{"id":"77901e66-7007-4b98-b860-21809b253507","arxiv_id":"2607.14728","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"VQ-Touch applies VQGAN with deformable convolutions and discrete diffusion to generate tactile images across sensors with few-shot mixed training.","lead":"VQ-Touch is a framework for generating tactile images from visual, label, or touch inputs using a VQGAN-based encoder and a discrete diffusion decoder, with a few-shot mixing scheme meant to transfer across sensor families. If it works, it could reduce the need for large per-sensor tactile datasets in robotics, but the experiments as written contain an unfair baseline and a zero-shot claim contradicted by the training protocol.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-sensor generalization claim lacks quantitative support: Fig.5 is qualitative, mixes target-sensor images into training, and clustering is unvalidated.","rationale":"After reading the paper, the most load-bearing assumption is that the sensor-family clustering and few-shot mixed training actually enable transfer to a new sensor. The only experiment purporting to show this is Fig.5, which is qualitative and trains on target-sensor images, so it cannot distinguish the contribution of the family data from simple fine-tuning. No quantitative metric is provided, and the clustering itself is not examined. If this mechanism fails, the paper's main novelty—cross-sensor data efficiency—disappears, leaving a standard combination of VQGAN and discrete diffusion. The reader's verdict identified the same area, and I agree that this is a load-bearing gap. The MAE comparison confound is also concerning, but the cross-sensor validation gap is more fundamental to the paper's identity. Given that the paper as written does not support its central cross-sensor claim, rejection is appropriate, though the authors could address this with additional experiments.","tokens_in":7760,"tokens_out":9294,"duration_ms":82240,"concrete_test":"Run the Fig.5 cross-sensor reconstruction experiment with three training conditions: (A) only the 50 target-sensor images; (B) only the family data; (C) family data + 50 target images (the proposed 1:5 mixed training). Evaluate on a held-out test set from the target sensor, reporting FID, SSIM, and LPIPS. Also report the unsupervised cluster assignments for the four datasets used (Touch and Go, FabricVST, VisGel, HCT) and compare them to known sensor labels (GelSight, GelSight Mini, DIGIT). If (C) is not significantly better than (A), or cluster assignments do not correspond to sensor types, the cross-sensor transfer claim is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty—sensor-family clustering and few-shot mixed training for cross-sensor transfer—is never quantitatively validated. Sec. III-B ('Validation of Transferability and Generalizability across Datasets and Sensors') refers only to Fig.5, which shows sample reconstructions but no FID/SSIM/LPIPS numbers. The caption states the dataset consists of 50 images from 'unseen' sensors (GelSight Derivative Version I and DIGIT) combined with family data at a 1:5 ratio, and the text says a small number of images from unseen sensors are combined to train DM-VQGAN. This means the transfer is few-shot, not zero-shot, and there is no ablation showing that the family data contributes beyond the 50 target-image samples. If training on only the 50 images (or on family data alone) yields similar reconstruction quality, the 'across-sensor' contribution collapses to standard fine-tuning. Additionally, the sensor-family clustering in Sec. II-B is described conceptually, but no experiment reports cluster assignments, cluster validity, or sensitivity to the clustering method. Without quantitative cross-sensor results and an ablation isolating the family-data contribution, the claim that VQ-Touch enables data-efficient generalization across sensors is unsupported. This is load-bearing because the title and abstract position cross-sensor compatibility as a primary contribution, while the framework's other components (VQGAN + discrete diffusion) are not new.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VQ-Touch, a tactile image generation framework combining a VQGAN-based encoder (DM-VQGAN) with deformable and dilated convolutions, a sensor-family clustering scheme, few-shot mixed training, and a discrete diffusion decoder for multimodal conditioning. The stated contributions are (i) DM-VQGAN as an efficient tactile representation learner, (ii) a unified multimodal conditional generation framework, (iii) sensor-family-level modeling for cross-sensor transfer, and (iv) state-of-the-art reconstruction and generation performance under limited data. Experiments are reported on four tactile datasets with comparisons to MAE, VQGAN, UniTouch, PixArt-α, and TextToucher.","tokens_in":8118,"tokens_out":2946,"duration_ms":28359,"significance":"If the claims were fully supported, VQ-Touch would be a practically valuable contribution to tactile sensing, where data collection is expensive and sensor diversity is high. The idea of sharing representations across sensor families via clustering and few-shot adaptation is interesting and potentially impactful. The paper also includes a useful unified conditioning interface for vision, tactile, and label inputs. However, the current evidence is not sufficient to establish these claims: the main reconstruction comparison is confounded, the cross-sensor generalization experiment is qualitative and trains on target-sensor data, and the sensor-family clustering is not validated at all. The architectural ingredients are largely borrowed from prior VQGAN and discrete diffusion work, so the novelty rests on the cross-sensor claims, which are currently unsupported.","major_comments":[{"comment":"The comparison between MAE and VQGAN/DM-VQGAN is not apples-to-apples. MAE is trained for masked-input reconstruction with mask ratios 0.25, 0.5, and 0.75, while VQGAN and DM-VQGAN are trained for full-image autoencoding. The conclusion that 'MAE fails' and that DM-VQGAN with 100 samples outperforms MAE with 10k samples does not follow from this table, because the two methods are optimizing different tasks. To support the claim, MAE should also be evaluated on full-image reconstruction (or otherwise matched to the same training objective), or the comparison should be reframed.","section":"Table I, Sec. III-B"},{"comment":"The cross-sensor generalization claim is not quantitatively supported. The text and caption state that 50 images from unseen sensors are combined with family data at a 1:5 ratio to train DM-VQGAN. This is a few-shot setup, not zero-shot, yet the abstract and introduction claim 'zero-shot transfer.' More importantly, Fig. 5 shows only qualitative reconstructions with no FID/SSIM/LPIPS numbers and no baseline. There is no ablation comparing (a) training only on the 50 target-sensor images, (b) training only on family data, and (c) the proposed mixed training. Without such an ablation, the contribution of the sensor-family representation to cross-sensor performance is not established.","section":"Sec. III-B, Fig. 5"},{"comment":"The sensor-family clustering is a central novelty of the paper, but it is never validated experimentally. No cluster assignments are reported, no cluster validity metric is given, and there is no sensitivity analysis with respect to the feature extractor, the clustering algorithm, or the number of clusters. The claim that sensors with 'shared intrinsic properties' form meaningful families is a hypothesis; the paper provides no evidence that the proposed clustering produces families that improve transfer over random grouping or over training directly on the new sensor's data.","section":"Sec. II-B, Eq. (5)"},{"comment":"The claim that VQ-Touch 'outperforms state-of-the-art models' is weakened by incomplete experimental description. The CTTP metric in Table II is not defined, and no variance or significance measures are reported even though the text states experiments were repeated three times. Table IV compares only against UniTouch, omitting the stronger recent baselines that appear in Table II. Without a clear protocol and a consistent set of baselines across all tasks, the SOTA claim is not fully substantiated.","section":"Tables II and IV"}],"minor_comments":[{"comment":"Typos and grammatical issues: 'm ulti-scale' in Sec. II-A, 'unimodal' appears in the conclusion where 'multimodal' seems intended, and several sentences are run-on. A careful proofread is needed.","section":"Throughout"},{"comment":"The discrete diffusion decoder is described at a high level, but no training objective, diffusion schedule, or algorithm for the discrete diffusion process is provided. The paper cites prior work, but since the unified conditioning interface is a claimed contribution, more technical detail is needed for reproducibility.","section":"Sec. II-C"},{"comment":"Figure 2 is dense and hard to parse, especially the 'Clustering' and 'Mixed Training' components. The relationship between sensor-specific codebooks and the shared codebook is unclear from the figure alone.","section":"Fig. 2"},{"comment":"The datasets are listed but no statistics (image counts per sensor, resolution, train/test splits) are given. This makes it difficult to assess the data-efficiency claims.","section":"Sec. III-A"}],"recommendation":"major_revision","confidential_remarks":"The core idea is worth exploring, but the current manuscript does not substantiate its central claims. The comparison in Table I is methodologically flawed, the cross-sensor evaluation is qualitative and not isolating the proposed mechanism, and the clustering is completely unvalidated. These are fixable with additional experiments, so I do not recommend rejection at this stage, but the revision must include those experiments and a re-framing of the zero-shot claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: VQ-Touch is a plausible combination of known generative components for tactile image generation, and the authors clearly know the problem area. But the central cross-sensor contribution is not backed by the evidence, and the headline comparison against MAE is apples-to-oranges.\n\nWhat's genuinely new here is the sensor-family clustering and few-shot mixed training idea, plus the use of deformable convolutions and multi-scale dilated fusion inside a VQGAN for tactile images. That's worth discussing. The unified discrete diffusion decoder with CLIP, label, or tactile conditioning is also reasonable. And the paper does show strong quantitative results in Tables II–IV for reconstruction and generation on standard benchmarks, even if the baseline set is limited.\n\nThe soft spots are load-bearing. First, Table I compares MAE under masked reconstruction (mask ratios 0.25–0.75) against VQGAN/DM-VQGAN doing full-image autoencoding. These are different tasks; of course MAE does worse when it has to reconstruct from a masked input. The takeaway that 'MAE fails' doesn't follow. Second, the cross-sensor experiment in Fig. 5 is qualitative only—no FID, SSIM, or LPIPS numbers—and the protocol trains on 50 images from the 'unseen' sensors mixed with family data at a 1:5 ratio. That's few-shot adaptation, not zero-shot transfer, even though the paper's Fig. 1 and introduction use 'zero-shot' language. Third, the sensor-family clustering is described but never validated: no cluster assignments, no cluster-quality metrics, and no ablation showing that the family data actually helps beyond the 50 target images. If training on just those 50 images gives similar reconstruction quality, the cross-sensor claim collapses to standard fine-tuning.\n\nSo I largely agree with the reader's verdict. The engineering is competent, but the paper as written overclaims its central novelty. With an ablation isolating the family-data contribution, quantitative cross-sensor metrics, and a corrected comparison against MAE, it could be a decent contribution. As it stands, the validation doesn't support the abstract's claims.\n\nI'd still send it to peer review—there's enough substance here that a serious referee could guide a major revision—but I wouldn't cite it or take its conclusions at face value.","headline":"Reasonable engineering, but the cross-sensor story is unvalidated: 'zero-shot' is actually few-shot, and the MAE comparison is confounded.","tokens_in":8572,"tokens_out":2211,"would_cite":false,"duration_ms":21489,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"VQ-Touch generates high-fidelity tactile images for new sensors using only about 50 real images per sensor, by sharing a learned codebook across sensor families.","keywords":["tactile image generation","vision-based tactile sensors","VQGAN","discrete diffusion","few-shot learning","cross-sensor generalization","multimodal generation","representation learning"],"falsifier":"Re-run the cross-sensor reconstruction with the same 50 images but replace the family training data with equal amounts of unrelated sensor data, or with no family data; if FID/SSIM do not worsen, sensor-family clustering is not the cause of the transfer. Alternatively, test zero-shot reconstruction with 0 target images; failure would contradict the zero-shot label.","tokens_in":7639,"feed_emoji":"🖐️","tokens_out":4531,"duration_ms":44377,"temperature":0.7,"pith_summary":"VQ-Touch asks whether tactile image generation can be made data-efficient enough to work across different tactile sensors with minimal new data. The paper argues yes: a VQGAN variant adapted to the low-variance, deformation-heavy structure of vision-based tactile images learns a compact codebook; sensors that share design are clustered into families; and a discrete diffusion model generates tactile images conditioned on visual images, other tactile images, or labels. With family data plus very few target images, the framework reconstructs and generates tactile images better than baselines, including a masked-autoencoder baseline trained on 100 times more data. If the claim holds, synthetic touch data becomes practical for training robotic perception where real tactile data is expensive or sensor-specific.","feed_headline":"Only ~50 images unlock tactile generation for a new sensor","feed_subtitle":"VQ-Touch reuses a sensor-family codebook and a discrete diffusion decoder, so synthetic touch data becomes practical.","key_machinery":"DM-VQGAN, a VQGAN encoder-decoder that uses deformable convolutions to learn spatial offsets for irregular deformation patterns and parallel dilated convolutions (rates 1, 2, 4) to fuse multi-scale features, producing discrete codebook representations tailored to tactile images. Sensor-family clustering assigns unseen sensors to a family by computing feature prototypes and nearest-cluster distance, after which few-shot mixed training combines a small number of target-sensor images with family images at a 1:5 ratio to adapt the codebook. A discrete diffusion decoder, operating in the discrete latent space and conditioned through a frozen CLIP encoder for visual inputs, the learned codebook fo","core_discovery":"The paper's central claim is that a single framework, VQ-Touch, can generate high-fidelity tactile images across sensors and scenarios while needing very little data. The supporting discovery is that representing tactile images with a deformable multi-scale VQGAN codebook, then aligning sensors into families by unsupervised feature clustering, allows a codebook trained on family data to be transferred to a new sensor with only around 50 real images from that sensor mixed at a 1:5 ratio with family data. A discrete diffusion decoder then converts visual images, other tactile images, or semantic labels into tactile images. Across four public datasets spanning GelSight, GelSight Mini, and DIGIT","pith_inferences":["The 'zero-shot transfer' label is optimistic: the reported cross-sensor experiments train on 50 images from the unseen sensor, so the contribution of the family codebook is not isolated; true zero-shot transfer remains untested.","If sensor-family clustering really is doing the work, its quality is measurable—clustering accuracy against known sensor lineage, or ablating with random family assignment, would quantify how much sharing matters. The paper does not provide this validation.","The conditioning interface is generic enough to extend beyond the three modalities tested; one could condition on force readings or audio, a step the paper does not take.","Data efficiency at 100 samples suggests the bottleneck is codebook coverage, not model capacity; a testable extension is scaling latent codebook size against sample count."],"forward_implications":["Tactile data collection for a new sensor could shrink from large multi-dataset pretraining to under a hundred real images plus family data, making synthetic data viable for robotics.","Vision-to-touch and label-to-touch generation allow tactile training data to be produced from scene images or semantic annotations when a physical sensor is unavailable.","Classifiers trained on VQ-Touch-generated images transfer to real recognition tasks with higher accuracy than prior generated images, so the synthetic data carries usable contact information.","The same codebook infrastructure supports multiple mainstream vision-based tactile sensor variants, suggesting one model can serve a fleet of sensors rather than per-sensor models.","Few-shot mixed training may reduce redundant data collection across overlapping sensor designs."],"fun_headline_variants":["Cross-sensor tactile with just ~50 real images","New sensor? 50 images generate tactile data","VQ-Touch: 50 real images train a codebook for new sensors","Few-shot tactile: ~50 images give you cross-sensor generation","Cross-sensor touch generation from just 50 images"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that sensors clustered into the same family share a tactile image representation, so a codebook trained on family data plus only 50 images of a new sensor transfers to that sensor; if family similarity is weak, the reported gains could simply come from the 50 target images.","fun_headline_variants_meta":{"raw":{"variants":["Cross-sensor tactile with just ~50 real images","New sensor? 50 images generate tactile data","VQ-Touch: 50 real images train a codebook for new sensors","Few-shot tactile: ~50 images give you cross-sensor generation","Cross-sensor touch generation from just 50 images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001097,"raw_usage":{"total_tokens":4391,"prompt_tokens":692,"completion_tokens":3699,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":436,"completion_tokens_details":{"reasoning_tokens":3615}},"tokens_in":436,"tokens_out":3699,"duration_ms":23867,"temperature":1.0,"reasoning_tokens":3615,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T01:11:05.102864+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the cross-sensor reconstruction with the same 50 images but replace the family training data with equal amounts of unrelated sensor data, or with no family data; if FID/SSIM do not worsen, sensor-family clustering is not the cause of the transfer. Alternatively, test zero-shot reconstruction with 0 target images; failure would contradict the zero-shot label.","supporting_citations":[],"review_version":1}