acceptodds
Under review as a conference paper at ICLR 2027

DanceOPD: On-Policy Generative Field Distillation

Abstract

Modern image generation demands a single model that unifies diverse capabilities, including text-to-image (T2I), local editing, and global editing. However, these capabilities are rarely naturally aligned and often conflict. For instance, editing tends to degrade T2I performance, while global and local editing interfere with each other. Consequently, effectively composing these capabilities has become a central challenge for image generation model training. To tackle this, we introduce DanceOPD, an on-policy generative field distillation framework for flow-matching models that routes each sample to one capability field, queries one low-noise student-induced state, and trains with a simple velocity MSE objective. With each capability source defined as a velocity field over the shared flow state space, the student learns from fields queried on its own rollout states to compose expert capabilities. This formulation also absorbs operator-defined fields such as classifier-free guidance. Comprehensive experiments on T2I, editing, realism-field absorption, and CFG absorption show that our approach improves multi-capability composition, strengthening target capabilities while preserving anchor generation quality. We believe this work establishes a practical route for generative field distillation in flow-matching models.

Primary area:
generative models
Keywords:
Image Generation, Distillation

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 3 reviews, above 85% of the others.
Reviewer_dMkYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DanceOPD formulates multi-capability composition as on-policy distillation of compatible generative velocity fields into one flow-matching student. Each sample is routed to one frozen field, which supervises the student through velocity MSE at a single low-noise state from a detached student rollout. The formulation accommodates both learned expert fields and operator-defined fields such as classifier-free guidance. Validation covers generation/editing composition, local/global editing composition, realism absorption, and guidance absorption, with ablations of routing, query placement and count, objectives, and initialization.

Critical strengths
  1. An insightful perspective on capability composition through velocity fields. The paper represents specialized generative capabilities as compatible velocity fields and organizes their composition around which field to query, where to query it, and how many states to supervise. This provides a coherent formulation encompassing both learned experts and operator-defined fields such as classifier-free guidance. Its conceptual value is the unified treatment of these capability sources; its practical value is distilling their behaviors into one deployable student.

  2. Encouraging experimental results and informative ablations. Across both composition settings, the reported results show improved average editing performance while preserving text-to-image generation capability. Systematic ablations examine routing, query placement and count, objectives, rollout length, and initialization, supporting the main design choices and providing useful guidance for applying the method.

Critical weaknesses

The T2I off-policy baseline uses random-noise endpoints, limiting the evidence for the advantage of on-policy querying. Appendix B.4 constructs off-policy states as

ztoff=αtx0+σtϵ.z_t^{\mathrm{off}}=\alpha_t x_0+\sigma_t\epsilon.

For editing, x0x_0 is an encoded target image. For T2I, however, Equation 50 sets x0=ϵ0x_0=\epsilon_0, with ϵ0\epsilon_0 and ϵ\epsilon independent standard Gaussian noise. Consequently,

\sim\mathcal{N}\left(0,(\alpha_t^2+\sigma_t^2)I\right).$$ At low noise, $\alpha_t\approx1$ and $\sigma_t\approx0$, so the query remains approximately random noise rather than a lightly noised image latent. The teacher still supplies a velocity target, but receives a query latent with no image-derived structure at a near-clean timestep. This may disadvantage the baseline precisely where the query distribution concentrates. The comparison therefore does not distinguish the benefit of current-student states from the benefit of replacing these random endpoints with image-derived endpoints. A control using prompt-matched teacher-generated images would materially strengthen the on-policy claim. This concern applies to the T2I bucket; the editing buckets already use image-derived endpoints.
Questions for authors
  1. Assessment-critical controlled comparison: Does the T2I/editing advantage over off-policy distillation persist when its T2I endpoint is x0=E(xteacher)x_0=E(x_{\mathrm{teacher}}), using teacher-generated images for the same prompts? Please keep the other training settings matched and account for endpoint-generation cost.

  2. Clarification: Which checkpoint initializes each method in Table 2? Appendix B.5 identifies DanceOPD's diagnostic initialization, but the initialization used for its main composition results is unclear.

7 comments
AuthorsOpenReview ↗

Follow-Up to Your Review

Dear Reviewer dMkY,

Thank you again for the time and effort you put into reviewing our work!

We understand that this might be a super busy period for you. We wanted to briefly let you know that we have posted point-by-point responses to your questions and concerns and uploaded a revised manuscript with the changes highlighted.

Whenever you have a chance, we would be grateful for your feedback on whether the responses address your concerns. We would be happy to clarify anything that remains open.

Thank you!

Best,

The Authors of Submission 4626

AuthorsOpenReview ↗

Follow-Up to Your Review (Second Reminder)

Dear Reviewer dMkY,

Thank you again for the time and care you put into reviewing our work!

We know this is a busy period, so we wanted to follow up briefly on our point-by-point responses and the revised manuscript, where we highlighted the changes for easy reference.

If you have a chance to look over our responses, we would appreciate hearing whether they address your main concerns. If anything remains unclear, we would be happy to clarify it here.

Thanks again for your support!

Sincerely yours,

The Authors of Submission 4626

AuthorsOpenReview ↗

Follow-Up to Your Review (Third Reminder)

Dear Reviewer dMkY,

We are following up on our point-by-point responses and the revised manuscript, which we posted several days ago. We believe these materials further address your concerns and highlight the changes we made in response to your feedback.

As the end of the Author-Reviewer Discussion phase is fast approaching, please let us know whether our responses address your main concerns or if any points still need clarification. We are available to discuss them here.

Thank you again for your time and effort devoted in this review!

Best,

The Authors of Submission 4626

Reviewer_UV1HOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DanceOPD represents compatible generation and editing teachers as velocity fields, routes each sample to one task-specific teacher, queries a single low-noise-biased state on the student's own trajectory, and applies local velocity-MSE distillation. It evaluates T2I/editing and local/global editing composition, plus realism and CFG absorption.

Critical strengths

(1) Combining generation and editing capabilities in one model is practically valuable. The method is simple and does not require retaining multiple teachers at deployment. (2)The systematic comparisons of query location, supervision density, initialization, and objectives offer useful guidance for practical distillation. (3)Both main composition settings demonstrate strong average editing performance while retaining compositional T2I ability. The qualitative results also show meaningful transfer of editing behavior.

Critical weaknesses

(1)B.4's Eq. (50) uses a Gaussian random latent as the offline T2I "clean" endpoint, querying mostly low-noise states around it—a substantial departure from real or teacher-generated image endpoints. Matched-initialization OPD comparisons and an offline control with teacher-generated endpoints would help; Flow-OPD should also be labeled as an adaptation without MAR. (2)Table 2's local/global composition raises average editing performance from the local teacher's 5.095 to 5.498, but lowers subject removal from 6.017 to 4.348 (approximately 27.7%), trading off against background and style gains. Section 3.1 disclaims global Pareto optimality, but Section 4.1's claim of avoiding compromise still needs to address this regression. GenEval retention covers only the evaluated compositional T2I abilities. Explicit preservation targets and an explanation of the largest local-edit failure would prevent average gains from implying comprehensive capability retention. (3)Table 8 shows that local-edit initialization achieves 5.751, compared with 4.193 for merged initialization and 1.889 for T2I initialization. Yet Flow-OPD uses merged initialization, raising concerns about whether the comparison isolates algorithmic improvements. Matched-initialization comparisons are needed to establish the method’s advantage and clarify its ability to acquire capabilities absent from the initial student.

Questions for authors

(1)Please specify each Table 2 method's initialization and training budget, and provide matched-initialization OPD comparisons and an offline T2I control using teacher-generated image endpoints. (2)Do single- and dense-query variants share query-time marginals and loss scales? Can direct gradient-correlation measurements separate correlation from state-distribution and update-budget effects?

5 comments
AuthorsOpenReview ↗

Follow-Up to Your Review

Dear Reviewer UV1H,

Thank you again for the time and effort you put into reviewing our work!

We understand that this might be a super busy period for you. We wanted to briefly let you know that we have posted point-by-point responses to your questions and concerns and uploaded a revised manuscript with the changes highlighted.

Whenever you have a chance, we would be grateful for your feedback on whether the responses address your concerns. We would be happy to clarify anything that remains open.

Thank you!

Best,

The Authors of Submission 4626

AuthorsOpenReview ↗

Follow-Up to Your Review (Second Reminder)

Dear Reviewer UV1H,

Thank you again for the time and care you put into reviewing our work!

We know this is a busy period, so we wanted to follow up briefly on our point-by-point responses and the revised manuscript, where we highlighted the changes for easy reference.

If you have a chance to look over our responses, we would appreciate hearing whether they address your main concerns. If anything remains unclear, we would be happy to clarify it here.

Thanks again for your support!

Sincerely yours,

The Authors of Submission 4626

AuthorsOpenReview ↗

Follow-Up to Your Review (Third Reminder)

Dear Reviewer UV1H,

We are following up on our point-by-point responses and the revised manuscript, which we posted several days ago. We believe these materials further address your concerns and highlight the changes we made in response to your feedback.

As the end of the Author-Reviewer Discussion phase is fast approaching, please let us know whether our responses address your main concerns or if any points still need clarification. We are available to discuss them here.

Thank you again for your time and effort devoted in this review!

Best,

The Authors of Submission 4626

Reviewer_YXH4OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents DanceOPD, an on-policy distillation framework for consolidating multiple image-generation and editing capabilities into a single model. Each training sample is routed to a predefined teacher, which provides velocity supervision at a low-noise state visited by the current student. Training uses a single-query velocity MSE objective. Experiments cover text-to-image and editing composition, local and global editing composition, realism absorption, and classifier-free guidance absorption.

Critical strengths

(1) Industrial development often produces multiple checkpoints with complementary strengths, while deployment benefits from maintaining a single model. The paper addresses this practically relevant problem by transferring capabilities from specialized teachers while seeking to preserve existing functionality. (2) The method combines task-based teacher routing, student-generated query states, and direct velocity regression into a straightforward procedure. The formulation also accommodates operator-defined fields such as classifier-free guidance, extending its usefulness beyond transferring capabilities from separately trained experts. (3) The evaluation considers several capability-composition and field-absorption settings and reports encouraging improvements in the balance between generation and editing performance. The experiments provide practical guidance for implementing multi-teacher OPD.

Critical weaknesses

(1) The task-to-teacher correspondence is assumed rather than sufficiently validated as DanceOPD relies on predefined capability buckets and a fixed teacher assignment for each bucket. However, task-level specialization does not establish that the assigned teacher provides the most appropriate supervision for every sample. Teachers may have overlapping strengths, and individual prompts may require multiple capabilities. Although the authors acknowledge ambiguous task boundaries as a limitation, additional evidence would clarify the applicability of predefined routing beyond clearly separated expert roles. (2) The reported ablations support low-noise single-query supervision in the tested setting, but it remains unclear whether the shared Beta(5,2) sampling distribution is equally suitable across different capabilities. Furthermore, because query states are generated by the current student, earlier updates can affect subsequent supervision, making sample and task ordering relevant. (3) The distinction from Flow-OPD could be clarified more precisely. Flow-OPD already employs task-based teacher routing and teacher velocity-field supervision on student-generated states. DanceOPD’s more specific contributions therefore appear to lie in its low-noise single-query strategy, simplified training recipe, and extension to editing and field absorption. These are useful developments, but the paper would benefit from a clearer separation between the shared framework and its new contributions.

Questions for authors

(1) How are the task-to-teacher assignments validated? Could the authors report cross-teacher performance on shared samples and evaluate a small set of overlapping or compound tasks where the preferred teacher is less obvious? (2) Does the low-noise sampling advantage hold across individual task categories, and how stable are the results across training schedules? Could the authors provide per-task sampling comparisons and a limited study of multiple random orderings or balanced task schedules under matched initialization and training budgets? (3) Could the authors clarify the methodological differences from Flow-OPD and isolate the effects of query location, query count, and optimization objective under matched backbones, teachers, initialization, and training budgets?

5 comments
AuthorsOpenReview ↗

Follow-Up to Your Review

Dear Reviewer YXH4,

Thank you again for the time and effort you put into reviewing our work!

We understand that this might be a super busy period for you. We wanted to briefly let you know that we have posted point-by-point responses to your questions and concerns and uploaded a revised manuscript with the changes highlighted.

Whenever you have a chance, we would be grateful for your feedback on whether the responses address your concerns. We would be happy to clarify anything that remains open.

Thank you!

Best,

The Authors of Submission 4626

AuthorsOpenReview ↗

Follow-Up to Your Review (Second Reminder)

Dear Reviewer YXH4,

Thank you again for the time and care you put into reviewing our work!

We know this is a busy period, so we wanted to follow up briefly on our point-by-point responses and the revised manuscript, where we highlighted the changes for easy reference.

If you have a chance to look over our responses, we would appreciate hearing whether they address your main concerns. If anything remains unclear, we would be happy to clarify it here.

Thanks again for your support!

Sincerely yours,

The Authors of Submission 4626

AuthorsOpenReview ↗

Follow-Up to Your Review (Third Reminder)

Dear Reviewer YXH4,

We are following up on our point-by-point responses and the revised manuscript, which we posted several days ago. We believe these materials further address your concerns and highlight the changes we made in response to your feedback.

As the end of the Author-Reviewer Discussion phase is fast approaching, please let us know whether our responses address your main concerns or if any points still need clarification. We are available to discuss them here.

Thank you again for your time and effort devoted in this review!

Best,

The Authors of Submission 4626

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank our reviewers for the time and effort in providing thorough and constructive feedback on our submission!


We are encouraged to see that our reviewers recognize this work as a practical and effective framework for multi-capability generative field distillation.

  • Reviewers YXH4, UV1H, and dMkY recognize that consolidating generation and editing capabilities into a single model is "practically valuable" and "does not require retaining multiple teachers at deployment".
  • Reviewers YXH4 and dMkY find that our velocity-field formulation "provides a coherent formulation encompassing both learned experts and operator-defined fields such as classifier-free guidance".
  • Reviewers YXH4, UV1H, and dMkY appreciate our "encouraging experimental results and informative ablations" that "offer useful guidance for practical distillation".

We highlight our contributions as follows

  • DanceOPD formulates multi-capability composition as Hard-Routed On-Policy Velocity-Field Distillation in flow-matching models.
  • By combining Sample-Wise Teacher Routing with a Single Low-Noise Student-State Query and Plain Velocity MSE Regression, our framework consolidates text-to-image generation, local and global editing, realism enhancement, and Classifier-Free Guidance Absorption into a single deployable student model with high training efficiency.

As suggested by the reviewers, we revised our manuscript and Appendix Section F, where Table A through Table F below correspond to Table 10 through Table 15 in Appendix Section F.

Additional Results & Experiments

  • Added shared-sample cross-teacher evaluations across separated, overlapping, and compound tasks alongside routing applicability comparisons beyond clearly separated expert roles in Table A and Appendix Section F.1, as suggested by Reviewers YXH4 and UV1H.
  • Added per-category query-location, step-damping, and unit-normalized multi-query comparisons in Table B and Appendix Section F.2, as suggested by Reviewers YXH4 and UV1H.
  • Added the solver-step and LoRA-gradient amplification probe alongside frozen-trajectory velocity interventions in Table C and Appendix Section F.2, as suggested by Reviewers YXH4 and UV1H.
  • Added strictly matched-initialization OPD comparisons across starting checkpoints and training budgets in Table D and Appendix Section F.3, as suggested by Reviewers YXH4, UV1H, and dMkY.
  • Added the controlled off-policy T2I baseline using prompt-matched teacher-generated image endpoints and the complete Paper Table 2 initialization and GPU-hour breakdown in Table E and Appendix Section F.4, as suggested by Reviewers UV1H and dMkY.
  • Added within-rollout and independent-rollout gradient correlation measurements and the matched-marginal four-query control in Table F and Appendix Section F.5, as suggested by Reviewers UV1H and YXH4.

Other Changes & Modifications

  • We have clarified the methodological distinction from Flow-OPD in Section 2 and Paper Table 1 and detailed our official Flow-OPD implementation under matched capability teachers in Appendix Section B.3, as suggested by Reviewers YXH4 and UV1H.
  • We have specified the Paper Table 2 student initializations and compute budgets in Appendix Section B.5 and Table E, as suggested by Reviewers UV1H and dMkY.
  • We have revised Section 4.1 to state the category-level Subject Removal teacher asymmetry in Local and Global Edit composition alongside the scope of GenEval capability preservation, as suggested by Reviewer UV1H.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.