acceptodds
Under review as a conference paper at ICLR 2027

Dep2C: Dependency-Aware Credit Assignment and Coordinated Decoding in Masked Image Generation

23 views

Abstract

Masked generative models (MGMs) enable efficient text-to-image generation through iterative parallel decoding. However, coarse-grained credit assignment may overlook the unequal downstream importance of committed tokens, while confidence-based decoding may jointly commit strongly dependent tokens. We propose Dep2C, a dependency-aware framework for fine-grained credit assign- ment and coordinated decoding. During RL post-training, Dep2C-Credit traces dependencies from future, prompt-relevant predictions back to previously com- mitted tokens and accumulates them along the generation trajectory to estimate to- ken importance. These estimates guide hierarchical credit allocation across decod- ing steps and individual tokens. During inference, Dep2C-Decoding combines prediction confidence and dependency degree to prioritize candidates, while using pairwise dependencies to constrain joint commitment. Experiments on Show-o and Lumina-DiMOO demonstrate improvements in both compositional genera- tion and complex prompt following, achieving GenEval scores of 0.86 and 0.92 and DPG-Bench scores of 84.37 and 83.69, respectively.

Primary area:
generative models
Keywords:
Masked Generative Models, Text-to-Image Generation, Reinforcement Learning, Credit Assignment, Dependency Modeling, Parallel Decoding

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 4 reviews, above 85% of the others.
Reviewer_gt3VOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper uses averaged upper-layer attention as a proxy for token dependency in masked generative models. It applies this proxy in two ways: (i) Dep2C-Credit, which accumulates discounted, prompt-relevant attention from future masked queries to each committed token and uses it to reweight the image-level GRPO advantage across steps and tokens; and (ii) Dep2C-Decoding, which reweights confidence by attention degree and greedily keeps strongly co-attending masked tokens out of the same parallel commit. The method is simple, needs no extra forward passes, and gives consistent gains on Show-o and Lumina-DiMOO.

Critical strengths
  • This paper presents a clear and intuitive idea. The paper frames token dependency as a single concept that plays opposite roles: emphasize strong inter-step dependency during training, and avoid strong intra-step dependency during parallel decoding. The framing is easy to follow, and both components reuse attention that is already computed.
Critical weaknesses
  • The credit signal Fk=∑s>tγs−t−1fs,kF_k = \sum_{s>t} \gamma^{s-t-1} f_{s,k} sums over more future steps for early-committed tokens, so it systematically assigns higher weights to early tokens regardless of their actual dependency. This makes the step-level allocation similar in effect to timestep-reweighting methods such as MaskFocus [1], yet the paper does not compare against them. In addition, a simple baseline that assigns higher weights to earlier timesteps should also be included, to show how the different weights for tokens within the same timestep matter.

[1] G. Zhang et al., "MaskFocus: Focusing Policy Optimization on Critical Steps for Masked Image Generation," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026.

Questions for authors
  • As mentioned in the weaknesses, the authors should add comparisons with (1) MaskFocus and (2) a timestep-only baseline that assigns higher weights to earlier timesteps.
  • Figure 6 shows the training collapses. Does Dep2C-Credit still collapse at ≈4.4K steps using other seeds? Can the authors give some explanation on what causes the collapse in both methods?
  • The main results compare Dep2C at 4K steps with Mask-GRPO at 3K steps (Fig. 6). Can the authors report the main tables under a matched training budget (like 3K for both)?
3 comments

Response — Part 1/2

We thank the reviewer for recognizing the clarity and intuition of Dep2C's unified dependency perspective. We address the concerns about timestep weighting, training degradation, and training budgets using supplementary comparisons and the existing ablations.

Q1. MaskFocus and the role of within-step credit allocation

Question: As mentioned in the weaknesses, the authors should add comparisons with (1) MaskFocus and (2) a timestep-only baseline that assigns higher weights to earlier timesteps.

Earlier-committed tokens have a longer interval over which future influence is accumulated, so the accumulation horizon can affect the distribution of credit across steps. However, credit is not determined solely by commitment time: it also depends on attention-derived dependency strength and the relevance of future masked positions to prompt entities. We use a MaskFocus comparison and hierarchical allocation ablations to examine the effects of step-level optimization and within-step token credit, respectively.

Comparison with MaskFocus at the same training step count. We reproduce MaskFocus on Show-o with K=2 and K=6 critical steps. The original MaskFocus setting selects six critical steps from a 64-step generation trajectory, approximately 10% of the total. Since our training rollouts use 20 decoding steps, K=2 approximately preserves this selection ratio, whereas K=6 preserves the original number of critical steps. These configurations examine proportional adaptation and retention of the original count, respectively.

Both configurations are evaluated at 400 training steps, matching the checkpoint step count in the main Show-o experiments. For each MaskFocus configuration, evaluation uses 64 decoding steps, CFG=3.5, and seed=42, with 2,212 images generated and evaluated across 553 prompts. The table combines the main-table results with the supplementary MaskFocus results.

MethodSingleTwo Obj.CountingColorsPositionAttr.Overall
Mask-GRPO1.000.870.740.830.600.630.78
MaskFocus, K=20.980.880.750.840.340.620.73
MaskFocus, K=60.980.870.730.810.380.620.73
Dep2C0.990.940.850.870.810.700.86

Scores are rounded to two decimal places for consistency with the main tables; the unrounded overall scores for the two MaskFocus configurations are 0.7343 and 0.7310. This comparison provides additional method-level evidence at the same training step count. Dep2C's observed advantages are particularly apparent in compositional categories such as Two Objects, Counting, Position, and Attribute Binding. We next use controlled ablations to examine the two levels of credit allocation.

Separating step-level weighting from within-step credit allocation. All variants in Table 8 use a matched 1K-step training budget and are evaluated with 16 decoding steps at CFG=0:

Allocation(βt,βs)(\beta_t,\beta_s)GenEval ↑
Equal-step baseline(0,0)0.4266
Step budget only(0.25,0)0.4604
Within-step allocation only(0,0.5)0.4607
Hierarchical(0.25,0.5)0.4738

The step-budget-only variant retains dependency-derived step weights while using uniform allocation within each step. The within-step-only variant instead fixes equal budgets across steps and varies token credit within each step, directly testing whether token-level differentiation contributes beyond step-level reweighting.

Table 10 further shows that, with equal step budgets, both moderate within-step allocation strengths (βs=0.25,0.5\beta_s=0.25,0.5) outperform uniform allocation, whereas excessive concentration at βs=1\beta_s=1 reduces the score to 0.4074. Table 9 likewise shows that increasing step-budget concentration does not continually improve performance, supporting the role of moderate credit allocation.

Together, these experiments address whether the gains arise solely from upweighting earlier timesteps: with the total budget of each step fixed, within-step allocation increases GenEval from 0.4266 to 0.4607, a gain of 3.41 percentage points; adding within-step allocation to step-level allocation further increases the score from 0.4604 to 0.4738, a gain of 1.34 percentage points. The former variant does not increase the total weight of earlier steps. Thus, under this evaluation setting, the gains cannot be fully explained by early-timestep reweighting; distinguishing tokens within the same step according to downstream dependencies provides additional value.

Response— Part 2/2

Q2. Training degradation across seeds and possible causes

Question: Figure 6 shows the training collapses. Does Dep2C-Credit still collapse at approximately 4.4K steps using other seeds? Can the authors give some explanation on what causes the collapse in both methods?

The decline at approximately 4.4K steps in Figure 6 is an observation from that particular training run, not a fixed failure threshold of the method. Across our Show-o experiments covering two datasets and two reward settings, as well as the Lumina-DiMOO experiments, Dep2C delays degradation relative to the corresponding baselines. This supports consistency across the tested settings. These experiments are not, however, repetitions with different training seeds under an identical configuration; variation in the degradation onset across training seeds has not yet been systematically quantified.

One possible explanation is the interaction between noisy credit assignment under trajectory-level feedback and continued policy updates. Tokens within the same trajectory differ in their contributions to downstream predictions and final reward. Applying trajectory-level feedback to these decisions may reinforce some actions that did not actually contribute to the high reward. As updates accumulate, the resulting changes can propagate through the contexts of subsequent generation steps and affect training stability.

Dep2C refines credit assignment using downstream dependency information and may mitigate this process. Credit weighting itself, however, does not directly constrain the magnitude of policy updates and therefore cannot guarantee the elimination of long-run training instability. This is a plausible mechanism rather than an experimentally established causal explanation. Our empirical conclusion is that Dep2C delays degradation in the tested settings, not that it completely prevents collapse.

Q3. Comparison under a matched training budget

Question: The main results compare Dep2C at 4K steps with Mask-GRPO at 3K steps (Fig. 6). Can the authors report the main tables under a matched training budget (like 3K for both)?

Yes. Table 7 already reports results for all variants at 3,000 training steps on Lumina-DiMOO. The exact checkpoints used in the main tables are 3,100 steps for Mask-GRPO and 3,800 steps for Dep2C. These were selected by inspecting the online training success curves and choosing saved checkpoints from the high-success phase preceding clear degradation. The main tables report performance under this selection criterion, whereas Table 7 specifically evaluates a fixed training step count. Selected metrics from Table 7 are reproduced below:

Configuration (all trained for 3K steps)GenEvalCountingPositionAttr.UniPercept Mean
Mask-GRPO0.880.860.870.7253.42
+ Dep2C-Credit0.900.930.920.7553.42
+ Dep2C-Decoding0.880.890.870.7453.65
+ Both components0.910.930.910.7753.68

At the same 3K-step training budget, Credit alone increases GenEval from 0.88 to 0.90; the complete Dep2C method reaches 0.91, improving over the baseline by 3 percentage points. Counting, Position, and Attribute Binding improve by 7, 4, and 5 percentage points, respectively. Thus, Dep2C's gains on these metrics do not depend solely on the additional training steps used in the main tables.

We will highlight these matched-budget results in the revision and distinguish curve-guided checkpoint selection in the main Lumina-DiMOO tables, the matched 3K-step comparison in Table 7. This will allow readers to assess performance under a matched training budget separately from performance with curve-guided checkpoint selection.

Dear Reviewer gt3V,

As the discussion period draws to a close, we would like to briefly follow up on our responses above. Thank you again for the time and care you have put into reviewing our paper. Your questions helped us better examine whether the gains come from timestep weighting, within-step credit allocation, or differences in training budget.

In our responses, we provided additional MaskFocus comparisons, discussed the ablations separating step-level and within-step allocation, and highlighted the matched 3K-step results. We also clarified the checkpoint selection procedure and the limits of our evidence regarding training degradation across seeds.

We would greatly appreciate hearing whether these responses address your concerns. If any explanation or comparison remains unclear, please let us know. We will be monitoring the discussion closely and will respond as promptly as possible to any follow-up questions.

Thank you again for your thoughtful feedback and consideration.

Reviewer_o76VOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Dep2C for masked text-to-image generation. Exploiting the attention matrices already computed in the existing pipeline, it adds two components: Dep2C-Credit assigns credit to image tokens according to attention and uses this importance estimate to allocate weights in the RL stage; Dep2C-Decoding combines confidence with attention-derived dependency and prevents strongly dependent tokens from being committed in the same step, so that conflicting generation results are avoided. Component ablations, a matched-budget training comparison, and fixed-state token-remasking analyses examine the respective contributions of the two components and the interpretation of the importance signal.

Critical strengths

1. Both components have measurable effects that are verified in isolation. On Lumina-DiMOO, enabling credit assignment raises GenEval from 0.88 to 0.91. On the frozen pretrained model, replacing only the decoder raises DPG-Bench from 81.87 to 82.54, with the compatibility constraint alone reaching 82.35. The two components play different roles: credit assignment improves compositional correctness, the decoder improves complex prompt following, and neither adds forward passes.

2. The paper rigorously validates the assumption on which its method depends. Section 4.4 re-masks committed tokens in future states and observes how much the subsequent predictions change. The results show that tokens with higher importance scores lead to clearly larger changes in downstream predictions; when tokens are binned into five groups by score, the average change rises from bin to bin; the Spearman correlation between score and change is 0.719. This shows that the score does measure a token’s influence on downstream predictions, rather than being an arbitrarily assigned weight.

Critical weaknesses

1. A substantive independent decoding contribution beyond DAPD has not been established. The decoder in Section 3.4 shares DAPD’s core mechanisms: a symmetric attention-derived dependency graph, confidence-weighted degree ordering, and greedy selection that avoids committing strongly coupled positions together. These mechanisms appear in §§3.1 and 4.3 of Kim et al., DAPD, arXiv:2603.12996v1, which the paper already cites in Section 2. Dep2C introduces a fixed reveal quota with threshold relaxation and modifies the scoring details for masked image generation. These are concrete adaptations, but their significance as an independent algorithmic advance remains unsubstantiated. Table 5 supports the role of degree ordering and compatibility constraints relative to confidence decoding, but does not establish an advance over DAPD’s mechanism. On the current evidence, the decoder is merely a useful adaptation, and the main methodological novelty rests on Credit.

Questions for authors
  1. Beyond DAPD’s graph construction, confidence-weighted degree ordering, and greedy compatibility selection, does Dep2C-Decoding add a substantive algorithmic contribution? What reasoning or existing evidence supports its significance?
2 comments

Question: Beyond DAPD's graph construction, confidence-weighted degree ordering, and greedy compatibility selection, does Dep2C-Decoding add a substantive algorithmic contribution? What reasoning or existing evidence supports its significance?

Response.

We thank the reviewer for recognizing the independent effects of both components and the validation of downstream influence. Dep2C connects credit assignment during training with coordinated token selection during inference through a unified dependency perspective. Regarding the decoding contribution, the significance of the fixed reveal quota and adaptive threshold is that they separate control of the generation schedule from dependency coordination, rather than simply applying DAPD to image generation.

What changes in the selection procedure.

In threshold-driven selection, the current dependency structure affects the commitment size, which in turn changes subsequent mask states and the realized number of decoding iterations. Our design instead obtains the reveal quota at each step from the original schedule and searches for a compatibility threshold to coordinate the positions selected under that quota. The schedule determines how many positions to reveal; dependency information guides which positions to commit jointly.

When the current threshold does not allow the prescribed quota to be met, we adjust the threshold and reconstruct the commitment set rather than reduce the commitment size. This preserves the scheduled progression of the number of revealed tokens while allowing dependency information to determine their identities. The change therefore concerns the overall selection procedure, not merely its objective.

Why this matters and what the evidence supports.

This design allows dependency-constrained token selection to operate within a prescribed decoding budget, without adding decoding iterations when compatibility constraints are too restrictive. The following supplementary experiments provide preliminary support. The table below summarizes results on the frozen pretrained Lumina-DiMOO model, without LoRA, at CFG=3.5.

Decoding methodThreshold rangeFixed stepsMean iterationsDPG ↑
Confidence decoding—64—81.63 ± 0.43
Dep2C-DecodingAdaptive search64—82.62 ± 0.24
DAPD-Direct0.01–0.05—80.3281.09
DAPD-Direct0.05–0.15—34.8475.32

The first two rows provide a paired comparison at a fixed decoding budget; the last two report quality versus realized decoding iterations for DAPD-Direct under two threshold settings. The two groups use different evaluation protocols. For the first two rows, ± denotes the sample standard deviation across three generation seeds.

At the same 64-step budget, Dep2C-Decoding improves mean DPG by 0.99 points over confidence decoding, with improvements for all three generation seeds. For the two tested DAPD-Direct configurations, relaxing the threshold reduces the realized number of decoding iterations but also lowers generation quality. The fixed-budget paired comparison supports the effectiveness of the complete decoding procedure; the DAPD-Direct results provide complementary observations about the trade-off among the compatibility threshold, decoding iterations, and quality.

These results support the practical value of coordinating dependencies within a fixed reveal schedule. The comparison to confidence decoding evaluates the complete decoder, while the DAPD-Direct comparison is preliminary; neither isolates the contribution of quota adjustment alone.

Dear Reviewer o76V,

As the discussion period draws to a close, we would like to briefly follow up on our response above. Thank you again for your careful review and for recognizing the separate contributions of the two components and the validation of the dependency signal. Your question about DAPD helped us make the contribution of our decoding procedure more precise.

In our response, we explained how quota-driven threshold adjustment coordinates dependency-based token selection under the original reveal schedule. We also provided additional comparisons and clarified what they support, including the limitations of the preliminary DAPD-Direct comparison. We have revised Related Work and Sec. 3.4 to explicitly distinguish the mechanisms shared with DAPD from our quota-driven selection design, with changes highlighted in blue.

We would greatly appreciate hearing whether these explanations and revisions address your concern. If any aspect remains unclear or needs further discussion, please let us know. We will be monitoring the discussion closely and will respond as promptly as possible to any follow-up questions.

Thank you again for the time and care you have devoted to reviewing our work.

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

The paper targets masked generative models (a class of text-to-image generation models) to be enhanced at both the post-training and inference stages by exploiting token dependencies derived from the model’s attention maps. During training, the relevancy between the token and the prompt-relevant future mask predictions is measured and computed as per-token weight. During decoding, a graph is built on masked positions for scoring candidates and the candidates are greedily selected with compatibility check.

The paper's contribution lies in assigning more accurate token credit through their dependency computation for masked generative models. The validation on various benchmark shows consistent improvement on compositional correctness or image quality, but the statistical significance is doubtful.

Critical strengths
  1. The problem of token dependency in masked image generation is well-explained, and the idea of calculating relevancy between token and the future masked predictions sounds new.
Critical weaknesses
  1. Similarity with DAPD [1] on Dep2C-Decoding The backbone structure of the Dep2C-Decoding highly matches with DAPD in the sense of using masked-to-masked attention edge scores on the same ratio (30%) final layers with greedy selection for token selection. Despite the difference that DAPD is for texts and Dep2C is for images, there is no significant difference.

  2. The gains are small and statistical significance is not provided. The gains on the experiments are 0.01~0.04 points and without error bars so their significance cannot be determined. The large gains for decoding at low CFG scale isn't practically meaningful because 0 and 1 aren't really the values used for generating images in general.

[1] Kim, B., Jeon, D., Jeon, M., & No, A. (2026, ICML). DAPD: Dependency-Aware Parallel Decoding via Attention for Diffusion LLMs

Questions for authors
  1. What is the substantial difference betweem DAPD and Dep2C-Decoding? The intra-step dependency mechanism appears to be the same as DAPD's, differing only in objective (quality at a fixed schedule rather than acceleration). Could the authors clarify what is new, and compare directly against DAPD?

  2. Were the experiments run only once? If so, how could you show improvements of 0.01~0.04 on GenEval and ~1 on DPG are statistically significant?

7 comments

Response— Part 1/2

We thank the reviewer for recognizing the motivation for modeling token dependencies and the idea of assigning credit based on influence on downstream predictions. Below, we address the distinction from DAPD, the reliability of the experimental results, and effectiveness across CFG settings.

Q1. Substantive distinction from DAPD and direct comparison

Question: What is the substantial difference between DAPD and Dep2C-Decoding? The intra-step dependency mechanism appears to be the same as DAPD's, differing only in objective (quality at a fixed schedule rather than acceleration). Could the authors clarify what is new, and compare directly against DAPD?

Dep2C connects credit assignment during training with coordinated token commitment during inference through a unified dependency perspective: inter-step dependencies identify tokens that should receive greater training credit, whereas intra-step dependencies guide which positions to commit jointly. Within this framework, the new decoding design integrates dependency constraints with a prescribed reveal schedule through a quota-driven adaptive threshold selection mechanism.

The common elements with DAPD are attention-based dependency estimation over masked positions and confidence-weighted dependency degree ranking. We do not claim the use of attention from the final 30% of layers, or the transfer from text to image generation itself, as a new contribution. The distinction lies in the overall commitment selection procedure, not merely the optimization objective.

DAPD determines the commitment set at each iteration using a compatibility threshold and other conditions; the commitment size and realized number of decoding iterations can vary with the dependency structure. Dep2C-Decoding first obtains the reveal quota ntn_t from the original schedule and then performs 16 bounded bisection-style threshold updates. At each iteration, greedy selection is rerun under the current threshold with a fixed priority order: the threshold is tightened when the quota can be met and relaxed otherwise. The selected threshold and fallback rule ensure that the quota is satisfied.

Thus, our greedy selection is part of a quota-driven constraint search that constructs a commitment set of the prescribed size. The key change is from threshold-controlled commitment size to quota-driven constraint adjustment, incorporating dependency coordination into a fixed generation schedule without introducing additional decoding iterations when compatibility constraints are too restrictive.

We additionally conduct a direct comparison using the Lumina-DiMOO base model, without LoRA, at CFG=3.5 and a resolution of 512×512:

Decoding methodDecoding iterationsDPG ↑
Confidence decodingFixed: 6481.63±0.43
DAPD-Direct, thresholds 0.01–0.05Mean: 80.3281.09
DAPD-Direct, thresholds 0.05–0.15Mean: 34.8475.32
Dep2C-DecodingFixed: 6482.62±0.24

Relative to the stricter DAPD threshold configuration, Dep2C achieves an observed score 1.53 points higher with approximately 20.3% fewer decoding iterations, providing preliminary evidence for its quality–iteration trade-off. These results support the effectiveness of the complete decoding procedure, although the full difference cannot be attributed solely to quota adjustment.

Response — Part 2/2

Q2. Repeated evaluations and statistical reliability

Question: Were the experiments run only once? If so, how could you show improvements of 0.01–0.04 on GenEval and approximately 1 point on DPG are statistically significant?

We have supplemented the original results with repeated generation evaluations using fixed model checkpoints. Below, we report individual runs, means, and sample standard deviations, distinguishing repeatability across generation seeds from statistical significance of the differences.

DPG-Bench: paired comparison of decoding methods. On the frozen pretrained Lumina-DiMOO model, confidence decoding and Dep2C-Decoding use the same three generation seeds and prompts, with 64 decoding steps, CFG=3.5, and otherwise identical sampling settings:

Generation seedConfidence decodingDep2C-DecodingPaired gain
081.3582.89+1.54
10000081.4282.53+1.11
20000082.1382.44+0.31
Mean ± sample SD81.63±0.4382.62±0.24+0.99

Dep2C improves over the baseline for all three seeds, with a mean gain of 0.99 points, showing that the improvement is not confined to a single generation run. We will report the number of runs, means, and standard deviations in the revision, explicitly noting that these repeated evaluations quantify generation randomness rather than variation across independent training runs.

Repeated GenEval evaluations and comparison with the baseline. To assess the consistency of the gains across generation seeds, we fix the Mask-GRPO and Dep2C checkpoints at 3,100 and 3,800 training steps, respectively, and keep the test prompts, inference parameters, and evaluation pipeline identical. Both models use Dep2C-Decoding, with 64 decoding steps and CFG=3.5, generating four images per prompt. Thus, this supplementary comparison evaluates the two trained checkpoints under the same inference decoder; its Mask-GRPO row is not the confidence-decoding baseline from the main tables. The results are as follows:

Evaluation runMask-GRPO + DecodingDep2C
Run 10.895520.92087
Run 20.887790.91333
Run 30.893590.91628
Mean ± sample SD0.89230±0.004020.91683±0.00380

For the first two runs, which use matched generation seeds, Dep2C improves over Mask-GRPO by 2.54 and 2.55 percentage points, respectively, with consistent directions and magnitudes. Across the three evaluations, the mean scores are 0.89230 and 0.91683, giving Dep2C a mean advantage of 2.45 percentage points.

These results show that the selected Dep2C checkpoint's advantage over the selected Mask-GRPO checkpoint is repeatedly observed under the tested generation seeds. The evaluations quantify generation variability for fixed checkpoints, rather than variability across independent training runs. The matched-training-budget results are reported separately in Table 7.

We will include the individual results, means, and sample standard deviations in the revision. These repeated evaluations show consistent improvements across the tested generation seeds and quantify the variability of the reported scores.

W2. Practical relevance beyond low CFG

The benefits of Dep2C-Decoding are not limited to CFG=0 or 1. Figure 5(b) shows improvements over confidence decoding at every tested CFG value. Beyond low CFG, DPG-Bench improves by 1.45, 0.81, and 1.06 points at CFG=2, 3, and 3.5, respectively. Thus, the larger gains at low CFG are not the sole evidence for the effectiveness of our decoder; improvements also hold at higher CFG values.

To further assess consistency, we conduct the additional three-seed comparison at CFG=3.5 with 64 decoding steps, changing only the decoding method on the frozen base model. As shown in the paired comparison above, all three seeds yield improvements, with a mean gain of 0.99 points. The single-run CFG sweep in the figure and the additional multi-seed mean are distinct summaries.

The CFG sweep in Figure 5(b) and the additional multi-seed evaluations provide complementary evidence of consistency across CFG settings and stability across generation seeds at a fixed setting, supporting the conclusion that decoding improvements are not restricted to low CFG.

One further point on Dep2C-Decoding. The paper states that it requires no additional forward passes, but it also introduces “auxiliary computation for attention extraction, graph construction, and token selection” (Sec. 3.4), including the repeated greedy passes of the threshold search. Since no wall-clock results are reported, could you provide the time per image (or per decoding step) and peak memory for Dep2C-Decoding against confidence decoding under the same settings? For Dep2C-Credit, the additional training time and memory for caching attention maps would also be helpful.

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

This paper proposes Dep2C for masked text-to-image generation, comprising dependency-aware credit assignment during RL post training and coordinated token selection during inference. Dep2C-Credit accumulates attention from masked positions in later decoding steps to previously committed tokens, weighted by the masked positions’ relevance to prompt entities, to estimate token importance. These estimates guide policy update weighting across steps and within each step. Dep2C-Decoding combines confidence with attention-derived dependency degree and uses pairwise dependency constraints to guide greedy token selection while preserving the reveal schedule. Experiments on Show-o and Lumina-DiMOO report improvements on GenEval and DPG-Bench over pretrained models and matched RL baselines. Component ablations, fixed-state token-remasking experiments, and frozen model decoding comparisons provide additional validation. The remasking analysis measures downstream prediction sensitivity rather than changes in final reward.

Critical strengths

S1. Consistent improvements over matched RL baselines on two image generation backbones. Compared with Mask-GRPO, Dep2C improves GenEval from 0.78 to 0.86 on Show-o and from 0.88 to 0.92 on Lumina-DiMOO, with corresponding DPG-Bench gains of 0.92 and 1.18 points, respectively (Tables 1 and 2). The component ablations, including the comparison at a matched 3k steps training budget, provide useful evidence for the benefit of the proposed weighting mechanism over matched Mask-GRPO training (Tables 4 and 7). These comparisons are important because they assess the contribution of the added mechanism rather than only comparing against pretrained models.

S2. The inference time component is evaluated separately from training. On frozen pretrained Lumina-DiMOO, the decoder ablation distinguishes the effects of dependency degree prioritization and compatibility constraints (Table 5). At each matched decoding budget, the two decoding strategies use identical prompts, random seeds, and sampling settings (Section 4.5). With CFG fixed at 3.5, the reported 8 steps DPG-Bench score of 82.35 exceeds the 64 steps confidence-based baseline of 81.87 (Figure 5(a)).

Critical weaknesses

W1. The relationship to the closest prior work needs a more detailed discussion. Dep2C-Decoding shares several key steps with DAPD, including attention-based dependency estimation, confidence-weighted degree ranking, and greedy selection that avoids strongly dependent positions. Although Dep2C preserves the original reveal schedule, the paper does not explain in detail which parts of the decoder are new relative to DAPD and what these changes contribute. For the training component, it would also help to distinguish the proposed importance-based weighting from RebRL’s masking-frequency-based reweighting. The two use different signals, and a direct discussion would make the contribution clearer.

W2. Some implementation and evaluation details are not fully specified. The paper does not clearly explain how prompt entities are extracted and aligned with text tokens, what value is used for the decoder’s degree exponent, or the search bounds and stopping rule for the threshold search (Sections 3.2–3.4). It is also unclear how the checkpoints in the main tables were selected, particularly given the decline in Mask-GRPO’s training success rate after about 3K steps (Appendix A.2). These details are needed to reproduce the comparisons and interpret the reported gains.

Questions for authors
  1. Apart from preserving the original reveal schedule, which parts of Dep2C-Decoding are new relative to DAPD, and which parts are retained or adapted? In particular, please clarify any differences in the priority score and dependency-constrained selection.

  2. How are prompt entities extracted and aligned with text tokens? What value is used for the decoder’s degree exponent, and what are the bounds and stopping rule for the threshold search in Section 3.4?

  3. Which training checkpoints are used in the main results, and how were they selected for Mask-GRPO and Dep2C? Table 7 provides a matched 3K steps comparison, but it would help to clarify the training budgets and selection rule for the main tables, given the decline in Mask-GRPO’s training success rate reported in Appendix A.2.

2 comments

Response Part 1/2: Relationship to DAPD and RebRL

We thank the reviewer for recognizing our results and controlled ablations. Part 2 covers Q2–Q3.

Q1. Retained and new components relative to DAPD

Dep2C connects training credit assignment and coordinated decoding through a unified dependency perspective: inter-step dependencies guide training credit, while intra-step dependencies guide joint commitment. We distinguish the retained attention-based dependency representation and confidence-weighted degree ranking from the quota-driven dependency-constrained selection procedure. The latter preserves the schedule by adapting the compatibility threshold to the prescribed reveal quota, rather than allowing the threshold to determine commitment size.

Dependency representation and priority score.

We retain the symmetric attention-derived dependency graph and joint ranking by confidence and dependency degree. Our priority is:

si=ci(didˉ)βd,di=∑jDij.s_i=c_i\left(\frac{d_i}{\bar{d}}\right)^{\beta_d},\qquad d_i=\sum_j D_{ij}.

Here, cic_i is the predicted probability of the sampled candidate token and dˉ\bar d is the mean dependency degree. We use βd=1\beta_d=1. Mean-degree normalization does not change the deterministic ranking, and we do not claim it as a new ranking mechanism.

Dependency-constrained selection.

DAPD determines each commitment set using a compatibility threshold and other conditions; commitment size and realized decoding iterations can vary with the dependency structure. Dep2C first obtains the reveal quota ntn_t from the original schedule, then uses bounded threshold search and greedy selection to satisfy it. The threshold is tightened when the quota can be met and relaxed otherwise; Part 2 gives the search details.

Both methods check compatibility using dependency information. Dep2C embeds candidate selection within a fixed-quota-driven threshold search and reconstructs the commitment set after each update. Thus, the overall procedure differs from selection under a given threshold. The change is from threshold-controlled commitment size to quota-driven constraint adjustment: dependency information coordinates which positions to commit jointly at the prescribed size, without adding decoding iterations when constraints are too restrictive.

Evidence for the decoding design.

Table 5 isolates dependency degree prioritization and compatibility constraints relative to confidence decoding. We complement these ablations with a DAPD comparison on frozen pretrained Lumina-DiMOO, without LoRA, at CFG=3.5 and 512×512:

DecoderDecoding iterationsDPG
Confidence decodingFixed 6481.63 ± 0.43
DAPD-Direct, thresholds 0.01–0.05Mean 80.3281.09
DAPD-Direct, thresholds 0.05–0.15Mean 34.8475.32
Dep2C-DecodingFixed 6482.62 ± 0.24

Dep2C improves over confidence decoding for all three generation seeds, with a 0.99-point mean gain, supporting the complete decoder at a fixed step count. Relative to the stricter DAPD configuration, its observed score is 1.53 points higher with 20.3% fewer iterations, providing preliminary quality–iteration trade-off evidence.

Distinction from RebRL (W1).

RebRL addresses imbalanced gradient update strength caused by masking-frequency differences, compensating for differences in update opportunities and cumulative gradient contributions. Dep2C-Credit addresses a different issue: trajectory-level feedback cannot distinguish individual tokens' downstream influence. Even with equal update opportunities, tokens can affect subsequent predictions differently; balancing opportunities does not resolve dependency-based credit assignment.

Dep2C-Credit accumulates attention from future masked positions to previously committed tokens, weighted by those masked positions' relevance to prompt entities. These estimates guide allocation across steps and within each step, providing finer-grained credit for trajectory-level reward.

RebRL balances update opportunities; Dep2C-Credit distinguishes downstream dependency influence. These weights account for prompt-relevant downstream influence, providing a finer-grained basis for assigning trajectory-level reward. Equalizing masking-related update opportunities does not remove the need to distinguish tokens by their effects on subsequent predictions.

Response — Part 2/2: Implementation details and checkpoint selection

Q2. Entity extraction, token alignment, and decoding parameters

Entity extraction and token alignment.

Entities are phrases describing objects and their attributes, such as "a red car." We prioritize entity information in data annotations. When unavailable, we construct phrases from object categories and color attributes. For prompts without structured annotations, we apply rule-based splitting using conjunctions and spatial relation expressions.

We use the tokenizer's character-offset mapping to locate all subword tokens corresponding to each entity in the actual conditional text supplied to the model, including the input template. These text-token positions aggregate attention from masked image positions to prompt entities, estimating the masked positions' entity relevance for credit assignment. If no entity can be aligned, we fall back to the text tokens of the full prompt. We will include these implementation details in the revision.

Degree exponent.

We use βd=1.0\beta_d=1.0. Part 1 provides the priority score and variable definitions.

Threshold search bounds and stopping rule.

The threshold is determined adaptively from the current dependency matrix to control dependencies among jointly committed tokens while satisfying the commitment quota. The interval is [0,max⁡i,jDij][0,\max_{i,j}D_{ij}], covering the strictest to fully relaxed compatibility constraints without manually specifying thresholds for different prompts.

With priorities fixed, if greedy selection fills the quota at the current threshold, we lower the upper bound to attempt a stricter constraint; otherwise, we raise the lower bound. Search stops after 16 bisection-style updates, reducing the interval width to 2−162^{-16} of its initial range. Selection is rerun at the final upper bound. If the quota is still not met, we fall back to the maximum dependency value, ensuring that dependency constraints do not change the commitment size prescribed by the reveal schedule.

Q3. Main-table checkpoints, training budgets, and selection criteria

For the Lumina-DiMOO experiments corresponding to Figure 6, the main tables use Mask-GRPO and Dep2C checkpoints at 3,100 and 3,800 training steps, respectively. For both methods, we inspected online GenEval training success curves and manually selected a saved checkpoint from the high-success phase before clear degradation, rather than defaulting to the final checkpoint.

Mask-GRPO's training success declines noticeably after approximately 3.3K–3.5K steps, so we use its pre-degradation checkpoint at 3.1K. Dep2C maintains high success for longer, and we use its pre-degradation checkpoint at 3.8K. Thus, the main-table Mask-GRPO result is from before degradation, while the selected checkpoints have different training step counts.

We distinguish the full training run length from the evaluated checkpoint:

MethodActual stopping stepMain-table checkpointTable 7 checkpoint
Mask-GRPO350031003000
Dep2C450038003000

To separate training duration from method effects, Table 7 provides component ablations at a matched budget of 3,000 training steps on Lumina-DiMOO:

ConfigurationGenEvalUniPercept Mean
Mask-GRPO baseline0.8853.42
Credit only0.9053.42
Decoding only0.8853.65
Credit + Decoding0.9153.68

At matched training steps, Credit alone raises GenEval from 0.88 to 0.90; both components yield 0.91, a gain of 3 percentage points over the baseline. Counting improves from 0.86 to 0.93, Position from 0.87 to 0.91, and Attribute Binding from 0.72 to 0.77. These results show that compositional improvements persist under a matched training budget and cannot be attributed solely to Dep2C's additional training steps in the main tables.

The complete method also increases UniPercept Mean from 53.42 to 53.68. This metric is evaluated on images generated for GenEval and provides complementary evidence at matched training steps.

We will report the backbone, training step count, and checkpoint selection criterion explicitly in the relevant tables and experimental settings, distinguishing the main tables' curve-guided selection from Table 7's matched-budget comparison.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.