acceptodds
Under review as a conference paper at ICLR 2027

Semantically Steerable Image Tokenization

Abstract

Adaptive image tokenizers compress an image into a variable number of tokens matched to its complexity, and recent work makes this allocation controllable: conditioned on a target reconstruction error, the tokenizer itself decides how many tokens an image needs. However, the fidelity measure guiding this allocation shapes which aspects of the image are preserved under compression. The pixel-wise error used in prior work is blind to the spatial structure that carries an image's meaning, and under heavy compression such tokenizers are prone to abruptly discarding tokens that encode essential content. We introduce GIST (**G**uided **I**mage **S**emantic **T**okenizer), an adaptive tokenizer that can be steered toward a requested *semantic* fidelity. Its token allocation is jointly controlled by pixel-level and semantic target errors, with semantic error measured in the feature space of a frozen DINOv3 encoder. This joint control enables *semantic protection*: with the semantic target fixed at zero, the reconstruction target can be relaxed to discard pixel detail while preserving what the image depicts. In a controlled comparison on ImageNet-100, GIST outperforms its pixel-controlled counterpart KARL at every token count, with up to % lower LPIPS and % lower DreamSim. On average, lowering its semantic target reduces LPIPS faster per token than lowering the reconstruction target, and up to in the low-token regime. Semantic fidelity is thus an effective control signal for adaptive tokenization: it lets a tokenizer see, and so preserve, what an image depicts.

Primary area:
generative models
Keywords:
image tokenization, adaptive tokenization, variable-length tokenizers, semantic fidelity, image compression, image reconstruction, perceptual quality, generative models, DINOv3

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2z8gOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces GIST (Guided Image Semantic Tokenizer), an adaptive tokenizer that exceeds prior designs that are constraint by pixel reconstruction targets and achieves semantic protection via adding constraints related to semantic reconstruction. Controlled comparison on ImageNet-100 indicates that GIST outperforms pixel-controlled baseline (KARL) with better LPIPS and DreamSim.

Critical strengths
  1. The observation that pixel-level target is not always suited for adaptive tokenization design is a valuable insight, and that semantic protection provides new insight into designing better adaptive tokenizers.
Critical weaknesses
  1. Limited Novelty: The contribution of this paper is straightforward, adding a semantic reconstruction target and semantic conditioning module/signal to control semantic levels. However, adding semantic targets in one or another way has already became the de facto standard practice in tokenizer designs with many prior works (e.g. [1-3]). The authors should further clarify the novelty of this design beyond introducing semantic loss/conditioning based on vision foundation models.

  2. Limited evaluation. This limitation is folded. (i) the authors only compared against KARL in the main paper; (ii) the main experiments are on ImageNet-100, the authors should at least perform experiments on larger-scale ImageNet-1k, which is relatively standard for developing visual tokenizers.

  3. Unclear generation performance. As the authors stated in related works, the major aim for developing image tokenizers should be to better serve downstream generative models. As discovered by previous work, better reconstruction performance doesn't necessarily translates to better generation performance. Therefore the lack of results on generative models significantly undermine the authors' contribution of the new proposed tokenizer.

[1] Xiong, Tianwei, et al. "Gigatok: Scaling visual tokenizers to 3 billion parameters for autoregressive image generation." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

[2] Zheng, Anlin, et al. "Vision foundation models as effective visual tokenizers for autoregressive image generation." arXiv preprint arXiv:2507.08441 (2025).

[3] Fu, Zixuan, et al. "Improving Flexible Image Tokenizers for Autoregressive Image Generation." arXiv preprint arXiv:2601.01535 (2026).

Questions for authors

See weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 2z8g (Part 1/3)

We thank the reviewer for the detailed feedback and for highlighting the value of semantic protection when pixel-level targets alone are insufficient for adaptive tokenization. Below, we clarify GIST’s contribution relative to prior uses of semantic information in tokenizers, summarize the broader tokenizer comparisons, and present additional ImageNet-1k and downstream generation experiments.

Semantic fidelity targets guide adaptive token allocation

[…] The authors should further clarify the novelty of this design beyond introducing semantic loss/conditioning based on vision foundation models.

We agree that incorporating semantic information from pretrained vision models is established practice in image tokenization, as discussed in our related work. We believe our contribution concerns a fundamentally different use of this information: using semantic fidelity to control adaptive token allocation. GigaTok and ReTok use semantic alignment to shape the representations learned during training; VFMTok builds its representation on foundation-model features and semantic reconstruction. GIST instead lets the user specify a desired semantic fidelity, which guides the tokenizer’s decision about how many tokens to retain and what information those tokens preserve. Methodologically, we also operate differently: GIST introduces no DINO feature-alignment loss; rather, a frozen DINO model measures the semantic error used for this conditioning. To make this distinction easier to locate, we will also state it explicitly next to the training objective in Section 3.2.

The distinction therefore goes beyond introducing another semantic loss objective: semantic fidelity becomes a controllable criterion for token allocation. The cited methods do not provide this control. ReTok supports variable-length representations, but its semantic regularization depends on the manually selected sequence length; GIST uses the requested semantic fidelity to guide the allocation itself. Our contribution is this use of semantic fidelity for adaptive allocation, together with the experimental characterization of its behavior and benefits.

AuthorsOpenReview ↗

Response to Reviewer 2z8g (Part 2/3)

Broader tokenizer comparisons and ImageNet-1k evaluation

(i) Comparing GIST with additional tokenizer families

[…] the authors only compared against KARL in the main paper […]

We agree that the broader comparisons should be more visible in the main text. Our main experiments focus on KARL† because our central question is whether adding semantic control improves adaptive token allocation under a matched tokenizer backbone and training recipe. Comparing GIST with KARL† isolates the effect of our contribution, because they share the base tokenizer, training data, optimization schedule, and evaluation protocol. Comparisons between independently trained tokenizer families cannot isolate this contribution because several of these factors differ simultaneously. For broader context, Appendix C compares GIST with ALIT, three FlexTok variants, three One-D-Piece variants, and the released KARL checkpoint. The table below presents a compact summary of these results at an input budget of 128 tokens.

TokenizerTraining dataLPIPS ↓DreamSim ↓
KARL†IN1000.190.15
GISTIN1000.190.15
Released KARLIN1000.190.16
ALITIN1000.200.14
FlexTok, three variantsIN1k0.25–0.270.16–0.17
One-D-Piece, three variantsIN1k0.20–0.210.12–0.14

At an input budget of 128 tokens, GIST matches the LPIPS of KARL† and the released KARL checkpoint (0.19), and achieves lower LPIPS than ALIT (0.20), FlexTok (0.25–0.27), and One-D-Piece (0.20–0.21). The DreamSim comparison is mixed: GIST (0.15) performs better than FlexTok (0.16–0.17), while ALIT (0.14) and One-D-Piece (0.12–0.14) achieve lower errors. These results place GIST competitively among existing tokenizers for fixed-budget reconstruction. Appendix C contains a fuller discussion of these results.

We will bring a compact summary of these comparisons into the main text.

(ii) Reconstruction gains extend to ImageNet-1k

[…] the main experiments are on ImageNet-100, the authors should at least perform experiments on larger-scale ImageNet-1k, which is relatively standard for developing visual tokenizers.

Our main reconstruction experiments follow the ImageNet-100 evaluation setting used in KARL’s main reconstruction comparisons (Tables 1–2), our closest predecessor. We have (i) extended this evaluation to the full 50,000-image ImageNet-1k validation set, and (ii) continued stage-2 training of GIST and KARL† on ImageNet-1k under matched optimization settings. Both experiments are reported in Appendix “Reconstruction Quality on ImageNet-1k” (Figure 30 and Table 8).

First, we evaluate the ImageNet-100-trained checkpoints without additional training. GIST retains its reconstruction advantage on ImageNet-1k: at 128 mean active tokens, it achieves LPIPS of 0.232 versus 0.297 for KARL†, closely matching the corresponding ImageNet-100 results of 0.231 versus 0.296. Although DreamSim increases for both methods on ImageNet-1k, the absolute gap remains similar: 0.221 versus 0.298, compared with 0.200 versus 0.278 on ImageNet-100.

Second, we continue stage-2 training of both checkpoints for 20 epochs on ImageNet-1k using matched optimization settings. GIST’s advantage increases at 128 mean active tokens: it achieves LPIPS of 0.215 and DreamSim of 0.194, compared with 0.294 and 0.292 for KARL†, corresponding to reductions of 26.9% and 33.5%, respectively. The advantage also persists at 192 mean active tokens. These comparisons use linear interpolation at matched mean active token counts.

The table summarizes the three settings at 128 mean active tokens (IN100 = ImageNet-100; IN1k = ImageNet-1k).

Tokenizer trainingEvaluationLPIPS KARL† ↓LPIPS GIST ↓DreamSim KARL† ↓DreamSim GIST ↓
IN100IN1000.2960.2310.2780.200
IN100IN1k0.2970.2320.2980.221
IN100 + 20 IN1k epochsIN1k0.2940.2150.2920.194

These results show that GIST’s reconstruction advantage at reduced token counts extends to ImageNet-1k, both without additional training and after matched continuation.

AuthorsOpenReview ↗

Response to Reviewer 2z8g (Part 3/3)

Initial generation results show benefits and trade-offs

[…] As discovered by previous work, better reconstruction performance doesn't necessarily translates to better generation performance. Therefore the lack of results on generative models significantly undermine the authors' contribution of the new proposed tokenizer.

We agree that improved reconstruction does not establish improved generation, and we do not interpret our results as evidence of better generative performance. Our contribution concerns whether semantic fidelity can serve as an explicit control signal for adaptive token allocation and preserve semantic content under compression. Our controlled token–fidelity comparisons and semantic probing directly evaluate this contribution.

Adaptive token allocation and the token–fidelity trade-off are established research objectives that have been studied independently of downstream generation. Closely related examples include ALIT (ICLR 2025), ElasticTok (ICLR 2025), KARL (NeurIPS 2025), and InfoTok (ICLR 2026 oral), which evaluate allocation, reconstruction, or representation properties without downstream image/video-generation experiments. This scope also extends to quantization research: the Rotation Trick (ICLR 2025) and SimVQ (ICCV 2025) explicitly distinguish improvements to tokenizer properties from demonstrated generation benefits. As SimVQ states, “evaluating VQ models primarily through downstream generation tasks may not provide the most insightful assessment of their fundamental properties.” Our contribution similarly concerns semantic control of adaptive allocation, which our token–fidelity comparisons and semantic probing evaluate directly.

Nevertheless, to investigate downstream applicability, we have also added a controlled class-conditional generation study in the appendix, “Generation Quality across Active Token Counts”, using matched 91.6M-parameter MaskGIT-style generators trained for 50,000 updates each on ImageNet-100. The generators predict both code values and activity from class labels alone. Our initial generation study provides encouraging evidence of downstream benefits. Over the shared token range, the interpolated curves show lower gFID without classifier-free guidance and higher class consistency and precision, but lower recall, at both evaluated guidance settings. For example, without guidance, the measured points near 37 mean active tokens (36.67 for GIST; 36.92 for KARL†) give gFID of 69.03 versus 71.25 and class accuracy of 15.99% versus 14.13%, respectively. At these points, precision is higher (0.488 versus 0.430), but recall is lower (0.422 versus 0.468). At the tested guidance scale g=3g=3, GIST achieves better gFID than KARL† at some token counts and worse gFID at others. We present these findings as an initial study of downstream use, while our central claim remains the semantic control and token-fidelity benefits evaluated directly in the paper.

We would appreciate your feedback on whether our responses address your main concerns, and would be happy to clarify any remaining points.

Reviewer_j817OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces GIST, an adaptive image tokenizer that extends KARL with semantic fidelity control. Motivated by the loss of semantic content under pixel-error-based compression, it proposes:

  • Dual-target conditioning: Controls token retention using both pixel reconstruction error and DINOv3-based semantic error.
  • Dual-pass training: Measures both errors in an initial reconstruction and uses them to train conditional token halting. Experiments on ImageNet-100 and MS COCO show better reconstruction than a retrained KARL baseline at matched average token counts, while classification probes indicate improved encoder representations.
Critical strengths

The experiments are thorough and provide convincing support for the dual-target design. The reconstruction examples are also compelling: with the same number of tokens, the method better preserves the main subjects and key semantic content of the original images, consistent with the quantitative results.

Critical weaknesses

The methodological novelty is relatively limited. The architecture and training procedure largely follow KARL, with the main addition being a separate semantic target. While the extensive experiments demonstrate the value of this change, the method remains a fairly incremental extension of the existing framework.

Questions for authors
  1. How sensitive are the reported gains to the choice of semantic encoder? It would be helpful to know whether similar improvements hold beyond DINOv3, as the current results leave some uncertainty about how broadly the approach applies.

  2. Does GIST retain its quality advantage at matched average token counts when the halting threshold is varied? Since 0.75 is a fixed design choice, it is unclear how much the reported gains depend on this setting.

2 comments
AuthorsOpenReview ↗

Response to Reviewer j817

We thank the reviewer for the constructive feedback and for highlighting the experimental support for the dual-target design and its preservation of key semantic content. Below, we clarify the contribution and provide additional experiments testing the choice of semantic encoder and halting threshold.

Separate pixel and semantic controls improve reconstruction under compression

The methodological novelty is relatively limited. The architecture and training procedure largely follow KARL, with the main addition being a separate semantic target. While the extensive experiments demonstrate the value of this change, the method remains a fairly incremental extension of the existing framework.

The reviewer is right that GIST builds closely on KARL’s architecture and training procedure, as explicitly acknowledged in Section 3. Our central contribution is demonstrating that separately controlling pixel and semantic fidelity enables adaptive token allocation to preserve image content more effectively under compression. This separation introduces a capability absent from KARL’s pixel-only control: users can relax pixel fidelity while still targeting high semantic fidelity.

Our ablations reveal why this separation matters: replacing KARL’s pixel target with a semantic target improves semantic informativeness but does not reproduce GIST’s reconstruction gains. Our isocurve analysis further clarifies the distinct roles of the two targets in allocating tokens and preserving image content. Despite the incremental nature of the method, we believe this new capability and the understanding established by our experiments constitute a meaningful contribution.

Reconstruction gains also hold with SigLIP 2

How sensitive are the reported gains to the choice of semantic encoder?

Thank you for this suggestion. To investigate the dependence on the semantic encoder, we trained an additional GIST variant using frozen SigLIP 2-L/16 in place of DINOv3-L/16. We kept the tokenizer architecture and training recipe unchanged. We have added this experiment to the appendix, “Generalization Across Semantic Encoders.”

On ImageNet-100 validation, GIST with SigLIP 2 reduces LPIPS by 21.1% and DreamSim by 25.4% relative to KARL† at a matched mean of 64 active tokens. The corresponding reductions with DINOv3 are 28.4% and 34.9%, respectively. Both semantic encoders improve reconstruction quality at reduced token counts, with larger gains for DINOv3 under stronger compression. Thus, GIST’s reconstruction gains also hold with SigLIP 2, although they are smaller than with DINOv3.

Reconstruction gains persist across the tested halting thresholds

Does GIST retain its quality advantage at matched average token counts when the halting threshold is varied?

Thank you for raising this question. We have added an appendix, “Reconstruction Quality across Halting Thresholds”, which shows that GIST retains its reconstruction quality advantage at matched average token counts across the tested thresholds.

We evaluate the existing GIST and KARL† checkpoints at inference halting thresholds τ∈0.50,0.75,0.90\tau \in \\{0.50,0.75,0.90\\} on the ImageNet-100 validation set. As in the main experiments, we vary the reconstruction target over its full range for both methods and fix GIST’s semantic target at zero. Figure 25 presents the LPIPS and DreamSim Pareto fronts separately for each threshold.

GIST achieves lower LPIPS and DreamSim than KARL† at matched mean active token counts for all three thresholds. Its reconstruction advantage therefore persists when the inference halting threshold changes from the default value of 0.75.

Thank you for your response. I don't have any further questions.

Reviewer_R8unOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GIST is a semantically conditioned adaptive image tokenizer for allocating tokens according to pixel and semantic reconstruction targets. It extends KARL's single-pass adaptive tokenization by jointly conditioning token representations and halting decisions on both error targets. Semantic error is measured using frozen DINOv3 CLS features of the original and reconstructed images and supplies a conditioning signal for the second training forward pass rather than a feature-alignment loss for backpropagation. At matched mean active token counts, experiments on ImageNet-100 reduce LPIPS and DreamSim by up to 30% and 37% relative to retrained KARL, respectively, with reconstruction gains also observed on COCO.

Critical strengths
  • The main comparison controls base tokenizer capacity and training protocol

  • Reconstruction improvements extend beyond ImageNet examples with a single main subject

Critical weaknesses
  • Reliability of dual-target control remains unverified

Better average reconstruction does not establish that user-specified error bounds are met. Section 3.1 (p.4) describes the targets as upper bounds, but the experiments do not report how actual pixel and DINOv3 errors track them. This is particularly important when independently setting a zero semantic target and a relaxed pixel target. Please report actual errors and the frequency and magnitude of target violations, and clarify what a zero target means operationally.

  • Relaxing a target can unexpectedly increase token use

Conditional halting does not ensure predictable token allocation. In Figure 12 (p.27), with the semantic target fixed at zero, relaxing the black swan's pixel target from 0.18 to 0.24 to 0.32 increases active tokens from 41 to 42 to 58. Because the targets change both token representations and halting scores, the fixed threshold does not ensure monotonic token counts. An alternative is to predict a nonnegative score for each token representing its estimated error contribution if discarded, with scores fixed for a given image and semantic setting. Sort tokens by increasing score and discard the longest prefix whose summed score does not exceed the user-specified error budget. This guarantees nested retained sets and a non-increasing retained-token count as the budget is relaxed, although satisfaction of the actual reconstruction-error bound would still require validation. Please explain the benefit of this design trade-off and report how frequently and by how much such reversals occur across the evaluation set.

Questions for authors
  • Is pixel decoding performed during Stage 1?

Appendix A.1 (p.14, L713–717) says latent-distillation pre-training avoids pixel decoding, whereas Section 3.2 (p.5, L228–234) decodes predictions to measure pixel and DINOv3 errors for the second forward pass. Is decoding performed during Stage 1, and how are the two conditioning errors obtained?

  • How are the 2×/5× efficiency gains calculated?

Section 4.3 (p.8, L405–407) reports that semantic control reduces LPIPS 2× faster per token on average and up to 5× in the low-token regime. Please specify the calculation, curves and intervals compared, averaging procedure, and low-token range.

3 comments
AuthorsOpenReview ↗

Response to Reviewer R8un (Part 1/2)

We thank the reviewer for the careful assessment and for recognizing the controlled comparison and reconstruction improvements beyond images with a single main subject. Your questions helped us identify where the reliability and operational meaning of the targets needed clearer evaluation and explanation. Below, we quantify target violations and token-count reversals, clarify how conditioning errors are obtained during Stage 1, and explain how the reported efficiency gains are calculated.

Target reliability and observed violations

Better average reconstruction does not establish that user-specified error bounds are met. […] Please report actual errors and the frequency and magnitude of target violations, and clarify what a zero target means operationally.

We agree that better average reconstruction does not establish that each image meets the requested targets. GIST uses the targets to condition its encoder and token allocation, but does not check or enforce the resulting errors at inference. We will revise the upper-bound wording to make this distinction explicit. To quantify the actual errors and frequency and magnitude of target violations, we added an appendix, “Achieved Errors and Target Violations”.

Our findings can be summarized as follows. The actual-error curves show that raising the semantic target generally raises semantic error. Pixel error initially rises as we relax the pixel target, but levels off or decreases at the more relaxed settings. The violation-frequency heatmaps show that (i) strict targets are frequently missed, and (ii) relaxing the pixel target makes pixel violations less common, but does not ensure that a strict semantic target is met. The violation-magnitude heatmaps show that, for both error measures, the average violation-magnitude generally decreases as the corresponding target is relaxed. The largest semantic violations occur when the semantic target is strict and the pixel target is relaxed. Together, these findings show that while the targets provide learned steering of reconstruction quality, they do not guarantee that each image’s reconstruction errors remain below the requested values at inference.

Operational meaning of zero. Operationally, each input is mapped to a discrete conditioning bin. For the main GIST checkpoint, the smallest semantic bin is 0.04: a semantic input of zero selects this embedding and does not certify zero DINOv3 error. We will clarify that the targets provide learned control of reconstruction quality without guaranteeing per-image error bounds.

Causes, frequency, and magnitude of token-count reversals

Conditional halting does not ensure predictable token allocation. […] Please explain the benefit of this design trade-off and report how frequently and by how much such reversals occur across the evaluation set.

We agree that conditional halting does not guarantee monotonic token counts. Changing a target selects a different learned embedding, changing both the token representations and their halting scores. The halting loss supervises individual settings without imposing an ordering across targets, so scores can cross the fixed threshold in either direction. To quantify this behavior, we added an Appendix, “Frequency and Magnitude of Token-Count Reversals”.

We find that across adjacent target settings, reversals occur in 23.0% of comparisons when relaxing the pixel target and 6.8% when relaxing the semantic target. Among reversals, the mean increases are 13.4 and 5.1 tokens. The added appendix reports the full transition-level results, including the transitions highlighted by the reviewer.

The reviewer’s proposed selector is an interesting alternative: fixing nonnegative per-token scores and discarding tokens in increasing score order under a cumulative budget guarantees non-increasing token counts as that budget is relaxed. GIST instead learns target-conditioned token representations and keep/discard decisions directly. The main challenge in the proposed approach is learning scores whose sum reliably predicts reconstruction error. A token’s contribution depends on which other tokens remain: two redundant tokens may each be dispensable when removed individually, yet removing both can lose important information. Consequently, even accurate estimates of individual removal costs need not add up to the error caused by removing several tokens together.

GIST bypasses the need to predict how much reconstruction error discarding each token will cause. Instead, it jointly learns token representations and keep/discard decisions conditioned on the requested fidelity. We hypothesize that learning these decisions from observed reconstruction outcomes is easier than estimating per-token error contributions that remain reliable when several tokens are discarded together. However, this comes without a guarantee of monotonic token counts.

AuthorsOpenReview ↗

Response to Reviewer R8un (Part 2/2)

Is pixel decoding performed during Stage 1?

[…] Is decoding performed during Stage 1, and how are the two conditioning errors obtained?

Thank you for pointing this out. The wording in Appendix A.1 is imprecise. During Stage 1, we do decode the first-pass predictions using the frozen VQGAN decoder to obtain the two conditioning errors. Specifically, we measure the mean pixel-wise L1 error between the reconstruction and the input image, and the mean squared difference between their frozen DINOv3 CLS features. Both errors are computed without gradient tracking and discretized into the conditioning bins used by the second forward pass.

Stage 1’s reconstruction objective remains entirely in VQGAN token space. We use pixel decoding only to obtain the conditioning errors, not to compute the Stage 1 reconstruction loss. We do not backpropagate through the pixel decoder or DINOv3. The second forward pass does not require pixel decoding. We will correct Appendix A.1 and clarify Section 3.2 to distinguish this target-measurement step from pixel-space reconstruction training.

How are the 2×/5× efficiency gains calculated?

How are the 2×/5× efficiency gains calculated? […] Please specify the calculation, curves and intervals compared, averaging procedure, and low-token range.

The factors compare the two curve families in Fig. 5: varying the semantic target with the reconstruction target fixed, and vice versa. For each curve, we calculate the LPIPS reduction between its minimum-token and maximum-token points per added token:

s=Lˉlow−LˉhighTˉhigh−Tˉlow,s=\frac{{\bar{L}}_{low}-{\bar{L}}_{high}}{{\bar{T}}_{high}-{\bar{T}}_{low}},

where Lˉ\bar{L} and Tˉ\bar{T} denote mean LPIPS and mean retained tokens. We take the mean of the 19 curve slopes separately for semantic control and reconstruction control. The means are 0.00201 for semantic control and 0.00095 for reconstruction control, giving 2.12×, rounded to 2×.

For the low-token comparison, we evaluate both curve families over the same interval of 32.36-46.37 mean retained tokens. We include all curves whose observed token ranges cover this interval: six semantic-control curves and seven reconstruction-control curves. We linearly interpolate LPIPS at the interval boundaries and calculate the reduction per added token using the formula above. Averaging these slopes with equal weight within each family gives 0.004378 for semantic control and 0.000954 for reconstruction control, yielding 4.59×, or approximately 5× when rounded to one significant digit. Thus, “low-token” refers to approximately 32–46 mean retained tokens.

We have added an appendix, “LPIPS Reduction per Token”, which clarifies the calculation.

Thank you for the detailed responses and additional experiments. Your responses address most of my concerns. My remaining concern is the limitation in target control: requested error bounds are not guaranteed, and token allocation can be non-monotonic. Therefore, I maintain my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.