acceptodds
Under review as a conference paper at ICLR 2027

Cycle Alignment for Understanding-Guided Visual Editing in Unified Multimodal Models

Abstract

Image editing demands both fine-grained visual understanding and controllable generation. A model must comprehend objects and layouts in the source image, parse editing instructions, and modify specific regions while preserving identity and background. Unified multimodal models (UMMs) inherently possess both capabilities within a single architecture, making them well-suited for editing. However, in conventional UMM training, the understanding and generation branches are optimized independently. Despite sharing the same backbone, the potential for understanding to directly guide generation remains largely unexploited. To exploit this potential, we propose Cycle Alignment (CycleA), a post-training paradigm that closes the loop between understanding and generation. Specifically, CycleA routes the generation branch's output back into the frozen understanding branch, where source-image VQA constraints provide differentiable semantic reconstruction feedback to the generation branch. This feedback operates directly in its native feature space through a lightweight frozen projector, avoiding the decode-encode overhead. To prevent the mapped features from drifting, we further introduce a feature alignment loss that keeps them close to real visual features. Across the evaluated configurations, CycleA improves aggregate image-editing scores over BAGEL after a 2,000-step post-training schedule following 5,000-step projector pretraining, while text-to-image benchmark scores remain comparable.

Primary area:
generative models
Keywords:
Generative Model, United Multimodal Model, Post Training, Cycle Alignment

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_i4AJOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Cycle Alignment (CycleA), a post-training method that uses a unified model’s own understanding branch as semantic feedback for generation. A frozen projector maps the generation branch’s one-step clean latent into the understanding feature space, and the frozen understanding branch answers VQA questions about the image from these features. The VQA loss and a feature-alignment loss are backpropagated into the generation branch, alternating with standard text-to-image steps, and no editing data is used. On BAGEL, CycleA improves ImgEdit (3.42 to 3.88) and GEdit-Bench-EN (6.52 to 7.08), while text-to-image scores stay comparable.

Critical strengths

I like the starting point of this paper. As a follow-up to RecA, it directly uses the model’s own understanding branch as a differentiable critic. The latent-space path is cheap, and text-to-image ability is largely preserved after post-training.

Critical weaknesses
  • The method is not very intuitive. The paper uses no image editing data at all, so why does it improve image editing? RecA also uses only images and a reconstruction loss, but it has a clear intuition: the image input acts as a dense prompt, and dense, accurate supervision improves image generation. In RecA, the editing improvement is a by-product. In this paper, image generation does not improve while image editing does, which I find strange. I would also like to know how many questions the frozen understanding branch answers correctly when given the projected features of real latents. And what happens if the generated latent is decoded into an image and the understanding branch, or a stronger VLM, supervises it from there? The latter is clearly a more intuitive way to do post-training. Training a dedicated MLP to bridge the two feature spaces makes the method feel somewhat odd to me.
  • The method stacks many modules, and I cannot tell which component clearly does the work, so the insight is limited. CycleA includes a separately pretrained and frozen per-patch projector, a one-step clean-latent estimate restricted to t ≤ 0.3, a VQA loss through the frozen understanding branch, a feature-alignment loss, a RecA-style reconstruction condition with 30 templates, and alternating text-to-image steps. In Tab. 3, L_vqa only (3.77), L_feat only (3.71) and captioning (3.71) are within 0.06 of each other, and there is no GEN-only row under GPT-4.1, so I cannot judge how much these variants gain over plain continued training. I think the authors should work out where the core improvement actually comes from.
  • The paper does not explain how RecA is used. L269 only says the generation branch “follows the reconstruction formulation of RecA”, and the condition c is never defined. Judging from Fig. 2 and the 30 templates in Sec. 4.1, the generation branch seems to be conditioned on a reconstruction prompt plus the image’s own understanding features, and the generation stream directly reuses RecA’s data. If so, the ALIGN step is essentially RecA’s self-conditioned reconstruction, with the loss moved from latent velocity space into the feature and QA space of the frozen understanding branch. This change is reasonable in itself, but it makes RecA under the same setting the most important control, and that is exactly what the paper lacks. The RecA row in Tab. 1 is cited from the original paper, and settings such as the number of training steps may differ. In my experience, RecA-style post-training is fairly sensitive to the number of training steps, so the current comparison is not fair, and this is not necessarily to CycleA’s disadvantage. A matched comparison that replaces L_vqa + L_feat with RecA’s flow-matching loss in the same ALIGN step would show whether the semantic loss adds anything.
  • The writing needs work. For example, the introduction spends a full paragraph (L74–87) on related methods, which duplicates the Related Work section. Sec. 3.3 splits a short training procedure into six bold run-in paragraphs, two pairs of which say the same thing (L353 and L355, L346 and L361). Overall, the writing, layout and allocation of detail are not yet at the level of a top venue, and I think the paper needs careful polishing.
  • The Overall scores in Tab. 1 are not aggregated consistently. CycleA’s reported 3.88 is a sample-level score, and its category mean is 3.81, while for every other row cited in the table, Overall equals the category mean. Computed the same way, CycleA is 0.39 above BAGEL and only 0.05 above RecA. Could the authors check this and use one aggregation throughout?
Questions for authors

Please see Weaknesses.

8 comments
AuthorsOpenReview ↗

Response to Reviewer i4AJ’s Concerns (Part 1)

We thank the reviewer for the constructive comments. We clarify the reconstruction-based motivation and use the new QA and gradient measurements to explain the proposed feedback mechanism.

Weaknesses

[W1a] Why editing can improve without editing pairs

ALIGN and image editing both take a source image and a text instruction as conditions. BAGEL already supports editing. CycleA uses reconstruction QA and feature alignment to train the generation branch to better preserve objects, attributes, and relations from the source image. This can reduce semantic errors in its existing editing process without paired editing data. Text-to-image generation has no source image to preserve, so its benchmark scores need not improve in parallel.

Importantly, on GEdit-Bench-EN, CycleA raises semantic consistency from 7.36 to 8.03, achieving the highest score among all compared methods in Table 10 of Appendix A.1. This substantial improvement supports the effectiveness of our semantic guidance for editing.

[W1b] QA evaluation of the latent-to-understanding interface

We evaluated the QA answer-token accuracy on 200 samples before and after post-training, including direct real-image input as a reference. Both reconstruction routes show small numerical increases. This comparison quantifies the answer-token accuracy gap between projected features and decoded reconstructions.

Model stateReal image via ViTDecoded reconstructionsProjected predicted latents
Before post-training51.3%50.9%44.0%
After post-training51.3%51.7%44.4%

[W1c] Projector rationale and comparison with decoded-image feedback

We thank the reviewer for this suggestion. We tested decoded-image feedback using the same frozen understanding branch and otherwise identical training, inference, and evaluation settings. Decoding and re-encoding takes longer and yields lower editing performance.

StrategyEditing performance, sample averageTraining time per iteration
Projector3.8826.5 s
Decode and re-encode3.5130.4 s

The additional decoding and visual encoding increase computation and lengthen the gradient path, which may attenuate the training signal. A stronger external VLM supervising decoded images would still require these stages, so stronger understanding alone would not remove this overhead or shorten the feedback path. We have added the ablation to the main paper.

AuthorsOpenReview ↗

Response to Reviewer i4AJ’s Concerns (Part 2)

[W2a] Component roles and ablation evidence

We thank the reviewer for highlighting the need for clearer attribution. Our main contribution is to turn pretrained understanding into image-specific semantic feedback through a fixed, aligned latent interface. This design is supported by a 0.37 gain over decoded-image feedback using sample-level averaging under GPT-4.1, and category-average gains of 0.78 over shuffled QA and 0.22 over GEN-only training under Qwen-VL-Max. We summarize each component’s role and supporting evidence below.

ComponentRoleRelated evidence
Pretrained, frozen projectorProvides a fixed semantic interface without RGB decoding or visual re-encoding.(a) compares feedback paths. (i) measures the QA readout gap.
One-step clean-latent estimate with t ≤ 0.3Provides an estimated reconstruction for semantic feedback while limiting high-noise prediction errors.(b) tests the noise restriction.
VQA through the frozen readerChecks specific objects, attributes, and relations using existing understanding capabilities.(c) removes VQA. (e) replaces it with captioning. (g) breaks image–QA correspondence. (h) measures its gradient signal.
Feature alignmentConstrains projected features toward source-image understanding features.(d) removes this constraint. (h) measures its gradient signal.
RecA reconstruction conditions and 30 templatesDefine semantic-preserving reconstruction without editing pairs.Reused from RecA and not separately ablated.
Alternating GEN and ALIGN trainingRetains standard generation training alongside ALIGN updates.(f) tests the additional benefit of ALIGN over GEN-only training.

Performance ablations evaluated with GPT-4.1

IDVariantEditing score, category average
ReferenceFull CycleA3.82
(b)Expand ALIGN noise to t_max = 1.03.57
(c)Remove VQA and retain feature alignment3.67
(d)Remove feature alignment and retain VQA3.75
(e)Replace VQA with captioning3.73

The feedback-path comparison (a) retains sample-level averaging:

IDVariantEditing score, sample average
ReferenceProjector-based CycleA3.88
(a)Replace projector feedback with decoding and re-encoding3.51

In (a), decoding and re-encoding also increases iteration time from 26.5 s to 30.4 s.

Training controls evaluated with Qwen-VL-Max

All post-trained variants use the same 2,000-step budget.

IDVariantEditing score, category average
ReferenceBAGEL3.87
(f)GEN-only continued training3.96
(g)Shuffle image–QA correspondence3.40
ReferenceFull CycleA4.18

CycleA exceeds GEN-only by 0.22, while breaking image–QA correspondence reduces its score by 0.78. Gains are computed before rounding. These controls support a benefit from image-specific semantic feedback beyond continued training.

Additional diagnostics and repeated runs

IDExperimentResultsInterpretation
(h)Gradient diagnostic on eight images at two timestepsMedian gradient norms are 6.9 for VQA and 0.09 for feature alignment.VQA supplies the larger gradient signal within ALIGN on the tested inputs.
(i)QA answer-token accuracy on 200 samplesBefore → after: real images 51.3% → 51.3%, decoded reconstructions 50.9% → 51.7%, projected predicted latents 44.0% → 44.4%.Quantifies the QA readout gap between projected predicted latents and the decoded-image and real-image references.
(j)Three independent training seedsUsing sample-level averaging, BAGEL+CycleA achieves 3.85 ± 0.03, compared with 3.75 ± 0.04 for BAGEL+RecA.Supports the repeatability of the full-method comparison.
AuthorsOpenReview ↗

Response to Reviewer i4AJ’s Concerns (Part 3)

[W3] RecA conditioning and matched comparison

We thank the reviewer for identifying this ambiguity.

RecA reuse

In ALIGN, c combines source-image understanding features with a semantic-preserving reconstruction prompt reused from RecA. The VQA question is supplied separately to the frozen reader. GEN reuses RecA’s image–caption data and standard flow-matching objective. CycleA retains RecA’s reconstruction conditioning but changes how the generated content is supervised.

Matched comparison

We implemented the requested control by replacing only the VQA and feature-alignment losses in ALIGN with RecA’s flow-matching loss. The data streams, GEN/ALIGN schedule, timesteps, trainable parameters, inference settings, and evaluator remain identical. Both variants use 2,000 updates. Results across three independent training seeds are shown below.

ALIGN objectiveEditing performance, sample average, mean ± std
RecA flow matching3.75 ± 0.04
CycleA VQA + feature alignment3.85 ± 0.03

The 0.10 mean gain with sample-level averaging supports the additional value of semantic feedback under the same reconstruction conditions. We have added this control and clarified the conditioning in the revised paper.

[W4] Writing and organization

We thank the reviewer for these suggestions and hope the revisions below address these concerns and make the paper easier to follow.

  1. Introduction. We shortened the discussion of prior methods and retained the detailed comparisons in Related Work.
  2. Sec. 3.3, training schedule. We merged the repeated GEN/ALIGN descriptions and now state the objectives and strict alternation once.
  3. Sec. 3.3, gradient flow and freezing. We consolidated the repeated statements and clarified that frozen parameters still allow gradients to pass through to the generator.
  4. Allocation of detail. We brought the essential timestep, gradient-detachment, and freezing details into the main text. Appendix A.2 retains the supporting implementation details.
  5. Notation and layout. We unified the notation across Figure 2, equations, and text, reduced unnecessary bold run-in headings, and polished the language throughout.

[W5] Consistent Overall aggregation

We thank the reviewer for identifying this inconsistency. We have revised Table 1 by averaging the nine category scores with equal weights for every method.

MethodCategory-average Overall
BAGEL3.42
BAGEL + RecA3.76
CycleA3.82

We have specified equal-weight category averaging in the evaluation section and table caption, and revised the comparative claims accordingly.

Reviewer_T9XsOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

This paper proposes Cycle Alignment (CycleA), a post-training method for unified multimodal models (UMMs) that uses a model's visual understanding capability to improve image generation and editing. For an image-question-answer sample, the generation branch predicts a clean VAE latent from a lightly noised latent. A frozen per-patch projector maps this prediction into the understanding branch's visual feature space, where a frozen understanding pathway computes a VQA cross-entropy loss. A feature regression loss toward the real image's visual features is added as a shorter gradient path. The generation-side parameters are updated while the understanding-side parameters and projector remain frozen, and the VQA/feature objectives are alternated with ordinary flow-matching updates. On BAGEL, the method reports improved ImgEdit and GEdit-Bench-EN scores while retaining text-to-image performance.

Critical strengths
  1. A simple differentiable bridge within an existing UMM. Mapping the predicted clean latent directly into the understanding feature space avoids an external VLM and a decode–encode loop. This is attractive as a practical post-training recipe and makes the semantic gradient path explicit.

  2. A plausible combination of complementary objectives. The targeted VQA loss supplies semantic supervision while the feature-alignment loss provides a shorter, representation-level constraint; the reported single-loss ablations are consistent with the claim that these signals are complementary, and text-to-image scores remain largely unchanged.

Critical weaknesses
  1. The reconstruction condition and its compatibility with VQA supervision are not specified. The paper says that the reconstruction condition is drawn from 30 templates covering fidelity, style, color, geometry, texture, object count, composition, and environment, but does not provide the templates, their selection/instantiation rules, or any compatibility constraint between the reconstruction condition and the image VQA question-answer pair. The VQA target describes the source image, whereas the reconstruction condition is fed to the generation branch; if it changes a property queried by the VQA question, the VQA cross-entropy remains computable but can become a competing or noisy gradient rather than a valid semantic reconstruction signal. The paper also omits the scale, filtering, and reuse/sampling details of the two data streams. Because the post-training stage is only 2,000 steps, these choices can materially affect both the result and the interpretation of the alignment gain.

  2. The central method is poorly presented and incompletely positioned relative to prior understanding-supervised post-training. Figure 2, which should establish the paper's main computation graph, does not label the two text inputs as the reconstruction condition and the VQA question; it appears to route an image emitted by the generation decoder into the Projector and then into the Understanding Encoder, whereas the equations and Appendix A.2 specify a generated clean VAE latent mapped directly to an understanding-feature interface without decoding or re-encoding. The figure also labels only generic CE and MSE losses without identifying the VQA loss, feature-alignment loss, and combined ALIGN objective. This is a substantive problem because the latent-to-feature path is the central contribution, and the current overview can lead a careful reader to infer a different image-to-image method. The main text compounds the problem by giving only a high-level account of the strict Gen/ALIGN schedule, timestep choices, gradient-detachment exception, and freezing policy, with the operational description deferred to Appendix A.2. At the same time, the empirical positioning is incomplete: RecA uses dense real-image understanding embeddings as generation conditions for self-supervised reconstruction, while CycleA uses generated clean latents, a projector, VQA, and feature alignment; UNO is much closer in high-level mechanism because it freezes the understanding expert and backpropagates captioning and visual-regression objectives from generated representations. Although UNO is cited, it is absent from the quantitative baseline list despite reporting BAGEL-based results on the same benchmark family (e.g., GEdit-Bench-EN overall 7.17 versus CycleA's 7.08, subject to protocol differences). Without a corrected overview, a more self-contained method presentation, and a matched UNO comparison, it is difficult to assess both the claimed mechanism and the incremental novelty.

Questions for authors
  1. How are the 30 reconstruction-condition templates constructed and instantiated? Do they always specify semantic-preserving reconstruction, and is there any filtering or matching between cc and the image QA pair (q,a)(q,a)? Please report the templates, the sampling distribution, and the fraction of potentially conflicting examples.

  2. What are the sizes, filtering procedures, train/validation splits, and sampling/reuse rules for the generation and alignment streams? How many effective ALIGN and GEN updates are performed per dataset pass during the 2,000-step schedule?

  3. Why is UNO omitted from the quantitative comparison despite being cited and reporting BAGEL-based results on ImgEdit and GEdit-Bench-EN? Can the authors provide a matched comparison among BAGEL, GEN-only, RecA, UNO, and CycleA using the same checkpoint, data, training budget, sampler, evaluator, and random seeds?

  4. Please clarify Figure 2 and the actual interface: does the projector consume the estimated clean VAE latent and output features directly at the understanding connector/pathway, with no generated RGB image or second visual-encoder pass? Please also label qq, cc, LvqaL_{\mathrm{vqa}}, LfeatL_{\mathrm{feat}}, and LALIGNL_{\mathrm{ALIGN}} consistently in the figure and main text.

  5. Are the main results averaged over multiple random seeds, and are task-level variance or confidence intervals available? The paper's mechanism claim would also benefit from reporting the relative gradient contributions of LvqaL_{\mathrm{vqa}}, LfeatL_{\mathrm{feat}}, and the GEN updates under a matched training budget.

3 comments
AuthorsOpenReview ↗

Response to Reviewer T9Xs’s Questions

We thank the reviewer for the detailed questions. We clarify the computation graph and training protocol, then present the available quantitative evidence for the latent interface and optimization objectives.

Questions

[Q1] Reconstruction templates and QA compatibility

We would like to clarify that the 30 templates are prompts reused from RecA for semantic-preserving reconstruction, rather than instructions to modify image content. During training, one template is sampled uniformly (1/30) and used directly as the prompt, without placeholder filling. Selection is independent of the QA pair. No additional QA-based matching or filtering is applied because the original QA remains applicable to the reconstructed content. We provide all 30 templates in the supplementary material.

[Q2] Dataset sizes, splits, filtering, and reuse

We clarify the datasets, sampling procedures, and effective update counts for both streams, and add the details to the supplementary material, as summarized below:

(1) Data: GEN uses 6,000 image–caption pairs; ALIGN uses 6,000 image–QA pairs.

(2) Filtering and splits: Unreadable images and missing or empty captions/QA pairs are skipped; neither stream has a held-out validation split.

(3) Sampling and reuse: Files are shuffled once with fixed seeds, then read sequentially. Exhausted loaders repeat the same order; process restarts reset their cursors.

(4) Training exposure: Strict 1:1 alternation yields 1,000 GEN and 1,000 ALIGN updates. With a global batch of 6, each stream consumes 6,000 sample presentations, including repeats.

[Q3a] UNO and controlled baseline comparisons

We thank the reviewer for pointing out this omission. We have added controlled UNO comparisons on ImgEdit. UNO (Gen) uses reconstruction training to match CycleA’s training setting. UNO (Edit) follows UNO’s original paired-editing training task, with all other settings identical to CycleA’s. We further retrained RecA from the same BAGEL checkpoint, retaining CycleA’s data, training budget, schedule, sampler, evaluator, and random seeds, while replacing only the ALIGN loss with flow matching.

MethodEditing performance, sample average
RecA3.75 ± 0.04
CycleA3.85 ± 0.03

Table 8 in Appendix A.1 also compares BAGEL, GEN-only, and CycleA under the same evaluator, obtaining category-average Overall scores of 3.87, 3.96, and 4.18, respectively. GEN-only and CycleA use a matched 2,000-update post-training budget. Together, these controls support benefits beyond ordinary continued training and RecA-style reconstruction.

[Q4] Projector interface and Figure 2 notation

We thank the reviewer for pointing out this ambiguity in our original presentation. We have revised Figure 2 and the main text to clarify the interface and unify the notation. The feedback path is estimated clean VAE latent → frozen projector → understanding connector → frozen understanding branch, with no RGB decoding or second visual-encoder pass. Gradients pass through the frozen modules to the generation branch.

[Q5a] Random seeds and task-level uncertainty

To assess seed sensitivity, we repeated training and evaluation for BAGEL+CycleA, BAGEL+RecA, and Nexus-Gen+CycleA with three independent random seeds. We report mean ± standard deviation below. The small standard deviations support consistent performance across the tested seeds.

MethodBAGEL+CycleABAGEL+RecANexusGenNexusGen+CycleA
Sample-average editing performance (mean ± std)3.85 ± 0.033.75 ± 0.042.682.74 ± 0.04

[Q5b] Relative gradient contributions

We thank the reviewer for this suggestion. We report median gradient L2 norms measured on the same diagnostic inputs and trainable generation parameters. VQA provides the dominant gradient signal within ALIGN on the tested inputs, while feature alignment serves as an auxiliary constraint. GEN’s flow-matching objective is included as a reference.

VQA (CE loss)Feature alignment (MSE loss)GEN reference (flow matching)
6.90.099.15
AuthorsOpenReview ↗

Response to Reviewer T9Xs’s Concerns

Weaknesses

[W1a] Unspecified reconstruction condition

We clarify that reconstruction uses both image and text conditions: the source image supplies the content, while the text prompt requests semantic-preserving reconstruction. The original QA therefore remains valid supervision for the intended reconstruction target. The 30 prompts are reused from RecA, sampled uniformly, and used directly without placeholder filling or QA-based matching/filtering. We have provided the complete prompt list in the supplementary material.

[W1b] Consistency between reconstruction and QA supervision

We clarify that the conditions request semantic-preserving reconstruction; changes to queried properties are therefore reconstruction errors that VQA should correct. On the same 200 samples, both QA paths improve after post-training, as shown below. Matched QA also outperforms shuffled QA under otherwise identical training settings, as shown in Table 8 in Appendix A.1, supporting image-specific semantic supervision.

Model stateDecoded-image QA answer-token accuracyProjected-feature QA answer-token accuracy
Before post-training50.9%44.0%
After post-training51.7%44.4%

[W1c] Missing data scale and sampling details

We have added the filtering and sampling/reuse rules for both streams to the supplementary material. The 2,000-step schedule strictly alternates 1,000 GEN and 1,000 ALIGN updates, corresponding to 6,000 sample presentations per stream (12,000 in total). The budget-matched GEN-only and shuffled-QA controls in Table 8 in Appendix A.1 further support the additional benefit of semantically matched QA supervision.

[W2a] Method presentation and positioning

We thank the reviewer for highlighting these issues. We have revised Figure 2 and the main text to clarify:

  • Inputs: reconstruction condition and VQA question.
  • Feedback path: clean VAE latent → frozen projector → understanding connector → frozen reader; no RGB decoding or re-encoding.
  • Losses: VQA loss, feature-alignment loss, and combined ALIGN objective.
  • Positioning: CycleA’s pre-aligned, frozen interface and targeted QA feedback, compared with UNO’s noisy-state supervision.

[W2b] Operational details deferred to the appendix

We thank the reviewer for this suggestion. We have expanded the main text to specify:

  • Schedule: strict 1:1 GEN/ALIGN alternation, with separate updates.
  • Timesteps: GEN uses full-range logit-normal sampling; ALIGN uses uniform sampling over [0, 0.3].
  • Gradient flow: BAGEL’s default understanding-path detachment is disabled in CycleA, allowing VQA gradients to reach the generation branch.
  • Freezing: only generation-side parameters are updated; the projector and understanding pathway remain frozen but differentiable.

[W2c] Quantitative comparison with UNO

We thank the reviewer for highlighting this omission. We compare CycleA with RecA and two UNO configurations. UNO (Gen) uses reconstruction training to match CycleA’s training setting. UNO (Edit) uses paired editing data to follow UNO’s original training task, with all other settings identical to CycleA’s. All methods use 2,000 updates, identical inference settings, and the same evaluator. Results are mean ± standard deviation across three training seeds.

MethodSample-average editing performance (mean ± std)
RecA3.75 ± 0.04
UNO(Gen)3.39 ± 0.11
UNO(Edit)3.70 ± 0.05
CycleA3.85 ± 0.03

UNO (Gen) shifts UNO’s original training task from editing to reconstruction, whereas CycleA’s QA and feature alignment are designed for semantic-preserving reconstruction. Restoring editing-based training raises UNO’s score by 0.31 and narrows its gap to CycleA to 0.15. Notably, UNO additionally learns a meta-query-based visual readout during post-training, whereas CycleA enters this stage with an already aligned and frozen semantic interface. Consequently, UNO requires longer training to converge and exhibits inferior performance under the same training configuration as CycleA.

Thank you for the clarification and additional experiments. These have largely addressed my main concerns, and I am raising my score from 2 to 3.

Reviewer_tycLOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Cycle Alignment, a post-training method that uses a unified multimodal model's frozen understanding branch to supervise its generation branch. A pretrained, frozen projector maps estimated clean VAE latents into visual understanding features. Training alternates standard flow matching with image question-answer supervision and feature alignment, without additional paired editing data. Experiments report improved editing scores on BAGEL, broadly preserved text-to-image performance, and a small gain on Nexus-Gen. Training controls and loss ablations provide evidence that the semantic supervision contributes to the improvement.

Critical strengths
  1. The method provides a differentiable path from the model's own understanding branch to its generation branch without decoding and re-encoding the predicted image. The added path is used only during training, leaving inference unchanged. This makes the approach practically attractive.
  2. CycleA improves BAGEL on both editing benchmarks. The GEN-only and shuffled-QA controls support a contribution from matched semantic supervision, and a second evaluator retains the overall improvement. The loss ablations also show gains from combining QA and feature alignment.
Critical weaknesses
  1. The added value over the closest methods remains unclear. UNO already uses a frozen understanding branch to supervise generative representations through captioning and visual regression. CycleA changes the feedback path and uses targeted QA, but provides no direct comparison with UNO. Its captioning ablation does not reproduce the full UNO method. The RecA result in Table 1 is cited from prior work, so the 3.88 versus 3.75 comparison does not establish an advantage under matched data, training budget, and evaluation settings. These omissions limit the evidence for the significance of the specific design.
  2. The understanding branch evaluates projected latent features rather than features extracted from the decoded reconstruction. Freezing the projector and matching real-image features are sensible constraints, but the paper does not quantify whether the proxy preserves the QA behaviour of actual reconstructed images. The editing gains therefore do not establish how faithfully this path evaluates generated visual content. A comparison with decoded-image feedback would also test the claimed efficiency trade-off. Repeated runs or uncertainty estimates are absent, which further limits confidence in the smaller gains over RecA and on Nexus-Gen.
Questions for authors
  1. Can you provide comparisons with UNO and RecA using matched training budgets, inference settings, and the same evaluator? These comparisons would clarify whether the proposed path and QA objective offer a meaningful advantage.
  2. How closely do QA predictions from projected features agree with those from the decoded reconstructions, before and after post-training? Do you have a decode-and-re-encode control that measures both editing performance and training cost?
  3. How stable are the improvements across training and generation seeds, and what uncertainty accompanies the benchmark differences? This is particularly relevant to the 0.13 gap over RecA and the 0.06 gain on Nexus-Gen.
3 comments
AuthorsOpenReview ↗

Response to Reviewer tycL’s Questions

We thank the reviewer for the constructive feedback. We have added matched comparisons, QA accuracy evaluations, a decode-and-re-encode ablation, and three-seed experiments.

Questions

[Q1] Matched UNO and RecA comparisons

We compare CycleA with RecA and two UNO configurations. UNO (Gen) uses reconstruction training to match CycleA’s training setting. UNO (Edit) uses paired editing data to follow UNO’s original training task, with all other settings identical to CycleA’s. All methods use 2,000 updates, identical inference settings, and the same evaluator. Results are mean ± standard deviation across three training seeds.

MethodSample-average editing performance (mean ± std)
RecA3.75 ± 0.04
UNO(Gen)3.39 ± 0.11
UNO(Edit)3.70 ± 0.05
CycleA3.85 ± 0.03

UNO (Gen) shifts UNO’s original training task from editing to reconstruction, whereas CycleA’s QA and feature alignment are designed for semantic-preserving reconstruction. Restoring editing-based training raises UNO’s score by 0.31 and narrows its gap to CycleA to 0.15. Notably, UNO additionally learns a meta-query-based visual readout during post-training, whereas CycleA enters this stage with an already aligned and frozen semantic interface. Consequently, UNO requires longer training to converge and exhibits inferior performance under the same training configuration as CycleA.

[Q2a] QA accuracy before and after post-training

We selected 200 samples from the dataset and evaluated QA answer-token accuracy before and after post-training. The results are shown below. Both paths improved after post-training, suggesting that our training enhances the generation branch’s ability to preserve semantic information.

Model stateFrom decoded resultsFrom projected features
Before post-training50.9%44.0%
After post-training51.7%44.4%

[Q2b] Decode-and-re-encode Ablation

We compare projector-based and decode-and-re-encode feedback under otherwise identical training, inference, and evaluation settings, as shown below. Decode-and-re-encode increases per-iteration training time and reduces editing performance, with the latter possibly attributable to attenuation of the training signal along the longer gradient path. We have added this ablation to the main paper.

StrategyEditing performance, sample averagePer-iteration Training Time
Projector3.8826.5 s
Decode-and-re-encode3.5130.4 s

[Q3] Stability across training seeds

To assess seed sensitivity, we conduct full training and evaluation for BAGEL+CycleA, BAGEL+RecA, and Nexus-Gen+CycleA with three different random seeds. As shown below, the small standard deviation indicates that CycleA’s performance is consistent across seeds.

MethodBAGEL+CycleABAGEL+RecANexusGenNexusGen+CycleA
Sample-average editing performance (mean ± std)3.85 ± 0.033.75 ± 0.042.682.74 ± 0.04
AuthorsOpenReview ↗

Response to Reviewer tycL’s Concerns

Weaknesses

[W1a] The distinction from UNO is unclear

Beyond improved performance, CycleA offers three advantages over UNO:

  1. Better reuse of pretrained capabilities: CycleA refines generation within an already aligned semantic space, directly strengthening existing capabilities. UNO must additionally make noisy generative states interpretable through a new understanding route, adding an adaptation task during post-training.
  2. More targeted semantic supervision: VQA explicitly checks objects, attributes, and relations, targeting details that global captioning may overlook.
  3. More direct content preservation: Aligning the reconstruction features used for QA with source-image understanding features jointly encourages semantic correctness and source consistency.

[W1b] Fairness of the cited RecA result

CycleA follows RecA’s published training data and evaluation settings. Table 1 cites RecA’s original category scores following common benchmark reporting practice. We further retrained RecA with matched training budgets, inference settings, and evaluation protocols. Results are reported as mean ± standard deviation across three training seeds.

MethodSample-average editing performance (mean ± std)
RecA3.75 ± 0.04
CycleA3.85 ± 0.03

These results support CycleA’s added value and stability. These experiments have been added to the revised paper.

[W2a] Faithfulness of projected features

We would like to clarify that the projector is independently pretrained to map real-image VAE latents to their corresponding understanding features, and remains frozen during post-training. The feature-alignment loss further anchors the projected reconstruction features to source-image understanding features. Q2a evaluates QA answer-token accuracy through this interface against the decoded-image reference, while Q2b compares the two feedback strategies in editing performance and training time.

[W2b] Efficiency trade-off of the projector

We thank the reviewer for this helpful suggestion. Our additional ablation study shows that the projector is a key component for achieving both strong editing performance and efficient training. With sample-level averaging, replacing it with decoded-image feedback reduces the editing score by 0.37 and increases iteration time (see Q2b). Thus, the projector contributes to performance as well as computational savings, rather than merely offering an efficiency trade-off.

[W2c] Missing repeated runs and uncertainty estimation

Thanks for highlighting this issue. We have added three-seed training and evaluation experiments (see Q3). Using sample-level averaging, CycleA retains average gains of 0.10 over RecA and 0.06 on Nexus-Gen, with standard deviations of 0.03–0.04. These repeated runs strengthen the evidence beyond single-run comparisons and make the training variability explicit.

Thank you to the authors for their detailed rebuttal. The newly added experimental results have largely addressed my concerns. I am therefore willing to raise my score from 2 to 3.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to Reviewers and Summary of Revisions

We thank all three reviewers for their careful evaluation and constructive feedback. In response, we have made the following revisions. Changes in the revised manuscript are highlighted in blue:

  1. We added matched RecA comparisons and tested UNO under reconstruction training and editing training. We also report three-seed results for the controlled comparisons and the Nexus-Gen experiments.

  2. We measured QA accuracy from projected features and decoded reconstructions before and after post-training. We also compared the two feedback routes in editing performance and training time.

  3. We added gradient diagnostics and expanded the discussion of the loss ablations, GEN-only training, and shuffled-QA controls to explain the roles of the main components.

  4. We clarified the reconstruction conditions, where gradients flow, which parameters are frozen, and how GEN and ALIGN steps alternate. The supplement now includes all 30 reconstruction prompts and the data sampling and reuse rules.

  5. We revised Figure 2 and removed repetition from the Introduction and Section 3.3. Table 1 now uses the unweighted mean of category scores for every method. We also specify how scores are aggregated in the additional experiments.

We hope these revisions clarify CycleA’s contribution and strengthen its experimental support. We provide detailed responses to each reviewer’s comments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.