acceptodds
Under review as a conference paper at ICLR 2027

UniRect-CoT: Enhancing Generation in Unified Multimodal Models via Reflective Rectification with Inherent Understanding

Abstract

Unified Multimodal Models (UMMs) aim to integrate visual understanding and generation within a single structure. However, these models exhibit a notable capability mismatch, where their understanding capability often outperforms their generation capability. This mismatch suggests that the model's rich internal knowledge, while effective for understanding tasks, is not fully utilized during generation. To address this, we draw inspiration from the human “Thinking-While-Drawing” paradigm, where humans continuously reflect on intermediate results and revise them according to their understanding of the intended target. In this paper, we propose UniRect-CoT, a training-free unified rectification chain-of-thought framework. We regard the multi-step denoising process in UMMs as an intrinsic visual reasoning process, whose intermediate states provide opportunities for continuous reflection and rectification. By leveraging the UMM's inherent understanding to reflect on intermediate results in light of the target instruction and rectify them accordingly, our framework forms a reflective chain of thought within a single denoising trajectory, enabling the model's internal knowledge to guide generation throughout the process. Specifically, the UMM generates a state-conditioned reflection, whose autoregressive negative log-likelihood under the same model defines a model-native differentiable semantic loss that translates textual reflection into gradients for latent rectification. Extensive experiments demonstrate that UniRect-CoT generalizes across existing flow-based UMMs, consistently enhancing overall text-to-image generation performance.

Primary area:
generative models
Keywords:
Unified Multimodal Models, Image Generation

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_Ci3pOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a training-free inference framework for unified multimodal models. During denoising, it estimates a clean image, asks the same model to produce a state-dependent target description, and backpropagates the description's autoregressive negative log-likelihood into the latent representation. GITO performs multiple local updates, after which candidates are selected using the likelihood of the original user instruction before denoising resumes. The main contribution is the combination of model-native reflection, a differentiable semantic objective, and candidate selection, rather than a new flow-matching or gradient-guidance principle. Experiments cover four flow-based models on GenEval and DPG-Bench, with analyses of semantic targets, intervention windows, and inference cost.

Critical strengths
  1. The central design is supported by targeted controls, not just a reflection narrative. With selection based on the user instruction, Table 3 reports 79.1 for directly optimizing the original instruction, 78.1 for text-only expansion, 79.4 for static reflection, and 80.4 for dynamic reflection. Using the reflective target for selection instead reduces the score to 78.8. These comparisons support both image-conditioned feedback and the separation between reflection for proposing corrections and the original instruction for selecting candidates. The paper explicitly acknowledges its use of existing gradient-guidance mechanisms and generally positions the contribution appropriately (Section 5.4).
  2. The cross-model improvements and complementarity are practically meaningful. All four models improve in overall score on both benchmarks. GenEval gains range from 1.0 percentage point for OmniGen2 to 7.1 points for Janus-Flow, while BAGEL's counting score rises from 74.3 to 81.9 (Table 1). BAGEL Think further improves from 79.7 to 81.3, suggesting that trajectory rectification can complement rather than merely replace pre-generation reasoning (Table 2). The evidence spans different baseline strengths, although applicability should remain limited to the evaluated flow-based UMMs.
  3. The appendix meaningfully examines the repair–preservation trade-off. On paired samples, the paper reports 86 repaired failures and 28 regressions among previously correct outputs, yielding a net gain of 58 and a regression rate of 1.65%, rather than presenting only successful examples (Table 7). Efficiency is also reported in absolute terms: the method takes 43.2 seconds per image versus 24.7 for the baseline, but compares favorably with the evaluated Best-of-3 configuration at 76.9 seconds and a score of 80.1, versus 80.4 for UniRect-CoT (Table 6). This provides useful evidence for the quality–cost trade-off.
Critical weaknesses
  1. Evidence for proxy reliability is weaker than evidence for final performance. Figure 3 reports only aggregate accuracy when judging whether intermediate images satisfy the instruction, without the proportions of correct and incorrect examples, confusion matrices, or category-level results. It also does not directly test whether the candidate ranking in Equation (11) predicts improvements in the final image (Sections 4.1–4.3). The paper already acknowledges that likelihood is a proxy, and quantifies regressions, so the concern is not an absence of failure discussion. Rather, trajectory-level diagnostics are missing for incorrect reflections and cases where proxy scores improve while final outputs deteriorate. Moreover, BAGEL's position score decreases from 49.0 to 48.5 (Table 1), so the aggregate gain should not be interpreted as improving every compositional capability, much less overall visual quality based on alignment metrics alone.
  2. Statistical robustness and reproduction details limit the strength of comparative claims. Table 1 specifies four consecutive samples per prompt from a fixed initial seed, but Tables 1–3 do not report repeated-run variation or paired confidence intervals. The overall advantages over TiR and Best-of-3 are only 0.6 and 0.3 percentage points, respectively (Tables 2 and 6). The paper sweeps update counts and intervention windows, but does not explain whether the preferred settings were selected using independent validation prompts; this is a clarification request, not evidence of test-set leakage.
Questions for authors
  1. Could you provide prompt-paired confidence intervals or results across multiple seed sets for the key differences in Tables 1–3, and clarify whether the hyperparameters in Tables 4–5 were selected on an independent validation set?
  2. What are the label distributions, balanced accuracy, and error types at each timestep in Figure 3? How strongly does the candidate ranking in Equation (11) correspond to final GenEval correctness?
7 comments
AuthorsOpenReview ↗

Response to Reviewer Ci3p (Part 1)

We sincerely thank reviewer Ci3p for the thoughtful and constructive feedback, which will help improve our work. Our responses are as follows.

W1.1:  Figure 3 reports only aggregate accuracy when judging whether intermediate images satisfy the instruction, without the proportions of correct and incorrect examples, confusion matrices, or category-level results.

A: We provide additional quantitative and qualitative analyses to further examine the reliability of our proxy guidance and clarify its limitations.

Intermediate Judgment Reliability for Figure 3: We collect samples using GenEval prompts with different random seeds across the four UMMs. To avoid class imbalance, we construct a balanced evaluation set by selecting 500 samples that are correct across all four UMMs and 500 samples that are incorrect across all four UMMs, resulting in 1,000 samples with a 1:1 correct-to-incorrect ratio. At each timestep, we ask each UMM to predict whether the final generated image will satisfy the prompt, using its final GenEval flag as the ground truth at every timestep. We further report the confusion matrices across denoising intervals in Table 1. The accuracy increases from 34.1% in steps [0,5) to 73.3% in [5,10), showing that informative judgments emerge well before generation is complete.

Table 1: Confusion matrix and judgment accuracy across different denoising timesteps.

[0,5)[5,10)[10,20)[20,30)[30,49]
TP(%)17.236.639.440.639.8
FN(%)32.913.410.69.410.2
FP(%)33.113.310.59.310.1
TN(%)16.936.739.540.839.9
accuracy(%)34.173.378.881.379.8

W1.2: It also does not directly test whether the candidate ranking in Equation (11) predicts improvements in the final image (Sections 4.1–4.3).

A: To examine whether Eq. (11) predicts improvements in the final image, we conduct the following experiment. At each timestep within the rectification window, we independently continue denoising from each candidate state to obtain its final image. We then use two signals to evaluate each candidate: (1) its Eq. (11) likelihood score, computed on the estimated clean image at that timestep; and (2) the GenEval correctness of its final image, which serves as the ground-truth label.  This allows us to assess whether Eq. (11) reliably ranks intermediate candidates according to their eventual generation correctness. As shown in Tab. 2, candidates that eventually produce correct images consistently receive higher likelihood scores than those producing incorrect images across all evaluated timesteps, with average scores of -3.01 and -3.53, respectively (a mean gap of 0.52). These results demonstrate that Eq. (11) assigns higher scores to candidates leading to correct final images, supporting its reliability for ranking intermediate candidates during rectification.

Table 2: Candidate likelihood scores by final correctness within the rectification window.

t=5t=6t=7t=8t=9t=10Overall
Correct-3.25-3.08-3.00-2.94-2.91-2.87-3.01
Incorrect-3.72-3.56-3.52-3.46-3.47-3.45-3.53
Score gap (Δ)0.470.480.520.520.560.580.52
AuthorsOpenReview ↗

Response to Reviewer Ci3p (Part 2)

W1.3: The paper already acknowledges that likelihood is a proxy, and quantifies regressions, so the concern is not an absence of failure discussion. Rather, trajectory-level diagnostics are missing for incorrect reflections and cases where proxy scores improve while final outputs deteriorate.

A: We conduct a systematic trajectory-level failure analysis and identify three major failure modes, illustrating how errors emerge and propagate throughout the generation process:

(1) Reflection Hallucination: The UMM may be misled by intermediate images, producing inaccurate reflective targets or hallucinated descriptions, which drive subsequent gradient updates in incorrect directions. As shown in Fig. 8 of the revised manuscript, the UMM mistakenly treats the current image as the desired result and describes a spatial relationship that contradicts the prompt. This incorrect reflection provides little corrective guidance, ultimately leading to failure.

(2) Incomplete Rectification: Even with an accurate reflective target, the limited rectification capacity at each timestep may prevent errors from being fully corrected within the predefined update budget. As shown in Fig. 9 of the revised manuscript, although gradient updates turn most of the carrot blue, the limited rectification capacity leaves part of it orange, resulting in an incorrect final image.

(3) GSS Misranking: When correct and incorrect candidates are visually similar, the UMM may struggle to distinguish them, assigning higher scores to incorrect candidates and selecting suboptimal latent states. As shown in Fig. 10 of the revised manuscript, GSS selects an incorrect candidate despite the availability of a correct one, ultimately producing five bowls instead of the requested number.

We have included this trajectory-level failure analysis, together with the corresponding visualizations, in Appendix G of the revised manuscript.

W1.4: Moreover, BAGEL's position score decreases from 49.0 to 48.5 (Table 1), so the aggregate gain should not be interpreted as improving every compositional capability, much less overall visual quality based on alignment metrics alone.

A: We clarify the scope of our performance claims and provide additional multi-seed results. The original results in the paper use a single random seed 42, where BAGEL shows a slight Position decrease (49.0 to 48.5), while the other three backbones improve. To further assess robustness, we conduct experiments with four random seeds(42, 1234, 2026, 3407). As shown in Table 3, UniRect-CoT consistently improves nearly all compositional categories across the four backbones, with BAGEL's average Position score increasing from 48.4 to 49.5. These results provide stronger evidence of improvements across compositional capabilities beyond the aggregate score. Additionally, Figures 4 and 6 in the original manuscript provide qualitative visual comparisons, further illustrating the visual improvements achieved by UniRect-CoT.

Table 3: Multi-seed evaluation across UMM backbones.

MethodSingle Obj.Two Obj.CountingColorsPositionColor Attr.Overall
JanusFlow94.1±1.244.0±3.131.7±3.275.5±2.031.4±2.925.7±1.350.4±1.9
w/ Ours95.3±1.156.8±3.041.3±1.681.4±0.341.1±2.933.9±1.358.3±1.2
OpenUni-RecA98.9±0.491.1±0.946.7±1.487.4±0.340.0±1.667.9±0.272.0±0.3
w/ Ours99.2±0.393.4±0.955.7±1.689.6±0.345.0±2.471.1±0.975.7±0.7
OmniGen299.6±0.293.2±0.570.7±2.086.8±0.553.0±0.966.0±1.678.2±0.5
w/ Ours99.5±0.295.1±0.672.1±0.687.1±0.553.9±1.768.6±2.179.4±0.3
Bagel99.1±0.295.0±0.875.5±1.886.3±1.248.4±1.062.9±2.177.9±0.2
w/ Ours99.7±0.396.3±0.980.7±2.587.5±0.549.5±1.768.4±0.780.4±0.4
AuthorsOpenReview ↗

Response to Reviewer Ci3p (Part 3)

W2.1: Table 1 specifies four consecutive samples per prompt from a fixed initial seed, but Tables 1–3 do not report repeated-run variation or paired confidence intervals. The overall advantages over TiR and Best-of-3 are only 0.6 and 0.3 percentage points, respectively (Tables 2 and 6).

A: We provide additional multi-seed evaluations to assess statistical robustness, compare performance and efficiency against TiR and Best-of-3, and clarify our experimental settings for reproducibility. We have included the results in Appendix H and J of the revised manuscript.

Statistical Robustness. To assess statistical robustness, we conduct additional experiments for Tables 1-3 in the original paper using four common random seeds (42, 1234, 2026, and 3407), and report the results as mean ± standard deviation. As shown in Table 4-6, UniRect-CoT statistically improves generation performance across different UMM backbones, outperforms the compared training-free test-time methods, and validates the effectiveness of its key design components.

Comparison with Best-of-3 and TiR. As shown in the Table 5, UniRect-CoT achieves 80.4 ± 0.4, outperforming TiR (79.5 ± 0.5) and Best-of-3 (80.1 ± 1.1). It also exhibits lower variance, indicating more stable performance. Moreover, UniRect-CoT requires only 43.2 seconds per image, compared with 116.9s for TiR and 76.9s for Best-of-3, demonstrating a better balance between generation quality and inference efficiency.

Table 4: Multi-seed evaluation across UMM backbones.

MethodSingle Obj.Two Obj.CountingColorsPositionColor Attr.Overall
JanusFlow94.1±1.244.0±3.131.7±3.275.5±2.031.4±2.925.7±1.350.4±1.9
w/ Ours95.3±1.156.8±3.041.3±1.681.4±0.341.1±2.933.9±1.358.3±1.2
OpenUni-RecA98.9±0.491.1±0.946.7±1.487.4±0.340.0±1.667.9±0.272.0±0.3
w/ Ours99.2±0.393.4±0.955.7±1.689.6±0.345.0±2.471.1±0.975.7±0.7
OmniGen299.6±0.293.2±0.570.7±2.086.8±0.553.0±0.966.0±1.678.2±0.5
w/ Ours99.5±0.295.1±0.672.1±0.687.1±0.553.9±1.768.6±2.179.4±0.3
Bagel99.1±0.295.0±0.875.5±1.886.3±1.248.4±1.062.9±2.177.9±0.2
w/ Ours99.7±0.396.3±0.980.7±2.587.5±0.549.5±1.768.4±0.780.4±0.4

Table 5: Multi-seed comparison with test-time methods.

MethodSingle Obj.Two Obj.CountingColorsPositionColor Attr.OverallLatency (s)
Baseline99.1±0.295.0±0.875.5±1.886.3±1.248.4±1.062.9±2.177.9±0.224.7
Qwen Rewrite99.5±0.594.7±0.875.2±2.287.9±2.148.1±0.965.3±1.178.5±0.328.7
UiG99.4±0.493.8±1.476.1±1.686.7±1.149.8±2.362.8±1.078.1±0.156.4
TiR99.1±0.694.3±1.380.2±1.685.0±1.252.5±2.366.1±1.779.5±0.5116.9
Best-of-399.4±0.496.3±0.680.6±2.387.4±0.850.9±2.666.1±3.580.1±1.176.9
UniRect-CoT99.7±0.396.3±0.980.7±2.587.5±0.549.5±1.768.4±0.780.4±0.443.2

Table 6: Multi-seed ablation of semantic target design and GSS.

SettingUpdateITE TargetGSS TargetSingle Obj.Two Obj.CountingColorsPositionColor Attri.Overall ↑
Baseline–––99.1±0.295.0±0.875.5±1.886.3±1.248.4±1.062.9±2.177.9±0.2
ITE OnlyStaticUser–99.4±0.395.0±1.277.1±2.686.9±0.748.1±1.266.2±0.878.8±0.3
StaticExp.–99.6±0.593.6±0.576.9±1.686.3±0.847.9±2.664.2±3.278.1±0.3
StaticRef.–99.7±0.493.8±2.178.9±2.487.9±1.548.7±4.465.4±2.879.1±0.2
DynamicRef.–99.7±0.494.5±1.678.5±1.989.2±1.248.3±2.666.9±2.979.5±0.3
ITE+GSSStaticUserUser99.2±0.495.0±0.878.3±1.987.5±1.049.6±1.665.7±1.379.2±0.2
StaticExp.User99.5±0.294.2±0.777.7±2.086.8±0.848.9±2.063.7±2.878.4±0.3
StaticRef.User99.7±0.394.8±1.877.4±1.787.0±0.749.8±3.867.9±1.979.5±0.5
DynamicRef.Ref.99.4±0.494.3±2.178.4±2.587.4±1.849.9±3.364.6±2.779.0±0.5
DynamicRef.User99.7±0.396.3±0.980.7±2.587.5±0.549.5±1.768.4±0.780.4±0.4

W2.2: The paper sweeps update counts and intervention windows, but does not explain whether the preferred settings were selected using independent validation prompts; this is a clarification request, not evidence of test-set leakage.

A: We provide the details of the reproduction below. All four hyperparameters are shared across the four UMM backbones without model-specific tuning. Specifically, we use η = 200, δ = 0.001, W = [5, 10], and K = 3 for Janus-Flow, OpenUni-RecA, OmniGen2, and BAGEL. For each backbone, we follow its official inference configuration. This demonstrates the generalizability of our method. We have added these implementation details to Appendix J of the revised manuscript to facilitate reproducibility.

Reviewer_i7EaOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Unified multimodal models (UMMs) often understand images better than they generate them. Existing approaches use this understanding mainly before generation (e.g., prompt rewriting or planning) or across multiple full generation rounds, and overlook interaction between the two capabilities during the denoising process itself. This paper introduces UniRect-CoT, a training-free framework that interleaves reflection and rectification within a single denoising trajectory. At selected steps, the UMM reflects on the look-ahead image, and the likelihood of that reflection under the same model is used as a differentiable loss to rectify the latent. A greedy selection step then chooses among the rectified candidates according to the original prompt.

Critical strengths

The paper proposes a training-free, model-native rectification scheme. The same UMM supplies the reflection, the differentiable semantic loss (Eq. 6), and the selection score (Eq. 11), with no external reward model and no additional training.

UniRect-CoT consistently improves GenEval and DPG-Bench on four different flow-based UMMs (Janus-Flow, OpenUni-RecA, OmniGen2, BAGEL). It also shows complementarity with BAGEL's Think mode (79.7 → 81.3).

Beyond the main results, the paper includes informative analyses: the ablation on guidance and selection targets (Tab. 3), the performance–latency comparison with other test-time strategies (Tab. 6), and the repair/regression analysis (Tab. 7), which shows a low regression rate.

Critical weaknesses

Sec. 2.2 states that existing methods "mainly reason before generation or refine across multiple generation rounds." However, Thinking-while-Generating (TwiG, CVPR 2026) also interleaves textual reasoning and reflection throughout visual generation within a single trajectory. Likewise, using the UMM's likelihood of the original prompt as a score (Eq. 11) has been used as an intrinsic reward in Learning to Generate via Understanding (CVPR 2026). The paper should discuss these works and clarify its differences from them. It should also cover reward-guided latent optimization methods such as RewardFlow and ReNO.

In Tab. 1, several reproduced baselines are noticeably lower than the numbers reported in their original papers. Janus-Flow gets 0.530 GenEval / 73.0 DPG versus 0.63 / 80.09 reported; BAGEL gets 0.776 versus 0.82 reported. The improved results are still below the reported baseline numbers. Could the authors clarify how their inference settings differ (resolution, CFG, etc.)? Tab. 1 also mixes reported and reproduced numbers, which makes the comparison hard to read.

How exactly is the judgment accuracy in Fig. 3 computed, and what is the ratio of correct to incorrect samples? About 77% of BAGEL's final samples are already correct (1702/2212, Tab. 7), so a judge that always answers "yes" would score close to the plateau in Fig. 3. Reporting accuracy on failure cases would make this motivation more convincing.

The update rule (Eqs. 8–9) is somewhat unclear. With δ=0.001\delta = 0.001 and η=200\eta = 200, the update becomes a fixed-norm step of ηδ=0.2\eta\delta = 0.2 whenever ∣gt∣2>δ|g_t|_2 > \delta. Is the clipping effectively always active? In addition, Sec. 3 introduces guidance as a velocity modification with γ\gamma, but the method directly updates ztz_t (Eq. 9). The relation between the two should be clarified.

In Fig. 5(c) and Tab. 6, which verifier does Best-of-N use to select among its samples? If it differs from the score in Eq. 11, the comparison with GSS may not be like-for-like.

Questions for authors

How many rectification steps does W=[5,10]W=[5,10] contain? Are η\eta, δ\delta, WW and KK shared across all four UMMs, or tuned per model?

Have you tried a control in which the CSA gradient is replaced by a random direction of the same norm, with GSS kept? This would help separate the effect of the semantic gradient from the effect of candidate selection.

Could you provide a latency breakdown (reflection generation, backward passes, GSS scoring) and the peak GPU memory?

6 comments
AuthorsOpenReview ↗

Response to Reviewer i7Ea (Part 1)

We sincerely thank reviewer i7Ea for the thoughtful and constructive feedback, which will help improve our work. Our responses are as follows.

W1: Sec. 2.2 states that existing methods "mainly reason before generation or refine across multiple generation rounds." However, Thinking-while-Generating (TwiG, CVPR 2026) also interleaves textual reasoning and reflection throughout visual generation within a single trajectory. Likewise, using the UMM's likelihood of the original prompt as a score (Eq. 11) has been used as an intrinsic reward in Learning to Generate via Understanding (CVPR 2026). The paper should discuss these works and clarify its differences from them. It should also cover reward-guided latent optimization methods such as RewardFlow and ReNO.

A: We conduct a systematic comparison of UniRect-CoT with TwiG [1], GvU [2], RewardFlow [3], and ReNo [4]. We have added the corresponding discussion to Sec. 2 and the comparative experiments to Appendix I of the revised manuscript.

(1) Comparison with TwiG: Both TwiG and UniRect-CoT explore understanding capability during generation, but TwiG operates at the autoregressive regional level and relies on prompt optimization, whereas UniRect-CoT performs timestep-level and uses the UMM's native CE loss to rectify latent states. For a fairer comparison, we adapt TwiG to our evaluation setting. As shown in Table 1, UniRect-CoT outperforms TwiG (80.4 vs. 79.2), demonstrating the effectiveness of our approach.

(2) Comparison with GvU: GvU uses the CE loss of static user prompts to optimize model parameters under a training setting, whereas UniRect-CoT uses dynamically state-conditioned reflections to optimize latent states in a training-free manner. To provide a more direct comparison, we replace our reflective target with GvU's original user-prompt objective while keeping the latent optimization framework unchanged. As shown in Table 1, our method achieves 80.4, outperforming this variant (79.1). This highlights the importance of state-conditioned reflection for latent rectification.

(3) Comparison with RewardFlow and ReNO: Both RewardFlow and ReNO follow an external reward-driven paradigm, relying on external pretrained reward models to guide generation through reward gradients, overlooking the UMMs' inherent and powerful understanding capabilities. In contrast, UniRect-CoT follows a model-native understanding-driven paradigm, leveraging the UMM's inherent understanding capabilities to guide its own generation. As shown in Table 1, UniRect-CoT(80.4) outperforms ReNO (79.8) and RewardFlow (79.0), highlighting the potential of understanding-driven generation.

Table 1: More comparisons with related works.

MethodSingle Obj.Two Obj.CountingColorsPositionColor Attr.Overall
Baseline99.195.774.386.749.061.077.6
TwiG99.793.278.183.258.562.379.2
RewardFlow99.495.580.384.648.566.079.0
ReNO99.794.483.786.450.863.879.8
GvU Variant99.495.276.588.351.064.379.1
Ours99.797.081.987.248.568.080.4

[1] Thinking-while-Generating: Interleaving Textual Reasoning throughout Visual Generation, CVPR 2026.

[2] Learning to Generate via Understanding: Understanding-Driven Intrinsic Rewarding for Unified Multimodal Models, CVPR 2026.

[3] RewardFlow: Generate Images by Optimizing What You Reward, CVPR 2026.

[4] ReNO: Enhancing One-step Text-to-Image Models through Reward-based Noise Optimization, NeurIPS 2024.

AuthorsOpenReview ↗

Response to Reviewer i7Ea (Part 2)

W2: In Tab. 1, several reproduced baselines are noticeably lower than the numbers reported in their original papers. Janus-Flow gets 0.530 GenEval / 73.0 DPG versus 0.63 / 80.09 reported; BAGEL gets 0.776 versus 0.82 reported. The improved results are still below the reported baseline numbers. Could the authors clarify how their inference settings differ (resolution, CFG, etc.)? Tab. 1 also mixes reported and reproduced numbers, which makes the comparison hard to read.

A: We clarify our inference settings and provide independent reproduction evidence as follows. We have also added the corresponding implementation details to Appendix J of the revised manuscript

Inference settings. We follow the official inference implementations: BAGEL uses 1024×1024 resolution, CFG text scale 4; Janus-Flow uses 384×384 resolution, CFG weight 2.0. For both models, we fix the initial seed to 42 and generate four consecutive samples per prompt. Importantly, each baseline and its UniRect-CoT counterpart use identical inference settings and initial noise seeds, ensuring fair paired comparisons.

Independent reproduction evidence for Janus-Flow. The official repository does not provide a dedicated GenEval evaluation script or sufficient reproduction details. Our results (0.530 GenEval / 73.0 DPG) even surpass those independent reproductions reported by TorchUMM [1], which obtains 0.500 GenEval / 72.0 DPG.

Independent reproduction evidence for BAGEL. Similar discrepancies have been reported by the community. In the official BAGEL GitHub Issue #123 [2], multiple users reported GenEval scores around 0.78, with independent reproductions ranging from 0.775 to 0.779. Additionally, TorchUMM [1] and FlowCast [3] also report 0.78, similar to our reproduced result of 0.776(~0.78).

Table Presentation. Following your suggestion, we replaced the † markers for reproduced results in Table 1 of the revised manuscript with *, making the distinction clearer and improving readability.

[1] TorchUMM: A Unified Multimodal Model Codebase for Evaluation, Analysis, and Post-training. arXiv preprint arXiv:2604.10784, 2026.

[2] BAGEL, GitHub Issue #123, https://github.com/ByteDance-Seed/Bagel/issues/123.

[3] FlowCast: Trajectory Forecasting for Scalable Zero-Cost Speculative Flow Matching. International Conference on Learning Representations (ICLR), 2026

W3: How exactly is the judgment accuracy in Fig. 3 computed, and what is the ratio of correct to incorrect samples? About 77% of BAGEL's final samples are already correct (1702/2212, Tab. 7), so a judge that always answers "yes" would score close to the plateau in Fig. 3. Reporting accuracy on failure cases would make this motivation more convincing.

A: We clarify the evaluation protocol of Fig. 3 and provide additional failure-case accuracy results.

(1) Evaluation Set Construction and Sample Ratio. We collect samples using GenEval prompts with different random seeds across the four UMMs. To avoid class imbalance, we construct a balanced evaluation set by selecting 500 samples that are correct across all four UMMs and 500 samples that are incorrect across all four UMMs, resulting in 1,000 samples with a 1:1 correct-to-incorrect ratio.

(2) Judgment Accuracy Computation. For each generated image, we retain its complete denoising trajectory and final GenEval correctness label as the ground truth at every timestep. At each timestep t, we estimate the corresponding clean image and ask the UMM to predict whether the final generated image will satisfy the user prompt. We calculate judgment accuracy by comparing this prediction with the final GenEval correctness label of the final image from the same denoising trajectory.

(3) Always-Yes Baseline. Since our evaluation set contains equal numbers of correct and incorrect samples, a trivial judge that always predicts "yes" achieves only 50% accuracy, rather than the plateau observed in Fig. 3.

(4) Failure-Case Accuracy. We further report judgment accuracy specifically on incorrect final samples, as shown in Table 2. All four UMMs progressively improve their ability to identify eventual generation failures during denoising, achieving relatively high accuracy well before generation is complete. These results further demonstrate that UMMs can recognize generation failures from intermediate states, supporting the motivation of our approach.

Table 2: Failure-case judgment accuracy across denoising timesteps.

[0,5)[5,10)[10,20)[20,30)[30,49]
Janus-Flow37.368.273.378.673.0
OpenUni-RecA33.673.180.883.382.4
OmniGen238.078.080.181.782.6
BAGEL26.674.581.582.481.4
AuthorsOpenReview ↗

Response to Reviewer i7Ea (Part 3)

W4: The update rule (Eqs. 8–9) is somewhat unclear. With δ = 0.001 and η = 200, the update becomes a fixed-norm step of ηδ = 0.2 whenever ‖g‖₂ > δ. Is the clipping effectively always active? In addition, Sec. 3 introduces guidance as a velocity modification with γ, but the method directly updates z(Eq. 9). The relation between the two should be clarified.

A: We clarify the role of gradient clipping and the relationship between velocity guidance and our latent rectification update

(1) Is gradient clipping effectively always active? No. We measure the clipping activation rate using BAGEL at each rectification timestep, as shown in Table 3. Clipping is activated only when the gradient norm ‖gₜ‖₂ exceeds δ, accounting for 30.17%–53.34% of gradient updates. This confirms that clipping is applied selectively to limit excessive gradient updates while preserving smaller ones.

Table 3: Gradient clipping activation rate across rectification timesteps on BAGEL.

t=5t=6t=7t=8t=9t=10
Activation Rate53.34%45.26%40.42%35.94%32.08%30.17%

(2) What is the relationship between velocity guidance and our latent update? Section 3 introduces gradient-based velocity guidance as a general background, whereas Eq. (9) specifies our actual implementation. Specifically, our method is inspired by gradient-based velocity guidance, but does not directly follow its velocity-modification formulation. Instead, we adopt an explicit latent-space rectification strategy, directly updating the intermediate latent state using the CSA gradient at selected denoising timesteps.

W5: In Fig. 5(c) and Tab. 6, which verifier does Best-of-N use to select among its samples? If it differs from the score in Eq. 11, the comparison with GSS may not be like-for-like.

A: Thank you for pointing out this concern. We clarify the verifier used by Best-of-N and provide additional multi-seed comparisons to ensure a fair and robust evaluation as follows.

Verifier for Best-of-N. Best-of-N in Fig. 5(c) and Tab. 6 uses exactly the same likelihood-based score defined in Eq. (11) to rank and select generated samples as GSS. Therefore, both methods use the same verifier, ensuring a like-for-like comparison.

Multi-seed comparison. For a more robust comparison, we additionally evaluate Best-of-N and UniRect-CoT using four random seeds (42, 1234, 2026, and 3407). As shown in Table 4, UniRect-CoT achieves 80.4 ± 0.4, slightly outperforming Best-of-3 (80.1 ± 1.1), while exhibiting substantially lower variance. Moreover, UniRect-CoT requires only 43.2s per image, compared with 76.9s for Best-of-3 (~1.78× faster). These results demonstrate that UniRect-CoT achieves better generation performance and stability with substantially lower inference cost than Best-of-N selection.

Table 4: Multi-seed Comparison with Best-of-N

MethodSingle Obj.Two Obj.CountingColorsPositionColor Attr.OverallLatency (s)
Baseline99.1 ± 0.295.0 ± 0.875.5 ± 1.886.3 ± 1.248.4 ± 1.062.9 ± 2.177.9 ± 0.224.7
Best-of-299.2 ± 0.294.8 ± 1.178.3 ± 1.287.1 ± 0.949.4 ± 1.864.3 ± 1.878.9 ± 0.651.8
Best-of-399.4 ± 0.496.3 ± 0.680.6 ± 2.387.4 ± 0.850.9 ± 2.666.1 ± 3.580.1 ± 1.176.9
Ours99.7 ± 0.396.3 ± 0.980.7 ± 2.587.5 ± 0.549.5 ± 1.768.4 ±0.780.4 ± 0.443.2

Q1: How many rectification steps does W = [5, 10] contain? Are η, δ, W, and K shared across all four UMMs, or tuned separately for each model?

A: We clarify the number of rectification updates and the hyperparameter settings across the four UMMs as follows.

(1) How many rectification steps does (W=[5,10]) contain? The rectification window (W=[5,10]) covers 6 denoising timesteps. With K=3 ISR updates at each selected timestep, this results in 18 rectification updates in total.

(2) Are η, δ, W, and K shared across all four UMMs or tuned separately for each model? All four hyperparameters are shared across the four UMM backbones without model-specific tuning. Specifically, we use η = 200, δ = 0.001, W = [5, 10], and K = 3 for Janus-Flow, OpenUni-RecA, OmniGen2, and BAGEL.

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

This paper proposes a training-free framework that employs the comprehension ability of unified multimodal models to guide its image generation process. Specifically, at selected timesteps, the UMMs inspect the image decoded from the early denoising step and then prompt the same model to generate a description, which can be used to calculate the NLL loss and to update the latent state. Extensive experiments show that the proposed method is effective at improving the existing flow-based models.

Critical strengths
  1. This paper is well-organized and easy to follow.
  2. The evaluation is across four different models with semantic-target ablation, which comprehensively validates the generalization of the proposed method.
  3. This approach is training-free without any finetuning cost.
Critical weaknesses
  1. It seems that the proposed method can also employ better external VLMs to provide more accurate gradients for image latent rectification. It would be better to include this comparison.
  2. Failure case analysis. The proposed method will potentially over-rectify the image latent, which in turn generates corrupted images.
  3. The actual contributions of this paper are the adaptive semantic targets and candidate selection. The optimization is from existing gradient-based guidance. The idea of image latent rectification is also common in other literature.
  4. It is necessary to include uncertainty estimates for the performance, such as BAGEL's GenEval score (80.4) vs TiR score (79.8).
Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer SiTk (Part 1)

We sincerely thank reviewer SiTk for the thoughtful and constructive feedback, which will help improve our work. Our responses are as follows.

W1: It seems that the proposed method can also employ better external VLMs to provide more accurate gradients for image latent rectification. It would be better to include this comparison.

A: We replace BAGEL's native understanding branch with three external VLMs (InternVL3-8B, LLaVA-OneVision-7B, and Qwen3-VL-8B-Instruct) to provide reflection targets and semantic gradients for latent rectification. Their parameters are frozen, while gradients are propagated through the VLMs to the input latents. As shown in the table below, our method consistently outperforms the BAGEL baseline (77.6) when using any of the three external VLMs. Compared with BAGEL's native understanding branch (80.4), InternVL3 achieves a slightly higher score (80.6), while LLaVA-OneVision (79.8) and Qwen3-VL (80.2) achieve slightly lower scores. This experiment shows that understanding models with different capabilities lead to varying reflection and generation performance, validating both the effectiveness of our unified rectification Chain-of-Thought approach and the potential of leveraging UMMs' inherent understanding capabilities alone for effective generation improvement. We have included this experiment in Appendix F of the revised manuscript.

Table 1: Comparison of UniRect-CoT using different external VLMs.

MethodGen.Und.Single Obj.Two Obj.CountingColorsPositionColor Attr.Overall
BaselineBagel-99.195.774.386.749.061.077.6
UniRect-CoTBagelInternVL3-8B99.796.582.887.849.367.380.6
BagelLLaVA-OneVision100.096.978.886.948.068.079.8
BagelQwen3-VL-8B-Instruct99.796.281.686.449.068.280.2
BagelBagel99.797.081.987.248.568.080.4

W2: Failure case analysis. The proposed method will potentially over-rectify the image latent, which in turn generates corrupted images.

A: We conduct a systematic failure case analysis as follows.

Over-rectification: Over-rectification manifests in two forms: gradient explosion, which corrupts generated images, and excessive but visually plausible modifications, which degrade semantic alignment. In this paper, we introduce gradient clipping to mitigate gradient explosion and GSS to prevent excessive but visually plausible modifications. When both gradient clipping and GSS are removed, the generated images are severely corrupted, so we do not report GenEval scores for this setting. As shown in Table 2, using either Gradient Clipping or GSS alone remains insufficient: reflections may provide useful guidance, but excessive yet visually plausible modifications (79.7) or performance degradation below the baseline (75.0) can still occur, limiting performance gains. Only their combination effectively mitigates over-rectification and unlocks the potential of our method (80.4). These results also demonstrate the effectiveness of our design.

Additional failure case analysis: We conduct a systematic additional failure case analysis and identify three major failure modes: (1) Reflection Hallucination: The UMM may be misled by intermediate images, producing inaccurate reflective targets or hallucinated descriptions, which lead subsequent gradient updates in incorrect directions. As shown in Fig. 8 of the revised manuscript, the UMM mistakenly treats the current image as the desired result and describes a spatial relationship that contradicts the prompt, providing little corrective guidance and leading to failure. (2) Incomplete Rectification: Even with an accurate reflective target, the limited rectification capacity at each timestep may prevent errors from being fully corrected within the predefined update budget. As shown in Fig. 9 of the revised manuscript, although the gradient turns most of the carrot blue, limited rectification capacity leaves part of it orange, resulting in failure. (3) GSS Misranking: When correct and incorrect candidates are visually similar, the UMM may struggle to distinguish them, assigning higher scores to incorrect candidates and selecting suboptimal latent states. As shown in Fig. 10 of the revised manuscript, GSS selects an incorrect candidate even though a correct candidate is available, resulting in five bowls in the final image.

All the details have been added to Appendix G of the revised manuscript.

Table 2: Effect of gradient clipping and GSS.

MethodGradient ClippingGSSSingle Obj.Two Obj.CountingColorsPositionColor Attr.Overall
Baseline--99.195.774.386.749.061.077.6
UniRect-CoT✗✓98.894.470.985.445.055.575.0
✓✗100.095.580.387.846.069.079.7
✓✓99.797.081.987.248.568.080.4
AuthorsOpenReview ↗

Response to Reviewer SiTk (Part 2)

W3: The actual contributions of this paper are the adaptive semantic targets and candidate selection. The optimization is from existing gradient-based guidance. The idea of image latent rectification is also common in other literature.

A: The core contribution of this paper is the introduction of a novel Thinking-While-Drawing generation paradigm for UMMs, which enables UMMs to continuously reflect on and guide their own generation during the denoising process, rather than treating understanding and generation as separate stages. Unlike conventional gradient-based guidance that relies on predefined objectives or external supervision, our approach exploits the UMM's inherent understanding capabilities to provide state-conditioned semantic guidance throughout a single denoising trajectory. To realize this paradigm, we construct a model-native differentiable semantic loss based on the UMM's own autoregressive likelihood, translating its state-conditioned textual reflections into gradients that steer latent rectification. Thus, gradient-based guidance serves as an enabling mechanism for our proposed paradigm, rather than the primary contribution itself.

W4: It is necessary to include uncertainty estimates for the performance, such as BAGEL's GenEval score (80.4) vs TiR score (79.8).

(1) Multi-seed evaluation. To provide a more robust evaluation, we conduct additional experiments for Tables 1–3 using four common random seeds (42, 1234, 2026, and 3407), and report the results as mean ± standard deviation. As shown in Tables 3–5, the results across four random seeds are statistically consistent with those reported in the original manuscript, demonstrating that the performance gains of UniRect-CoT, the advantages over competing methods, and the effectiveness of its key design components are statistically reliable.

(2) Comparison with TiR. Across four seeds, UniRect-CoT achieves an average GenEval score of 80.4, outperforming TiR (79.5) by 0.9. Meanwhile, UniRect-CoT takes only 43.2s per image, reducing inference time by approximately 63% compared with TiR (116.9s). These results demonstrate the superiority of UniRect-CoT in both generation quality and efficiency.

Table 3: Multi-seed evaluation across UMM backbones.

MethodSingle Obj.Two Obj.CountingColorsPositionColor Attr.Overall
JanusFlow94.1±1.244.0±3.131.7±3.275.5±2.031.4±2.925.7±1.350.4±1.9
w/ Ours95.3±1.156.8±3.041.3±1.681.4±0.341.1±2.933.9±1.358.3±1.2
OpenUni-RecA98.9±0.491.1±0.946.7±1.487.4±0.340.0±1.667.9±0.272.0±0.3
w/ Ours99.2±0.393.4±0.955.7±1.689.6±0.345.0±2.471.1±0.975.7±0.7
OmniGen299.6±0.293.2±0.570.7±2.086.8±0.553.0±0.966.0±1.678.2±0.5
w/ Ours99.5±0.295.1±0.672.1±0.687.1±0.553.9±1.768.6±2.179.4±0.3
Bagel99.1±0.295.0±0.875.5±1.886.3±1.248.4±1.062.9±2.177.9±0.2
w/ Ours99.7±0.396.3±0.980.7±2.587.5±0.549.5±1.768.4±0.780.4±0.4

Table 4: Multi-seed comparison with test-time methods.

MethodSingle Obj.Two Obj.CountingColorsPositionColor Attr.OverallLatency (s)
Baseline99.1±0.295.0±0.875.5±1.886.3±1.248.4±1.062.9±2.177.9±0.224.7
Qwen Rewrite99.5±0.594.7±0.875.2±2.287.9±2.148.1±0.965.3±1.178.5±0.328.7
UiG99.4±0.493.8±1.476.1±1.686.7±1.149.8±2.362.8±1.078.1±0.156.4
TiR99.1±0.694.3±1.380.2±1.685.0±1.252.5±2.366.1±1.779.5±0.5116.9
Best-of-399.4±0.496.3±0.680.6±2.387.4±0.850.9±2.666.1±3.580.1±1.176.9
UniRect-CoT99.7±0.396.3±0.980.7±2.587.5±0.549.5±1.768.4±0.780.4±0.443.2

Table 5: Multi-seed ablation of semantic target design and GSS.

SettingUpdateITE TargetGSS TargetSingle Obj.Two Obj.CountingColorsPositionColor Attri.Overall ↑
Baseline–––99.1±0.295.0±0.875.5±1.886.3±1.248.4±1.062.9±2.177.9±0.2
ITE OnlyStaticUser–99.4±0.395.0±1.277.1±2.686.9±0.748.1±1.266.2±0.878.8±0.3
StaticExp.–99.6±0.593.6±0.576.9±1.686.3±0.847.9±2.664.2±3.278.1±0.3
StaticRef.–99.7±0.493.8±2.178.9±2.487.9±1.548.7±4.465.4±2.879.1±0.2
DynamicRef.–99.7±0.494.5±1.678.5±1.989.2±1.248.3±2.666.9±2.979.5±0.3
ITE+GSSStaticUserUser99.2±0.495.0±0.878.3±1.987.5±1.049.6±1.665.7±1.379.2±0.2
StaticExp.User99.5±0.294.2±0.777.7±2.086.8±0.848.9±2.063.7±2.878.4±0.3
StaticRef.User99.7±0.394.8±1.877.4±1.787.0±0.749.8±3.867.9±1.979.5±0.5
DynamicRef.Ref.99.4±0.494.3±2.178.4±2.587.4±1.849.9±3.364.6±2.779.0±0.5
DynamicRef.User99.7±0.396.3±0.980.7±2.587.5±0.549.5±1.768.4±0.780.4±0.4

Comments on the paper

2 comments
AuthorsOpenReview ↗

List of Revisions

We sincerely thank all reviewers for their constructive feedback. We have revised the manuscript in response to their comments. The revisions are highlighted in blue for easy reference. The main revisions are as follows:

  • In Section 2.2, we have clarified our contributions and discussed related work in response to reviewers SiTk and i7Ea.

  • In Table 1, we have clarified the presentation of reproduced baselines in response to reviewer i7Ea.

  • In Appendix J, we have added implementation details in response to reviewers i7Ea and Ci3p.

  • In Appendix F, we have added comparisons with external VLMs, as suggested by reviewer SiTk.

  • In Appendix G and Figures 8–10, we have added trajectory-level failure analyses to address the concerns raised by reviewers SiTk and Ci3p.

  • In Appendix H, we have added multi-seed evaluations to address the performance variability concerns raised by reviewers SiTk and Ci3p.

  • In Appendix I, we have added comparisons with related methods in response to reviewer i7Ea.

  • In Appendix A.3 and Section 4.1, we have revised the relevant descriptions and added a discussion of peak GPU memory usage in response to reviewer i7Ea's suggestions.

AuthorsOpenReview ↗

Summary of the Discussion Phase and Paper Revisions

Dear Area Chair,

We sincerely thank you and the reviewers for the time and effort devoted to assessing our submission. The constructive feedback has helped us improve the paper. Below, we summarize the strengths recognized by the reviewers and the revisions made in response.

First, the reviewers highlight the following strengths:

  • Training-free, model-native design: Reviewers recognize the use of the UMM’s inherent understanding to improve generation without additional training (SiTk, i7Ea, Ci3p).
  • Consistent improvements and thorough analyses: Reviewers recognize the improvements across four flow-based UMMs (SiTk, i7Ea, Ci3p). Reviewers i7Ea and Ci3p also highlight the guidance-and-selection target ablations, quality–latency comparisons, and repair–regression analysis, which examine the method’s design choices, computational trade-offs, and balance between repairing failures and preserving correct outputs.
  • Clear presentation and useful insights: All three reviewers recognize the paper’s clarity and informative value, with Ci3p reporting that the work provides “a substantially deeper understanding of the problem or question.”

Second, we have made the following clarifications and additions in response to the reviews and discussion:

Revisions to the paper:

1. Contribution and comparisons (SiTk, i7Ea)

  • Contribution clarification: In Section 2.2, we clarify the contribution of training-free, reflection-guided latent rectification and expand the related-work discussion (SiTk, i7Ea).
  • Related-method comparisons: In Appendix I, we add comparisons with related approaches (i7Ea).
  • External VLM comparisons: In Appendix F, we examine how different understanding models affect rectification (SiTk).

2. Ablation studies and statistical robustness (SiTk, i7Ea, Ci3p)

  • Semantic-gradient control: We provide the requested matched-norm random-direction control to isolate the contribution of semantic gradients from candidate selection (i7Ea).
  • Performance variability: In Appendix H, we report four-seed means and standard deviations for the main comparisons and ablations (SiTk, i7Ea, Ci3p).

3. Reliability and failure analysis (SiTk, i7Ea, Ci3p)

  • Proxy diagnostics: We provide additional judgment statistics and analyses relating candidate scores to final generation correctness (i7Ea, Ci3p).
  • Trajectory-level failures: In Appendix G and Figures 8–10, we add failure cases covering reflection hallucination, incomplete rectification, and candidate misranking (SiTk, Ci3p).

4. Reproducibility and presentation (i7Ea, Ci3p)

  • Implementation details: In Appendix J, we provide fuller inference settings, paired evaluation details, and shared hyperparameter configurations (i7Ea, Ci3p).
  • Baseline presentation: In Table 1, we distinguish reported and reproduced results more clearly (i7Ea).
  • Explanations and wording: We provide a more detailed explanation of Figure 3, revise the relevant descriptions following the reviewers’ suggestions, and disclose peak GPU memory usage in Section 4.1 and Appendix A.3 (i7Ea).

Discussion phase: Reviewer SiTk supported acceptance in the initial review. Following our rebuttal and additional experiments, both Reviewers i7Ea and Ci3p explicitly expressed support for acceptance. i7Ea characterizes the issues discussed as “limitations of interpretation and presentation rather than reasons to reject the method.” We have revised the manuscript accordingly. Ci3p states that the supplementary experiments “have addressed my previous concerns” and concludes: “Overall, I hold a positive recommendation for accepting this paper.”

All revisions are highlighted in blue, and detailed responses are provided in our replies to each reviewer.

Thank you for your time and consideration. We hope this summary assists in your decision-making process.

Sincerely, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.