{"id":"2ca176c9-65bd-4d23-a4e0-c812a4cae463","arxiv_id":"2607.16824","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":12,"one_line_summary":"Register tokens from frozen vision transformers can be converted into extra global prior tokens that modestly improve frozen tokenized image generation on ImageNet.","lead":"This paper shows that the 'attention sink' tokens vision transformers form during training can be pulled out at test time and reused as compact global clues for image generation. The trick gives small quality and speed gains on ImageNet benchmarks while leaving the network weights untouched.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 10's cross-model codebook alignment is unspecified and possibly fitted; same-image prior inflates the headline FID gain.","rationale":"The reader's weakest assumption identifies the same load-bearing point: the method assumes a register feature from one frozen model can be mapped into the discrete codebook of an unrelated frozen tokenizer/decoder via nearest-neighbor or Procrustes alignment, and that the inserted token still carries useful structure after the projection. My reading confirms this is the least secure step. The paper's own text names 'whitening followed by nearest-neighbor search' and 'linear Procrustes mapping' but gives no equations or fitting details, making the central 'training-free' claim impossible to verify. The same-image prior issue compounds this: Table 5 shows same-image priors outperform cross-image priors, so the headline FID improvement is not a clean demonstration of a transferable global prior. The causal controls in the supplement (low-frequency decomposition, shuffled controls, source sensitivity) are good-faith evidence that structure matters, but they do not resolve the underspecified alignment operator or the privileged-information confound. Given the reader already issued a CONDITIONAL verdict, my concern does not change that verdict; it identifies a specific condition—full specification and cross-image evaluation of the Eq. 10 alignment—that should be satisfied before the headline generation claim is accepted at face value.","tokens_in":23424,"tokens_out":6242,"duration_ms":69543,"concrete_test":"Obtain the exact alignment operator used in Eq. 10: dimensions of r^(ℓ) and c_j, whether a whitening or Procrustes matrix is fitted, and on which split. Then rerun Table 4 using only cross-image same-class priors with a completely fixed, hyperparameter-free alignment (e.g., a random projection normalized to the codebook dimension). If the RegToken-vs-[CLS] FID gap shrinks to <0.1, or if the method requires the fitted W to show any gain, the central transferable-register claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central generation claim depends on Eq. 10, which selects a codebook entry by cosine similarity between the extracted register feature r^(ℓ) and codebook entries c_j. This presupposes that these vectors live in the same space, but the paper never states the dimensionality of r^(ℓ) vs. c_j, nor does it specify the operator that maps between them. 'Whitening followed by nearest-neighbor search' and 'linear Procrustes mapping' are named but not defined: no equations, no data split for any fitted covariance/projection, and no parameters. If the alignment uses a linear map fit on ImageNet, the method is not training-free; if it is a fixed random projection, the gains may be coincidental. Independently, the headline number (FID 20.5 → 20.1) is obtained with a same-image prior: the register is extracted from the same real image used in the FID set, so part of the gain can reflect privileged target-image statistics. Table 5 itself shows a clear ordering (same-image 20.1 < same-class 20.6 < random 21.1), confirming the prior source matters. The cross-image controls help, but without the alignment operator specified and without error bars, the transferable-register interpretation of the headline gain is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies test-time register tokens in vision transformers and asks whether the global, low-frequency structure they encode can be operationalized for tokenized image generation without retraining. The authors first present diagnostic evidence (Table 1) that register-token features have stronger low-frequency spectral concentration than [CLS] or patch-mean features across DINOv2 and OpenCLIP. They then propose RegToken, a training-free method that localizes register structure via NFN-based layer selection, identifies register-relevant channels with TokenRank, and constructs a register token via a projection-and-conservation update. This token is mapped to the codebook of a frozen HCT-style decoder and used as a global prior, with optional optimization of the inserted token only. The paper reports improvements on ImageNet generation metrics (e.g., FID-5k 20.5 to 20.1, SigLIP 3.6 to 3.9) and faster test-time optimization, while keeping the backbone and decoder frozen.","tokens_in":1708,"tokens_out":1765,"duration_ms":49339,"significance":"If the claims hold, the paper makes a useful contribution by turning an often-dismissed attention-sink phenomenon into a plug-and-play generative prior. The diagnostic section is careful: the low-frequency analysis is statistically grounded, uses paired tests, and is complemented by a pixel-space DCT correlation in the supplement. The method is also extensively ablated, and the supplement includes causal controls (low- vs. high-frequency decomposition, shuffled priors, matched-norm random priors) that strengthen the interpretation. The main limitation is that the headline generation gain is obtained in a same-image prior setting, which is acknowledged in the supplement but not in the abstract, and the cross-model codebook alignment in Eq. (10) is underspecified. The value of the paper therefore depends on whether the transferable-prior interpretation can be established outside the same-image diagnostic and whether the alignment step is exactly defined.","major_comments":[{"comment":"The cross-model codebook alignment is underspecified. The text mentions 'whitening followed by nearest-neighbor search' and 'linear Procrustes mapping,' but gives no equations, no dimensionalities of r^(ℓ) vs. c_j, no data used to estimate any whitening covariance or Procrustes matrix, and no train/test split. If the mapping is fitted on ImageNet data, the method is not training-free; if it is fixed or random, the gains may be coincidental. Because Eq. (10) is the interface between the backbone and the decoder, the entire generation pipeline depends on this step. The authors should specify the exact operator, state whether any statistics are estimated and on which data, report sensitivity to the alignment choice, and provide reproducible code for this step.","section":"§4.4, Eq. (10) and §5.2"},{"comment":"The headline improvements (FID-5k 20.5→20.1, SigLIP 3.6→3.9) are obtained with a same-image prior: the register is extracted from the same real image whose class is used for generation. Table 5 shows a clear ordering (same-image 20.1 < same-class 20.6 < random 21.1), confirming that the prior source matters. The supplement calls this 'a source-conditioned diagnostic setting' rather than a zero-shot protocol, but the abstract and introduction present the gain without this caveat. Moreover, the Table 4 comparison against [CLS] prior is also same-image, so the claimed advantage over [CLS] may be specific to privileged target-image information. The authors should report cross-image same-class and random-source comparisons against the [CLS] prior under the identical protocol, and clearly label the same-image setting as an upper-bound diagnostic in the abstract and main text.","section":"§5, Table 4 and Table 5"},{"comment":"All FID-5k numbers are reported as point estimates with no error bars or repeated evaluations. Given that the reported gains are small (0.4 FID, 0.3 SigLIP) and FID-5k is known to have high variance, these differences may not be statistically meaningful. The supplement reports variance across seeds for CLIPScore trajectories but not for FID, IS, CLIP, or SigLIP. The authors should report mean ± std over at least 3–5 independent decoding runs for the main comparisons, or use a larger FID evaluation set, so the reader can judge whether the improvements exceed noise.","section":"Table 4 and general evaluation"},{"comment":"The projection-and-conservation update is stated qualitatively: α is set to 'approximately preserve the mean projection magnitude,' but the exact formula for α is not given. Since α is one of many hyperparameters (κ, n, k, w, Δ, s, γ, ∼α, λ, η, h), the method's 'training-free' status and reproducibility would benefit from an explicit definition of α in terms of the computed statistics. The robustness ablations are reassuring, but the conservation step should be specified exactly rather than left as an approximation.","section":"§4.3, Eq. (9)"}],"minor_comments":[{"comment":"The abstract states 'FID-5k 20.5 to 20.1' without noting that 20.5 is the [CLS]-prior baseline, not the no-prior baseline. This should be clarified.","section":"Abstract"},{"comment":"The caption says 'PCA-projected token features and their 1-D FFT spectra along the feature dimension.' It would be clearer to state explicitly that the FFT is applied after PCA projection and to define the number of PCA components.","section":"Figure 2"},{"comment":"The column header 'norm gap (Attention)' is ambiguous; it seems to list attention values but the units and definition should be stated in the caption.","section":"Table 2"},{"comment":"The sentence 'It reaches the same CLIPScore threshold in fewer steps (Steps@τ: 74→52)' refers to Table S2 but the τ definition appears only in Appendix A.3. A brief definition in the main text would help.","section":"§5.2"},{"comment":"The text says 'we report FID-5k, IS, CLIP, SigLIP, and trajectory-based metrics under identical settings for all compared priors,' but the main paper does not show error bars for these metrics; the supplement should state explicitly which metrics are computed once versus averaged over seeds.","section":"Appendix A.4"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid diagnostic core and a well-structured method, but the central generation claim is currently supported mainly by a same-image prior setting, which is an oracle-like condition. The cross-model alignment step (Eq. 10) also needs to be spelled out for the method to be reproducible and for the 'training-free' claim to be verifiable. If the authors can provide cross-image controls against [CLS] prior and specify the alignment exactly, the paper could become suitable for publication; without that, the headline result is difficult to interpret."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First: this paper is worth a serious referee, but the headline generation number should be read with the same-image caveat. The new idea is real — turning attention-sink registers into injectable global priors for tokenized image generation is, as far as I know, not done before. The diagnostics are strong: paired t-tests on 1000 images show register features concentrate low-frequency energy more than [CLS] or patch-mean in both DINOv2 and OpenCLIP, and the supplement includes sensible controls (matched-norm random, shuffled, low- vs. high-frequency decomposition). The ablations are careful, and the authors are honest that the same-image setting is source-conditioned, not zero-shot. The soft spots are where the reader and stress-test say they are. The FID gain reported in the abstract is 20.5 to 20.1, but that is against a [CLS] prior, not the no-prior baseline, and there are no error bars or repeated runs. Eq. 10 is genuinely underspecified: cosine similarity between a register feature and a codebook entry presupposes matched spaces, and the whitening/Procrustes alignment mentioned in Section 5.2 is never defined. If that linear map is fit on ImageNet, the training-free claim is overstated. If it is a fixed random projection, the gains need stronger evidence. The same-image prior doesn't sink the paper, because same-class and random controls show ordering and the authors are transparent about it, but it does keep the headline from being a clean zero-shot result. The large hyperparameter list bothered me less after reading the ablations; performance is stable across ranges. The paper also clearly lays out failure cases and scope limits, which I respect. Overall: the core observation — registers carry usable global structure — is well supported. The method is a composition of borrowed parts, but that's okay given the application is new. The main missing piece is a fully specified, reproducible alignment procedure. A serious referee should ask for that before the generation claim is accepted at face value. I'd bring this to reading group and would cite it if I worked on compact token generation.","headline":"Solid diagnostic core and an honest but modest generation experiment; the main gap is an underspecified cross-model codebook alignment that leaves the training-free claim unverifiable.","tokens_in":785,"tokens_out":1895,"would_cite":true,"duration_ms":43185,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Vision transformers' outlier 'register' tokens can serve, training-free, as global priors that improve tokenized image generation.","keywords":["attention sinks","registers","vision transformers","tokenized image generation","training-free","global priors","low-frequency structure","test-time optimization"],"falsifier":"Run the identical RegToken pipeline but replace the extracted register vector with one whose feature dimensions are randomly permuted while keeping its norm; if FID-5k and SigLIP stay at 20.1/3.9 instead of degrading toward the no-prior baseline (21.2/3.5), the claim that the structured content matters would be refuted. A second check: substitute a patch-mean vector through the same codebook projection; if it matches RegToken's gains, registers are not special.","tokens_in":23324,"feed_emoji":"🎨","tokens_out":4016,"duration_ms":38847,"temperature":0.7,"pith_summary":"RegToken asks whether the high-norm outlier tokens that vision transformers accumulate—often dismissed as attention sinks—actually carry usable global image structure. The paper argues they do: register-associated features concentrate energy in low frequencies more than either [CLS] readouts or patch means, and correlate with pixel-space low-frequency content. On that basis, RegToken extracts a compact register subspace from a frozen backbone and maps it into a frozen compact-token generation pipeline, improving ImageNet FID-5k from 20.5 to 20.1, SigLIP from 3.6 to 3.9, and cutting optimization steps to threshold from 74 to 52—all with no weight updates. The claim is that what look like artifacts are in fact lightweight global priors that a frozen decoder can exploit.","feed_headline":"Attention-sink tokens improve tokenized generation, no training","feed_subtitle":"RegToken turns register tokens into global priors, trimming FID to 20.1 and lifting SigLIP alignment on ImageNet.","key_machinery":"The load-bearing object is the test-time register token: a single auxiliary token constructed from a frozen vision transformer's register subspace. It is produced by (i) NFN-based layer selection, which locates layers whose outputs are dominated by a few feature directions; (ii) TokenRank-guided subspace extraction, which ranks channels by their activity on outlier tokens and by their stationary attention centrality; and (iii) a projection-and-conservation update that moves activation energy from outlier patches into the register token while preserving the mean projection magnitude. The resulting vector is mapped to the decoder's discrete codebook by cosine-similarity nearest neighbor or a l","core_discovery":"The central discovery is that test-time register structure—feature directions associated with high-norm outlier tokens in a frozen ViT—encodes smooth, scene-level statistics such as illumination, color tone, and coarse layout, and that this structure can be captured by a three-stage procedure: NFN-based layer localization, TokenRank-guided subspace extraction, and a projection-and-conservation update that transfers register-subspace energy from outlier patches to a dedicated token. Mapped into a discrete codebook via nearest-neighbor or Procrustes alignment and inserted as a global token into a frozen 1D-token decoder, this training-free prior improves generation quality and text-image align","pith_inferences":["If register features are truly global low-frequency summaries, a testable extension is to manipulate the register token directly for style transfer or illumination editing, rather than using it only as a fidelity prior.","The same-image prior in the main protocol is a privileged setting; the paper's own cross-image results suggest a class-level prior, so a follow-up could measure how much of the gain survives under purely text-conditional generation with no image-derived prior at all.","Because the projection relies on cosine similarity to a codebook trained for reconstruction, the effect's strength likely depends on how well the backbone's global statistics align with the tokenizer's codebook; a systematic study varying tokenizer training data could bound the method's portability.","The method can be interpreted as a soft, training-free form of conditioning: even 'artifactual' internal tokens carry a usable bottleneck of global information worth reusing in generation."],"forward_implications":["If registers really carry global low-frequency structure, the same extraction can serve as a cheap global prior for other frozen generative pipelines without retraining.","Inserting register priors improves convergence of test-time token optimization, suggesting they give the optimizer a better initialization.","Because the prior is extracted from a frozen backbone and mapped by codebook lookup, it can be applied to new decoders as long as a light alignment is available.","The low-frequency concentration diagnostic offers a principled way to choose where and what to inject, replacing hand-picked heuristics.","The gains transfer across backbone scales and token budgets, indicating the phenomenon is not peculiar to one architecture."],"fun_headline_variants":["Turn attention sinks into global priors for tokenized image generation","RegToken: train-free registers as global priors for generation","Registers from attention sinks become train-free global priors","No training: register tokens as global priors for generation","FID improves to 20.1 with train-free register tokens"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method assumes that a register feature vector extracted from a frozen vision backbone, after being projected into the discrete codebook of an unrelated frozen tokenizer, still carries the global structure that produced the generation gains.","fun_headline_variants_meta":{"raw":{"variants":["Turn attention sinks into global priors for tokenized image generation","RegToken: train-free registers as global priors for generation","Registers from attention sinks become train-free global priors","No training: register tokens as global priors for generation","FID improves to 20.1 with train-free register tokens"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001895,"raw_usage":{"total_tokens":7302,"prompt_tokens":814,"completion_tokens":6488,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":6403}},"tokens_in":558,"tokens_out":6488,"duration_ms":38216,"temperature":1.0,"reasoning_tokens":6403,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T19:49:30.826220+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical RegToken pipeline but replace the extracted register vector with one whose feature dimensions are randomly permuted while keeping its norm; if FID-5k and SigLIP stay at 20.1/3.9 instead of degrading toward the no-prior baseline (21.2/3.5), the claim that the structured content matters would be refuted. A second check: substitute a patch-mean vector through the same codebook projection; if it matches RegToken's gains, registers are not special.","supporting_citations":[],"review_version":1}