervenes at steps 2–4 of eight-step Z-Image-Turbo. Experiments report +17.60 percentage points NetWin over the frozen model on 574 complex-scene prompts, improvements in seven T2I-CompBench++ categories, and 0.84% additional end-to-end latency.

Critical strengths

The method offers a practically attractive adaptation mechanism: an 18.6M-parameter corrector improves instruction following while preserving the frozen backbone’s eight-step schedule, with only 26.9 ms additional end-to-end latency under the reported setting. Conditioning on the previous-step velocity removes dependence on the current backbone output and enables concurrent execution. The controlled ablations are particularly useful: they separately examine rollout-state selection, state-dependent target recomputation, target decomposition, velocity conditioning, and intervention timing. These experiments support the importance of training on corrected trajectories and applying corrections at selected early steps.

Critical weaknesses

The main concerns are evaluation validity and the strength of the baseline comparisons. The headline +17.60 pp NetWin relies on an automated GSB protocol that is not calibrated against human judgments; repeated judging of fixed image pairs establishes repeatability, but not correctness or uncertainty across generated samples and training runs. The external benchmark results are encouraging, although several gains are numerically modest and lack confidence intervals. In addition, Appendix A.2 specifies baseline architectures and learning rates but does not fully report training duration, tuning budgets, or checkpoint-selection rules for most methods, limiting confidence that the comparisons reflect well-optimized alternatives. DiffCon also differs in supervision and intervention frequency, so its comparison does not cleanly isolate the contribution of the parallel correction design.

Questions for authors
  1. How well do the GSB judgments agree with blinded human assessments of instruction following and visual quality? Can you provide paired confidence intervals for NetWin that capture prompt and generation variability, rather than only repeated judging of fixed outputs?

  2. What training budgets, hyperparameter searches, and checkpoint-selection criteria were used for each baseline? In particular, how sensitive are the negative D-OPSD result and the small DiffCon gain to these choices?

  3. Were the active-step set {2,3,4} and other design choices selected exclusively using the validation split?

  4. What is the latency of the same trained corrector under concurrent versus sequential execution, keeping the active steps fixed?

3 comments

W1&Q1: We sincerely thank the reviewer for raising these important concerns regarding the reliability of automated GSB judgments and the statistical robustness of our evaluation results. To address these concerns, we conducted additional human evaluations of instruction following and visual quality, as well as a paired bootstrap analysis to quantify the uncertainty in NetWin. We report the detailed results below.

(1) Human Evaluation.

Due to time constraints, we randomly sampled 200 prompts from the final evaluation set of 574 prompts, using sampling without replacement with a fixed random seed of 42. Each image pair was independently evaluated by three human annotators, and the majority vote was used to determine the Human GSB outcome. The automated GSB evaluation yielded a NetWin of +18.50 percentage points on this subset.

For visual quality, human annotators preferred our method in 47 cases, the baseline in 34 cases, and rated 119 cases as ties, yielding a net preference of +6.50 percentage points.

For instruction following, our method was preferred in 51 cases, the baseline in 16 cases, and the two methods were rated as tied in 133 cases, yielding a net preference of +17.50 percentage points.

These results provide additional human evaluation evidence supporting the effectiveness of our method, particularly in improving instruction following.

(2) Statistical Significance and Confidence Interval.

For the full set of 574 evaluation prompts, the Corrector achieves 228 wins, 217 ties, and 129 losses, yielding a NetWin of +17.25 percentage points. A paired prompt-level bootstrap with 10,000 resamples yields a 95% confidence interval of [+10.98, +23.52] percentage points. Each resampling unit preserves the paired baseline–Corrector comparison for the same prompt, including ties. Since the confidence interval excludes zero, the results provide statistical evidence of a positive improvement over the baseline.

W2&Q2:We sincerely thank the reviewer for raising these important concerns regarding the fairness and reliability of our baseline comparisons. To address these concerns, we provide further clarification of the implementation details, training budgets, hyperparameter configurations, and checkpoint-selection criteria for each baseline. We also conducted additional sensitivity analyses for DiffCon and D-OPSD to investigate whether their reported performance could be affected by inference-time hyperparameters, training duration, or checkpoint selection. Our detailed responses are presented below.

(1)Baseline Implementation, Training Budgets, and Checkpoint Selection.

We clarify the implementation sources, training configurations, and checkpoint-selection criteria for the reproduced baselines.

SFT+LoRA. We followed the Z-Image-Turbo training recipe from DiffSynth-Studio, adopting its learning rate of 1×10−41\times10^{-4}. We trained for 2,000 optimizer steps and evaluated the final checkpoint at step 2,000, matching the training budget and checkpoint-selection criterion of our method.

DPO+LoRA. We adapted the Hugging Face Diffusers Diffusion-DPO implementation to Z-Image-Turbo, retaining its reference learning rate of 1×10−51\times10^{-5}. We trained for 2,000 optimizer steps and evaluated the final checkpoint at step 2,000, matching the training budget and checkpoint-selection criterion of SFT+LoRA.

D-OPSD. We followed the publicly released Z-Image-Turbo training script, using a learning rate of 1×10−41\times10^{-4} and 2,000 optimizer steps. We evaluated the final checkpoint at step 2,000.

DiffCon. As no official implementation was publicly available to us, we independently adapted the method to Z-Image-Turbo following the formulations and implementation details in the original paper. We adopted its SFT-setting learning rate of 1×10−51\times10^{-5} and 1,000-step training budget, evaluating the final checkpoint at step 1,000. We plan to release our adapted implementation to facilitate reproducibility.

(2)Sensitivity to Baseline Configurations.

We did not conduct an exhaustive hyperparameter search for every baseline. To address the reviewer's concern regarding the negative D-OPSD result and the limited DiffCon improvement, we provide the following analysis.

DiffCon. Following the original DiffCon paper (Yang et al., ICML 2026), we additionally evaluated the inference-time guidance strength (\lambda_{\mathrm{model}}) over the reported range (1/\lambda_{\mathrm{model}}\in{0.1,0.2,\ldots,1.0}), while keeping the trained SFT DiffCon model and all other settings unchanged. Our results confirm that the original setting ((\lambda_{\mathrm{model}}=1)) achieves the best NetWin, whereas stronger guidance substantially degrades performance (e.g., −55.75 pp at (\lambda_{\mathrm{model}}=2) and −99.48 pp at (\lambda_{\mathrm{model}}=5)). Unlike the improvements reported on Stable Diffusion v1.4, stronger guidance is detrimental in our few-step Z-Image-Turbo setting. These findings indicate that the limited performance gain of DiffCon cannot be attributed to insufficient inference-time guidance tuning within the range explored in the original paper.

D-OPSD. To examine sensitivity to training duration and checkpoint selection, we additionally evaluated the checkpoints at steps 1,000 and 2,000, corresponding to the training budgets reported in the original paper and the publicly released Z-Image-Turbo script, respectively. Both checkpoints yielded negative performance gains on our target-domain evaluation, indicating that the negative result is not solely attributable to selecting the final checkpoint at step 2,000.

We believe this result may partly reflect a mismatch between D-OPSD's supervision mechanism and our target task. D-OPSD primarily focuses on adapting step-distilled models to new concepts, styles, and domains while preserving few-step generation capability. Its training aligns the text-conditioned student's velocity predictions with those of a multimodally conditioned teacher receiving both the prompt and target image. However, this objective does not explicitly optimize the correction of complex compositional instruction-following errors.

Consequently, improvements in our setting depend on whether the multimodally conditioned teacher provides sufficiently informative supervision for spatial relations, counting, and multi-subject interactions. The original D-OPSD paper also highlights this limitation in Section 4 (Requirements for Teacher Capability), noting that training may fail when the teacher cannot provide effective supervision even with multimodal conditioning.

Q3:We thank the reviewer for raising this important question regarding our experimental protocol. Yes, the active-step set {2, 3, 4} was selected exclusively using the validation split, and the test split was never used for hyperparameter tuning or design selection. We will clarify this protocol in the revised paper and release our code to facilitate reproducibility.

Regarding the correction steps, our analysis of z-image-turbo motivated applying corrections at early steps: these steps have a substantial influence on instruction following and the formation of the overall image structure, while few-step models have limited capacity for further denoising and adjustment in subsequent steps. As shown in Figures 2 and 6, some action-related contours already change at early steps when the corrector is introduced, while later steps primarily refine the image through denoising. Consequently, applying corrections at later steps may have limited effectiveness in adjusting established structures and may adversely affect final image quality. Guided by this motivation, we compared candidate step combinations using only the validation split and ultimately selected {2,3,4}.

Other design choices, such as On-Trajectory Flow-Bridge Supervision, were motivated by an analysis of the correction mechanism: corrections alter the model’s current state, potentially creating a mismatch between supervision targets constructed from states along the ideal trajectory and the corrected states actually visited by the model. We therefore introduced this design to mitigate that supervision mismatch.

W3&Q4:

We thank the reviewer for this insightful comment. To isolate the efficiency benefit of concurrent execution, we benchmarked the same trained corrector under sequential and concurrent execution, keeping the active steps fixed at {2, 3, 4}.

On an NVIDIA RTX PRO 6000 Blackwell GPU, concurrent and sequential execution introduced additional inference latencies of 9.86 ms (+0.39%) and 15.86 ms (+0.63%), respectively, compared with the Z-Image-Turbo baseline (2516.89 ms/image). Concurrent execution thus reduced the incremental overhead by 37.8% (6.00 ms/image) relative to sequential execution.

Furthermore, to enable a fairer comparison with DiffCon, we restricted its intervention steps to {2, 3, 4}, matching our method's correction frequency. Under this setting, DiffCon introduced an additional inference latency of 36.49 ms/image (+1.45% relative to Z-Image-Turbo baseline). This overhead is substantially higher than that of our lightweight corrector, even under sequential execution, provided the measurements are made under comparable hardware and inference settings.

These results demonstrate that our corrector remains computationally lightweight even without concurrent execution, while concurrency further reduces its inference overhead. The controlled sequential-versus-concurrent comparison directly isolates the efficiency benefit of our concurrent execution design.

Reviewer_gw8gOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a lightweight velocity corrector for improving instruction following in few-step text-to-image generation. The corrector operates alongside a frozen Z-Image-Turbo backbone, with corrections restricted to selected early sampling steps. The authors introduce On-Trajectory Flow-Bridge Supervision to construct state-dependent correction targets. The method achieves improvements in instruction following with minimal inference overhead.

Critical strengths
  1. The method is computationally efficient, preserving the original sampling schedule and backbone forward-pass count.
  2. The ablations provide meaningful evidence for the proposed supervision strategy, particularly the importance of corrected rollouts and dynamically recomputed targets.
  3. The authors provide useful insights into correction timing, showing that early-step intervention is effective while late-step corrections can substantially degrade performance.
Critical weaknesses

My main concern is that the experiments do not sufficiently isolate the contribution of the corrector architecture itself.

The proposed method combines three components: a lightweight sidecar corrector, On-Trajectory Flow-Bridge Supervision, and correction restricted to early sampling steps. While the authors provide convincing ablations for the latter two components, the advantage of the sidecar architecture over direct backbone adaptation remains unclear.

LoRA baselines are applied across a broad noise schedule or the whole on-policy trajectory, whereas the proposed method modifies only steps 2, 3, and 4. Fine-tuning distilled models can disrupt their learned few-step sampling behavior. Restricting LoRA adaptation to early steps could potentially alleviate this issue by leaving subsequent denoising steps unchanged.

Therefore, the performance gap against the existing LoRA baselines does not establish whether the improvement comes from the corrector architecture or from the more suitable supervision and intervention strategy.

One potential advantage of the corrector is its ability to incorporate previous-step velocity as an additional input. However, this does not fully establish whether the corrector is effective.

Questions for authors

Could the authors provide a step-restricted LoRA baseline, trained and applied only at steps 2, 3, and 4?

Ideally, this comparison would include LoRA trained with the same on-trajectory Flow-Bridge supervision and a comparable parameter budget. This would help isolate the architectural contribution from the effects of supervision and intervention timing.

2 comments

We thank the reviewer for the constructive feedback. To examine the contributions of the corrector architecture and the supervision strategy separately, we conducted two controlled LoRA experiments. Both LoRA baselines have 19.85M trainable parameters and are trained and applied only at sampling steps {2, 3, 4}, using a trainable parameter budget comparable to our corrector.

W1&Q1 — Contribution of the corrector architecture under matched supervision and intervention timing.

We compare our corrector with a step-restricted LoRA baseline trained using the same On-Trajectory Flow-Bridge supervision. The results are summarized below.

MethodSupervisionTraining / Intervention StepsGSB NetWin (pp)
Step-restricted LoRAOn-Trajectory Flow-Bridge{2, 3, 4} / {2, 3, 4}+12.38
OursOn-Trajectory Flow-Bridge{2, 3, 4} / {2, 3, 4}+17.25

Under matched supervision and intervention timing, our corrector achieves a 4.87 percentage-point higher NetWin than LoRA. This comparison supports an additional benefit from the corrector design beyond the supervision strategy and the choice of intervention steps.

Q2 — Contribution of On-Trajectory Flow-Bridge supervision.

We also compare the two step-restricted LoRA baselines while keeping their parameter budget and active steps unchanged.

MethodTrainable ParametersSupervisionTraining / Intervention StepsGSB NetWin (pp)
Step-restricted LoRA19.85MOriginal supervision{2, 3, 4} / {2, 3, 4}+2.79
Step-restricted LoRA19.85MOn-Trajectory Flow-Bridge{2, 3, 4} / {2, 3, 4}+12.38

Applying On-Trajectory Flow-Bridge supervision increases LoRA's NetWin by 9.59 percentage points, demonstrating that the supervision strategy is effective beyond the sidecar architecture.

Together, these controlled comparisons show that On-Trajectory Flow-Bridge supervision provides a substantial improvement, while the corrector architecture delivers an additional gain under matched supervision and intervention timing. We will clarify these contributions in the revised paper.

Thanks for the additional experiments! The matched LoRA comparison helps address my main concern. I will raise my score to Weak Accept.

Reviewer_ZJq3OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a lightweight velocity correction framework for improving instruction following in few-step text-to-image generation. It introduces a small parallel corrector trained with on-trajectory flow-bridge supervision while keeping the original generator frozen. Experiments demonstrate improved compositional instruction following with minimal inference overhead.

Critical strengths

The method is well-motivated and technically interesting, improving instruction following without modifying the pretrained generator or increasing its sampling steps.

The experimental results are promising, with substantial improvements over the baseline and minimal additional computational overhead.

Critical weaknesses

My main concern is the insufficient direct visual evidence of how the proposed velocity corrections affect the actual generation trajectory.

Although Figure 6 provides a step-wise qualitative example, a single example is insufficient to convincingly demonstrate that corrections at the selected steps consistently produce the intended improvements. More extensive visualizations of actual intermediate outputs are needed to establish the effectiveness and reliability of the proposed correction mechanism.

Questions for authors

Could the authors provide additional visualizations of the actual intermediate generation outputs, particularly before and after the correction steps (2, 3, and 4), across multiple prompts and random seeds?

It would be particularly helpful to show how the corrections change the intermediate representations and how these changes translate into improved instruction following in the final images.

1 comment

We sincerely thank the reviewer for highlighting the need for more extensive intermediate generation visualizations. To address this concern, we provide six additional visualization cases (Cases 1–6), covering three prompts evaluated at two random seeds each. These visualizations are available in the anonymized supplementary repository.

The examples illustrate improvements in following instructions about actions, poses, and spatial relations. For each case, we visualize the intermediate generation outputs of the frozen baseline and our corrector using the same initial noise, with particular attention to correction steps {2, 3, 4} and their effects on the final Step-8 outputs.

These examples provide qualitative evidence that early-step velocity corrections can influence subsequent generation trajectories and improve instruction following.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.