{"id":"b54d00d3-3404-4133-9c00-b55e633c5fcb","arxiv_id":"2506.08257","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A 32-token 1D image tokenizer, optimized at test time against CLIP or reconstruction losses, performs image editing, inpainting, and class-conditional generation without training a generative model.","lead":"A pretrained image tokenizer that squeezes a picture into 32 discrete codes can, with no generative model training, edit, inpaint, and generate images by optimizing those codes at test time. A generalist reader might care because it suggests highly compressed tokenizers already carry enough structure to act as lightweight generators.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Concern: FID-50k of 8.2 is achieved only after cherry-picking the best iteration count and adding tweaks whose interactions with seed association are not controlled; the reported advantage over the baseline is not robustly established.","rationale":"The reader's verdict is CONDITIONAL with a different weakest assumption: the reliability of straight-through gradients through VQ. I agree that is an unproven mechanism, but the paper's own ablations (Table 3: VQ vs no-VQ, 17.1 vs 15.1) show the method still works without VQ, just slightly worse, so that assumption is not load-bearing for the core existence claim. The load-bearing issue for the headline numbers is evaluation protocol: FID numbers are selected at the best iteration count, the best configuration mixes several tweaks and seed-association choices, and the comparison in Table 6 is against unconditional methods that do not receive 1000 selected training images. This is not an accusation of dishonesty; the authors disclose the selection (Section 5.3, C.2) and even note the impact statement disclaims SOTA. But a stress-test of the central claim should ask: if an independent lab re-runs the published pipeline, will they get 8.2 or 15.5? The paper gives no variance, no fixed protocol, and no seed release. There is independent support: external TiTok checkpoints, reproducible code link, and the qualitative copy-paste results are convincing. The mechanism question (why VQ helps) is interesting but does not threaten the empirical claim; the evaluation-protocol question directly threatens the reported headline. Conditional is the right verdict: accept only after fixing the protocol or clearly labeling the best-iteration/semi-parametric nature.","tokens_in":16597,"tokens_out":2813,"duration_ms":25669,"concrete_test":"Re-run the full 50k class-conditional generation protocol with a fixed, pre-registered number of iterations (e.g., 300 for VQ-LL-32) and with a fixed seed-association rule (e.g., top-1% CLIP), computing FID-50k for both the reported configuration and a variant that masks the seed-image pool (or uses a held-out pool of 1000 images from a disjoint class set). If FID-50k moves by more than ~2 points or the masked-pool variant approaches the random-token FID of 15.5, the claim that the tokenizer is inherently generative needs to be weakened to 'retrieval-assisted test-time optimization.'","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim is that a pretrained 1D tokenizer, with only test-time optimization, achieves FID 8.2 on ImageNet class-conditional generation (Table 6, backed by Table 4). This number is the strongest evidence that the tokenizer itself is generative. However, the number is not obtained by a fixed protocol. Table 4 reports FID at the best iteration count, and Table 2 explicitly says the iteration count was chosen as the one yielding the best FID-5k for each tokenizer. With FID/IS trade-off curves that cross as a function of iterations (Figure A5), selecting the iteration count by best FID is a form of post-hoc selection. The main comparison in Table 6 also mixes configurations: the 8.2/182 row uses the improved pipeline with token noise and token regularization, while semi-parametric baselines such as RDM-IN and IC-GAN are evaluated with their standard protocols. More importantly, the claim of 'generation without training' is semi-parametric: it needs a pool of 1000 ImageNet training images and, for the best FID, top-1% CLIP seed association. The paper's own Table 5 shows that without seed images but with the same tweaks, FID is 15.5. The jump from 15.5 (random-token, no training set access) to 8.2 (1000 selected training images plus seed-dependent optimization) is the entire margin that makes the headline result competitive. It is therefore not established that the tokenizer generates without a database; it is established that test-time optimization plus retrieval from a small curated training subset reaches FID 8.2 under best-iteration reporting. A single-number comparison in Table 6 overstates what is demonstrated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper demonstrates that a pretrained 1D image tokenizer with vector quantization (TiTok, 32 discrete tokens) can be used for image editing, inpainting, and class-conditional image generation without training any dedicated generative model. The main methods are heuristic copy-paste token manipulation for attribute editing and gradient-based test-time optimization of tokens using plug-and-play objectives such as CLIP similarity or reconstruction loss. On ImageNet class-conditional generation with 1000 seed images and CLIP-based seed association, the method achieves an FID-50k of 8.2 over 50k samples. Extensive ablations show that higher compression (fewer tokens, smaller codebook, discrete VQ, 1D structure) improves generation quality, and the paper honestly reports the adversarial vulnerability of CLIP guidance while adding SigLIP scores as a sanity check.","tokens_in":16916,"tokens_out":12448,"duration_ms":138973,"significance":"If the results hold, this paper offers a striking reframing of highly compressed 1D tokenizers as implicit generative models, with implications for understanding why heavy compression improves downstream generation. The paper is strong on empirical breadth: it ships code, ablates seed count, seed association, tokenizer type, VQ versus continuous latents, 1D versus 2D structure, optimization iterations, and noise/regularization tweaks, and it includes an adversarial sanity check for CLIP guidance. The central claim is an empirical existence proof rather than a derivation, so circularity concerns are minimal. The main limitations are the semi-parametric setting (the method needs a pool of seed images and CLIP-based association to reach the headline FID) and the post-hoc selection of optimization iterations in some comparisons; both are disclosed in the paper but deserve clearer treatment in the system-level comparison.","major_comments":[{"comment":"The reported FID-50k of 8.2 for 'Test-time optimization (VQ-LL-32)' is not tied to a single, fully described protocol. It appears to correspond to the fully tweaked pipeline from Table 4 (random crops, EMA, token noise, and token regularization) combined with CLIP top-1% seed association over a pool of 1000 seed images, yet Table 6 labels the data access only as '1000 random images' and the main text does not specify the iteration count or the tweaked configuration. Since this number is the strongest quantitative evidence that the tokenizer alone can approach trained generative models, the authors should state the exact configuration and include the random-association result (Table 1, FID-5k 20.7 for 2000 seeds) and the no-seed result (Table 5, FID-5k 15.5) in the same comparison so that the reader can see how much of the margin is due to CLIP-based seed selection and optimization tweaks.","section":"Table 6 / Section 5.6"},{"comment":"The tokenizer comparison uses the number of optimization iterations that yields the best FID-5k for each tokenizer. Because the FID-IS trade-off curves in Figure A5 cross as a function of iterations, the conclusion that more aggressive compression improves generation may be confounded by per-tokenizer optimization budgets rather than reflecting the tokenizer alone. Please report a fixed-iteration comparison (for example, 300 iterations for all models) and clarify whether the monotonic trend in Table 2 persists under a common budget.","section":"Table 2 / Section C.2 / Figure A5"}],"minor_comments":[{"comment":"The notation VQD(·) is used in Eq. (4) before it is defined; please define it in the main text or reference Algorithm A1 at first use.","section":"Section 4.1, Eq. (4)"},{"comment":"The rows '+token noise' and '+token reg' are not explicitly labeled as cumulative or independent additions on top of the (∗) configuration; please state the relationship so that the FID-50k value of 8.2 is reproducible.","section":"Section 5.4, Table 4"},{"comment":"The sentence 'achieving an FID of 8.6 for 50k samples' refers to the EMA configuration in Table 4, but Table 1 reports only FID-5k; consider reporting FID-50k for the seed-size ablations as well to avoid inconsistency.","section":"Section 5.2"},{"comment":"The definition of the importance metric g(k)_P_i is difficult to parse; the notation involving 'cov' with a subscript 2 should be made explicit, for example by writing out the covariance matrix and the norm used.","section":"Section 3.1, Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"This is a solid empirical paper with a clear central claim and extensive ablations. My main concern is that the headline number in Table 6 is not accompanied by a precise protocol description, and the per-tokenizer iteration selection in Table 2 could partially drive the compression trend. I do not see grounds for rejection, but the revision should make the evaluation protocol fully transparent, ideally by adding the no-seed and random-association baselines to the system-level comparison and reporting a fixed-iteration version of the tokenizer ablation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper's main claim—that a 32-token 1D VQ tokenizer can be used for generation via test-time optimization—is real and worth knowing. But the headline FID of 8.2 is a best-case number, not a fixed-protocol result. The fully seed-free version gets 15.5, and the 8.2 depends on 1000 seed images, CLIP-based seed selection, and iteration counts chosen post hoc for best FID.\n\nWhat is actually new: first, the demonstration that token positions in a 1D tokenizer are disentangled enough that simple token copy-paste transfers global attributes like blur, lighting, and image quality. Second, that gradient ascent through the VQ step with a CLIP objective works on 32 discrete tokens without the augmentation pipeline VQGAN-CLIP needed, and that 2D tokenizers fail here. The ablations are thorough: compression level, codebook size, VQ vs continuous, seed count and association, noise and regularization, plus SigLIP as a second check against CLIP adversarial overfitting. Code is released. This is solid empirical work.\n\nThe soft spots are real but mostly in reporting and framing. The FID numbers in Tables 2 and 6 are chosen at the iteration count that gives the best FID-5k for each tokenizer. That is post-hoc selection. No variance is reported, and the exact seed images aren't released. More importantly, the best number, 8.2, requires a pool of 1000 ImageNet training images and top-1% CLIP seed association. Without seeds, the same tweaks give FID 15.5 (Table 5). So the strong claim \"tokenizer generates on its own\" is not quite what is demonstrated; it's \"tokenizer plus a small curated database plus seed-dependent optimization reaches 8.2.\" The paper is transparent about the semi-parametric setup, but the Table 6 label \"1000 random images\" is misleading because the effective association is CLIP-selected.\n\nOn the straight-through gradient assumption: the paper doesn't explain why VQ regularization makes the optimization work, only that it does. That's a minor concern, not a flaw, and the empirical evidence is clear.\n\nWho this is for: anyone working on tokenizer design, semi-parametric generation, or test-time optimization. It deserves a serious referee. I'd send it to review, but ask the authors to adopt a fixed optimization schedule and report variance before publication.","headline":"The paper's core finding—that a 32-token 1D VQ tokenizer enables generation via test-time optimization—is real, but the headline FID 8.2 is a best-case number; the honest seed-free version is 15.5.","tokens_in":17559,"tokens_out":4150,"would_cite":true,"duration_ms":44849,"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":"One pretrained 1D tokenizer, with no dedicated generative model, can edit and generate images through test-time optimization of its 32 discrete tokens.","keywords":["1D tokenizer","test-time optimization","training-free generation","vector quantization","image editing","CLIP guidance","latent space semantics","image generation"],"falsifier":"Decode images from randomly sampled sequences of 32 tokens from the TiTok VQ-LL-32 codebook (with no optimization) and check whether the outputs are realistic and diverse; if they are not, the tokenizer alone is not generative and the reported results rely on the test-time optimization path rather than an emergent generative capacity.","tokens_in":16362,"feed_emoji":"🖼️","tokens_out":6514,"duration_ms":65191,"temperature":0.7,"pith_summary":"The paper sets out to show that a highly compressed 1D image tokenizer is itself a generative model. Using a pretrained TiTok tokenizer that represents an image as only 32 discrete tokens, the authors show that simple manipulations—copying a token from one image into another—produce coherent image edits, and that gradient-based test-time optimization of token features against objectives like CLIP similarity or reconstruction loss can perform text-guided editing, inpainting, and even text-to-image generation with no training of any dedicated generative model. On ImageNet class-conditional generation from 1000 seed images, this training-free procedure reaches FID 8.2 over 50k samples. The central, somewhat surprising claim is that compression itself—fewer tokens, smaller codebook, discrete vector quantization—strengthens these emergent generative abilities.","feed_headline":"32 tokens, no generator training: test-time optimization suffices","feed_subtitle":"No dedicated generative model is needed: test-time token optimization does editing, inpainting, and text-to-image.","key_machinery":"The load-bearing mechanism is test-time token optimization: the continuous feature vectors just before the vector-quantization step are treated as decision variables, and gradients of an image-level objective (CLIP similarity, L1 reconstruction, etc.) are backpropagated through the decoder and through the quantization step via a straight-through estimator. The paper emphasizes that the quantization step must be inside the optimization loop—optimizing the pre-quantization features directly gives poor results—and hypothesizes that VQ acts as a regularizer. A smaller set of tokens and a smaller codebook make the discrete search space easier to navigate and increase each token's global semantic responsibility.","core_discovery":"The paper's central discovery is that the latent space of a pretrained 1D tokenizer is semantically structured and generative enough that a decoder alone can be turned into an image generator. Concretely, each token position can encode a global attribute such as scene lighting, background blur, or subject type, so replacing one token with the corresponding token from another image transfers that attribute. Going further, the authors treat the tokenizer as an unconditional generator in a semi-parametric setting: initialize tokens from a seed image (or from noise), then optimize the pre-quantization features with straight-through gradients through the vector-quantization step to maximize a plug-and-play objective. This yields diverse, realistic images without training a generative model, and the paper finds that the more aggressive the compression—32 tokens over 64 or 128, discrete VQ over continuous latents, smaller codebooks—the better the generation quality.","pith_inferences":["If compression is what confers generative power, then pushing to even more extreme compression (e.g., 8 or 16 tokens) might further improve training-free generation, up to the point where the decoder must learn enough structure to be called a generator itself.","The straight-through gradient assumption could be tested directly by swapping in alternative discrete-optimization rules (such as Gumbel-Softmax or a randomized search over the codebook) and measuring whether generation quality survives; if it does, the claim's stated mechanism would need revision.","The finding suggests a new evaluation axis for tokenizer research: not just reconstruction fidelity but 'generative readiness' of the latent space, which could be measured by training-free FID.","The same recipe may transfer to other domains—video, audio, or 3D data—wherever a highly compressed 1D discrete tokenizer exists."],"forward_implications":["Generative modeling can be decoupled from training: any highly compressed VQ tokenizer can be converted into a generator by adding test-time optimization, without learning a distribution over tokens.","Tokenizer design should treat compression as a generative lever: reducing token count and codebook size improves training-free generation quality.","The same optimization loop handles editing, inpainting, and generation, so a single pretrained tokenizer replaces role-specific generative models for these tasks.","Class-conditioned generation from ImageNet with 1000 seed images and CLIP-based seed association reaches FID 8.2, competitive with trained unconditional and semi-parametric generators in the paper's comparison."],"supporting_citations":[{"why":"Supplies the pretrained TiTok 1D tokenizer on which all experiments and the claimed generative capability rest.","marker":"(Yu et al., 2024b)"},{"why":"Provides the CLIP text-image similarity objective that drives text-guided editing and class-conditional generation.","marker":"(Radford et al., 2021)"},{"why":"Provides the MaskGIT VQGAN, used both as the 2D-tokenizer baseline that fails at this task and as the reconstruction backend inside TiTok.","marker":"(Chang et al., 2022)"},{"why":"Defines the ADM FID evaluation protocol used to report generation quality, making the quantitative claims comparable.","marker":"(Dhariwal & Nichol, 2021)"},{"why":"Supplies SigLIP, the extra metric that shows prompt alignment is not merely a CLIP adversarial artifact.","marker":"(Zhai et al., 2023)"},{"why":"Provides the vector-quantization mechanism that the paper argues is essential to generative capability.","marker":"(van den Oord et al., 2017)"}],"fun_headline_variants":["No training needed: 32 tokens + test-time optimization generate images","32-token latent space turns decoder into image generator without training","Test-time token optimization: image generation with zero generative training","From 32 tokens to edited images: no generator training required","Image generation via token tweaks: no training, just optimization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method depends on straight-through gradient estimates through the vector-quantization step being a reliable optimization signal; the paper states this is essential (Section 4.1) but offers only a hypothesis that VQ provides regularization, not a mechanism.","fun_headline_variants_meta":{"raw":{"variants":["No training needed: 32 tokens + test-time optimization generate images","32-token latent space turns decoder into image generator without training","Test-time token optimization: image generation with zero generative training","From 32 tokens to edited images: no generator training required","Image generation via token tweaks: no training, just optimization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000641,"raw_usage":{"total_tokens":2914,"prompt_tokens":875,"completion_tokens":2039,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":1953}},"tokens_in":491,"tokens_out":2039,"duration_ms":15925,"temperature":1.0,"reasoning_tokens":1953,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:15:07.035141+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Decode images from randomly sampled sequences of 32 tokens from the TiTok VQ-LL-32 codebook (with no optimization) and check whether the outputs are realistic and diverse; if they are not, the tokenizer alone is not generative and the reported results rely on the test-time optimization path rather than an emergent generative capacity.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MaskGIT VQGAN, used both as the 2D-tokenizer baseline that fails at this task and as the reconstruction backend inside TiTok."},{"cited_title":"Sigmoid loss for language image pre-training","cited_arxiv_id":null,"evidence_quote":"Supplies SigLIP, the extra metric that shows prompt alignment is not merely a CLIP adversarial artifact."}],"review_version":1}