acceptodds
Under review as a conference paper at ICLR 2027

Teacher as Reference: Learning Real-Data Corrections for Few-Step Distillation

Abstract

Few-step distillation makes text-to-image generation efficient, but teacher matching alone does not specify how real observations should correct student predictions. We introduce Reference-Informed Non-adversarial Few-step Distillation (RINFD), which injects real-data information through two distinct training signals. Teacher-to-Data correction (T2D) learns teacher-referenced displacements from reconstruction–observation pairs on noisy bridge queries, then supplies prediction-level corrections at re-noised student outputs. Frequency-sensitive sliced Wasserstein distillation (FD-SWD) provides distribution-level supervision by aligning local, multi-scale statistics of predicted clean endpoints with real latents. Both signals act on the same clean prediction alongside teacher-guided distillation, without a discriminator. Controlled SDXL experiments support their complementarity: the default combination exceeds all tested FD-SWD-only weights on PickScore and ImageReward, and bridge sampling improves on reconstruction-only correction training. Weight sweeps reveal a balance between preference, color fidelity, and composition, with no single setting leading every metric. Across SDXL, SD3.5 Large, and FLUX.1-dev, the complete framework achieves the best average rank on two backbones and the second-best on the third among the evaluated four-step students, with the highest GenEval on both flow backbones.

Primary area:
generative models
Keywords:
Few-step Distillation, Text-to-Image, Non-adversarial Framework

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_WJ1gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RINFD, a non-adversarial framework for adding real-data supervision to few-step text-to-image distillation. First, Teacher-to-Data correction (T2D) constructs displacement targets from teacher reconstructions of real images, trains a correction adapter on noisy points sampled along bridges between the reconstruction and the real latent, and injects the predicted correction into student updates. Second, FD-SWD matches multi-scale spatial and frequency-sensitive patch statistics between student clean predictions and real latents using sliced Wasserstein distance. Combined with DMD and ISG, RINFD obtains the best reported average rank on SDXL and FLUX.1-dev and the second-best on SD3.5 Large among the evaluated four-step systems.

Critical strengths
  1. The paper asks a relevant question: how to use real images in DMD-style distillation without relying on a discriminator. The separation between prediction-level correction and distribution-level statistical matching is intuitive.
  2. The SDXL ablations are more informative than a simple full-versus-base comparison. In particular, the paper examines plain versus frequency-sensitive SWD, the two components individually, several loss weights, and bridge sampling versus reconstruction-only queries.
Critical weaknesses
  1. Although RINFD introduces no inference-time overhead, training requires an online generator-field estimator, a T2D correction adapter, and FD-SWD, with alternating generator, field-regression, and real-pair correction updates. T2D further requires teacher reconstructions, bridge queries, and additional frozen-network evaluations. The resulting wall-clock time, memory footprint, and number of teacher calls may therefore be substantially higher than those of the DMD/ISG baseline. The paper reports the hardware and batch size but does not provide training time, peak memory, or a compute-matched comparison with DMD, SenseFlow, or SwD, making the practical cost difficult to assess.
  2. The main component ablations are conducted only on SDXL and are reported as single-run point estimates. Relative to FD-SWD alone, the full method improves PickScore, ImageReward, and CFM, but decreases HPSv2 and GenEval. Thus, the current results may reflect a shift to a different metric operating point rather than a consistent synergistic improvement.
  3. In the first example of Figure 7, the prompt asks for “a dog with a disc on the beach.” Compared with SwD, the RINFD outputs on both SD3.5 Large and FLUX.1-dev place the disc in visually implausible positions. Its relation to the dog’s mouth or body is unclear, and the FLUX result resembles a ring attached around the dog’s neck rather than a disc being carried or placed naturally in the scene. This example suggests that RINFD may improve aggregate preference metrics while still producing locally plausible but spatially inconsistent objects.
Questions for authors

see weaknesses

4 comments

We thank the reviewer for recognizing the relevance of non-adversarial real-data supervision and the informative SDXL ablations. We address the concerns about computational cost, metric trade-offs, and spatial consistency below.

1、Training cost and comparison fairness

MethodOverhead (H20 GPU hours)
Base (DMD + ISG)45–55
SenseFlow90–100
RINFD60–70

Having no additional inference-time modules does not imply negligible training overhead. The generator-field estimator is already part of the DMD/ISG baseline, while T2D and FD-SWD introduce additional training computation. As shown in the Table above, our training-cost comparison shows that Base (DMD + ISG) requires approximately 45–55 GPU-hours, SenseFlow requires 90–100 GPU-hours, and full RINFD requires 60–70 GPU-hours. Using the midpoints of these ranges, RINFD requires approximately 30% more training GPU-hours than Base and 32% fewer than SenseFlow. We consider the additional training cost over Base worthwhile for performance development. As shown in Table 2 in our manuscript, RINFD improves ImageReward from 0.4273 to 0.6517 (+52.5%), PickScore from 21.27 to 21.77, and HPSv2 from 0.2800 to 0.2916. CLIP and CFM also increase from 0.3297 to 0.3336 and from 13.01 to 13.15, respectively. The additional training computation yields gains across five of the six reported metrics, with the largest relative improvement in ImageReward. This cost is incurred during training, while the deployed student retains the same four-NFE inference procedure without additional inference overhead.

2、Additional component ablations on FLUX.1 dev

We agree that the SDXL results do not establish a consistent improvement. Relative to FD-SWD alone, the full method improves PickScore, ImageReward, and CFM but reduces HPSv2 and GenEval. Our intended claim is that T2D can provide additional preference gains when combined with FD-SWD, with trade-offs that depend on the backbone. To examine whether this behavior extends beyond SDXL, we have now conducted a component ablation on FLUX.1-dev (see our response to reviewer cAUz). Adding T2D improves HPSv2 from 0.2907 to 0.2939, PickScore from 21.87 to 21.93, and ImageReward from 0.9562 to 0.9935. GenEval increases from 0.7010 to 0.7024, while CFM remains unchanged. Thus, the FLUX estimates show additional preference gains without the HPSv2 or GenEval decreases observed on SDXL. This additional ablation provides evidence of complementary benefits on another backbone. We will narrow our claim in the revised manuscript accordingly and explicitly acknowledge both the SDXL trade-offs.

3、Spatial inconsistency in Figure 7

The first example in Figure 7 reveals a failure in object relationships. The dog-and-disc example highlights an ambiguity in the spatial relationship between the objects. Similar ambiguity is also visible in the SwD and SenseFlow outputs on SD3.5 Large, where the disc appears to overlap the dog’s mouth–neck region. This issue is not unique to RINFD. RINFD does not include an explicit objective for aligning or constraining spatial relationships between objects and therefore cannot guarantee relational correctness in individual outputs. Our reported improvements in aggregate preference metrics should be interpreted within this scope: they do not imply uniformly improved spatial consistency or object interactions. We will explicitly discuss this limitation in the revised manuscript, using this example to distinguish visual appearance quality from relational correctness.

AuthorsOpenReview ↗

Further Discussion with the Reviewer

Dear Reviewer WJ1g,

Thank you for your time and constructive feedback. We have carefully considered your comments and provided detailed responses in our previous official comment. We summarize the main points below:

  1. We quantified the training cost, reporting approximately 30% more GPU-hours than Base and 32% fewer than SenseFlow. We also clarified that the additional computation is confined to training, with no additional inference overhead..

  2. We provided a new component-ablation results on FLUX.1-dev, showing additional preference gains from combining T2D with FD-SWD, with GenEval essentially maintained. We also acknowledged the SDXL trade-offs and will narrow our claims to reflect these backbone-dependent results.

  3. We discussed the spatial inconsistency in Figure 7, acknowledging that RINFD does not explicitly constrain object relationships. We will clarify this limitation and distinguish improvements in aggregate preference metrics from relational correctness in individual outputs.

We hope these clarifications and new results help address your concerns. We would appreciate hearing whether any points remain unresolved and would be happy to provide further clarification.

Thank you again for your thoughtful feedback and for helping us improve our work.

Best regards,

The Authors

Thank you for the authors’ response. Nevertheless, my concern has not been fully addressed. As also noted by other reviewers, T2D itself does not appear to provide a sufficiently substantial improvement. Although the newly added FLUX experiments show that T2D can improve HPSv2 and PickScore, the gains on these metrics appear smaller than those achieved by FD-SWD in Table 2. Moreover, given the motivation behind the Teacher-Constructed Correction loss, I would expect its introduction to improve the perceptual realism of the generated images. However, based on the qualitative examples presented in the paper, I do not observe a clear improvement in realism compared with SwD-FLUX. Therefore, I am inclined to maintain my current score.

Dear Reviewer WJ1g,

Thank you for your follow-up and for clarifying your remaining concerns. Your comments helped us recognize that our previous response focused too narrowly on metric improvements and did not sufficiently clarify the central contribution of our work.

1. Our main contribution: non-adversarial real-data supervision for few-step distillation

Our work investigates how real-data supervision can be incorporated into few-step distillation without introducing an adversarial discriminator. We implement this through two complementary forms of supervision: FD-SWD matches local statistics, while T2D learns a teacher-to-data correction. Within this framework, T2D provides a learned correction toward real-data targets beyond statistical matching. Our contribution is this approach to incorporating real-data information, rather than a claim that T2D produces a large improvement on all metric.

2. What the additional ablation establishes

We agree that FD-SWD contributes more to the FLUX PickScore and ImageReward gains. Nevertheless, T2D provides further improvements when added to FD-SWD: HPSv2 increases from 0.2907 to 0.2939, PickScore from 21.87 to 21.93, and ImageReward from 0.9562 to 0.9935, while GenEval is essentially maintained and CFM remains unchanged. For HPSv2, the additional gain from T2D (+0.0032) is larger than the gain from adding FD-SWD to Base (+0.0017). These results support a complementary contribution, although its magnitude varies across metrics and backbones. We will make this scope explicit in the revised manuscript.

3. The expectation of improved perceptual realism

We understand your expectation that real-data correction should improve perceptual realism. However, learning a conditional-mean correction does not directly optimize human judgments of realism, and preference-score improvements alone cannot establish such an improvement. We acknowledge that the presented qualitative examples do not clearly demonstrate a realism advantage over SwD-FLUX. Accordingly, we will distinguish our methodological contribution—incorporating real-data supervision without adversarial training—from the claim of visibly superior realism. We will clarify the observed benefits, trade-offs, and limitations in the revised manuscript.

Thank you again for helping us clarify both the contribution and the evidence supporting it.

Best regards,

The Authors

Reviewer_Q4B2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a Reference-Informed Non-adversarial Few-step Distillation method. T2D learns displacements between teacher reconstructions and their corresponding real observations using noisy bridge queries, and subsequently applies the learned corrections to student predictions. FD-SWD matches multi-scale local statistics of predicted clean latents and real latents. RINFD achieves the best average metric rank on SDXL and FLUX.1-dev, and the highest GenEval scores among the evaluated students on SD3.5 Large and FLUX.1-dev.

Critical strengths
  1. T2D provides a concrete way to incorporate real-data feedback without introducing an adversarial loss, which is an appealing design.
  2. The experiments provide empirical support for the method through both controlled ablations and evaluation across multiple backbones.
Critical weaknesses
  1. Evidence for correction transfer remains indirect. Appendix A.6 demonstrates improved displacement prediction at low-noise interior and real-endpoint queries. However, at the reconstruction endpoint (a=0a=0), the bridge-trained adapter has NMSE above the zero-correction baseline and near-zero cosine similarity at all tested query timesteps. The student-query RMS measurements characterize response magnitude rather than directional accuracy or usefulness, as the authors acknowledge. Thus, while the downstream ablations support the complete T2D procedure, they do not directly establish that the learned reconstruction–observation corrections remain useful when queried on student predictions.

  2. Attribution of the T2D gains. As discussed in Appendix A.5, the applied correction combines a learned increment with a guidance-reference difference. The current component ablation evaluates the combined procedure but does not isolate the contribution of learning from real-data pairs. A guidance-reference-only control would strengthen the attribution of the observed gains to the learned correction.

Questions for authors
  1. Training schedule: Since the correction adapter’s objective depends on real data and a frozen teacher rather than the evolving student, why update it throughout distillation instead of pretraining and freezing it?

  2. Correction scale: What value or schedule is used for γγ in Eq. (10)?

  3. Effectiveness of T2D: In Table 2, FD-SWD alone already provides substantial improvements. Adding T2D improves PickScore and ImageReward but reduces HPSv2 and GenEval. This raises concerns about the overall effectiveness of T2D. See Weaknesses 1 and 2 for further discussion.

2 comments

We thank the reviewer for the careful and constructive questions regarding (1) the training schedule, (2) the correction scale (γ), and (3) the effectiveness of T2D. We address them below.

1、The effectiveness of training schedule.

MethodNFECLIP ↑HPSv2 ↑PickScore ↑ImageReward ↑CFM ↑GenEval ↑
Base40.32970.280021.270.427313.010.5717
Pretrain & Frozen40.33720.2891‾\underline{\text{0.2891}}21.64‾\underline{\text{21.64}}0.5563‾\underline{\text{0.5563}}13.150.5644‾\underline{\text{0.5644}}
Joint Training40.3336‾\underline{\text{0.3336}}0.291621.770.651713.150.5551

For the question of whether the pretrained components need to be jointly optimized. To address this point, we compare three settings under the same four-NFE protocol: the Base model, a Pretrain-and-Frozen variant, and the full Joint-Training variant. Pretrain-and-Frozen already improves substantially over Base on five of the six metrics. In particular, ImageReward increases from 0.4273 to 0.5563 (+30.2%), while CLIP, HPSv2, PickScore, and CFM also improve. This confirms that the gains of RINFD do not arise solely from jointly updating the pretrained components. Joint Training provides additional preference-oriented gains beyond the frozen variant. HPSv2 increases from 0.2891 to 0.2916, PickScore from 21.64 to 21.77, and ImageReward from 0.5563 to 0.6517. CFM remains unchanged. These gains come with a trade-off: CLIP decreases from 0.3372 to 0.3336 and GenEval from 0.5644 to 0.5551. Thus, Joint Training is not uniformly superior; it is beneficial primarily for the preference metrics targeted by our method, whereas Pretrain-and-Frozen performs better on CLIP and GenEval. We will include this ablation in the revised manuscript and clarify that freezing the pretrained components is already an effective configuration, while joint training offers additional preference gains at the cost of a different metric trade-off.

2、Correction scale: What value or schedule is used for in Eq. (10)?

We use a predefined annealing schedule for the correction scale γ in Eq. (10), gradually increasing it from 0 to 1 as training progresses. This schedule progressively introduces the bridge correction signal into the training of the few-step generator. We will clarify this schedule in the revised manuscript.

3、Effectiveness of T2D: new FLUX.1-dev evidence

We agree that the SDXL results involve a preference–composition trade-off. We have now conducted a component ablation on FLUX.1-dev (refer to the response to reviewer cAUz). Relative to FD-SWD alone, adding T2D changes: HPSv2: 0.2907 → 0.2939; PickScore: 21.87 → 21.93; ImageReward: 0.9562 → 0.9935; GenEval: 0.7010 → 0.7024; CFM: 12.91 → 12.91. Thus, additional preference gains are also observed on FLUX, without the HPSv2 or GenEval degradation observed on SDXL. These point estimates support the usefulness of the combined procedure in another backbone.

AuthorsOpenReview ↗

Further Discussion with the Reviewer

Dear Reviewer Q4B2,

Thank you for your time and constructive feedback. We have carefully considered your comments and provided detailed responses in our previous official comment. We summarize the main points below:

  1. We discussed the performance trade-offs on SD3.5 Large and SDXL, including the counting deficit, and outlined possible explanations while acknowledging the limitations of these interpretations.

  2. We conducted a training-schedule ablation, comparing Base, Pretrain-and-Frozen, and Joint Training. The results show that freezing the pretrained components already provides substantial improvements, while joint training offers additional preference gains with trade-offs in CLIP and GenEval.

  3. We provided a new component-ablation results on FLUX.1-dev, showing additional preference gains from adding T2D to FD-SWD, with GenEval essentially maintained and CFM unchanged. These results extend the evidence for T2D’s usefulness to another backbone.

We hope these clarifications and new results help address your concerns. We would appreciate hearing whether any points remain unresolved and would be happy to provide further clarification.

Thank you again for your thoughtful feedback and for helping us improve our work.

Best regards,

The Authors

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

The paper adds two non-adversarial real-data supervision signals on top of DMD /ISG few-step distillation.

T2D: the frozen teacher reconstructs real latents, which gives a displacement Delta = z_r - z_T. A correction adapter learns this displacement on bridge queries between z_T and z_r, and the learned correction is injected into the student clean prediction during training.

FD-SWD: generated and real endpoint patches are matched with multi-scale patch statistics using sliced Wasserstein distance.

Critical strengths

The motivation is reasonable. Real-data supervision without a GAN is attractive, and keeping all auxiliary branches training-only is a good design.

Experiments cover three backbones: SDXL, SD3.5 Large, and FLUX.1-dev.

Critical weaknesses
  • Performance gains show a backbone/metric-dependent trade-off and lack mechanistic explanation: RINFD has the best average rank on SDXL and FLUX.1-dev, but is lower than SwD on SD3.5 Large; on SDXL, GenEval and counting are lower than DMD2/SwD, and the advantages are mainly concentrated in some preference and color metrics. The paper only identifies the counting gap on SDXL, does not explain why it underperforms SwD on SD3.5, and does not systematically analyze why effects differ across backbones (diffusion vs. flow-based).
  • Additional branches and forward passes are introduced during training, but there is no training-cost analysis. Although there is no extra overhead at inference, multiple auxiliary branches and extra forward computation are added during training. The paper gives training configurations and steps but does not report actual training time, GPU hours, or memory increase.
  • Ablations of the core mechanism are completed only on SDXL; SD3.5 Large and FLUX are evaluated only with the full framework. Therefore, whether conclusions such as “T2D and FD-SWD are complementary” and “bridge queries outperform endpoint training” hold on flow backbones has not been verified.
Questions for authors

Can you explain why the average rank on SD3.5 Large is lower than SwD, and why counting on SDXL lags significantly behind? Is there a mechanism-level analysis? Can you perform ablation on at least one flow backbone to verify conclusions across backbones? Can you supplement actual training time, GPU hours, or memory increase so that the cost of the additional training branches can be evaluated?

2 comments

We thank the reviewer for recognizing the motivation of non-adversarial real-data supervision and our evaluation across three backbones. We address the questions on cross-backbone ablations, performance trade-offs, and training cost below.

1、SD3.5 ranking and the SDXL counting gap

SwD and RINFD impose different forms of supervision. SwD progressively increases the spatial resolution during generation and applies patch-level MMD in intermediate teacher features, whereas RINFD operates on full-resolution clean endpoints and combines local statistical matching with a teacher-to-data conditional-mean correction. This difference is consistent with the observed metric trade-off in Table 1. On SD3.5 Large, RINFD is better on CFM (13.16 vs. 13.06) and GenEval (0.7336 vs. 0.7215), while SwD is better on CLIP, HPSv2, PickScore, and ImageReward. A plausible interpretation is that SwD’s coarse-to-fine generation and intermediate feature matching better preserve structural details, while RINFD’s real-data correction and endpoint statistics improve color and local fidelity. This remains a hypothesis because the two methods also differ in their training data. The same limitation is visible on SDXL. RINFD is better than SwD on two-object composition (78.54% vs. 72.73%), but substantially worse on counting (37.50% vs. 52.50%). We attribute it to T2D learns a conditional-mean displacement, and FD-SWD matches local statistics; neither explicitly enforces object cardinality or relational structure. Conditional-mean regression may also average ambiguous details. Thus, the counting deficit appears category-specific rather than a uniform composition failure. We will add this mechanism-based discussion in the revised manuscript.

2、New component ablation on FLUX.1-dev

Following the reviewer’s suggestion, we conducted a component ablation on FLUX.1-dev at four NFE, the results are as follows.

MethodNFECLIP ↑HPSv2 ↑PickScore ↑ImageReward ↑CFM ↑GenEval ↑
Base (DMD + ISG)40.33570.289021.410.812812.360.6731
Base + FD-SWD40.33200.290721.870.9652‾\underline{\text{0.9652}}12.910.7010‾\underline{\text{0.7010}}
Base + T2D40.32960.2926‾\underline{\text{0.2926}}21.87‾\underline{\text{21.87}}0.880912.990.6797
Full ((\lambda_S=0.4,=0.4, \lambda_T=0.5)=0.5)40.3324‾\underline{\text{0.3324}}0.293921.930.993512.91‾\underline{\text{12.91}}0.7024

Both components individually improve HPSv2, PickScore, ImageReward, CFM, and GenEval over Base, while reducing CLIP. More importantly, adding T2D to FD-SWD increases HPSv2 from 0.2907 to 0.2939, PickScore from 21.87 to 21.93, and ImageReward from 0.9562 to 0.9935. CFM remains nearly unchanged, and GenEval is essentially maintained, changing from 0.7010 to 0.7024. These results extend the evidence for complementary preference gains to a flow backbone. They also show that the SDXL compositional trade-off is not reproduced in this FLUX setting. The new experiment tests component contributions beyond diffusion backbone.

3、Training cost

We agree that training cost should be quantified separately. For the reported SDXL runs on four NVIDIA H20 GPUs, Base (DMD + ISG) requires approximately 45–55 GPU-hours, whereas the full RINFD framework requires 60 - 70 GPU-hours. Based on the midpoints of these ranges, this corresponds to approximately 30% more training GPU-hours. We consider this additional training cost worthwhile for performance development. As shown in Table 2, Full improves ImageReward from 0.4273 to 0.6517 (+52.5%), PickScore from 21.27 to 21.77, and HPSv2 from 0.2800 to 0.2916 over Base, alongside improvements in CLIP and CFM. More importantly, the additional computation is confined to training. The deployed student retains the same four-NFE inference procedure without additional inference overhead.

AuthorsOpenReview ↗

Further Discussion with the Reviewer

Dear Reviewer cAUz,

Thank you for your time and constructive feedback. We have carefully considered your comments and provided detailed responses in our previous official comment. We summarize the main points below:

  1. We discussed the performance trade-offs on SD3.5 Large and SDXL, including the counting deficit, and outlined possible explanations while acknowledging the limitations of these interpretations.

  2. We conducted a new component ablation on FLUX.1-dev at four NFE, providing evidence of complementary preference gains from FD-SWD and T2D beyond the diffusion backbone.

  3. We quantified the additional training cost, which amounts to approximately 30% more GPU-hours on SDXL than Base, and clarified that the student retains the same four-NFE inference procedure without additional inference overhead.

We hope these clarifications and new results help address your concerns. We would appreciate hearing whether any points remain unresolved and would be happy to provide further clarification.

Thank you again for your thoughtful feedback and for helping us improve our work.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers and Area Chairs for a brief summary

Dear Reviewers and Area Chairs,

We sincerely thank all three reviewers and the area chairs for their time and thoughtful feedback.

We appreciate the recognition of our motivation for non-adversarial real-data supervision and the evaluation across multiple backbones. Your questions have helped us better assess the component contributions, clarify the computational cost, and discuss the spatial inconsistency of the generated image.

We have responded to each reviewer individually and summarize the main clarifications and new results below.

  1. Component effectiveness beyond SDXL. We conducted a new component ablation on FLUX.1-dev at four NFE. Adding T2D to FD-SWD improves HPSv2, PickScore, and ImageReward, while GenEval and CFM are essentially maintained. These results provide additional evidence of complementary preference gains on a flow backbone. We also acknowledge that the benefits depend on the backbone and will narrow our claims accordingly.

  2. Training-schedule ablation. We compared Base, Pretrain-and-Frozen, and Joint Training under the same four-NFE protocol. Freezing the pretrained components already improves five of the six reported metrics over Base. Joint training provides further preference gains, with trade-offs in CLIP and GenEval.

  3. Correction-scale schedule. We clarified that the correction scale γ in Eq. (10) gradually increases from 0 to 1 during training, progressively introducing the bridge correction signal.

  4. Training cost and comparison fairness. We reported SDXL training costs of approximately 45–55 H20 GPU-hours for Base, 60–70 for RINFD, and 90–100 for SenseFlow. Based on the range midpoints, RINFD requires approximately 30% more training computation than Base and 32% less than SenseFlow. The deployed student retains the same four-NFE inference procedure without additional inference overhead.

  5. Performance trade-offs and spatial inconsistency. We discussed the SD3.5 ranking differences and the SDXL counting deficit, distinguishing observed results from possible mechanism-based explanations. We also acknowledged the relational failure in Figure 7. Since RINFD does not explicitly constrain object cardinality or spatial relationships, aggregate preference gains do not imply uniformly improved counting or relational correctness.

We will incorporate all newly reported experimental results and discussions into the revised manuscript. Thank you again for your constructive feedback and for helping us improve the clarity and rigor of our work.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.