{"id":"ff8824c7-9267-4faf-b304-f4df102c42a0","arxiv_id":"2505.03667","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DisTok maps arbitrary class-distribution vectors to learned creative tokens, letting a diffusion model blend three or more concepts into one image in a single pass.","lead":"This paper proposes DisTok, a method that turns a list of percentages, such as 50% cat, 30% bird, 20% fish, into a single learned token and uses it to generate hybrid images blending those concepts. If it works as claimed, it makes creative, out-of-distribution image generation more controllable and faster than prior sampling-based approaches.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The distribution-consistency loop uses BLIP as both teacher and examiner; without independent annotation of class proportions, 'superior text-image alignment' on DCG may only reflect alignment to BLIP's biases.","rationale":"The reader's weakest assumption points to the same load-bearing concern: the predicted class distribution from a vision-language model is treated as ground truth for training and for the KL-consistency evaluation. I agree that this is the most fragile part of the central claim. The paper's other gaps, such as missing thresholds theta1 and theta2 and the absence of a direct quantitative DCG baseline, are real but secondary; they affect reproducibility and completeness rather than the validity of the core argument. The circularity, however, directly threatens the claim of 'superior text-image alignment' in the distribution-conditional setting. If BLIP is biased or under-specifying for out-of-distribution hybrids, DisTok may simply be reward-hacking the VLM: producing images whose BLIP softmax matches the input distribution while not actually reflecting the intended class proportions. The GPT-4o and user studies provide some independent support for overall creativity and aesthetics, but they do not test distribution fidelity, so they cannot break the circularity. Because the reader already conditioned acceptance on independent evaluation of distributional consistency, no verdict change is needed. The specific check I propose, independent human or second-VLM annotation of class proportions on the same images, would settle whether the concern lands.","tokens_in":14178,"tokens_out":6106,"duration_ms":67099,"concrete_test":"Take the 30 DCG distributions (or a fresh held-out set) and the images generated by DisTok and by the strongest prompt-based baseline (e.g., Kandinsky with explicit percentages). Have human annotators, or a second independently trained open-vocabulary VLM such as GPT-4o, judge for each image which of several candidate class-distribution vectors best matches the visual content, or directly estimate the proportion of each class present. Then compare these independent judgments with BLIP's predicted distribution and with the KL values in Table 4. If DisTok's consistency advantage over the w/o-Cons ablation disappears or reverses under independent labeling, the shared-VLM circularity is confirmed and the distribution-alignment claim must be substantially weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that BLIP's softmax over the fixed 60-class CangJie set is a faithful and sufficient ground truth for the semantics of a generated image. This assumption enters the method twice. In Sec. 3.3, the VLM's predicted distribution p_vlm(c|x_crt) is used as structured supervision; in Sec. 3.4, L_cst aligns the decoded token with a previously VLM-labeled token; and the KL ablation in Table 4 measures the same BLIP distribution against the input distribution. The novelty gate (max_c p_vlm < tau) also uses BLIP. The loop is therefore closed: DisTok is trained to produce images whose BLIP softmax matches the input distribution, and is then evaluated by how well the BLIP softmax matches the input distribution. Because the label set is coarse and the images are deliberately out-of-distribution hybrids, BLIP's predictions can reflect its own confusion rather than true visual composition, e.g., dominant-class bias or arbitrary assignment for novel blends. If that is the case, the distributional-consistency improvement in Table 4 may be an improvement in matching BLIP's biases, not in semantic fidelity. The GPT-4o and user studies measure global creativity and aesthetics, not class-proportion fidelity, so they do not resolve this circularity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Distribution-Conditional Generation (DCG), a formulation in which image synthesis is conditioned on soft class distributions rather than discrete labels or concept pairs, and proposes DisTok, an encoder-decoder architecture that maps such distributions to learned text tokens. Training alternates between recursive concept-pair fusion and a VLM-supervised consistency objective, with a dynamically growing concept pool. The authors report qualitative and quantitative results on DCG, text-pair-to-object generation, unconditional creation, style adaptation, and human/GPT-4o evaluations, claiming state-of-the-art performance and large speedups over prior creative-generation methods.","tokens_in":14391,"tokens_out":7296,"duration_ms":70544,"significance":"The DCG formulation is a meaningful step beyond pairwise concept fusion: it offers a continuous, controllable interface for creative synthesis and, if validated, would let a single token encode a multi-class distribution and be reused across styles without per-concept optimization. Strengths of the paper include a self-contained training loop, no per-concept test-time optimization, a clear 3s inference advantage, and extensive qualitative demonstrations. The paper also ships a user study and multiple automated metrics. However, the empirical support for the central SOTA claim is currently incomplete: Table 1 lacks quantitative DCG baselines, Table 4's KL evaluation is both circular and statistically thin, and the GPT-4o protocol appears to supply the target composition to the evaluator. These gaps must be addressed before the main claims can be accepted.","major_comments":[{"comment":"The central claim that DisTok 'consistently outperforms' diffusion baselines on the DCG task is not backed by the main quantitative table: Table 1 reports VQAScore/PickScore/ImageReward for DisTok (DCG) but omits the corresponding rows for SD3, SD3.5, Kandinsky, and FLUX on the same distribution-conditioned inputs, so no baseline comparison is possible on that task. Please add those baseline rows or clearly state why they are unavailable; the current comparison rests only on qualitative Fig. 3 and on the GPT-4o scores in Table 2.","section":"Section 5.1, Table 1"},{"comment":"The training loop uses BLIP-predicted class distributions as supervision: the novelty gate (max_c p_vlm < tau), the stored p_vlm labels for novel concepts, and the consistency objective all depend on BLIP's softmax over the CangJie classes. Table 4 then evaluates distributional consistency by computing KL divergence between the input distribution and the same BLIP model's predictions on generated images. This is a closed evaluation loop: the reported improvement (0.0732 vs. 0.0602) may measure how well DisTok matches BLIP's biases rather than true visual composition. The paper should evaluate consistency with independent annotations (e.g., human proportion judgments or a different VLM) and should include non-DisTok baselines; otherwise the 'distributional consistency' claim is under-determined.","section":"Sections 3.3, 3.4, Table 4"},{"comment":"The only quantitative evidence for the proposed distribution-consistency objective is a single pair of KL values with no variance, no statistical test, and no baseline comparison for standard T2I models. The claim that removing this objective 'significantly increases divergence' requires error bars over repeated evaluations and at least one independent baseline, especially since the effect size (0.013) is small relative to likely run-to-run variation.","section":"Table 4 and Section 5.6"},{"comment":"The GPT-4o evaluation prompt begins by asserting that 'the subject of this evaluation is an image that represents a mixture of' specific classes and percentages, and then asks GPT-4o to rate alignment and integration. This phrasing supplies the ground-truth composition to the evaluator, which can bias the scores toward any image that superficially matches the description and likely explains the near-ceiling scores with near-zero standard deviations (e.g., 9.8±0.1 for Originality). The protocol should ask GPT-4o to infer the mixture from the image alone, or should be redesigned so the target composition is not embedded in the subject line.","section":"Appendix B.1, Table 2"}],"minor_comments":[{"comment":"The notation for L_reg is ill-defined: it is unclear whether the expectation is over sampled latent vectors and how mu(z) and sigma(z) are computed (element-wise mean/std of z, or statistics of the sample distribution). Please clarify, since the claim that latent vectors can be sampled from any zero-mean, unit-variance distribution rests on this regularizer.","section":"Equation (3), Section 3.3"},{"comment":"The abbreviation 'Kandin' should be 'Kandinsky' for consistency with the rest of the text.","section":"Table 2"},{"comment":"References [19] and [20] are the same paper (Zhiqiu Lin et al., ECCV 2024) and should be consolidated into one entry.","section":"References"},{"comment":"The claim that CreTok produces 'near-identical outputs' for (Lion, Snake) and (Lion, Shrimp) is only supported qualitatively; please provide a diversity metric or embedding similarity to substantiate the 'greater expressiveness' claim.","section":"Section 5.2"},{"comment":"The text says 'Each training iteration consists of n sampling steps' and later mentions 'gradient accumulation over n = 8 steps'; please clarify the relationship between iterations, sampling steps, and gradient accumulation.","section":"Section 3.5"},{"comment":"The user study reports aggregate vote counts only (e.g., 312:188) without per-participant proportions, inter-rater agreement, or significance testing; please provide these details to assess the reliability of the preference results.","section":"Section 5.5, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is fresh and the qualitative results are compelling, but the empirical support is currently insufficient: Table 1 lacks DCG baselines, Table 4's KL evaluation is circular and statistically thin, and the GPT-4o protocol appears leading. If the authors can add direct quantitative comparisons and an independent consistency evaluation, the contribution would be substantially stronger. The user study is a point in favor but does not resolve the circularity. I do not see a fundamental flaw in the method itself, so major revision rather than rejection seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is worth your time: conditioning image generation on a full class-distribution vector, not just a label or a concept pair, is a real generalization over BASS, CreTok, and ConceptLab. The encoder-decoder design, the dynamic concept pool, and the iterative composition of generated tokens into richer distributions are all sensible and reasonably novel. The 3s single-pass inference is a genuine practical advantage over BASS’s 40s sampling and ConceptLab’s 120s optimization. The qualitative results, at least from the figures, do show coherent hybrids that respect proportions better than the diffusion baselines. That part is credible.\n\nWhere the paper needs work is in how it supports the distribution-consistency claim. The stress-test note is right: BLIP is used as the teacher in Sec. 3.3 and as the examiner in Table 4. So the KL improvement (0.0732 → 0.0602) only shows that the model got better at fooling BLIP, not that it got better at matching true class proportions. That circularity is load-bearing for the distribution-fidelity claim, though it does not sink the whole paper—PickScore, ImageReward, GPT-4o, and the user study are all BLIP-independent, and those still favor DisTok. The circularity is a serious flaw in one specific claim, not evidence of general incompetence.\n\nOther gaps are smaller but still real. Table 1 reports metrics for DisTok on the DCG task but no baseline rows for SD3, SD3.5, Kandinsky, or FLUX on that task; the GPT-4o scores in Table 2 and user votes in Table 3 partly fill that hole, but not for the standard alignment/preference metrics. The loss in Eq. (2) uses thresholds θ1 and θ2, but the paper never says what they are. No code or data is released, so reproducibility is currently nil. The 60-class CangJie set and 30 constructed distributions are also a fairly small playground, and I would not call the generation “semantically unconstrained” when it is working inside that fixed label set.\n\nNone of this is fatal to the concept. The formulation is a contribution, and the architecture is simple enough that I could see it being replicated. What is missing is a cleaner evaluation protocol: independent annotation of class proportions, direct baseline rows on the DCG task, and the missing hyperparameters. If those are provided, this becomes a solid paper.\n\nWho is this for? People working on creative T2I, concept composition, or evaluation of out-of-distribution generation. It deserves a serious referee, but my recommendation would be major revision, not accept.\n\nBottom line: send it to review, ask the authors to fix the circularity and the missing baselines, and you have a paper worth publishing.","headline":"A genuinely new conditioning paradigm for creative T2I with a real circularity in its evaluation; worth refereeing but not accepting as-is.","tokens_in":14967,"tokens_out":2367,"would_cite":false,"duration_ms":26838,"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":"The paper claims that creativity for text-to-image models can be formulated as conditioning on a class distribution: one learned token encodes the whole distribution, and a diffusion model renders it as a single coherent hybrid in about…","keywords":["distribution-conditional generation","creative text-to-image synthesis","concept tokens","class distribution conditioning","diffusion models","vision-language supervision","concept pool","zero-shot concept generation"],"falsifier":"Take images DisTok generates for a known input distribution (say 55% pig, 45% sheep), then have independent human annotators—or a second vision-language model with a different class vocabulary—identify which classes are visible and how strongly. If the independent judge cannot recover the stated proportions, or systematically disagrees with the BLIP predictions used for training and for the KL metric in Table 4, then the claimed distributional fidelity is inherited from the supervisor, not a property of the images.","tokens_in":13933,"feed_emoji":"🎨","tokens_out":10936,"duration_ms":100985,"temperature":0.7,"pith_summary":"Most text-to-image systems can only recombine what they have already seen, and even dedicated creativity methods fuse at most two describable concepts at a time. This paper proposes distribution-conditional generation: creativity is modeled as synthesis from a class distribution—percentages over known classes—and a single learned token is asked to carry the whole mixture. The framework, DisTok, trains that token-maker by fusing concept pairs and by checking its own outputs against a vision-language model that predicts which classes a generated image actually shows, then feeds the prediction back as supervision. If the claim holds, a user can steer creation by choosing proportions rather than writing prompts, and novel hybrids with no linguistic description—e.g. a blend of five animals at uneven weights—appear in about three seconds with no per-concept optimization.","feed_headline":"One learned token packs a multi-class distribution into a single image","feed_subtitle":"DisTok turns class proportions into a reusable token, blending several known concepts in about three seconds.","key_machinery":"The load-bearing object is DisTok's encoder-decoder pair: a Distribution Encoder $E_{\\text{dis}}$ that maps a class distribution $p_c \\in \\Delta^K$ into a 20-dimensional latent $z = E_{\\text{dis}}(p_c)$, and a Creative Decoder $D_{\\text{tok}}$ that decodes $z$ into a token $t_{\\text{crt}}$ which is dropped into a prompt like \"a photo of a $\\langle t_{\\text{crt}}\\rangle$\" for the Kandinsky 2.1 diffusion model. Around this pair, the machinery is a growing Concept Pool $\\mathcal{P}$ and two complementary objectives: a mixing loss $L_{\\text{mix}}$ that fuses sampled concept pairs while a threshold caps any single concept's dominance, and a consistency loss $L_{\\text{cst}}$ that renders images from Gaussian-sampled latents, asks a vision-language model \"What animal is in the photo?\", and aligns the predicted class distribution with the input distribution. A latent regularizer $L_{\\text{reg}}$ pins samples to zero mean and unit variance so that any such distribution can drive exploration, and a novelty gate admits a sampled token into the pool only when its max predicted class probability stays below $\\tau = 0.85$.","core_discovery":"DisTok treats creative generation as the inverse of classification: just as a classifier answers an ambiguous input with a soft distribution over familiar classes, DisTok starts from that distribution and produces an image. The Distribution Encoder collapses the distribution into a low-dimensional latent, the Creative Decoder turns the latent into a token the diffusion model can render, and the Concept Pool grows as newly sampled tokens pass the novelty threshold, so later rounds can fuse richer mixtures. Two supervision channels train the pair: continuous concept combination, which fuses pairs of existing tokens under adaptive and restrictive prompts, and distribution consistency, where the vision-language model's prediction about a rendered image supervises the alignment between the input distribution and the visual outcome. The paper's central claim is that this single architecture covers the whole spectrum of creative generation—fine-grained multi-class composition with controlled proportions, two-concept fusion as its uniform-distribution special case, unconditional sampling from any zero-mean unit-variance latent, and style transfer via the same reusable token—and does so in about three seconds per concept without per-concept optimization.","pith_inferences":["Beyond the paper: the method's \"novelty\" is measured only against the fixed 60-class vocabulary, so an image that shares nothing with those classes would be rejected by the novelty gate rather than celebrated; extending the same recipe to an open or expanding class set is the natural next step.","Beyond the paper: the distributional-consistency loop is self-referential, since the same vision-language model supplies the supervision that shapes the tokens and then the metric (KL divergence) that certifies their fidelity; an independent judge would settle how much of the reported proportion fidelity is real.","Beyond the paper: a cheap stress test would freeze the trained encoder-decoder and rerun the pipeline with a different vision-language supervisor; large changes in the resulting tokens would reveal how much of DisTok's compositional behavior is inherited from the supervisor's biases."],"forward_implications":["A single token can encode three or more classes at once with controlled proportions, so multi-concept composition needs no new machinery beyond the one trained encoder-decoder.","Two-concept fusion (the TP2O task) falls out as the special case of a uniform distribution over two classes, letting one model serve both the new task and the existing benchmark.","Because latents can be drawn from any zero-mean, unit-variance distribution, the model generates novel concepts unconditionally, without reference images, prompts, or iterative optimization.","The same token stays recognizable under style prompts such as watercolor, line drawing, or pop art, separating the conceptual content from its rendering style.","Inference cost is roughly three seconds per concept, against about forty seconds for BASS and about one hundred twenty seconds for ConceptLab, which is what makes the token-level approach practically deployable."],"supporting_citations":[{"why":"Defines the CangJie dataset of 60 base concepts and 30 TP2O pairs, supplies the adaptive/restrictive prompt combination strategy that DisTok inherits, and serves as the token-based baseline DisTok must beat.","marker":"[9]"},{"why":"Establishes the text-pair-to-object baseline and the balance swap-sampling approach that DisTok replaces with single-pass encoding, and provides the speed comparison (about 40 s per concept).","marker":"[17]"},{"why":"Introduces the Creative Text-to-Image task and the iterative VLM-guided token optimization that DisTok's unconditional sampling is compared against (about 120 s per concept).","marker":"[34]"},{"why":"Kandinsky 2.1 is the diffusion backbone that renders tokens into images and supplies the CLIP text encoder used in the losses.","marker":"[33]"},{"why":"CLIP-L/14 text embeddings define the cosine-similarity losses L_mix and L_cst that align adaptive, restrictive, and generated-token prompts.","marker":"[31]"},{"why":"Stable Diffusion 3 is the standard T2I baseline that DisTok outperforms on the distribution-conditional task in Figure 3 and Table 2.","marker":"[7]"},{"why":"VQAScore is the text-image alignment metric used to report DisTok's superiority in Table 1.","marker":"[20]"},{"why":"PickScore is the human-preference metric used in the quantitative comparison of Table 1.","marker":"[14]"},{"why":"ImageReward is the second human-preference metric in Table 1, trained on human feedback for text-to-image outputs.","marker":"[40]"},{"why":"Supplies the GPT-4o model used for the creativity scoring (integration, alignment, originality, aesthetics) in Table 2.","marker":"[1]"}],"fun_headline_variants":["Turn any class distribution into a reusable image token","One token encodes an entire class distribution for image synthesis","Inverse classification: from class probabilities to creative images","DisTok maps class distributions to creative tokens in seconds","A single token casts a distribution across an image"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything hinges on accepting the vision-language model's softmax distribution over the fixed 60 CangJie classes as a faithful and complete description of what a novel hybrid image actually depicts; if that predictor mislabels or under-specifies out-of-distribution hybrids, both the training signal from Section 3.3 and the consistency metric in Table 4 are measuring the model's bias rather than the true visual composition.","fun_headline_variants_meta":{"raw":{"variants":["Turn any class distribution into a reusable image token","One token encodes an entire class distribution for image synthesis","Inverse classification: from class probabilities to creative images","DisTok maps class distributions to creative tokens in seconds","A single token casts a distribution across an image"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1381,"prompt_tokens":1004,"completion_tokens":377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":303}},"tokens_in":620,"tokens_out":377,"duration_ms":3934,"temperature":1.0,"reasoning_tokens":303,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:45:22.046582+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take images DisTok generates for a known input distribution (say 55% pig, 45% sheep), then have independent human annotators—or a second vision-language model with a different class vocabulary—identify which classes are visible and how strongly. If the independent judge cannot recover the stated proportions, or systematically disagrees with the BLIP predictions used for training and for the KL metric in Table 4, then the claimed distributional fidelity is inherited from the supervisor, not a property of the images.","supporting_citations":[{"cited_title":"Redefining <creative> in dictionary: Towards an enhanced semantic understanding of creative generation","cited_arxiv_id":null,"evidence_quote":"Defines the CangJie dataset of 60 base concepts and 30 TP2O pairs, supplies the adaptive/restrictive prompt combination strategy that DisTok inherits, and serves as the token-based baseline DisTok must beat."},{"cited_title":"Tp2o: Creative text pair-to-object generation using balance swap-sampling","cited_arxiv_id":null,"evidence_quote":"Establishes the text-pair-to-object baseline and the balance swap-sampling approach that DisTok replaces with single-pass encoding, and provides the speed comparison (about 40 s per concept)."},{"cited_title":"Conceptlab: Creative concept generation using vlm-guided diffusion prior constraints.ACM Transactions on Graphics, 43(3):1–14, 2024","cited_arxiv_id":null,"evidence_quote":"Introduces the Creative Text-to-Image task and the iterative VLM-guided token optimization that DisTok's unconditional sampling is compared against (about 120 s per concept)."},{"cited_title":"Kandinsky: An improved text-to-image synthesis with image prior and latent diffusion","cited_arxiv_id":null,"evidence_quote":"Kandinsky 2.1 is the diffusion backbone that renders tokens into images and supplies the CLIP text encoder used in the losses."},{"cited_title":"Learning transferable visual models from natural language supervision","cited_arxiv_id":null,"evidence_quote":"CLIP-L/14 text embeddings define the cosine-similarity losses L_mix and L_cst that align adaptive, restrictive, and generated-token prompts."},{"cited_title":"Scaling rectified flow trans- formers for high-resolution image synthesis","cited_arxiv_id":null,"evidence_quote":"Stable Diffusion 3 is the standard T2I baseline that DisTok outperforms on the distribution-conditional task in Figure 3 and Table 2."},{"cited_title":"Evaluating text-to-visual generation with image-to-text generation","cited_arxiv_id":null,"evidence_quote":"VQAScore is the text-image alignment metric used to report DisTok's superiority in Table 1."},{"cited_title":"Pick-a-pic: An open dataset of user preferences for text-to-image generation","cited_arxiv_id":null,"evidence_quote":"PickScore is the human-preference metric used in the quantitative comparison of Table 1."},{"cited_title":"Imagereward: Learning and evaluating human preferences for text-to-image generation","cited_arxiv_id":null,"evidence_quote":"ImageReward is the second human-preference metric in Table 1, trained on human feedback for text-to-image outputs."}],"review_version":1}