acceptodds
Under review as a conference paper at ICLR 2027

Dense-RAM: Edge-Level Credit Assignment for Reinforcement Learning in Flow-Based UV Unwrapping

Abstract

UV unwrapping is a fundamental step in creating textured 3D assets. Recent generative methods learn artist-like seam layouts from data, and reinforcement learning can further refine these layouts using UV-quality feedback. However, scalar feedback assigns a shared reward coefficient to all edges in a sampled layout, overlooking differences in local UV quality. To address this limitation, we introduce Dense-RAM, a reinforcement learning framework for post-training flow-matching seam generators. Built on Reinforce Adjoint Matching (RAM), Dense-RAM computes a candidate-level advantage from chart- and layout-level UV feedback, then distributes it across edges according to local quality. The resulting coefficients adapt the reward correction to each edge while preserving the candidate’s mean advantage. Experiments show that Dense-RAM improves chart topology and reduces UV fragmentation. Across RAM, DiffusionNFT, and AWM, edge-level allocation consistently improves final UV quality and reaches comparable quality in fewer optimization steps than scalar feedback.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
geometric modeling, surface parameterization, surface cutting

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_isL4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Dense-RAM, a reinforcement-learning post-training method for a flow-based UV seam generator. It combines chart- and layout-level UV feedback into edge-aligned scores, then distributes each candidate’s advantage across mesh edges while preserving its mean. The edge-level coefficients modify RAM’s velocity regression target. Paired experiments with RAM, AWM, and DiffusionNFT report better UV topology, less fragmentation, and higher atlas utilization than scalar feedback.

Critical strengths
  1. The edge-level credit allocation is a novel, well-motivated extension of scalar RAM. It uses structured UV feedback while preserving each candidate’s mean advantage, without adding trainable parameters.
  2. The paper is clearly written and well argued. Its theoretical analysis, matched scalar–dense comparisons, and reward ablations provide substantial support for the proposed method.
Critical weaknesses
  1. The edge scores reflect chart-level outcomes, not counterfactual seam decisions. Edges within the same distorted chart may receive indistinguishable feedback, limiting the method’s ability to identify where a missing seam should be added.
  2. Preserving the mean advantage does not preserve the update’s magnitude or direction. The paper lacks shuffled-weight or update-matched controls, multi-seed results, and paired distortion and failure-rate measurements.
Questions for authors
  1. For a high-distortion chart that would benefit from an additional seam, how does the method distinguish edges whose cutting would reduce distortion from other uncut edges in the same chart? Could you provide a targeted analysis?
  2. Could the authors report results on other publicly available test sets or established UV benchmarks to assess generalization beyond the current Objaverse split?
4 comments

Thank you for the positive assessment and the concrete requests. We ran every control you listed, and each one supports the allocation: with random weights of identical magnitude the all-disk rate falls from 64.98% to 23.49%, and scalar updates of matched or larger magnitude stay at 38.24–42.19%. All numbers are measured on the 500 held-out test meshes, and all tables will be added to the revised version.

W1 / Q1. Edges inside one high-distortion chart. Within a single candidate they are not distinguished, and the method does not rely on it: every uncut edge interior to the chart receives the same coefficient. Dense-RAM is edge-resolution allocation of candidate-level credit using region-level feedback, not per-edge attribution, and we will state this in Section 4.1.

The distinction comes from on-policy variation across the G=24G=24 candidates of a mesh. Candidates that leave the chart intact have lower bgb_g, and that region carries their largest negative coefficients. Candidates that cut it at a useful location produce passing sub-charts, obtain higher bgb_g, and those sub-charts carry their largest positive coefficients. A cut at an unhelpful location leaves a failing sub-chart and is not reinforced. Which edge to cut is thus learned from which sampled cuts succeed, and Appendix A.2 shows that all these cases push in the same parameter direction.

Targeted analysis (Table A). We sample SeamFlow, scalar RAM and Dense-RAM from identical initial noise and compare region by region. In charts that fail the distortion test under SeamFlow, 40.3% of the seams added by Dense-RAM fall in the top quartile of per-face distortion, against 31.0% for scalar RAM and 25% for a random edge; 55% of these charts are repaired, against 30%. Over the whole test set Dense-RAM repairs 60.5% of the failing area and damages 4.2% of the passing area; scalar RAM repairs 35.2% and damages 8.1%. A regional signal is therefore enough: the policy places new seams where distortion is concentrated, with less collateral change to valid regions.

Table A. Localization (paired sampling). Repair: failing under SeamFlow, passing after. Damage: the converse.

ModelFailing area (%) ↓Repair (%) ↑Damage (%) ↓New seams in top quartile (%)
SeamFlow30.4———
Scalar RAM22.135.28.131.0
Dense-RAM13.360.54.240.3

Q2. Other test sets. We evaluated the same checkpoints, without further training, on Toys4K. The ordering is preserved: Dense-RAM improves all-disk over scalar RAM by 20.66 percentage points and reduces islands by 10.56, with lower distortion.

Toys4KAll-disk (%) ↑Angle ↓Area ↓Islands ↓
SeamFlow27.590.130.1650.28
Scalar RAM43.270.130.1542.41
Dense-RAM63.930.110.1331.85

W2. Magnitude and direction; missing controls. Mean preservation alone does not fix the magnitude or direction of the update, so we tested both directly. None of the controls accounts for the gain.

  • Random weights (the shuffled-weight control). We keep each candidate's weight values but assign them to its edges uniformly at random. The weights are then independent of local UV quality, while their mean, second moment and clipped fraction equal those of Dense-RAM exactly. All-disk falls from 64.98% to 23.49%.
  • Update-matched controls. With σd\sigma_d the within-candidate standard deviation of dged_{ge}, Vare[w]≤α2\mathrm{Var}_e[w]\le\alpha^2, so the RMS coefficient of Dense-RAM is at most 1+α2≈1.118\sqrt{1+\alpha^2}\approx1.118 times that of scalar RAM. Reward coefficient ×1.12\times1.12, ×1.5\times1.5, ×2\times2: 39.72%, 40.17%, 38.24%. Learning rate ×1.5\times1.5, ×2\times2: 42.19%, 41.36%.
  • Direction. On identical batches at steps 0/100/250/500, the dense correction has a component γ∈\gamma\in [0.02, 0.06] along the scalar reward gradient and an orthogonal component ρ∈\rho\in [0.52, 0.64]. It is almost entirely a change of direction. Training with the orthogonal component alone recovers 89% of the gain; with the parallel component alone it matches scalar RAM.
  • Seeds. Three training seeds for RAM (Table B): all-disk 40.6 ± 1.9 vs. 64.7 ± 2.1, islands 54.5 ± 1.8 vs. 37.5 ± 1.7.
  • Paired statistics. Bootstrap over the 500 test meshes: +24.5 points all-disk (95% CI [21.0, 28.0]). Island counts are heavy-tailed, so we report the mean difference −17.4 (CI [−24.6, −10.5]) and the median difference −9 (CI [−12, −7]). Dense is better in all-disk on 41% of meshes, tied on 49%, worse on 10%.
  • Distortion. Dense allocation lowers angular and area distortion under all three objectives.

Table B. Controls, distortion and seed variation (500 steps). Visibility and utilization are as in Table 3.

ObjectiveVariantAll-disk (%) ↑Angle ↓Area ↓Islands ↓
—SeamFlow21.940.130.1666.42
DiffusionNFTScalar47.980.1210.14049.78
DiffusionNFTDense59.580.1140.12942.18
AWMScalar44.240.1230.14452.20
AWMDense56.180.1160.13244.38
RAMScalar (1.12β1.12\beta)39.720.1250.14655.1
RAMScalar (2β2\beta)38.240.1300.15056.09
RAMRandom weights23.490.1300.16065.3
RAM, 3 seedsScalar40.6 ± 1.90.125 ± 0.0040.145 ± 0.00554.5 ± 1.8
RAM, 3 seedsDense64.7 ± 2.10.111 ± 0.0030.121 ± 0.00437.5 ± 1.7

Every control you requested is thus in place, and each supports the same conclusion: the gain comes from aligned allocation. The revised version will add these controls, seed variation, distortion, the decomposition measurements, the localization analysis with two figures, and the Toys4K evaluation.

Dear Reviewer isL4,

We would appreciate your feedback on our rebuttal. Please let us know if any concerns remain. Thank you for your time and consideration.

The rebuttal substantially strengthens the empirical evidence and addresses most of my experimental concerns. My remaining reservation concerns the significance of the methodological contribution: the approach redistributes candidate-level feedback using chart-level quality, rather than distinguishing the contributions of individual edges within a chart. Although the localization analysis and additional evaluations demonstrate practical benefits, I remain unconvinced that this extension provides a sufficiently substantial advance in credit assignment. I therefore retain my weak-rejection recommendation, while acknowledging the authors’ substantial efforts and the improved experimental support.

Reviewer_Y9fPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Dense-RAM for reinforcement learning in flow-based UV unwrapping, using local UV quality to distribute candidate-level advantages across mesh edges. Post-training SeamFlow with this mechanism improves chart topology and reduces UV fragmentation compared with the pretrained baseline. Paired comparisons across multiple RL objectives also show consistent improvements over scalar feedback on the reported UV-quality metrics.

Critical strengths
  1. The paper applies local quality feedback to flow-based UV unwrapping, where seam decisions jointly affect chart topology and parameterization quality. The reported results provide useful evidence that spatially structured feedback is effective in this geometric setting, extending its applicability beyond visual generation.
  2. The proposed edge-level allocation is simple and requires no additional trainable modules. It can be incorporated into multiple flow-based RL objectives, making it a practical approach to exploiting regional UV-quality feedback.
  3. The reported experiments show consistent improvements over scalar feedback across RAM, DiffusionNFT, and AWM on the UV task, particularly in chart topology and fragmentation. The reward ablations also provide useful insights into the trade-offs between topology, distortion, and the number of UV islands.
Critical weaknesses
  1. Limited scope and impact. The core edge-level allocation rule is not inherently specific to flow models, yet all experiments use the same SeamFlow backbone. Testing it with an alternative policy, such as an autoregressive policy paired with a suitable RL algorithm, would strengthen the evidence for its generality and broader impact.
  2. The allocation mechanism is insufficiently validated. Prior work has demonstrated the benefits of local information for image generation post-training [1, 2]. Although the Scalar/Dense comparisons show improvements in UV unwrapping, they do not sufficiently isolate the value of Dense-RAM’s specific allocation design. First, Appendix A.2 acknowledges the trade-off between candidate-level sign consistency and independent edge-level reinforcement (L774–L777), but the benefit of enforcing sign consistency remains experimentally unclear; an independent per-edge advantage baseline would help. Second, shuffling edge weights within each candidate would help determine whether gains require correctly aligned local feedback, rather than nonuniform weighting alone. Third, Table 3 and Figure 5 omit angular and area distortion. These paired metrics are needed to assess potential trade-offs.
  3. Baseline adaptations need justification. Original RAM uses CFG during training and evaluation, but the guidance settings are unspecified here. Appendix B.5 explicitly uses a stop-gradient copy of the current policy without EMA (L914–L915), although original DiffusionNFT identifies EMA sampling-policy updates as important for stability. The authors should specify the reasons for these changes.
  4. Key training settings are missing. For example, the values of β\beta (L254) and τD\tau_D (L827–L833), the computation of σd\sigma_d (L857–L859), and essential normalization, optimization, and sampling settings are missing. A complete configuration table would improve reproducibility.
  5. Definitions and presentation need clarification. Define Table 1’s symbols (L207–L211) at first use and specify the evaluated chart set (L205) and the stage at which the reward island count is computed (L842–L844). The sentence AO denotes ambient occlusion. L203 appears out of place, as term AO is not used in the surrounding equations or Table 1. The reference at L316 should point to Appendix B.4 rather than Appendix C, and “twelve selected meshes” at L366 should be corrected to fourteen, matching Figure 4.

[1] Ziqi Ni et al. Seeing What Matters: Visual Preference Policy Optimization for Visual Generation. CVPR, 2026.

[2] Jiayang Sun et al. Towards Fine-Grained Attribution: Instance-Aware Preference Optimization for Aligning Diffusion Models. CVPR, 2026.

Questions for authors

Please see Weaknesses.

3 comments

Thank you for the detailed review. We ran both controls you proposed, and both support the allocation design: random edge weights collapse the all-disk rate from 64.98% to 23.49%, and independent per-edge advantages reach only 45.2%. All numbers are measured on the 500 held-out test meshes, and all tables will be added to the revised version.

W1. Scope: an autoregressive policy. We ran the experiment you suggested. The allocation rule is not specific to flow models: it only requires the policy's outputs to be aligned with the units at which the evaluator reports feedback. We apply it to SeamGPT, an autoregressive seam generator, post-trained with DPO as in SeamCrafter. Each seam token is mapped to the charts that its segment borders; the chart-level scores give mean-one token weights as in Eq. (7); these weights multiply the per-token log-likelihood-ratio terms of the DPO loss, and uniform weights recover standard DPO. With the same reward and test meshes, the allocation changes the all-disk rate from 39.2% to 44.7% and the island count from 54.37 to 47.88 relative to standard DPO (Table B). The allocation therefore improves four objectives across two policy families: RAM, AWM and DiffusionNFT on a flow model, and DPO on an autoregressive model. Along the data axis, the ordering Dense > Scalar > SeamFlow also holds on Toys4K (Part 2).

Table B. Autoregressive policy: SeamGPT with DPO.

VariantAll-disk (%) ↑Angle ↓Area ↓Islands ↓
SeamGPT (pretrained)18.50.170.2261.41
Standard DPO39.20.150.2154.37
DPO + dense allocation44.70.140.2047.88

W2. Validation of the allocation design.

Relation to [1, 2]. We will cite and discuss both. They show that spatially or instance-resolved preference signals improve post-training of image generators. Our setting differs: (i) the local signal comes from the task evaluator itself, with no extra perceptual or attribution model; (ii) the output domain is the edge set of an irregular mesh; (iii) the weights are nonnegative and mean-one, so each candidate's advantage is preserved and α=0\alpha=0 recovers the scalar objective exactly. Our contribution is not local feedback as such, but this advantage-preserving allocation.

Sign consistency vs. independent per-edge advantages. The baseline Cge=β(qge−1G∑hqhe)/σ^C_{ge}=\beta(q_{ge}-\tfrac1G\sum_h q_{he})/\hat\sigma has the same edge-mean as our rule, 1m∑eCge=Ag\tfrac1m\sum_e C_{ge}=A_g, so the comparison isolates the nonnegative-weight constraint. It reaches 45.2% all-disk; rescaled to the coefficient magnitude of Dense-RAM, 46.0%. Sign consistency is therefore the better design: charts are not aligned across candidates, so per-edge signs are noisy.

Random edge weights (the shuffling control you suggested). We keep each candidate's weight values but assign them to its edges uniformly at random. The weights are then independent of local UV quality, while the full coefficient distribution of Dense-RAM is preserved. All-disk drops from 64.98% to 23.49%. The gains require correctly aligned local feedback; non-uniform weighting alone is harmful. A larger scalar update does not explain the gain either (39.72% at the matched magnitude).

Distortion in Table 3 and Figure 5. Reported below; Table 3 and Figure 5 will be extended accordingly. Dense allocation lowers both measures under every objective, so the gains come without a distortion penalty.

Table A. Controls and distortion (500 steps). Visibility and utilization are as in Table 3.

ObjectiveVariantAll-disk (%) ↑Angle ↓Area ↓Islands ↓
—SeamFlow21.940.130.1666.42
DiffusionNFTScalar47.980.1210.14049.78
DiffusionNFTDense59.580.1140.12942.18
AWMScalar44.240.1230.14452.20
AWMDense56.180.1160.13244.38
RAMScalar40.500.1250.14554.62
RAMScalar (1.12β1.12\beta)39.720.1250.14655.1
RAMRandom weights23.490.1300.16065.3
RAMPer-edge advantage45.20.1250.14551.3
RAMDense64.980.110.1237.24

W3. Baseline adaptations.

CFG. SeamFlow is conditioned on mesh geometry through cross-attention and is trained without condition dropout, so it has no unconditional branch; no guidance is used for any method. We will state this in Section 5.1.

DiffusionNFT without EMA. With 500 optimizer steps, an EMA old policy stays close to the initial checkpoint for most of training, nearly equivalent to the frozen-reference option; we therefore used the current policy. The choice does not affect the conclusion. With the EMA old policy of the original method (decay 0.99), scalar reaches 46.1% all-disk / 51.0 islands and dense 57.2% / 43.9. Both settings will be reported.

W4. Missing settings (to be added to Appendix B).

SettingValue
β\beta20
σ^\hat\sigmaEMA (decay 0.9) of the std of bgb_g pooled over the batch
σd\sigma_dstd of dged_{ge} over the valid edges of candidate gg
α\alpha; reward weights0.5; 1 / 1 / 1 / 0.5
τD\tau_D3.0 on the chart MIPS energy σ1/σ2+σ2/σ1\sigma_1/\sigma_2+\sigma_2/\sigma_1 (equal to 2 without distortion)
Island reference / scale; seam threshold26 / 30; Se>0.5S_e>0.5
GG / KK / meshes per step / steps24 / 8 / 10 / 500
OptimizerAdamW, lr 3×10−53\times10^{-5}, weight decay 0.05
ODE steps; time density25 rollout / 50 evaluation, shift 3; p(t)=2(1−t)p(t)=2(1-t)
AWM; DiffusionNFTβKL=1.0\beta_{\mathrm{KL}}=1.0; spos=1.0s_{\mathrm{pos}}=1.0, bNFT=0.5b_{\mathrm{NFT}}=0.5
Evaluation3 candidates per mesh, fixed seed

The submission calls this quantity a Dirichlet energy. The implementation uses the MIPS energy, i.e. the Dirichlet energy normalized by area, and the revised version will use this name and state the formula.

W5. Presentation. All accepted; the revised version will:

  • define the symbols of Table 1 where the table is introduced;
  • unify Tgeval\mathcal T_g^{\mathrm{eval}} with Tgcut\mathcal T_g^{\mathrm{cut}}, the seam-induced charts before parameterization;
  • state that ngrewardn_g^{\mathrm{reward}} is counted on those charts, and the reported island metric on the final atlas;
  • move "AO denotes ambient occlusion" to the seam-visibility definition;
  • point L316 to Appendix B.4 and C.2;
  • correct "twelve" to "fourteen".

Additional test set: Toys4K (W1).

Toys4KAll-disk (%) ↑Angle ↓Area ↓Islands ↓
SeamFlow27.590.130.1650.28
Scalar RAM43.270.130.1542.41
Dense-RAM63.930.110.1331.85

The allocation design is thus validated by both controls you proposed, and the gains come without a distortion penalty. The revised version will add these results, the related-work discussion, the settings above, both NFT configurations and the scope statement.

Dear Reviewer Y9fP,

We would appreciate your feedback on our rebuttal. Please let us know if any concerns remain. Thank you for your time and consideration.

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

This paper proposes Dense-RAM for reinforcement-learning post-training of a flow-matching UV seam generator. It combines chart-level topology and distortion feedback with layout-level fragmentation and seam-visibility rewards. A group-centered candidate advantage is redistributed across mesh edges using nonnegative, mean-one weights derived from local quality, and the resulting coefficients modify the RAM regression target. Experiments on 500 held-out Objaverse meshes compare complete UV systems, isolate scalar versus edge-level allocation under RAM, DiffusionNFT, and AWM, and ablate the reward components.

Critical strengths
  • The paired Scalar/Dense comparisons provide useful evidence for the proposed allocation rule. Under RAM, all-disk rate increases from 40.50% to 64.98%, while UV islands decrease from 54.62 to 37.24; improvements also occur under DiffusionNFT and AWM. Holding the generator, reward, and downstream UV pipeline fixed within each pair makes these results more informative than only comparing against unrelated UV systems.
  • The method addresses a practical UV-layout trade-off rather than optimizing topology alone. The reward ablations show that removing the island penalty improves topology but increases fragmentation dramatically, from 37.2 to 312.4 islands. This helps explain why the full reward is useful for producing editable, efficiently packed layouts.
Critical weaknesses
  • Preserving the mean coefficient does not isolate credit allocation from a change in the effective optimization step. Appendix A.2 explicitly decomposes the update difference into a component along the scalar update and an orthogonal component, but the experiments do not quantify these components or compare against a scalar control with matched update magnitude or a learning-rate/reward-scale sweep. The empirical gains are substantial, but the current evidence does not establish how much comes from more informative spatial allocation versus stronger effective updates.
  • The edge-level attribution and its connection to practical layout quality remain insufficiently established. Local scores largely inherit chart-level outcomes, which do not identify which incident seam decisions caused a chart to be good or bad. The reward ablations demonstrate a strong trade-off between topology, distortion, and fragmentation, but there is no matched-quality analysis showing whether dense allocation remains preferable at comparable distortion or island count. The results support the full method under its chosen reward, while leaving the stronger credit-assignment interpretation and the quality trade-off less clear.
Questions for authors
  1. What evidence distinguishes better credit assignment from a different optimization direction or strength? Appendix A.2 shows that the dense correction can contain both a component parallel to the scalar update and an orthogonal component. The Scalar/Dense results establish that the full modification helps, but they do not tell us which component is responsible. Could the authors explain whether the improvement persists when these two effects are separated, for example by retaining only the parallel or orthogonal contribution, or by comparing against a scalar update of comparable magnitude?

  2. In what sense does chart quality identify the contribution of an individual seam edge? The topology and distortion signals are computed after cutting and parameterization, then assigned to incident edges. A bad chart may result from a small subset of its boundary decisions, while many other incident edges are already appropriate. Also, all nonzero edge coefficients retain the candidate's sign, and a candidate with zero group advantage supplies no local correction even if some charts are good and others are bad. Could the authors explain how the method handles a globally good layout containing a locally harmful seam, and the reverse case?

4 comments

Thank you for the precise questions. We ran every control you asked for. In short: scalar updates of matched or larger magnitude stay at 38.24–42.19% all-disk; a correction of identical magnitude with random allocation drops to 23.49%; Dense-RAM reaches 64.98%. All numbers are measured on the 500 held-out test meshes, and all tables will be added to the revised version.

W1 / Q1. Allocation vs. a stronger effective update. The two effects are separable, and we separate them here.

  • Magnitude bound. With σd\sigma_d the within-candidate standard deviation of dged_{ge}, Vare[w]≤α2\mathrm{Var}_e[w]\le\alpha^2 holds with or without clipping. By Prop. 2 the RMS coefficient of Dense-RAM is at most 1+α2≈1.118\sqrt{1+\alpha^2}\approx1.118 times that of scalar RAM (measured second-moment ratio: 1.18).
  • Scalar updates of matched or larger magnitude. Reward coefficient ×1.12\times1.12 (RMS-matched), ×1.5\times1.5, ×2\times2: 39.72%, 40.17%, 38.24%. Learning rate ×1.5\times1.5, ×2\times2: 42.19%, 41.36%.
  • Random weights. We keep each candidate's weight values but assign them to its edges uniformly at random. The weights are then independent of local UV quality, while their mean, second moment and clipped fraction equal those of Dense-RAM exactly. All-disk falls to 23.49%, close to the pretrained model.
  • Measured components (Table B). On identical batches we compute the scalar reward gradient ss and the dense correction c=g^D−g^Sc=\hat g^D-\hat g^S, and report γ=⟨c,s⟩/∥s∥2\gamma=\langle c,s\rangle/\|s\|^2 and ρ=∥c−γs∥/∥s∥\rho=\|c-\gamma s\|/\|s\| (split-half estimates over 20 batches). γ∈\gamma\in [0.02, 0.06] and ρ∈\rho\in [0.52, 0.64]: the correction is almost entirely orthogonal to the scalar update.
  • Retaining one component, as you suggested. Parallel only: 41.2%, i.e. scalar RAM. Orthogonal only: 62.3%, recovering 89% of the gain.

The improvement comes from the direction of the correction and its alignment with local quality. A stronger step does not reproduce it, and a correction of the same magnitude without alignment eliminates it.

Table A. Allocation controls under RAM (500 steps).

VariantAll-disk (%) ↑Angle ↓Area ↓Islands ↓
SeamFlow21.940.130.1666.42
Scalar (β\beta)40.500.130.1554.62
Scalar (1.12β1.12\beta)39.720.130.1555.13
Scalar (1.5β1.5\beta)40.170.130.1554.82
Scalar (2β2\beta)38.240.130.1556.09
Scalar (lr ×1.5\times1.5)42.190.130.1553.67
Scalar (lr ×2\times2)41.360.130.1554.15
Random weights23.490.130.1665.32
Per-edge advantage45.210.130.1551.31
Parallel only41.220.130.1454.07
Orthogonal only62.350.110.1239.15
Dense (α=0.5\alpha=0.5)64.980.110.1237.24

Table B. Decomposition of the dense correction.

Stepγ\gammaρ\rhocos⁡(g^D,g^S)\cos(\hat g^D,\hat g^S)
00.020.640.85
1000.030.610.86
2500.050.570.88
5000.060.520.90

W2 / Q2. What chart quality identifies. Dense-RAM does not rely on per-edge counterfactual attribution, and the results show that it does not need to. qgeq_{ge} is a regional label: all edges interior to one chart share one score. The method is edge-resolution allocation of candidate-level credit using region-level feedback, and we will state it in these terms. The allocation decides where a candidate's outcome is applied strongly; which edge changes is resolved by variation across the G=24G=24 candidates of a mesh, whose cuts differ.

Your cases follow from Eq. (7) and Remark 1:

  • Good layout, harmful seam (Ag>0A_g>0, dge<0d_{ge}<0). With the harmful region covering a fraction pp of the edges: for p<0.2p<0.2 its coefficient is exactly zero; for 0.2≤p≤0.50.2\le p\le0.5 its weight is 1−α(1−p)/p∈[0,0.5]1-\alpha\sqrt{(1-p)/p}\in[0,0.5]. Scalar RAM reinforces it with the full AgA_g. The negative signal comes from candidates with Ah<0A_h<0 in which that region also fails, where it carries the largest negative coefficient.
  • Poor layout, good region (Ag<0A_g<0, dge>0d_{ge}>0). The region receives a reduced or zero negative coefficient and is shielded from the uniform penalty of scalar RAM.
  • Ag=0A_g=0. No local correction. This is the deliberate cost of preserving each candidate's advantage, and the alternative that removes it performs worse: independent per-edge advantages, with local signal at Ag=0A_g=0 and mixed signs, reach 45.2% (Table A). This is expected, since charts are not aligned across candidates.

These cases are common: at initialization 78% of positive-advantage candidates contain a failing chart, with mean weight 0.48 on its edges vs. 1.16 elsewhere.

This reduces interference in practice. Sampling all models from identical initial noise and comparing region by region (Table C), Dense-RAM repairs more failing area than scalar RAM and damages about half as much passing area.

Table C. Localization (paired sampling). Repair: failing under SeamFlow, passing after. Damage: the converse.

ModelFailing area (%) ↓Repair (%) ↑Damage (%) ↓
SeamFlow30.4——
Scalar RAM22.135.28.1
Dense-RAM13.360.54.2

W2. Trade-off and matched-quality analysis. Dense allocation is lower in angular and area distortion under all three objectives (RAM: Table A). The gains are not bought with distortion.

ObjectiveAlloc.Angle ↓Area ↓
DiffusionNFTScalar / Dense0.12 / 0.110.14 / 0.13
AWMScalar / Dense0.12 / 0.120.14 / 0.13

Table D. Matched fragmentation (RAM).

Model @ stepIslands ↓All-disk (%) ↑Angle ↓Area ↓
Scalar @ 50054.6240.500.130.15
Dense @ 15054.7245.810.120.14
Scalar @ 100048.2650.120.120.14
Dense @ 50037.2464.980.110.12

At the same island count Dense-RAM has a higher all-disk rate and comparable distortion, reached in 150 steps instead of 500. Doubling the scalar schedule does not reach Dense-RAM at 500 steps.

Both concerns are thus answered with direct measurements: the gain is not a step-size effect, and a regional signal is sufficient to reduce interference. The revised version will add these controls and measurements, distortion in Table 3 and Figure 5, and the statements on the regional nature of qgeq_{ge} and the Ag=0A_g=0 case.

Dear Reviewer egbB,

We would appreciate your feedback on our rebuttal. Please let us know if any concerns remain. Thank you for your time and consideration.

Thanks the authors for the detailed rebuttal. These additional results addressed my main concerns. However, the overall approach appears to build upon existing RL-based post-training and reward-weighted regression frameworks, with the main contribution is somewhat limited. Therefore, I lean to maintain my current score.

Comments on the paper

1 comment

We thank all reviewers for their constructive feedback. The reviewers recognized the consistent improvements across post-training objectives (egbB, Y9fP), the simple allocation mechanism without additional trainable modules (Y9fP, isL4), and the method’s novelty and motivation (isL4). The main questions concerned the source of these gains, how regional feedback improves seam decisions, and generalization beyond the original setting. We have added targeted experiments and clarifications to address these questions. All changes are highlighted in red in the revised PDF.

The key revisions are summarized below:

  1. Allocation controls and update decomposition (Apps. C.4–C.5, Tabs. 6–7): We add reward-scale and learning-rate sweeps, shuffled weights, independent per-edge advantages, and parallel/orthogonal correction controls. Dense-RAM achieves 64.98% all-disk, compared with at most 42.19% in the scalar sweeps and 23.49% with shuffled weights. Retaining only the orthogonal correction recovers 89% of the improvement over scalar RAM, supporting the importance of the direction and spatial alignment of the correction.

  2. Regional feedback and local improvements (Sec. 4.1, App. C.6, Tab. 8): We clarify how chart-level quality guides edge-level allocation and add paired localization analysis. Dense-RAM repairs more initially failing area than scalar RAM (60.5% vs. 35.2%) while damaging less initially passing area (4.2% vs. 8.1%), demonstrating the practical benefit of regional feedback.

  3. UV-quality trade-offs and repeatability (Tab. 3; Apps. C.7–C.8, Tabs. 9–10): We add paired angular and area distortion, comparisons at matched fragmentation, longer scalar training, three training seeds, and paired bootstrap intervals. Dense allocation lowers both distortion measures under all three flow objectives, achieves higher all-disk rate at comparable fragmentation, and maintains its gains across training seeds.

  4. Additional policy and test set (Apps. C.10–C.11, Tabs. 12–13): Applying token-level allocation to SeamGPT with DPO improves all-disk from 39.2% to 44.7%. On Toys4K, the same RAM checkpoints without further training achieve 63.93% all-disk with dense allocation versus 43.27% with scalar allocation. These experiments extend the evidence to four objectives across two policy families and an additional public test set.

  5. Implementation details and presentation (Sec. 2, App. B, App. C.9; Tabs. 5, 11): We provide the complete training configuration, clarify the MIPS reward and chart definitions, explain the guidance settings, add a DiffusionNFT EMA comparison, and include the related ViPO and IAPO studies.

Detailed responses and supporting results are provided in the corresponding reviewer discussions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.