acceptodds
Under review as a conference paper at ICLR 2027

LightMGT: A Compact Masked Generative Transformer for Visual Generation

Abstract

Masked generative transformers (MGTs) synthesize images by predicting discrete visual tokens in parallel, but existing models offer limited efficiency advantages over diffusion models. We therefore propose LightMGT, a compact masked generative transformer for image generation and editing. Its 148M backbone reduces model size through a compact double-stream design and factorized token prediction. We introduce MGT-OPD to combine generation and editing capabilities in one model by matching specialist teachers’ token predictions on student-generated masked states. Antithetic Posterior Matching (APM) then distills iterative decoding into one step using a pair of masks and an updated auxiliary model, yielding LightMGT-Turbo. LightMGT achieves 0.70 on GenEval, 78.32 on DPG-Bench, and 6.49 on GEditBench-EN. On GenEval, it scores 0.70 versus 0.67 for the 12B-parameter FLUX.1-dev, while delivering 8.8× higher throughput. With a single decoding step, taking about 0.235 s, LightMGT-Turbo achieves scores of 0.58 on GenEval, 71.30 on DPG-Bench, and 6.21 on GEditBench-EN. MGT-OPD combines text-to-image generation and editing capabilities. In the LightMGT experiments, it improves GenEval and GEditBench-EN by 16.7% and 4.5% relative to weight merging, and by 11.1% and 7.1% relative to joint training, respectively. APM also improves one-step distillation. On native MaskGIT ImageNet-256, it reduces FID by 1.2% and increases IS by 4.8% relative to Di[M]O. Together, this work contributes to the community a compact backbone, capability composition, and one-step distillation for image generation and editing.

Primary area:
generative models
Keywords:
Generative AI, Text to Image, Image Editing

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_A2pjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LightMGT is a masked generative transformer for text-to-image generation and editing with a 148M backbone (MM-JiT double-stream blocks, two 512-way heads over 18-bit Open-MAGVIT2 codes, frozen Qwen3-0.6B text encoder).

The paper claims three contributions: the compact pre-train model LightMGT; MGT-OPD, which composes a T2I and an editing specialist through a token-level reverse KL loss; and APM, a one-step distillation built on Di[M]O. Validation: GenEval 0.70 and GEditBench-EN 6.49 (Turbo: 0.58 / 6.21), throughput against FLUX.1-dev, SD3-Medium and Meissonic, and FID 6.80 vs. 6.91 for Di[M]O on MaskGIT ImageNet-256.

Critical strengths
  1. A useful recipe and baseline. A single 148M-backbone model covers both generation and editing (GenEval 0.70, GEditBench-EN 6.49), with documented open data (Tables 9–10). This is the main contribution of this paper for the community.
  2. Solid Experimental reselts. With a 148M backbone, one model reaches GenEval 0.70 and the best GEditBench-EN average in Table 2 (6.49), above OmniGen2 (4B, 6.31) and FLUX.1 Kontext-dev (12B, 6.26). MGT-OPD keeps most of both specialists (0.70 vs. 0.73; 6.49 vs. 6.64) and beats joint training (0.63 / 6.06).
Critical weaknesses
  1. The novelty of this paper is limited. For the three main contributions listed in the paper, they are all based on existing recipes with very incremental modification. For example, APM's Eq. (5) is vβ(pT,pF)=∇ℓF[(1−β)KL(pT∥pF)+βKL(pF∥pT)],v_\beta(p_T,p_F)=\nabla_{\ell_F}\big[(1-\beta)\mathrm{KL}(p_T\Vert p_F)+\beta\mathrm{KL}(p_F\Vert p_T)\big], the generalized Jeffrey divergence of Di[M]O (arXiv 2503.15457, Eq. 8) with the explicit logit gradients of its Eqs. (16)-(17); Eq. (6) is the surrogate whose gradient is Di[M]O's Eq. (7). And the experimental comparison of APM and Di[M]O on MaskGIT also shows that these two are similar in terms of performance.

  2. The paper is motivated by MGTs offering limited efficiency advantages over diffusion models (l. 053–054). SANA-Sprint (arXiv 2503.09641) reports SANA-0.6B at 0.9 s (GenEval 0.64) and one-step SANA-Sprint-0.6B at 0.21 s (GenEval 0.72) on an A100, whereas LightMGT needs 1.42 s (0.70) and Turbo 0.235 s (0.58).

  3. "FLUX.1-dev-level performance" also rests mainly on GenEval Position (0.62 vs. 0.20), while DPG-Bench is lower (78.32 vs. 82.56). This matters because efficiency is the paper's main motivation.

Questions for authors
  1. what is the motivation for introducing eq(2) for KL loss calculation?
  2. What is Monetico mentioned in this paper, not clear without citing the corresponding paper.
  3. It's not clear why the authors construct two masked versions in APM. There is no ablation against equal-cost alternatives.
3 comments
AuthorsOpenReview ↗

Follow-Up to Your Review

Dear Reviewer A2pj,

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 11907

AuthorsOpenReview ↗

Follow-Up to Your Review (Second Reminder)

Dear Reviewer A2pj,

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 11907

AuthorsOpenReview ↗

Follow-Up to Your Review (Third Reminder)

Dear Reviewer A2pj,

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 11907

Reviewer_Ru4jOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents LightMGT, a compact masked generative transformer for text-to-image generation and image editing. It trains a text-to-image and editing model, uses MGT-OPD to combine the two capabilities, and applies APM to obtain a one-step variant. The resulting system achieves promising results with a 148M-parameter generative backbone.

Critical strengths
  • The paper addresses a useful practical goal: combining generation and editing in a relatively small model.
Critical weaknesses
  • The central methodological contribution is unclear. The on-policy composition principle closely follows DanceOPD; MGT-OPD adapts it to masked discrete tokens. APM is an incremental modification of Di[M]O. The paper describes these adaptations, but does not convincingly establish which new design choices are essential.

  • One-step distillation causes a substantial text-to-image drop: GenEval falls from 0.70 to 0.58 and DPG-Bench from 78.32 to 71.30. The performance of the most existing T2I models Turbo-version (i.e., Z-Image turbo, Ernie-Image turbo, Boogu turbo) will not drop this much. The paper does not analyze this degradation in much detail.

  • The compact-model comparison is incomplete. MiniT2I is cited but absent from the main comparison. Moreover, the “FLUX-level” characterization is supported by GenEval but less clearly by DPG-Bench (78.32 versus 82.56) or GenEval counting (0.40 versus 0.79). Qualitative examples alone do not establish comparable visual quality.

Questions for authors
  1. Can the authors ablate student-generated versus fixed states, shared versus independent randomness, and all-masked-position versus committed-token supervision on the full LightMGT generation and editing benchmarks, with matched training budgets?

  2. Can APM and Di[M]O be compared using the same LightMGT checkpoint and compute budget? What are the separate contributions of paired masks and the auxiliary-model re-query?

  3. Can the authors include a recent compact baseline such as MiniT2I under a consistent evaluation protocol, or otherwise clarify the limits of the comparison? Please also report the inference footprint including the frozen text encoder and tokenizer.

5 comments
AuthorsOpenReview ↗

Follow-Up to Your Review

Dear Reviewer Ru4j,

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 11907

AuthorsOpenReview ↗

Follow-Up to Your Review (Second Reminder)

Dear Reviewer Ru4j,

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 11907

AuthorsOpenReview ↗

Follow-Up to Your Review (Third Reminder)

Dear Reviewer Ru4j,

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 11907

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

LightMGT combines a 148M-parameter masked generation backbone with factorized LFQ outputs, on-policy distillation for generation and editing, and Antithetic Paired Masks (APM) for one-step distillation. The main model reports GenEval 0.70, DPG-Bench 78.32, and GEditBench-EN 6.49; the Turbo variant reports 0.58, 71.30, and 6.21, respectively. The compact multitask system is practically interesting. However, a duplicated row of editing results, incomplete descriptions of several training configurations, and insufficiently isolated APM ablations prevent a confident assessment of the individual contributions.

Critical strengths
  1. A practically useful compact generation-and-editing system. Tables 1–2 report competitive results from a 148M-parameter generation backbone, with factorized LFQ outputs and a one-step variant. Achieving useful quality across generation and editing with a small backbone is a meaningful systems contribution, even without an entirely new architecture.

  2. Capability composition is tested beyond a single model setting. Table 3 covers both Meissonic/EditMGT and LightMGT and includes specialist, weight-merging, and joint-training comparisons. The reported generation/editing tradeoff is promising and provides a concrete motivation for the proposed on-policy distillation approach.

Critical weaknesses
  1. An entire editing-result row is identical for two apparently different methods. LightMGT-Turbo in Table 2 and WeightMerge in Table 3B share all 11 GEditBench-EN category scores and the average of 6.21. The methods appear to use different checkpoints and sampling protocols. Please explain whether the same evaluation was reused, the table contains an error, or there is another reason, and identify the corresponding checkpoints and evaluation records. This is a reporting-consistency issue, not evidence of misconduct, but it prevents this row from serving as independent evidence for both methods.

  2. The claimed distillation contributions cannot yet be reproduced and attributed reliably. Section 3.2 does not sufficiently define the cross-tokenizer transport T_ctx, while the LightMGT-specific OPD and Turbo training configurations remain incomplete. For APM, Table 5 combines antithetic masks and querying the auxiliary model after its update, without isolating either change or matching their computation. The small FID change from 6.91 to 6.80 also lacks uncertainty across training runs. These gaps make it difficult to determine which proposed component accounts for the improvement over the existing distillation approach.

Questions for authors
  1. Reporting clarification: Why are all editing scores for Turbo and WeightMerge identical? Which checkpoint and evaluation run produced each row, and can the corrected or independently verified results be supplied?
  2. Method clarification: How is T_ctx implemented and trained, and what are the exact LightMGT OPD and Turbo configurations? What subset and sampling protocol were used in Table 4?
  3. Targeted validation: In a small matched-compute experiment, what are the separate effects of antithetic masks and a fresh auxiliary prediction? A controlled ablation and uncertainty for the key comparison would materially affect my assessment of APM.
6 comments
AuthorsOpenReview ↗

Follow-Up to Your Review

Dear Reviewer eYAz,

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 11907

AuthorsOpenReview ↗

Follow-Up to Your Review (Second Reminder)

Dear Reviewer eYAz,

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 11907

AuthorsOpenReview ↗

Follow-Up to Your Review (Third Reminder)

Dear Reviewer eYAz,

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 11907

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 compact unified masked visual generation and editing.

  • Reviewers eYAz, Ru4j, and A2pj think our unified 148M backbone is "a meaningful system contribution", pursues "a useful practical goal", and provides "a practical recipe and baseline model".
  • Reviewers eYAz and A2pj find that our capability composition and distillation results are "strong" and "solid", noting that LightMGT "outperforms larger baselines including OmniGen2 and FLUX.1 Kontext-dev" while MGT-OPD "mostly retains the performance of individual specialist models" and is "tested beyond a single model setting".
  • Reviewers eYAz and A2pj appreciate our transparent architecture and training recipe, highlighting the "148M generative backbone, factorized LFQ outputs, and a one-step variant" together with our documented public data provenance.

We would like to highlight and emphasize again our key contributions as follows:

  • LightMGT unifies high-resolution text-to-image generation and instruction-based editing within a 148M Double-Stream MM-JiT Backbone with AdaLN and Factorized LFQ Output Heads.
  • Furthermore, we propose MGT-OPD to consolidate task specialists via On-Policy Token Matching under shared exponential-race randomness, and Antithetic Posterior Matching to enable One-Step Distillation via complementary mask pairing and lookahead auxiliary re-querying.

As suggested by the reviewers, we revised our manuscript and Appendix Section E, where Table A through Table E below correspond to Appendix Table 11 through Appendix Table 15.

Additional Results & Experiments

  • Updated the Weight Merge row in Paper Table 3B and added the per-category GEditBench-EN and GenEval breakdown alongside the matched-backbone Di[M]O comparison in Table A and Appendix Section E.1, as suggested by Reviewers eYAz and Ru4j.
  • Added compute-matched APM ablations and three-seed uncertainty results on ImageNet-256 and LightMGT-Turbo in Table B and Appendix Section E.2, as suggested by Reviewers eYAz, Ru4j, and A2pj.
  • Added the matched-budget MGT-OPD component and objective ablation on the main LightMGT generation and editing benchmarks in Table C and Appendix Section E.3, as suggested by Reviewers Ru4j and A2pj.
  • Added the few-step scaling comparison between LightMGT-Turbo and undistilled LightMGT in Table D and Appendix Section E.4, as suggested by Reviewers Ru4j and A2pj.
  • Added MiniT2I to Paper Table 1 and added the full-system parameter, memory, and latency comparison with MiniT2I, SANA-0.6B, and SANA-Sprint-0.6B in Table E and Appendix Section E.5, as suggested by Reviewers Ru4j and A2pj.

Other Changes & Modifications

  • We have clarified the cross-tokenizer bridge Tctx\mathcal{T}_{\mathrm{ctx}}, the MGT-OPD and APM training protocols, and the Monetico reference in Section 3.2, Section 4.2, Appendix Section B, and Appendix Section D, as suggested by Reviewers eYAz and A2pj.
  • We have refined the metric-level comparisons in the abstract, Section 1, and Section 4.1, as suggested by Reviewers Ru4j and A2pj.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.