{"id":"855cf1e6-4747-4135-a3d1-440713962592","arxiv_id":"2508.13515","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"VGQ integrates 2D Gaussians into visual codebook quantization and reports an rFID of 0.556 on ImageNet reconstruction, a claimed state-of-the-art.","lead":"This paper proposes VGQ, an image tokenizer that replaces patch-based codebook quantization with 2D Gaussian distributions to capture position, rotation, and scale. It reports state-of-the-art reconstruction on ImageNet, with an rFID of 0.556, and claims better structural modeling for autoregressive image generation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim may be an artifact of uncontrolled Gaussian density: gains could come from added capacity, not structural modeling.","rationale":"The reader's weakest assumption identifies exactly the most load-bearing concern: the reported gains likely stem from increasing Gaussian density, which increases representation capacity independently of the geometric structural modeling. The abstract's own numbers show a large jump from rFID 1.00 to 0.556 when density is increased, confirming that capacity is a major factor. Without controlled comparisons at matched bitrate or parameter count, the 'state-of-the-art' claim is unsubstantiated. Since the full text is unavailable, the verdict remains unverified; our concern does not change the reader's UNVERDICTED verdict but sharpens the reason: the central causal claim needs a matched-capacity evaluation before it can be assessed. We therefore keep the verdict unchanged and propose a concrete test that would resolve the attribution question.","tokens_in":732,"tokens_out":2671,"duration_ms":27667,"concrete_test":"Obtain the released code and run a controlled ablation: keep the total number of Gaussians per image fixed (or match bitrate) while varying structure; in particular, evaluate VGQ with one Gaussian per token against VQ-GAN at the same codebook size, token count, and training compute. If VGQ-1G does not beat VQ-GAN on rFID/PSNR by the same margin as the high-density variant, the reported state-of-the-art result is attributable to capacity, not the 2D Gaussian representation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"VGQ's headline claim of state-of-the-art reconstruction (rFID 0.556, PSNR 24.93) is supported only by comparing a high-density variant against baselines whose per-token capacity is not matched. The abstract states that increasing the number of 2D Gaussians per token yields a significant boost, from rFID 1.00 to 0.556. That boost means the SOTA configuration uses many more parameters per token than the low-density version; if the comparison baselines (e.g., VQ-GAN) use a single vector per token, the improved rFID may reflect increased bitrate or model capacity rather than the proposed 2D Gaussian structural prior. The paper also does not report the trade-off curve at matched token counts or bitrates, making the 'substantially outperforming' claim untestable from the abstract. Additionally, it is unclear whether PSNR 24.93 is reported for the high-density variant alone; without a full evaluation protocol, the causal attribution to geometric modeling is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Visual Gaussian Quantization (VGQ), a visual tokenizer that integrates 2D Gaussian distributions into codebook-based quantization, explicitly modeling geometric properties such as position, rotation, and scale. The abstract claims state-of-the-art reconstruction on ImageNet 256x256 with an rFID of 0.556 and a PSNR of 24.93, achieved by increasing the density of 2D Gaussians per token, and reports that a lower-density variant already reaches an rFID of 1.00. The abstract is the only content available for review; no derivation, evaluation protocol, or code is provided.","tokens_in":904,"tokens_out":2101,"duration_ms":21058,"significance":"The idea of adding explicit geometric structure to visual tokenizers addresses a recognized limitation of patch-based quantization methods such as VQ-GAN. If the reported gains are real and attributable to the proposed Gaussian parameterization rather than to increased per-token capacity, the work could be a meaningful step for AR image generation and representation learning. However, in its current abstract-only form, the significance cannot be assessed quantitatively: the headline numbers are unsupported by any methodological detail, and the central causal claim is confounded by the admitted capacity increase.","major_comments":[{"comment":"The state-of-the-art claim is untestable from the manuscript: no evaluation protocol, dataset split, training configuration, baseline list, or statistical significance measures are provided, so the reported rFID of 0.556 and PSNR of 24.93 cannot be checked or reproduced.","section":"Abstract"},{"comment":"The abstract itself states that increasing the density of 2D Gaussians within tokens leads to a significant boost (from rFID 1.00 to 0.556), which suggests that the improved reconstruction may be driven by increased per-token parameter capacity rather than by the explicit geometric structure modeling; the comparison against methods that use a single vector per token is therefore not capacity-matched, and the 'substantially outperforming' claim conflates capacity with structure.","section":"Abstract"},{"comment":"The manuscript does not report the trade-off curve at matched token counts, codebook sizes, or bitrates, nor does it clarify whether the PSNR of 24.93 is reported for the high-density variant alone; without this information, the flexible trade-off claim and the attribution of gains to the Gaussian parameterization are unsupported.","section":"Abstract"},{"comment":"The statement 'Codes will be released soon' is a limitation: no code is currently available, and the absence of any implementation details prevents independent verification of the reported results, which is a load-bearing issue for a paper whose central contribution is an empirical state-of-the-art claim.","section":"Abstract"}],"minor_comments":[{"comment":"The abstract switches between 'we explored' and 'we proposed'; the past tense 'explored' undercuts the definitiveness of the contribution statement.","section":"Abstract"},{"comment":"The term 'rFID' is used without definition; since the metric may be unfamiliar to some readers, the full name (Fréchet Inception Distance) should be given at first use.","section":"Abstract"},{"comment":"The phrase 'visual codebook quantization frameworks' is vague; specifying the exact base framework (e.g., VQ-GAN or VQ-VAE) would improve clarity.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The review is based solely on the arXiv abstract; no full text was provided. The editor may wish to obtain the complete manuscript before making a final decision, but even from the abstract the capacity-confound concern is serious enough to require substantial revision and controlled experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is real: representing image latents as 2D Gaussian distributions inside a vector-quantized tokenizer, with explicit position, rotation, and scale parameters. That is a sensible departure from patch-based appearance-only tokenizers like VQ-GAN, and it could matter for autoregressive image generation where geometric structure is compressed away. The density trade-off described in the abstract is also genuinely useful — the paper reports rFID 1.00 at low density and 0.556 at high density, which is more transparent than most abstracts, since you can see the gain is tied to adding more Gaussians per token.\n\nWhere it gets soft is the headline claim. The abstract compares the high-density variant against existing methods without saying whether the baselines are matched for per-token capacity or bitrate. If VQ-GAN uses a single codebook vector per token and VGQ uses many Gaussians, the rFID improvement may simply be extra capacity, not the Gaussian structural prior. The same confound applies to the PSNR number, which isn't clearly tied to a specific density setting. There are no error bars and no evaluation protocol, so none of these numbers can be checked from the abstract. These are all expected limitations for an abstract-only submission — I'm not calling foul — but they mean the 'substantially outperforming' claim is untestable as stated.\n\nThe paper deserves a serious referee. The idea is new enough, and the target task (image tokenization for AR generation) is active enough, that a full version with matched-capacity baselines and an ablation separating the number of Gaussians from the Gaussian parameterization would be a useful contribution. If the full draft has those, I'd lean positive; if not, the SOTA claim collapses into a capacity argument. I'd send it to peer review with a request for capacity-matched comparisons and a clearer breakdown of where the gain comes from. I wouldn't cite it yet, but I'd want to read the full version when it's available.","headline":"Plausible architectural idea—2D Gaussians as quantized tokenizer entries—but the abstract alone can't support the SOTA claim, and the density/capacity confound is real.","tokens_in":1355,"tokens_out":1316,"would_cite":false,"duration_ms":14917,"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":"A new image tokenizer encodes image latents as 2D Gaussian distributions with explicit position, rotation, and scale, and reports reconstruction rFID 0.556 and PSNR 24.93 on ImageNet 256x256, outperforming prior codebook quantizers.","keywords":["visual tokenizer","2D Gaussians","vector quantization","image reconstruction","autoregressive image generation","structural modeling","ImageNet"],"falsifier":"Run a control tokenizer that uses the same number of Gaussian components per token but ignores or fixes the position, rotation, and scale parameters (for example, a fixed grid). If it matches rFID 0.556 and PSNR 24.93 on ImageNet 256x256, the improvement is capacity, not explicit structure.","tokens_in":586,"feed_emoji":"🖼️","tokens_out":4883,"duration_ms":41777,"temperature":0.7,"pith_summary":"This paper proposes Visual Gaussian Quantization (VGQ), a visual tokenizer that replaces patch-level appearance features with 2D Gaussian distributions, each carrying explicit position, rotation, and scale parameters. The authors argue that standard codebook tokenizers like VQ-GAN capture texture and color but lose geometric structure, and that explicit 2D Gaussian modeling restores that structure. On ImageNet 256x256, VGQ reports a reconstruction Fréchet inception distance (rFID) of 0.556 and a PSNR of 24.93, improving over the same model's rFID of 1.00 at lower Gaussian density. If these results hold, autoregressive image generators would get discrete tokens that preserve both appearance and spatial layout.","feed_headline":"Gaussian tokens reconstruct ImageNet at rFID 0.556","feed_subtitle":"A new visual tokenizer encodes position, rotation, and scale per token, beating prior codebook quantizers on 256x256 images.","key_machinery":"The central object is the 2D Gaussian distribution used as the token representation. Each token maps a local image latent to a Gaussian with explicit position, rotation, and scale parameters, and these parameters are quantized through the visual codebook. This carries the argument by making geometric layout directly quantizable, so the tokenizer can represent structure that appearance-only patch quantizers miss. The paper also treats the number of Gaussians per token as a tunable density, showing that higher density gives a significant boost in reconstruction fidelity and creates a flexible trade-off between token efficiency and visual richness.","core_discovery":"The central discovery is that a codebook quantization tokenizer can explicitly model geometric structure by encoding each visual latent as a 2D Gaussian distribution, parameterized by position, rotation, and scale, rather than as an appearance-only patch feature. These structure-related parameters are quantized through the codebook, so the discrete codes carry layout information that patch-based quantizers tend to discard. The paper reports that increasing the density of 2D Gaussians per token substantially improves reconstruction fidelity, and at the higher density VGQ reaches rFID 0.556 and PSNR 24.93 on ImageNet 256x256, a result the authors describe as substantially outperforming existing quantization-based tokenizers. The intended consequence is more faithful discrete representations for autoregressive image generation.","pith_inferences":["My inference: the large gain from higher Gaussian density suggests that part of the improvement may come from increased representation capacity rather than from the explicit geometric parameterization; a controlled comparison with equal per-token capacity would separate the two effects.","My inference: if used in autoregressive generation, VGQ's structure-carrying tokens could reduce layout drift across long token sequences, since position and scale are stated explicitly at each step.","My inference: a natural testable extension is to apply VGQ to video or multi-view imagery, where rotation and scale parameters correspond to camera motion, and check whether reconstruction consistency across frames improves."],"forward_implications":["Autoregressive image generators that consume VGQ tokens should reproduce geometric structure such as object position and orientation more faithfully than with appearance-only tokens.","The Gaussian density per token gives a controllable trade-off between token efficiency and reconstruction fidelity, letting users choose a compact code for generation or a richer code for reconstruction.","The reported rFID of 0.556 on ImageNet 256x256 sets a new reference point for quantization-based tokenizers to match or beat.","Because position, rotation, and scale are explicit parameters, downstream models may be able to manipulate these properties directly in token space rather than through pixel-level edits."],"supporting_citations":[],"fun_headline_variants":["2D Gaussian tokenizer hits rFID 0.556 on ImageNet","Geometric tokens beat patch quantizers: rFID 0.556","Visual tokenizer encodes shape, not just texture","2D Gaussians reshape image tokenization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains come from the explicit 2D Gaussian structural modeling rather than from the larger per-token capacity that more Gaussians provide.","fun_headline_variants_meta":{"raw":{"variants":["2D Gaussian tokenizer hits rFID 0.556 on ImageNet","Geometric tokens beat patch quantizers: rFID 0.556","Visual tokenizer encodes shape, not just texture","2D Gaussians reshape image tokenization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000377,"raw_usage":{"total_tokens":2029,"prompt_tokens":991,"completion_tokens":1038,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":967}},"tokens_in":607,"tokens_out":1038,"duration_ms":7767,"temperature":1.0,"reasoning_tokens":967,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:11:38.958635+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a control tokenizer that uses the same number of Gaussian components per token but ignores or fixes the position, rotation, and scale parameters (for example, a fixed grid). If it matches rFID 0.556 and PSNR 24.93 on ImageNet 256x256, the improvement is capacity, not explicit structure.","supporting_citations":[],"review_version":1}