acceptodds
Under review as a conference paper at ICLR 2027

UniOwl: Bridging Multimodal Understanding and Generation with Adaptive Conditioning

Abstract

Unified multimodal models aim to support visual understanding and image generation within a single architecture, but diffusion-based generation introduces a key trade-off: while semantic guidance should adapt across denoising steps, invoking a large language model at each step is computationally prohibitive. Existing methods rely on either static conditions or costly step-wise language computation. We introduce UniOwl, a unified multimodal framework for joint understanding and generation with Adaptive Dual-Channel Conditioning. UniOwl first extracts step-invariant semantic condition tokens from LLM hidden states, computed once and shared across all denoising steps, then refines them at each step with lightweight latent- and timestep-aware residuals. In parallel, visual condition tokens are derived from the current denoising latent via a shared visual encoder. The fused semantic and visual tokens provide dynamic conditioning signals for the diffusion model, while generation supervision on the visual pathway ensures consistency between generation and understanding. To further encourage generation-understanding consistency, we introduce a cycle consistency objective that feeds generated images back into the understanding pathway and encourages their output distributions to be consistent with the original predictions. Extensive experiments show that UniOwl achieves superior generation quality and semantic consistency over existing unified models, while maintaining competitive multimodal understanding performance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multimodal learning, unified multimodal models, multimodal understanding, visual generation, diffusion models, vision-language models

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

This paper presents UniOwl, a unified framework for multimodal understanding and image generation with adaptive dual-channel conditioning. The proposed method extracts semantic condition tokens from LLM hidden states and dynamically refines them at each diffusion step using the current denoising features and timestep information. In parallel, it introduces generation-driven visual conditioning, which extracts visual features from intermediate denoising results through a shared visual encoder. These two types of conditions are fused to guide image generation. Additionally, a cycle consistency loss is introduced to encourage semantic alignment between understanding and generation. Experiments show that UniOwl improves image generation performance while maintaining competitive visual understanding capabilities compared with existing unified multimodal models.

Critical strengths
  1. Effective adaptive semantic conditioning. The proposed method dynamically refines LLM-derived semantic conditions using the current denoising features and timestep information, without repeatedly running the LLM. The ablation results show a substantial improvement in GenEval from 0.67 to 0.81, demonstrating the effectiveness of the proposed conditioning mechanism.
  2. Strong image generation performance. UniOwl achieves competitive results among compact unified multimodal models, improving GenEval from 0.74 (Mobile-O) to 0.83 while maintaining comparable visual understanding performance (62.3 vs. 62.1). These results demonstrate the practical value of the proposed approach.
Critical weaknesses
  1. Limited Novelty. Prior works such as ELLA [1] and DATE [2] have explored timestep-aware and generation-state-dependent conditioning during diffusion. However, the paper neither discusses nor compares with these closely related methods, leaving the novelty and advantages of its adaptive conditioning mechanism unclear.

  2. High Computational Cost of Visual Conditioning. As shown in Table 3, visual conditioning only improves GenEval from 0.81 to 0.83, while Table 6 shows a substantial increase in inference latency from 771.9 ms to 2725.0 ms, raising concerns about its cost-effectiveness.

[1] Hu et al., ELLA: Equip Diffusion Models with LLM for Enhanced Semantic Alignment, 2024. https://arxiv.org/abs/2403.05135

[2] Na et al., Diffusion Adaptive Text Embedding for Text-to-Image Diffusion Models, 2025. https://arxiv.org/abs/2510.23974

Questions for authors

Please clarify the concerns raised in the weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer fyfg

Dear Reviewer fyfg,

We thank you for your constructive assessment and helpful suggestions. We appreciate your recognition of the effectiveness of our adaptive semantic conditioning mechanism and its generation performance. We have added experiments and analyses, revised the manuscript, and provided point-by-point responses below.

W1: Novelty of Adaptive Dual-Channel Conditioning

Thank you for suggesting these relevant comparisons. ELLA [1], DATE [2], and our UniOwl adapt diffusion conditioning through different information sources and update mechanisms. We clarify these distinctions and the specific contributions of UniOwl below.

Differences from ELLA and DATE. ELLA extracts timestep-dependent conditions from frozen LLM features through a trained semantic connector. DATE uses the current generation state to optimize text embeddings through gradients of a text-conditioned objective during inference. UniOwl’s Diffusion-Adaptive Semantic Conditioning instead learns residual cross-attention from cached semantic conditions to the current noisy latent and timestep. It therefore uses sample-specific state feedback beyond timestep information and updates conditions by forward computation, without test-time embedding optimization or an external reward evaluator. Our timestep-only control achieves 0.70 on GenEval, compared with 0.81 for semantic-only UniOwl. This supports the value of state feedback within our mechanism; it is not a direct experimental comparison with ELLA or DATE.

Contributions of UniOwl. Our Adaptive Dual-Channel Conditioning combines state-dependent semantic refinement with Generation-Driven Visual Conditioning. The former refines semantic conditions using latent-state feedback. The latter extracts visual feedback from intermediate generations through the shared visual encoder. Cycle Consistency Learning further encourages agreement between understanding outputs for original and generated images. These mechanisms connect dynamic generation conditioning with shared visual representations and understanding–generation consistency. We have discussed ELLA and DATE in the revised manuscript.

MethodAdaptation inputUpdate mechanismTest-time optimization
ELLAText features + timestepTimestep-aware connectorNo
DATEText embeddings + predicted imageGradient-based embedding updatesYes
UniOwl (semantic)Cached conditions + noisy latent + timestepLearned residual cross-attentionNo

W2: Generation Performance and Inference Efficiency

We thank the reviewer for pointing out the computational cost of visual conditioning. We have added GenEval evaluations of UniOwl-Latent and UniOwl-Lite alongside the timing measurements below. Timing measurements cover the instrumented generation-stage GPU operations, including final decoding, and exclude the initial LLM prompt-encoding pass. Latent achieves 0.81 without intermediate-image visual feedback. Lite refreshes visual conditions every four steps and achieves 0.82, compared with 0.83 for Full. Lite reduces the measured generation time by 54.8% relative to Full, with a 0.01 lower GenEval score. Full provides a further gain at substantially higher computational cost because it decodes and visually encodes intermediate latents at every step. These results quantify the performance–efficiency trade-off and show that dense visual feedback is optional. We have included these results and their analysis in the revised manuscript, and clarified that the low-overhead claim applies to semantic refinement. We have also discussed the computational cost of visual feedback in the Limitations section.

ConfigurationVisual feedback frequencyIntermediate VAE decodes (20 steps)Measured generation time (ms) ↓GenEval ↑
UniOwl-LatentNone0771.9 ± 9.50.81
UniOwl-Lite (k=4)Every four steps51230.4 ± 0.50.82
UniOwl-FullEvery step202725.0 ± 24.80.83

Reference

[1] Hu X, Wang R, Fang Y, Fu B, Cheng P, Yu G. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024.

[2] Na B, Park M, Sim G, Shin D, Bae H, Kang M, Kwon SJ, Kang W, Moon IC. Diffusion Adaptive Text Embedding for Text-to-Image Diffusion Models. Advances in Neural Information Processing Systems. 2025.

Thanks for the authors' response, which solved my main concerns. I still think running VAE decode multiple times at inference time is quite heavy, but the trade-off in Table 7 makes the analysis complete in my opinion, and the discussion with ELLA and DATE help clarify the methodological differences. Regarding the approach is reasonable and empirically effective, I will keep my score of 3.

Dear Reviewer fyfg,

Thank you for your constructive feedback and for acknowledging that our response addressed your main concerns. We agree that repeated VAE decoding remains costly and have discussed this limitation in the revised manuscript.

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

This paper presents UniOwl, a unified multimodal framework for visual understanding and image generation based on Adaptive Dual-Channel Conditioning. The authors identify a conditioning dilemma in existing diffusion-based unified models: static conditions cannot adapt to evolving denoising states, while repeatedly invoking an LLM at each denoising step introduces substantial computational overhead.

To address this challenge, the authors propose Diffusion-Adaptive Semantic Conditioning, which caches semantic representations from LLM hidden states and dynamically refines them using lightweight cross-attention residuals conditioned on the current DiT features and timestep. In parallel, Generation-Driven Visual Conditioning extracts visual feedback from the current noisy latent through a shared visual encoder and fuses it with the semantic conditions. The authors further introduce Cycle Consistency Learning, which encourages semantic alignment between original and generated images by minimizing the KL divergence between their output distributions from the understanding pathway.

UniOwl builds upon FastVLM-0.5B and SANA-600M-512 and adopts a three-stage training strategy. Experiments cover multimodal understanding, text-to-image generation, and component ablations. UniOwl achieves a GenEval score of 0.83, improving over Mobile-O's 0.74, while maintaining competitive visual understanding performance.

Critical strengths

1. A practical and efficient approach to diffusion-adaptive semantic conditioning.

The paper introduces a practical approach to dynamic semantic conditioning. By caching LLM-derived semantic representations and refining them through lightweight modules during denoising, UniOwl avoids repeatedly executing the LLM at each diffusion step. This provides a meaningful trade-off between adaptive conditioning and computational efficiency.

The experimental results support the effectiveness of this design. Dynamic Semantic Conditioning improves GenEval from 0.67 to 0.81, while the semantic-only configuration introduces only approximately 6.3% additional inference latency compared with static conditioning. These results demonstrate the practical value of the proposed lightweight semantic refinement mechanism.

2. Competitive generation performance and useful ablation studies.

UniOwl achieves competitive results among the selected compact unified multimodal models, reaching 0.83 on GenEval compared with 0.74 for Mobile-O. The improvements are particularly notable on compositional generation tasks, including Counting, Position, and Color Attribution. Meanwhile, the model maintains strong visual understanding performance.

The authors also provide ablation studies covering Dynamic Semantic Conditioning, Visual Conditioning, Cycle Consistency Learning, and the three-stage training strategy. These experiments offer useful evidence for the effectiveness of the overall architecture and its major components, helping to clarify their contributions to model performance.

Critical weaknesses

1. Further validation of the claimed denoising-state awareness.

Table 3 demonstrates the effectiveness of Dynamic Semantic Conditioning, particularly in improving generation performance. However, a key claim of the proposed mechanism is that it can adapt semantic guidance based on the current DiT denoising state. Since the refinement module uses both DiT features and timestep information, the existing ablation does not clearly distinguish the contribution of actual generation-state feedback from timestep awareness alone.

A simple timestep-only or shuffled-DiT-feature comparison would help clarify whether the improvement specifically comes from accessing the evolving DiT state, providing more direct evidence for the claimed state-aware adaptation.

2. Baseline comparability and training-budget differences.

The paper reports competitive results against the selected baselines, although the coverage and comparability of these evaluations could be further clarified. For example, OpenUni-B-512 also combines an MLLM with SANA for unified understanding and generation and reports a GenEval score of 0.84. Its component-level model scale is broadly comparable to UniOwl, yet it is not included in the comparison. Given potential differences in training and evaluation protocols, clarification of the baseline selection would be helpful.

Furthermore, Mobile-O is the most directly relevant baseline, as both methods use the same main understanding and generation backbones and similar training data. However, UniOwl adopts a longer training schedule. Although the existing component ablations support the effectiveness of the proposed mechanisms, clarifying the differences in training budgets would help better interpret the performance gains over Mobile-O.

Questions for authors

Q1. Does the semantic refinement mechanism specifically benefit from the current DiT state?

A central motivation for Diffusion-Adaptive Semantic Conditioning is to adjust semantic guidance according to the evolving denoising state. Table 3 demonstrates the effectiveness of the full module, but does not separate the contributions of DiT features and timestep information. Could the authors clarify whether the improvements depend on the actual DiT state, rather than timestep awareness alone? If feasible, a timestep-only or shuffled-DiT-feature comparison would help support this claim.

Q2. Baseline selection and evaluation comparability

OpenUni-B-512 reports a GenEval score of 0.84 and shares several architectural characteristics with UniOwl. Could the authors briefly explain why this relevant baseline was not included and clarify the evaluation protocols and sources of the reported baseline results?

In addition, since UniOwl and Mobile-O adopt different training schedules, could the authors clarify these training-budget differences and their potential influence on the performance comparison? This would help better understand the practical gains attributable to the proposed conditioning mechanisms.

1 comment
AuthorsOpenReview ↗

Response to Reviewer wVnQ

Dear Reviewer wVnQ,

We thank you for your thoughtful evaluation and constructive suggestions. We appreciate your recognition of the practical value of our semantic refinement mechanism. We have added experiments and analyses, revised the manuscript, and provided point-by-point responses below.

W1 & Q1: Contribution of the Denoising-State Input

We thank the reviewer for this suggestion. We have added two controls to distinguish denoising-state feedback from timestep awareness. The timestep-only control retains the timestep input but removes latent-state information. The cross-sample intervention replaces the current latent with another sample’s latent at the same timestep during inference. The results are reported below. Timestep-only conditioning achieves 0.70 on GenEval, compared with 0.81 for semantic-only UniOwl. Cross-sample state replacement achieves 0.69, though distribution shift may contribute to the decrease relative to semantic-only UniOwl. These results support the value of current-state feedback beyond timestep awareness in our Diffusion-Adaptive Semantic Conditioning mechanism. We have included these results and their analysis in the revised manuscript.

ConfigurationLatent-state inputGenEval ↑
Timestep-onlyNone0.70
Cross-sample replacementAnother sample’s latent0.69
UniOwl (semantic-only)Current sample’s latent0.81

W2 & Q2: Baseline Comparability and Training Settings

We thank the reviewer for the suggestions on baseline coverage and comparison fairness. We address the comparability of OpenUni-B-512 and the training-setting differences from Mobile-O below.

Discussion of OpenUni-B-512. Thank you for this suggestion. We agree that OpenUni-B-512 [1] is closely related to our work. Although OpenUni-B-512 and UniOwl use the same SANA generator family, their understanding backbones and training data differ. OpenUni-B-512 uses InternVL3-1B and approximately 23M alignment image–text pairs, while UniOwl uses FastVLM-0.5B and 9M pairs from different sources. OpenUni fine-tunes on BLIP3-o-60K with a frozen MLLM. UniOwl additionally uses ShareGPT-4o-Image and performs joint understanding–generation training. These differences should be considered when interpreting the reported results. We have added a discussion of OpenUni in the revised manuscript.

Component / training dataOpenUni-B-512UniOwl
Understanding backboneInternVL3-1BFastVLM-0.5B
Image generatorSANA-0.6B-512SANA-600M-512
Alignment dataApproximately 23M pairs9M pairs
Alignment sourcestext-to-image-2M; LAION-Aesthetic-6M; Megalith-10M; RedCaps-5MJourneyDB (4M); BLIP3o-Short-Caption (5M)
Generation fine-tuning dataBLIP3-o-60KBLIP3-o-60K + ShareGPT-4o-Image (45K)
Joint understanding–generation trainingNone; MLLM remains frozenApproximately 105K quadruplets from the later-stage image sources

Mobile-O Training Settings. Thank you for pointing out the potential influence of training-budget differences. We agree that the published Mobile-O result alone does not isolate this factor. We have therefore retrained Mobile-O under the training and evaluation protocol of our semantic-only configuration. We retained Mobile-O’s original connector, objectives, and stage-specific trainable modules. The training and evaluation settings are matched between Mobile-O and the semantic-only UniOwl variant. The UniOwl variant excludes visual conditioning and cycle consistency learning. Retrained Mobile-O achieves 0.75 on GenEval, compared with 0.81 for semantic-only UniOwl. This 0.06 advantage supports the benefit of our semantic conditioning design beyond the original training-setting differences. The table below documents those original differences. Global batch sizes include gradient accumulation.

StageModelTraining durationGlobal batch sizeLearning rate
Stage 1Mobile-O5 epochs5122e-4
UniOwl60k steps5122e-4
Stage 2Mobile-O20 epochs3842e-4
UniOwl25 epochs5122e-4
Stage 3Mobile-O7 epochs1281e-4
UniOwl10 epochs1281e-6

Reference

[1] Wu S, Wu Z, Gong Z, Tao Q, Jin S, Li Q, Li W, Loy CC. OpenUni: A Simple Baseline for Unified Multimodal Understanding and Generation. arXiv:2505.23661, 2025.

Reviewer_DE6AOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents UniOwl, a compact unified multimodal model connecting VLM and DiT generator through adaptive dual-channel conditioning. At each diffusion step, a semantic channel refines cached LLM condition tokens via lightweight cross-attention over current DiT features, while a visual channel decodes the noisy latent to extract visual feedback through the shared vision encoder. Combined with a cycle consistency loss during joint training, UniOwl maintains competitive VQA performance on seven understanding benchmarks while improving text-to-image accuracy on GenEval.

Critical strengths
  1. The motivation is clear, addressing the trade-off in unified diffusion generation between static text conditions that cannot adapt across denoising steps and the high cost of re-running the LLM at every step.

  2. Caching base semantic tokens once and refining them against DiT features at each step is simple and effective, accounting for almost the entire GenEval gain while adding only 6.3% inference latency.

  3. Within a compact 1.7B parameter budget, the framework supports both visual understanding and generation, preserving FastVLM's VQA performance while substantially improving compositional generation on GenEval.

Critical weaknesses
  1. The relationship to Mobile-O could be made clearer. UniOwl builds on Mobile-O's backbone, connector, data, and training recipe, and adds three modules. It would help to state this in the main text. The current experiments do not yet separate the effect of the new modules from that of the inherited recipe and longer training. For example, the ablation baseline is 0.67 on GenEval, while Mobile-O reports 0.74.

  2. The design does not match the stated mechanisms well. The method is motivated by efficiency, but the visual channel adds VAE decoding and vision encoding at every step. As a result, the full model is slower than the per-step LLM baseline it argues against.

Questions for authors
  1. How does UniOwl compare with Mobile-O retrained under the same training budget?

  2. What GenEval scores do UniOwl-Latent and UniOwl-Lite achieve?

I am willing to raise my score if the authors can convincingly address.

4 comments
AuthorsOpenReview ↗

Response to Reviewer DE6A (Reply 1/2)

Dear Reviewer DE6A,

We thank you for your careful evaluation and constructive feedback. We appreciate your recognition of the motivation and effectiveness of our diffusion-adaptive semantic conditioning mechanism. We have added experiments and analyses, revised the manuscript, and provided point-by-point responses below.

W1 & Q1: Methodological Differences and Training-Budget Comparability

Thank you for raising this concern. We clarify below which components UniOwl shares with Mobile-O and how our method and training settings differ. We then explain why the Base configuration in Table 3 is retrained under the UniOwl ablation protocol, and separately address the requested matched-budget comparison with Mobile-O.

Methodological Differences. UniOwl uses the same FastVLM-0.5B understanding backbone, SANA-600M-512 generator, and training-data sources as Mobile-O. Both methods employ three training stages, but differ in their stage-specific trainable components and objectives. UniOwl introduces Adaptive Dual-Channel Conditioning, comprising Diffusion-Adaptive Semantic Conditioning and Generation-Driven Visual Conditioning, together with Cycle Consistency Learning. Diffusion-Adaptive Semantic Conditioning refines cached LLM-derived semantic conditions using the current denoising state and timestep. Generation-Driven Visual Conditioning extracts visual feedback from intermediate generations through the shared visual encoder. Cycle Consistency Learning encourages semantic agreement between the understanding outputs for original and generated images. We have added a discussion of these methodological differences in the revised manuscript.

SettingMobile-OUniOwl
Understanding backboneFastVLM-0.5BFastVLM-0.5B
Image generatorSANA-600M-512SANA-600M-512
Alignment dataJourneyDB (4M) + BLIP3o-Short-Caption (5M)Same sources
Later-stage image sourcesBLIP3-o-60K + ShareGPT-4o-Image (45K)Same sources
Semantic conditioningStatic MCP conditionsDiffusion-Adaptive Semantic Conditioning
Generation-driven visual conditioningNoneShared-encoder feedback from intermediate generations
Cycle consistency objectiveNoneCycle Consistency Learning

Training-Budget Differences. UniOwl and Mobile-O use the same understanding backbone, image generator, and training-data sources. However, their training budgets and optimization settings differ. UniOwl uses 60k steps for pre-training, 25 epochs for SFT, and 10 epochs for unified post-training. Mobile-O uses 5, 20, and 7 epochs. The SFT global batch size is 512 for UniOwl and 384 for Mobile-O. During unified post-training, both models use a global batch size of 128 and LoRA rank/alpha 16/32. The learning rate is 1e-6 for UniOwl and 1e-4 for Mobile-O. Global batch sizes include gradient accumulation.

StageModelTraining durationGlobal batch sizeLearning rate
Stage 1Mobile-O5 epochs5122e-4
UniOwl60k steps5122e-4
Stage 2Mobile-O20 epochs3842e-4
UniOwl25 epochs5122e-4
Stage 3Mobile-O7 epochs1281e-4
UniOwl10 epochs1281e-6

Clarifying the Base Configuration in Table 3. The 0.67 result in Table 3 is obtained from UniOwl-Base, an internal ablation configuration retrained under the same training and evaluation protocol as the other UniOwl variants. It is not a reproduction of Mobile-O, whose published GenEval score is 0.74. Specifically, even without diffusion-adaptive semantic refinement, UniOwl-Base uses a cross-attention-based extractor to obtain conditioning tokens from LLM hidden states, rather than Mobile-O’s original MCP connector. To further address the concern that the improvement over Mobile-O may arise from differences in training budget, we have retrained Mobile-O under the training and evaluation protocol of our semantic-only configuration, retaining its original connector, training objectives, and stage-specific trainable modules. The training and evaluation settings are matched between Mobile-O and the semantic-only UniOwl variant. The UniOwl variant excludes visual conditioning and cycle consistency learning. The retrained Mobile-O achieves 0.75 on GenEval, while UniOwl with Diffusion-Adaptive Semantic Conditioning achieves 0.81. Thus, the semantic-only configuration retains a 0.06 GenEval advantage under the matched protocol, supporting that its improvement over Mobile-O is not explained by the training-setting differences alone. Table 3 separately evaluates the contributions of the proposed components within UniOwl. We have added a description of UniOwl-Base in the revised manuscript.

Our responses to W2 & Q2 are provided in Reply 2/2.

AuthorsOpenReview ↗

Response to Reviewer DE6A (Reply 2/2)

W2 & Q2: Generation Quality and Inference Efficiency We thank the reviewer for highlighting the computational cost of visual feedback and requesting the generation results of UniOwl-Latent and UniOwl-Lite. We first clarify the scope of our efficiency claim and the cost of visual feedback, then report the GenEval scores of these configurations alongside their inference times. The details are provided below.

Efficiency of semantic refinement. UniOwl-Latent refines cached semantic conditions using the current denoising state and timestep, without per-step LLM execution or pixel-space visual feedback. Its measured generation time is 771.9 ms, compared with 725.9 ms for the fixed-condition configuration, an incremental overhead of 6.3%. It is also lower than the 1201.0 ms measured for step-wise LLM conditioning. These results support the low incremental cost of the semantic refinement mechanism. UniOwl-Full takes 2725.0 ms, exceeding the 1201.0 ms of step-wise LLM conditioning because it additionally decodes and visually encodes intermediate latents at every step. The low-overhead claim therefore applies to semantic refinement, not the full visual-feedback configuration.

Visual feedback and generation quality. We thank the reviewer for this suggestion. We have added GenEval evaluations of UniOwl-Latent and UniOwl-Lite, which achieve 0.81 and 0.82, respectively. The table below pairs these scores with their inference times. Timing measurements cover the instrumented generation-stage GPU operations, including final decoding, and exclude the initial LLM prompt-encoding pass. Compared with Full, Lite reduces the measured generation time by 54.8%, with a 0.01 lower GenEval score, demonstrating the trade-off between visual feedback frequency, compositional accuracy, and computational cost. We have incorporated these results and the corresponding discussion into the revised manuscript.

ConfigurationVisual feedbackGenEval ↑Measured generation time (ms) ↓
UniOwl-LatentNone0.81771.9 ± 9.5
UniOwl-Lite (k=4)Every four steps0.821230.4 ± 0.5
UniOwl-FullEvery step0.832725.0 ± 24.8

Thank you for the additional experiments and clarifications. The comparison with Mobile-O under matched settings supports adaptive semantic conditioning, and the Latent/Lite results clarify the trade-off between quality and computational cost. My main concerns are addressed. I encourage a clearer explanation of how cycle consistency affects generation, and I am raising my score to 3.

Dear Reviewer DE6A,

Thank you for your constructive feedback and positive reassessment. Following your suggestion, we have further clarified the role of cycle consistency in generation in Section 3.4 and the Training Schemes analysis of the revised manuscript. We appreciate your careful evaluation.

Comments on the paper

1 comment

We sincerely thank all reviewers for their time and constructive feedback. We appreciate their recognition of the following contributions of our work:

  • Practical adaptive semantic conditioning (DE6A, wVnQ, fyfg). UniOwl caches LLM-derived semantic representations and refines them during denoising through lightweight state- and timestep-conditioned updates. All three reviewers recognized the effectiveness of this mechanism, which avoids repeated LLM execution.
  • Competitive generation and understanding within a compact model (DE6A, wVnQ, fyfg). UniOwl achieves competitive performance in compositional generation and visual understanding within a compact unified model.
  • Useful architectural and training analysis (wVnQ). The component and training-stage ablations provide evidence for the overall design and help explain the contributions of its major components.

The reviewers’ suggestions also helped us strengthen the experimental evidence and improve the manuscript. We summarize the main additions and revisions below:

  • Matched-training comparison (DE6A, wVnQ; Appendix D). Retrained Mobile-O achieves 0.75 GenEval versus 0.81 for semantic-only UniOwl under matched training and evaluation settings, supporting an advantage beyond the original training-setting differences.
  • Denoising-state controls (wVnQ; Appendix D). Timestep-only conditioning achieves 0.70, compared with 0.81 using the current sample’s state. Cross-sample replacement yields 0.69, providing complementary dependency evidence subject to possible distribution shift.
  • Quality–cost analysis (DE6A, fyfg; Appendix E). Latent, Lite, and Full achieve 0.81, 0.82, and 0.83 GenEval, respectively. Lite reduces measured generation time by 54.8% relative to Full with a 0.01 lower score. We clarify the scope of the efficiency claim and discuss the cost of dense visual feedback in Appendix F.
  • Clearer explanations. We expanded the discussion of ELLA, DATE, and OpenUni (Section 2), clarified UniOwl-Base (Section 4.3), and further explained cycle consistency in the method and training-stage analysis (Sections 3.4 and 4.3).

These revisions have been incorporated into the manuscript, with changes marked in blue. All three reviewers have confirmed that their main concerns were addressed. We thank the reviewers and the Area Chair for their careful consideration.

Best,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.