acceptodds
Under review as a conference paper at ICLR 2027

DynamicTok: Dynamic Tokenization for Flexible Image Representation and Generation

Abstract

Existing image tokenizers typically use a fixed number of tokens at fixed spatial locations, overlooking the substantial variation in visual complexity both across images and within individual images. To address this limitation, we introduce DynamicTok (Dynamic Image Tokenizer), which adaptively allocates tokens in both quantity and spatial location according to content complexity. DynamicTok first predicts a heatmap that estimates regional importance, assigning more tokens to visually informative regions and fewer tokens to redundant ones. To process these spatially non-uniform tokens effectively, we propose a hybrid 1D–2D tokenizer in which latent tokens capture global context while receiving localized cues from image tokens. DynamicTok can be readily integrated with existing generative models. Moreover, we introduce a two-stage generation strategy that predicts a heatmap before image synthesis, enabling flexible generation under different token budgets. On ImageNet-256 reconstruction, DynamicTok achieves rFID scores of 0.57 and 0.39 with 64 and 128 tokens, respectively, outperforming prior methods under the same token budgets. For ImageNet-256 generation, DynamicTok attains a gFID of 1.40 with 64 tokens, establishing a new state of the art among methods using fewer than 256 tokens. Moreover, both our tokenizer and generator can dynamically allocate tokens in terms of number and spatial locations as needed.

Primary area:
generative models
Keywords:
Image Tokenization, Image Generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_PFgQOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DynamicTok is an image tokenizer that combines variable token counts with content-dependent spatial allocation. A Sobel-gradient heatmap determines rectangular image partitions, and encoder/decoder fusion layers connect global latent tokens with their associated local regions. The method supports fixed-grid generation and a two-stage generator that first synthesizes a grayscale image to derive a partition. On ImageNet-256, the reported reconstruction rFID reaches 0.65 with the base model at an average of 64 tokens, and 0.39 at an average of 128 tokens. Fixed-grid 64-token generation achieves gFID 1.40, while two-stage generation with variable counts achieves 1.63.

Critical strengths
  1. The hybrid design provides a practical way to combine spatial and token-count flexibility while retaining a conventional latent generator. The internal ablations support the importance of its components: spatially adaptive training/inference improves reconstruction rFID from 0.96 to 0.71, and adding variable counts reaches 0.65 (p. 8, Table 3); encoder/decoder fusion improves rFID from 1.36 without fusion to 0.65 (p. 14, Table 8). These controls make the architectural contribution more persuasive than benchmark comparisons alone.

  2. The tokenizer is useful for both reconstruction and generation, with a credible quality/efficiency tradeoff. Dynamic training improves the matched fixed-grid generation baseline from gFID 1.66 to 1.40 with CFG (p. 9, Table 6). The two-stage variant supports multiple budgets: 48 tokens achieve gFID 1.69 at 108.4 ms versus SoftVQ-VAE's 1.78 at 143.6 ms (p. 9, Table 7), with both generation stages included in the reported costs. The paper clearly separates these results from its stronger fixed-grid one-stage result.

Critical weaknesses
  1. Key experimental details are missing, limiting reproducibility and assessment of the benchmark comparisons. Section 4.1 (p. 6) and Appendix A (pp. 13–14) describe architectures and loss weights but do not specify training duration, optimizer/learning-rate schedule, or training budgets for the tokenizer and generators. Section 4.1 also omits the FID sample count/reference split and diffusion sampler/step count; the latency protocol names a batch size but no GPU model (p. 8). These omissions matter when attributing improvements over prior systems and interpreting the efficiency claims.

  2. The incremental benefit of content-dependent token-count allocation is less established than that of spatial adaptivity and dynamic training. Table 3 (p. 8) changes both the training count distribution and the assignment policy when moving from 64 tokens to [32,96], without a matched random-allocation control. In Table 7 (p. 9), variable-count generation achieves almost the same gFID as fixed 64-token generation (1.63 versus 1.64) but has higher latency (217.1 versus 153.6 ms). This does not invalidate flexible budget control, but leaves the practical advantage of the automatic per-image allocation policy insufficiently demonstrated.

Questions for authors
  1. Clarification: Please provide the tokenizer/generator training schedules and compute budgets, the FID evaluation sample count/reference split, and the sampler/step count and GPU used for Table 7. Were the internal uniform baselines trained and evaluated with the same schedules and sampling settings?

  2. Clarification/targeted control: How much does gradient-based token-count assignment improve over randomly assigning the same count distribution while retaining gradient-based spatial partitions? For Table 7, how are variable-length batches processed, and what explains the latency increase from 153.6 to 217.1 ms despite similar average counts and FLOPs?

3 comments
AuthorsOpenReview ↗

Reponse to Reviewer PFgQ

W1: Experimental Details

Thank you for pointing this out. We provide the full tokenizer and generator training configurations, hardware specifications, and evaluation protocol below. All internal uniform baselines use the same training schedules and sampling settings.

Tokenizer & Generator Training Configuration

Tokenizer ConfigurationValueGenerator ConfigurationValue
Base ModelViT-B
enc/dec hidden dimension768hidden dimension1152
enc/dec #heads12#heads16
enc/dec #layers12#layers28
enc/dec patch size16patch size1
enc/dec position embedding2D Absolute PEposition embedding2D sinusoidal
optimizerAdamoptimizerAdamW
weight decay0.0weight decay0.0
base learning rate1e−4base learning rate2e−4
optimizer momentumβ1,β2\beta_1,\beta_2=0.9,0.95optimizer momentumβ1,β2\beta_1,\beta_2=0.9,0.999
global batch size256global batch size1024
learning rate schedulecosinelearning rate scheduleconstant
warmup steps10K
training steps300Ktraining steps500K
augmentationhorizontal flip, random cropaugmentationcenter crop
discriminatorDINOv2-Sdiffusion lossv-prediction
discriminator weight0.2 with adaptive weightdiffusion samplerEuler-Maruyama
discriminator start20Kdiffusion steps250
discriminator LeCAM0.001
perceptual weight1.0
evaluation metricFID-50k from ADMevaluation metricFID-50k from ADM
Estimated Training Budget284 H100 GPU HoursEstimated Training Budget432 H100 GPU Hours

Hardware, Latency and Evaluation Protocol

Table 7 was measured on a single NVIDIA H100 80GB HBM3 GPU with batch size 256. For latency computation, we used 3 warmup runs and 5 timed runs.

Sampling uses the Euler–Maruyama SDE solver with 250 steps and CFG. For FID evaluation, we generate 50,000 images and use the ADM FID-50K reference .npz file with the official ADM evaluation suite [1].

Higher latency of variable-length batches

When handling variable token counts, we currently pad all sequences to the maximum length of 96 tokens, which explains the increased latency. The reported FLOPs are calculated based on a single model evaluation rather than the full denoising trajectory; for the [32, 96] setting, the reported 32.3 GFLOPs is the theoretical average based on valid token counts and excludes padding overhead, including in the padded row. Below, we provide a comparison of the averaged inference time (without padding overhead) between the [32, 96] setting and the fixed 64-token setting. As shown, their FLOPs and actual latencies are highly comparable. In real-world deployment, service providers can implement length-based grouping (i.e., batching sequences with the same token count) or packed attention to eliminate padding overhead and effectively reduce computation.

MethodTokensGen. FLOPsGen. Latency (bs=256)
DynamicTok[32, 96], Padded32.3 G217.1 ms
DynamicTok[32, 96], Average32.3 G151.9 ms
DynamicTok6430.9 G153.6 ms
AuthorsOpenReview ↗

Reponse to Reviewer PFgQ

W2: Targeted Control on Token Allocation

To isolate the incremental benefit of gradient-based token-count allocation, we conducted an ablation study on image reconstruction. All settings below use a variable token count of [32, 96] and retain gradient-based spatial partitions; the only difference is whether the number of tokens is assigned based on gradients or randomly. Random token-count assignment preserves the same token-count distribution as gradient-based assignment. As shown in the table, using gradient-based allocation during both training and inference yields the best performance (rFID 0.65). Furthermore, regardless of the training mode, switching from random to gradient-based allocation during inference consistently improves reconstruction quality (e.g., improving from 0.74 to 0.65, and from 0.73 to 0.70).

Regarding image generation, we hypothesize that the marginal gain in gFID is limited by the first-stage generator. In reconstruction, the token count is deterministically derived from the ground-truth image complexity. In generation, however, the first-stage model must predict the token count, introducing a prediction variance that may dilute the benefits of content-dependent allocation. We will attribute the primary contribution of dynamic training to enabling flexible spatial and computational budget control, rather than content-dependent generation quality in the revised manuscript.

Training ModeInference ModerFID
GradientGradient0.65
GradientRandom0.74
RandomGradient0.70
RandomRandom0.73

The analysis of variable-length latency can be found in our response to Weakness 1 above.

Q1

For detailed information regarding the training schedules and computational budgets, please refer to our response to Weakness 1 above.

Q2

For the targeted control experiment comparing gradient-based and random token allocation, please refer to our response to Weakness 2 above. Regarding the processing of variable-length batches and the detailed explanation of the latency increase, please see our response to Weakness 1 above.

[1] Dhariwal P, Nichol A. Diffusion models beat gans on image synthesis[J]. Advances in neural information processing systems, 2021, 34: 8780-8794.

I have read the rebuttal and decided to maintain my score.

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

This paper introduces DynamicTok that assigns different numbers of tokens to different images and image regions. It targets two limitations of existing methods. 2D tokenizers use fixed grids, while compact 1D tokenizers may lose local spatial structure. The paper aims to provide both adaptive token counts and adaptive spatial locations.

Critical strengths
  1. Fixed grids cannot adapt to image content, while a compact 1D sequence may lose local spatial information. The paper addresses a meaningful problem in image tokenization.

  2. The paper reports strong reconstruction and one-stage generation results with relatively few tokens. The ablations also examine token range, heatmap type, and partitioning strategy. These results suggest a potentially useful quality–compute trade-off.

Critical weaknesses
  1. The paper states that one-stage generation uses a fixed grid. Dynamic token counts and spatial layouts are mainly supported by the two-stage pipeline. However, the best reported generation result (gFID 1.40 with 64 tokens) comes from the one-stage setting. Two-stage dynamic generation performs worse (1.63 in Table2). The current evidence supports the reconstruction results and fixed-token generation more strongly than the claim that dynamic allocation enables high-quality, controllable generation.

  2. The experiments focus mainly on 256×256 images from ImageNet-1K and class-conditional generation. The paper does not evaluate the method across datasets, resolutions, or broader generation tasks. It is therefore unclear whether the reported benefits generalize to different data distributions, image scales, and task settings. This limits the evidence for the method’s broader applicability.

  3. Gradient heatmaps may not reliably reflect semantic importance or representational need. Regions with high gradients may contain noise or repetitive textures and receive too many tokens. Conversely, smooth but semantically important regions may receive too few.

Questions for authors

How does varying the token count affect reconstruction quality and generation quality? Please report reconstruction results at different token counts and compare DynamicTok with FlexTok and GPSTok under matched token budgets and evaluation settings.

5 comments
AuthorsOpenReview ↗

Response to Reviewer WSHg

W1: One-stage vs. Two-stage Generation Performance

We acknowledge that the two-stage dynamic generation trails the one-stage fixed generation in absolute image quality. However, the primary objective of the two-stage pipeline is to introduce generation flexibility, specifically enabling explicit spatial control (Figure 7) and computational control (Table 7) by dynamically adjusting the token budget. Furthermore, despite the slight performance drop compared to our own one-stage model, the two-stage dynamic generation still outperforms strong baselines (e.g., MAETok, SoftVQ, TiTok) under equivalent or lower token budgets. Ultimately, both the one-stage and two-stage frameworks are core contributions of our work that focus on different aspects: the one-stage model pushes the upper bound of generation quality, while the two-stage model achieves a highly competitive balance between quality and dynamic, controllable generation. We will clarify this positioning in the revised manuscript.

W2: Generalization Across Datasets and Resolutions

Thank you for the valuable suggestion. While we initially trained our tokenizer and generator on ImageNet 256×256 as it is a standard benchmark, we have additionally trained them on ImageNet 512×512 to address your concern. The results are reported below. As shown, our method achieves strong performance in both image reconstruction and generation at this higher resolution. Furthermore, we provide reconstruction results on the FFHQ dataset, where DynamicTok also demonstrates excellent performance. Due to the limited time of the rebuttal period, we are currently unable to evaluate on additional generation tasks, but we plan to explore these extensions in future work.

MethodTokensGeneratorStagesrFIDgFID (w/o CFG)gFID (w/ CFG)
TiTok128MaskGIT-ViT-1.524.172.13
MAETok128SiT-XL/1-0.622.791.69
SoftVQ-VAE64SiT-XL/1-0.717.962.21
DynamicTok[32, 96]SiT-XL/11-stage0.632.181.52
DynamicTok[32, 96]SiT-XL/12-stage0.632.401.65
MethodrFID
VQGAN-LC3.81
SoftVQ2.12
DynamicTok1.93

W3: Complexity Measurement

We appreciate the reviewer’s insightful observation. We agree that gradient heatmaps might theoretically over-allocate tokens to repetitive textures while under-allocating them to semantically important regions. However, we ultimately chose gradient maps because they empirically demonstrate strong and stable performance in both image reconstruction and generation. To explicitly address this concern, we have already explored semantic-aware token allocation in our original manuscript. Specifically, in Table 4 and Table 9, we compared gradient maps with saliency maps (extracted via Visual Saliency Transformer), which prioritize foreground salient objects. We found that saliency maps perform significantly worse than gradient maps in image generation. We fully agree that finding a better complexity metric—one that optimally balances high-level semantic importance and low-level texture details, or perhaps learning this metric end-to-end—is an important open problem. We will expand the discussion in our revised manuscript to highlight the limitations of the current metric.

Complexity MeasurementrFIDgFID (w/o CFG)
Gradient Map0.652.57
Saliency Map0.674.77
AuthorsOpenReview ↗

Response to Reviewer WSHg

Q1: Varying Token Counts and Baseline Comparisons

We thank the reviewer for this question. First, we note that GPSTok does not support dynamic token counts, as it relies on a fixed budget of 128 tokens for both reconstruction and generation. Therefore, we cannot evaluate it under different matched token budgets. For reference, at its fixed budget of 128 tokens, GPSTok achieves an rFID of 0.65 and a gFID of 1.50. Instead, we report the image reconstruction performance of FlexTok and One-D-Piece across a range of token counts from 32 to 96. For image generation, we compare our method with FlexTok at 32 and 64 tokens, as its generator is not open-sourced and the original paper only reports gFID for these two specific settings. As shown in the tables below, DynamicTok scales excellently with the token budget. When the token count surpasses 48, DynamicTok consistently outperforms FlexTok in both reconstruction and generation. While FlexTok achieves better performance at an extremely low budget of 32 tokens, this advantage is primarily attributed to its heavy diffusion-based decoder. As detailed in the FLOPs comparison table, decoding a single image with FlexTok requires approximately 800× more computational FLOPs than DynamicTok.

Reconstruction Quality (rFID)

Method \ Tokens3248648096
One-D-Piece4.453.112.422.061.84
FlexTok1.421.371.341.281.25
DynamicTok1.620.890.650.570.54

Generation Quality (gFID (w/ CFG))

Method \ Tokens3264
FlexTok1.711.76
DynamicTok2.181.64

Computational Cost at 32 Tokens

MethodEnc. FLOPsDec. FLOPs
One-D-Piece48.6 GFLOPs120.5 GFLOPs
FlexTok157.8 GFLOPs23210.3 GFLOPs
DynamicTok26.8 GFLOPs29.7 GFLOPs
AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer WSHg,

Thank you again for your constructive feedback. We have posted our detailed response above and would be happy to provide any further clarification if needed. Thank you for your time and consideration.

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

DynamicTok adapts both token count and spatial allocation to image complexity. A Sobel-gradient heatmap determines the token budget and partitions an image into rectangular regions; regional pooling and expansion connect spatial image features with latent tokens in a hybrid tokenizer. The tokenizer supports fixed-grid one-stage generation and adaptive two-stage generation, where a lightweight model first generates a grayscale image to derive the layout. On ImageNet-256, the paper reports guided FID 1.40 for fixed 64-token generation and 1.63 for the adaptive 32-96-token setting.

Critical strengths
  1. The contribution goes beyond allocating more tokens to difficult test images. Table 3 shows that dynamic training improves reconstruction even with fixed-count, fixed-grid inference, from rFID 0.96 to 0.69. Table 6 likewise improves one-stage guided generation FID from 1.66 to 1.40 against the uniform baseline. These comparisons support a benefit of the training/representation design itself, while retaining compatibility with a conventional fixed-token generator.

  2. Joint control over token count and spatial allocation provides useful flexibility, and the ablations investigate the main components. Table 3 separates spatial and quantity adaptivity, while Appendix B shows the benefit of encoder/decoder fusion. The two-stage pipeline additionally enables spatial control. Table 7 explicitly includes both generation stages and shows a useful 48-token operating point: FID 1.69 at 108.4 ms versus SoftVQ's 1.78 at 143.6 ms under the reported timing protocol. These results support weak acceptance for flexible visual representation and generation.

Critical weaknesses
  1. Token count is not the complete representation budget for the adaptive mode. The decoder's expansion operation requires image-specific partition information in addition to latent values. Section 4.1 specifies 32 latent channels, but the storage/transmission or reconstruction of region boundaries and positions is not fully accounted for in the token-budget comparison. This matters when interpreting the dynamic reconstruction results as equally compact representations. The issue is specific to image-dependent layouts; a shared fixed grid does not have the same per-image side-information requirement. Cross-method counts of continuous and discrete tokens also do not, on their own, establish equal information budgets.

  2. The practical efficiency comparison needs a more reproducible protocol and careful interpretation across operating modes. The paper already distinguishes one-stage FID 1.40 from two-stage FID 1.63, and Table 7 includes both stages; I do not regard the first stage as an omitted cost. However, the variable-length 32-96 setting reports 217.1 ms, compared with 153.6 ms for the two-stage fixed-64 setting and 143.6 ms for SoftVQ. The cause of the additional variable-length overhead is not explained. Timing is reported at batch size 256 without naming the GPU or fully specifying the sampling and normalization conventions. The FLOP caption refers to TFLOPs while the entries use G, and it is unclear whether these counts cover one model evaluation or the complete trajectory. Training durations and optimization settings also need to be stated explicitly to assess how closely budgets are matched. These gaps limit conclusions about deployment efficiency rather than negating the quality improvements.

Questions for authors
  1. What exact partition information must accompany the latent tokens in adaptive reconstruction, how is it represented, and how large is this overhead relative to the latent representation? Please distinguish token-count comparisons from equal storage or information budgets.

  2. What GPU, sampling settings, and timing convention produce Table 7, and are partition construction and final decoding included? Please clarify the FLOP units, the handling of variable-length batches (including padding or grouping), and why the 32-96 setting is slower than fixed 64 tokens. Batch-size-1 measurements and explicit matched training budgets would help assess the practical benefit.

5 comments
AuthorsOpenReview ↗

Response to Reviewer GGxb

W1: Partition Information and Storage Overhead

Thanks for pointing this out. To reconstruct the image, the decoder indeed requires position information. In the fixed grid setting, token positions are predefined, requiring no overhead. In the dynamic setting, the exact partition information needed is simply the 2-dimensional spatial coordinate (x,y) for each token. The spatial partition can be recovered from these token positions, so no additional region-boundary information needs to be stored. Since the latent representation has 32 channels, storing this 2D position adds exactly 2 scalar values per token, increasing the per-token information budget from 32 to 34 channels. This represents a minimal relative overhead of 6.25%. We will explicitly clarify this distinction between token counts and actual information budgets in the revised manuscript.

W2: Hardware, Sampling Protocol and Efficiency Comparison

Thank you for pointing out these issues. We clarify the inference protocol and provide additional training and inference details below.

Hardware, sampling, and timing. The results in Table 7 were measured on a single NVIDIA H100 80GB HBM3 GPU with a batch size of 256. Sampling used the Euler–Maruyama SDE solver with 250 steps and classifier-free guidance (CFG). We performed three warm-up runs followed by five timed runs. The timed region includes both partition construction and final decoding.

Variable-length overhead. In the implementation used for Table 7, variable-length batches containing 32–96 tokens are padded to 96 tokens without length-based grouping, whereas the fixed-64 setting uses sequences of exactly 64 tokens. Because the [32, 96] setting is padded to the maximum length of 96, its actual computational cost and latency are comparable to the fixed 96-token setting, which explains why it is slower than the fixed 64-token setting.

FLOPs: The caption should read GFLOPs, not TFLOPs. The reported values correspond to a single model evaluation per image, rather than the complete sampling trajectory. For the variable-length setting [32, 96], the reported 32.3 GFLOPs is the theoretical average based on valid token counts and excludes padding overhead.

Training Budgets: We report estimated training times for DynamicTok and SoftVQ-VAE under the same training protocol. For DynamicTok, token sequences are padded to 96 tokens during both tokenizer and generator training. Training costs are measured using eight NVIDIA H100 80GB HBM3 GPUs. The training configurations are as follows:

  • Tokenizers: Both methods use the Adam optimizer with a batch size of 256 and are trained for 60 epochs (approximately 300,000 steps) on ImageNet at a resolution of 256 × 256 with 1,000 classes.

  • First-stage generator: We use LightningDiT-S with the AdamW optimizer and a batch size of 1,024, trained for 500,000 steps on ImageNet at a resolution of 256 × 256.

  • Image generators: The image generators for both SoftVQ-VAE and DynamicTok use the AdamW optimizer with a batch size of 1,024 and are trained for 500,000 steps.

Although DynamicTok requires more training time, it achieves better reconstruction performance and supports flexible token placement and token counts in both reconstruction and generation. We therefore consider the additional training overhead a reasonable trade-off.

MethodTokensTokenizer  CostGenerator Cost (1-stage)Generator Cost (2-stage)
SoftVQ-VAE64248 GPU Hours-244 GPU Hours
DynamicTok[32, 96]284 GPU Hours56 GPU Hours432 GPU Hours
AuthorsOpenReview ↗

Response to Reviewer GGxb

Batch-Size-1 Measurement: We measure inference latency at batch sizes of 1 and 64 using the protocol described above. The "[32, 96], Average" row is obtained by measuring inference latency separately at different token counts and then averaging these measurements. For the latency of the padded [32, 96] setting, one can refer to the 96-token setting, as their computational costs are comparable. The results are shown in the table below. All reported times for DynamicTok include the first-stage generation. The first-stage generation times at batch sizes of 1, 64, and 256 are 705.0 ms, 14.6 ms, and 9.6 ms, respectively. The higher latency at batch size 1 is largely attributable to the first-stage generator, which takes approximately 700 ms; this overhead becomes much smaller at larger batch sizes. At a batch size of 1, inference is heavily dominated by fixed per-request overheads, such as kernel launches. Therefore, we believe this setting may not faithfully reflect real-world inference latency, making measurements at larger batch sizes more meaningful. At batch sizes of 64 and 256, DynamicTok achieves lower latency than the listed baselines when operating with 32 tokens.

MethodTokensGen. FLOPsGen. Latency (bs=1)Gen. Latency (bs=64)Gen. Latency (bs=256)
DynamicTok[32, 96], Average32.3 G3203.1 ms170.8 ms151.9 ms
DynamicTok3215.4 G3206.8 ms98.6 ms83.0 ms
DynamicTok6430.9 G3206.6 ms172.7 ms153.6 ms
DynamicTok9645.52 G3197.3 ms260.5 ms216.5 ms
SoftVQ-VAE6429.1G2240.2 ms153.8 ms143.6 ms
MAETok12858.4 G2238.3 ms297.3 ms291.5 ms
SD-VAE1024118.7 G2498.4 ms805.4 ms799.7 ms

Q1

Please refer to our responses in Weakness 1 above.

Q2

Please refer to our responses in Weakness 2 above.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer GGxb,

Thank you again for your constructive feedback. We have posted our detailed response above and would be happy to provide any further clarification if needed. Thank you for your time and consideration.

Comments on the paper

1 comment

Dear reviewers,

The author-reviewer discussion period is ending soon. If you have not done so yet, please read the authors’ rebuttal and reply to it. Tell the authors whether it addresses your concerns, and update your score if your assessment has changed. With the high number of submissions this year, your input matters more than ever.

Thank you for your time and help.

Best regards,


Your 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.