acceptodds
Under review as a conference paper at ICLR 2027

GIF: A Conditional Multimodal Generative Framework for IR Drop Imaging in Chip Layouts

Abstract

IR drop analysis is a critical step in physical chip design, ensuring the power integrity of on-chip power delivery networks. As transistor density scales, the computational cost of IR drop analysis grows substantially, motivating scalable ML-based alternatives. Existing ML approaches formulate IR drop analysis as a spatial regression problem using power map representations, which fail to jointly model the geometric and topological factors that govern voltage drop distribution across the layout. We propose GIF (Generative IR drop Framework), to our knowledge the first framework to reformulate IR drop analysis as a conditional image generation problem. GIF conditions a diffusion model on two complementary inputs: a 34-channel geometry-aware feature image capturing local power consumption, cell density, and PDN resistance factors; and graph tokens encoding logical connectivity from the gate-level netlist, injected via gated cross-attention. This allows GIF to jointly capture local geometric variation and long-range structural dependencies that regression-based approaches do not model. On CircuitNet-N28 (28 nm), GIF achieves 0.791 SSIM and 0.9406 Pearson correlation. As a generative model, GIF produces calibrated uncertainty estimates as a natural byproduct of sampling. GIF also reports IR drop results on CircuitNet-N14 (14 nm), achieving 0.9106 Pearson and 0.8284 Spearman correlation across diverse design styles and establishing baselines for future work at this technology node.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computer Vision, Applied learning, Chip design

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

The paper proposes GIF for IR-drop prediction, combining layout feature maps and gate-level netlist representations within a conditional diffusion model. Experiments on CircuitNet-N28 and N14 evaluate prediction accuracy, hotspot detection, and uncertainty estimated from multiple generated samples.

Critical strengths
  1. GIF adapts established multimodal conditional diffusion ideas to combine complementary layout and netlist information for IR-drop prediction. This offers a useful application framework that produces both spatial predictions and uncertainty estimates.
Critical weaknesses
  1. The paper does not sufficiently distinguish its contribution from related topology–geometry encoding and multimodal fusion methods, such as DeepLayout[1] and LMM-IR[2]. Its empirical competitiveness also remains unclear without comparisons against more recent IR-drop predictors.

[1] DeepLayout: Learning Neural Representations of Circuit Placement Layout. ICML 2025.

[2] LMM-IR: Large-Scale Netlist-Aware Multimodal Framework for Static IR-Drop Prediction. DAC 2025.

Questions for authors
  1. Appendix Tables 12–13 report GIF’s sampling times. How does its inference latency, including four-sample averaging, compare with the regression baselines under matched hardware and input resolution?
  2. Beyond the graph/no-graph comparisons in Table 4, can the authors isolate the contribution of logical connectivity by retaining node features and model capacity while removing or randomizing edges?
3 comments

We thank the reviewer for the careful reading and the valuable comments. Below are our responses to each concern; we ran both suggested experiments. All changes in the revised PDF are marked in blue.

1. "The paper does not sufficiently distinguish its contribution from ... DeepLayout and LMM-IR." These two works solve different problems from GIF:

  • LMM-IR (DAC 2025) is for static IR drop (its title), on the ICCAD 2023 contest data. Its netlist is the power-grid SPICE netlist (resistors and current and voltage sources). It outputs one IR drop map.
  • DeepLayout (ICML 2025) learns general layout representations by self-supervised pretraining, for layout quality tasks. It could be added to GIF as another input.
  • GIF is for dynamic IR drop on CircuitNet, the same task as MAVIREC, PDNNet, and DuST. It uses the logical gate-level netlist (which cells are connected), not the power-grid netlist. It is a generative model, so one trained model gives the IR drop map and an uncertainty map.

We added both papers to Table 1 and the related work.

2. "Its empirical competitiveness also remains unclear without comparisons against more recent IR-drop predictors." The paper already compares GIF with recent IR drop methods; in both comparisons, GIF has the highest Pearson and is the only method with an uncertainty map. Table 3 (the same test layouts as PDNNet; GIF uses one sample):

SSIMPearsonUncertainty map
MAVIREC (DATE 2021)0.680.91No
PDNNet (2024)0.720.92No
GIF0.790.95Yes

Table 5 (regressors trained on the same 34-channel input):

PearsonUncertainty map
Restormer (CVPR 2022)0.936No
MAVIREC0.941No
GIF, N=40.941Yes
GIF, N=80.945Yes

For the recent DuST (DATE 2025), its own evaluation code gives MAVIREC an SSIM of 0.761 instead of 0.68, so its numbers are on a different scale (Section 4), and it gives no uncertainty map.

3. Q1: "How does its inference latency, including four-sample averaging, compare with the regression baselines?" One RTX A6000, same input, one test layout:

TimeOutput
Regressors (U-Net / Restormer / MAVIREC)17 / 25 / 55 msIR drop map
GIF, DDIM 100 steps, N=45.7 sIR drop map + uncertainty map
GIF, DDPM 1,000 steps, N=456.7 sIR drop map + uncertainty map

4. Q2: "Isolate the contribution of logical connectivity by retaining node features and model capacity while removing or randomizing edges." GIF is robust to edge changes, by design. Test on the trained N28 model (96 test layouts, same noise in all rows except the last):

Graph tokens computed fromSSIMPearsonChange of the IR drop map
Full graph0.7750.91750
Edges removed (node features kept)0.7880.91650.0040
Edges randomly rewired0.7760.91750.0012
Full graph, different noise0.7860.91700.0286

The graph tokens are one summary of the whole netlist (mean pooling), so single edge changes move the IR drop map 7x to 24x less than the sampling noise does. The contribution of the netlist is learned in training: models trained with graph tokens have higher SSIM (0.780 vs. 0.765) and 9x more stable Pearson across seeds (Appendix).

Thank you for the detailed response, the clarification regarding DeepLayout and LMM-IR, and the additional experiments. Two concerns remain:

  1. Recent baselines. My concern in W1 was the lack of direct comparisons with representative methods published in 2025–2026 at major EDA or AI venues that address the same task or can reasonably be adapted using comparable inputs. Another reviewer 7htC raised a related concern about the baseline coverage.

  2. Contribution of logical connectivity. Rewiring the edges at inference leaves Pearson unchanged, and removing them improves SSIM with little change in Pearson. The training comparison with and without graph tokens measures the effect of the entire graph branch. A matched training comparison using real versus randomized edges, with node features and architecture held fixed, would better isolate the contribution of logical connectivity.

Given these remaining concerns, I maintain my rating of 2.

Thank you for for reading and the follow-up. We answer the concerns from the follow up below.

1. "Direct comparisons with representative methods published in 2025-2026 ... that address the same task or can reasonably be adapted using comparable inputs."

The recent methods solve a different problem or give a different output:

MethodWhat it estimatesPower-grid inputLogical netlistOutput
LMM-IR (DAC 2025)static IR dropSPICE netlist of the power gridnoone IR drop map
PDNNet (2024)dynamic IR dropgraph of the power gridnoone IR drop map
DuST (DATE 2025)dynamic IR dropnonenoone IR drop map
GIFdynamic IR dropnoneyesIR drop map and uncertainty map
  • LMM-IR solves static IR drop and needs the power-grid SPICE netlist. It is built for a different problem and a different input.
  • DuST solves the same task and uses the same training and test layouts as Table 3. However, its own evaluation code gives MAVIREC an SSIM of 0.761, while Table 3 lists 0.68. Its numbers are therefore on a different scale (Section 4).
  • GIF has the highest SSIM (0.79) and Pearson (0.95) among the methods in Table 3, with one sample. It is the only method that gives an uncertainty map; this uncertainty is 2.0x higher on missed hotspots.

2. "A matched training comparison using real versus randomized edges ... would better isolate the contribution of logical connectivity."

GIF's claim is about the netlist as a whole, not about single edges. By design, the graph tokens are one summary of the whole netlist (mean pooling, Section 3). The test that matches this claim is the matched training comparison we ran, with the same architecture and the same features, with and without graph tokens, five training runs:

  • SSIM: 0.765 → 0.780 with graph tokens.
  • SSIM is higher on 90% of the test layouts.
  • Pearson varies 9x less across seeds.

We hope these clarifications address your concerns. We would be grateful if you would consider updating your score, and we are happy to answer any further questions during the discussion.

Reviewer_7htCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents GIF, a conditional multimodal generative framework for IR drop analysis in chip layouts. It reformulates IR drop prediction as a conditional image generation task using a diffusion model, conditioned on a 34-channel geometry-aware feature image and graph tokens extracted from the gate-level netlist. Experiments on CircuitNet-N28 and CircuitNet-N14 demonstrate state-of-the-art performance in image quality, correlation, and hotspot detection, along with per-pixel uncertainty estimation without requiring additional models.

Critical strengths

The reformulation of IR drop analysis as a conditional image generation problem is novel and well-motivated. Leveraging a diffusion model allows for natural uncertainty estimation, which is a significant practical advantage over deterministic regression methods.

The construction of the 34-channel geometry-aware feature image is physics-grounded, explicitly decomposing IR drop into current demand and effective PDN resistance. The additional layout-aware channels provide complementary signals missed by power maps alone.

The multimodal fusion design is thoughtful: geometry conditioning via adaptive normalization at all resolutions captures local variation, while graph tokens via gated cross-attention at the bottleneck capture global connectivity. The zero-initialized gate is a clever addition for stable training.

Comprehensive evaluation on two datasets (N28 and N14) with detailed ablation studies. The method establishes the first IR drop baselines on CircuitNet-N14 and shows that ensemble sampling improves performance.

Critical weaknesses

On CircuitNet-N28, the learned gate value (tanh(α)=0.036) indicates graph tokens contribute very little when designs share a common architecture. On CircuitNet-N14, graph conditioning is often unavailable due to incomplete metadata, and where available, it shows limited or negative impact on pixel-level metrics. This raises concerns about the practical robustness of graph-based conditioning in real-world scenarios with imperfect data.

The paper compares with prior methods using a single generated sample (Table 3) but reports its best results using the mean of N=4 samples (Table 4). This makes it difficult to isolate whether improvements over baselines come from the generative formulation or simply from ensembling.

Despite strong Pearson correlation, Spearman correlation remains relatively low, indicating that the ranking of IR drop tiles is not well preserved. Since hotspot detection heavily relies on rank ordering, this limits the practical utility for identifying the most critical hotspots.

Questions for authors
  1. For CircuitNet-N14, since many graphs are missing, how does the model handle samples with zero graph tokens during training and inference? Is there a fallback mechanism, or is it purely geometry-conditioned in those cases?

  2. The paper mentions that generating multiple samples is computationally expensive (12.6s for DDPM). Have the authors explored distillation or faster samplers (like DDIM with 100 steps, which Table 12 shows has lower SSIM) to balance speed and uncertainty quality?

  3. Can the authors provide more insight into why Spearman correlation remains low even with graph conditioning? Would a ranking loss or post-processing step help improve hotspot detection performance?

5 comments

We thank the reviewer for the careful reading and the valuable comments. Below are our responses to each concern. All changes in the revised PDF are marked in blue, and the new results are in the Appendix, "Additional Results (Discussion Period)".

1. "The learned gate value (tanh(α)=0.036) indicates graph tokens contribute very little." A small gate does not mean a small contribution. The gate scales the cross-attention output, whose size is also learned: in the trained N28 model, the gated output is 23% of the size of the bottleneck features. Training with and without graph tokens shows the contribution (same 48 test layouts, same noise, mean of N=4 samples):

SSIMPearsontanh(α)
With graph tokens (3 seeds)0.780 ± 0.0120.9416 ± 0.00020.036, 0.037, -0.042
Without graph tokens (2 seeds)0.765 ± 0.0200.9405 ± 0.0018n/a

The gate is small in every seed, and graph tokens still raise SSIM by the same amount as in Table 4 and make Pearson 9x more stable across seeds.

2. "On CircuitNet-N14, graph conditioning is often unavailable ... and where available, it shows limited or negative impact." N14 does not measure graph conditioning. The netlist graph files were not linked to the N14 layouts, so all N14 models are geometry-only; the revised PDF states this in Section 4, Table 7, and the Appendix. N14 therefore shows no negative graph effect. It shows that the geometry image alone gives the first N14 results (Pearson 0.91).

3. "Practical robustness of graph-based conditioning in real-world scenarios with imperfect data." GIF is robust when the netlist is missing. On 96 N28 test layouts (same noise, one sample):

Graph tokens at inferenceSSIMPearson
From the netlist0.7750.9175
Zero (no netlist)0.7880.9167

4. "Difficult to isolate whether improvements over baselines come from the generative formulation or simply from ensembling." The comparison with the baselines in Table 3 uses one sample (N=1), so it contains no ensembling; the Table 3 caption now says this. With one sample, GIF improves over the best baseline, PDNNet: SSIM 0.72 → 0.79 and Pearson 0.92 → 0.95. Averaging is an extra step on top (Pearson 0.916 → 0.941 from N=1 to N=4). In addition, the generative formulation gives the uncertainty map from one trained model (all 1,122 test layouts):

Variance-error correlationTrained models
GIF0.681
Regressor ensemble0.543
MC-dropout0.481

On missed top-10% hotspots, the uncertainty is 2.0x higher than on found hotspots, so the uncertainty map shows where to re-simulate.

5. "Spearman correlation remains relatively low ... this limits the practical utility for identifying the most critical hotspots." The low Spearman comes from the IR drop maps, not from GIF. In 102 N28 test layouts, 46% of the tiles have zero IR drop, and Spearman over the full map mostly ranks these tiles. The U-Net and MAVIREC regressors show the same value, and on tiles with IR drop, Spearman is much higher:

Full mapTiles with IR drop ≥ 0.01
GIF, N=40.540.70
U-Net regressor0.540.75
MAVIREC regressor0.560.77

On N14, where fewer tiles have zero IR drop, GIF reaches Spearman 0.84 (N=1) and 0.87 (N=4). For hotspots, the direct measure is top-10% precision, which is 0.80 (Table 6), so hotspot detection is measured by this precision, not by the full-map Spearman.

Q1. With zero graph tokens, GIF uses the geometry image only; no fallback is needed (point 3).

Q2. DDIM with 100 steps is 9x faster (1.4 s per sample) and keeps Pearson (0.930) and the variance-error correlation of the uncertainty map (0.582) (Table 12). Distillation (Salimans & Ho, ICLR 2022) can reduce the time further; we added it to the discussion.

Q3. Since the low full-map Spearman comes from tiles without IR drop (point 5), we report hotspot detection directly with top-10% precision and recall (Table 6).

Reviewer_7htCOpenReview ↗

Motivation and Baseline

Dear authors, Thank you for your reply. However, I have several critical concerns about the motivation and baseline.

  • A multimodal method has been proposed in Ma et al. (2025). However, the authors do not provide any experimental comparison with it. Although the authors claim that the previous work uses spatial layout features but not the logical connectivity of the gate-level netlist. I am not sure whether the logical connectivity of the gate-level netlist is necessary and how much improvement would be achieved if the feature is considered.
  • The authors mention previous work applying them to node-level voltage estimation or global PDN analysis rather than generating full-resolution IR-drop maps. However, it is possible that a super-resolution module or PINN cascaded with the output can generate the full-resolution IR-drop maps.

Thank you for the reading the response and follow up. We answer concerns raised in follow up below.

1. "A multimodal method has been proposed in Ma et al. (2025) ... no experimental comparison."

LMM-IR and GIF solve two fundamentally different problems.

LMM-IR (DAC 2025)GIF (ours)
What it estimatesstatic IR dropdynamic IR drop (the CircuitNet task, as in PDNNet and DuST)
Power-grid inputrequired: "the PDN model is provided as a SPICE netlist"none
Logical netlist (which cells are connected)noyes
Cell density and routing featuresnoyes
Outputone IR drop mapIR drop map and uncertainty map
DataICCAD 2023 contestCircuitNet-N28 and N14

So a direct comparison would test LMM-IR on a problem it was not built for.

Among methods for the same task, GIF is also the only one that uses the logical netlist and gives an uncertainty map:

MethodTaskPower-grid inputLogical netlistUncertainty map
PDNNet (2024)dynamicgraph of the power gridnono
MAVIREC (DATE 2021)dynamic (vectored)resistance mapsnono
DuST (DATE 2025)dynamicnonenono
GIFdynamicnoneyesyes

"I am not sure whether the logical connectivity ... is necessary and how much improvement would be achieved."

The improvement is measured, with the same architecture and the same features:

  • Table 4: graph tokens raise SSIM by +0.014.
  • Training seeds (Appendix): graph tokens raise SSIM by +0.015 (0.765 → 0.780), SSIM is higher on 90% of the test layouts, and Pearson varies 9x less across seeds.

2. "A super-resolution module or PINN cascaded with the output can generate the full-resolution IR-drop maps."

This cascade does not exist as a method for this task, so it cannot be a baseline. If it were built, it would:

  • need a second model after the first one, and
  • still give only one deterministic map.

GIF gives the full-resolution IR drop map and its uncertainty map from one model. This matters in practice. A deterministic method gives no warning where it is wrong, so a missed hotspot stays hidden. In GIF, the uncertainty on missed top-10% hotspots is 2.0x higher than on found hotspots, and the uncertainty tracks the error better than a regressor ensemble and MC-dropout (variance-error correlation 0.68 vs. 0.54 and 0.48). The engineer sees where to check before routing.

We hope these facts address the remaining concerns, and we would be grateful if the reviewer would consider them in the rating.

Reviewer_WPYiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GIF to model chip IR-drop prediction as conditional image generation. It uses a 34-channel layout feature map as the spatial condition and netlist-encoded graph tokens for connectivity. Feature modulation and gated cross-attention jointly control the diffusion model to generate IR-drop maps. The authors evaluate prediction quality on CircuitNet-N28 and CircuitNet-N14. They include ablations on feature channels, graph conditioning, ControlNet, and sampling, and compare GIF with deterministic regressors, MC-dropout, and regressor ensembles.

Critical strengths

1.The 34-channel input is backed by solid results: In Table 4, without ControlNet, expanding the input from 24 to 34 channels raises SSIM from 0.699 to 0.739 and lowers MAE from 0.0378 to 0.0352. This proves that layout and routing features add real value beyond standard power maps. It is a practical way to improve IR-drop prediction.

2.The uncertainty estimation shows real promise: Across the 72 test layouts, GIF reaches a variance-error correlation of 0.68, beating both MC-dropout (0.50) and the regressor ensemble (0.55). Scale calibration on the N28 validation set also drops the test miscalibration area from 0.37 to 0.11. These results show that sample variance is a good way to find unreliable regions.

Critical weaknesses
  1. The gains from diffusion and graph conditioning over simple deterministic models are small, and key baselines are missing. a. Single-sample GIF performs worse than the deterministic U-Net, and the paper only tests N=1 and N=4 without showing how accuracy changes across more sample counts (e.g., N=1, 2, 4, 8). b. Graph conditioning brings very small and unstable gains (it does not help the U-Net baseline or single-sample N14), and there are no mean or variance results across multiple random seeds. c. The paper misses two closely related works: it does not compare with DuST (DATE 2025, https://bunchgrape.github.io/docs/date25_irdrop_prediction.pdf), an open-source IR-drop predictor on N28.

  2. Some conclusions on the N14 dataset are not backed by the data. a. In Table 7, single-sample results are tested on all 3,456 samples, while the 4-sample mean is only tested on 100 samples; subtracting the two to claim a 1.97 dB gain without testing N=1 on the exact same 100 samples is unfair. b. Table 10 shows that 98% of all placements (10,585/10,795) and all 3,456 test samples have valid graphs, and the sample total (10,795) does not match Table 9 (10,444), so the data does not support the claim that ‘a large fraction of designs contain incomplete placement metadata, producing no valid graph tokens’.

Questions for authors
  1. Were the main models and key ablations trained independently with multiple random seeds? If so, please provide the results.

  2. With a fixed model checkpoint, sampler, and test set, can you report PSNR, SSIM, MAE, and RMSE for N=1, 2, 4, and 8 generated samples? I am curious whether the accuracy gain from sample averaging saturates and why N=4 was chosen.

  3. In Table 7, why do you directly subtract the 1-sample result on all 3,456 test samples from the 4-sample mean on just 100 samples to claim a 1.97 dB gain? Also, why does the sample total in Table 9 (10,444) not match Table 10 (10,795), and how could missing graphs hurt N14 performance when Table 10 shows valid graphs for 98% of all placements and 100% of the 3,456 test samples?

2 comments

We thank the reviewer for the careful reading and the valuable comments. Below are our responses to each concern. All changes in the revised PDF are marked in blue.

w1a/Q2. "Single-sample GIF performs worse than the deterministic U-Net" and N = 1, 2, 4, 8. A single sample is one random draw, while the U-Net outputs one average map; the GIF output that matches it is the mean of N samples. With N=4, GIF is already above the U-Net in Pearson, and it also gives an uncertainty map. Same checkpoint, same sampler, all 1,122 N28 test layouts:

SSIMPearsonUncertainty map (variance-error correlation)
GIF, N=10.7780.916n/a
GIF, N=20.7870.9320.53
GIF, N=40.7910.9410.64
GIF, N=80.7940.9450.68
U-Net regressor0.8360.937none
Regressor ensemble / MC-dropoutn/an/a0.54 / 0.48

The gain saturates at N=4: from N=4 to N=8, SSIM changes by only +0.003 and Pearson by +0.004, at twice the time. At N=4, the uncertainty map (0.64) is already better than the regressor ensemble (0.54) and MC-dropout (0.48). This is why we use N=4. PSNR, MAE, and RMSE for all N are in the Appendix, "Number of Generated Samples".

w1b/Q1. "Graph conditioning brings very small and unstable gains" and multiple random seeds. The gain is stable across training seeds. We trained GIF with graph tokens (3 seeds) and without graph tokens (2 seeds), 60 epochs each, and tested all models on the same 48 test layouts with the same noise (mean of N=4 samples):

SSIMPearson
With graph tokens (3 seeds)0.780 ± 0.0120.9416 ± 0.0002
Without graph tokens (2 seeds)0.765 ± 0.0200.9405 ± 0.0018

The SSIM gain is the same as in Table 4 (+0.015 here, +0.014 in Table 4), SSIM is higher on 90% of the test layouts, and Pearson varies 9x less across seeds. The two cases in the review do not test GIF with graph tokens: the U-Net is a regressor, not GIF, and the N14 models are geometry-only (point 5). For the feature ablation, the 24-to-34 channel gain in Table 4 (+0.040 SSIM) is more than twice the seed standard deviation. (Appendix, "Training Seeds and Graph Tokens".)

W1c:. "It does not compare with DuST." In Table 3, GIF has the highest SSIM (0.79) and Pearson (0.95), and it is the only method that gives an uncertainty map. DuST uses the same test layouts but its own evaluation code, and that code gives MAVIREC an SSIM of 0.761, while Table 3 has 0.68 for MAVIREC. DuST's numbers are therefore on a different scale from Table 3. We discuss DuST next to Table 3 (Section 4) and added it to Table 1 and the related work.

W2a. Table 7: "subtracting the two to claim a 1.97 dB gain ... is unfair." The gain from averaging holds when both rows use the same test layouts. N=1 and N=4 on the same N14 test layouts:

N14 test layoutsSSIM, N=1 → N=4Pearson, N=1 → N=4PSNR gain
The 100 in the paper0.512 → 0.5250.897 → 0.928+1.58 dB
380 (the 100 + 280 more)0.500 → 0.5110.906 → 0.933+1.51 dB

Table 7 now reports this same-layout comparison on the 380 test layouts.

W2b. Tables 9 and 10: "the data does not support the claim that a large fraction of designs contain incomplete placement metadata." The two tables count different things. Table 9 counts the samples used for training and testing (6,711 + 277 + 3,456 = 10,444). Table 10 counts the placement files read by our graph builder (10,795), including 352 that are not in our splits. On the N14 graph results: the netlist graph files were not linked to the N14 layouts, so all N14 models are geometry-only. The revised PDF states this in Section 4, Table 7, and the Appendix, and replaces the "incomplete placement metadata" sentence. The N14 results remain the first IR drop results on N14.

Dear authors, thank you for your reply. I still have several concerns about the motivation and the results on downstream tasks Solving the IR drop problems is not just about generating images. Although relying on the datasets, we still need practical results on the downstream tasks to prove the effectiveness, not only about the SSIM, PSNR, etc. In other words, these metrics are for the AI field, not for the EDA field.

Comments on the paper

1 comment

Hi Reviewers,

We are now near the end of the rebuttal period. If you haven't done so yet, please review the authors' responses and post a reply addressing whether their clarifications or new results resolve your concerns. Even if your score or recommendation remains unchanged, a brief note confirming what did or didn't shift your assessment—and why—is essential for the discussion and final decision. Thank you for your time and engagement.

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.