{"id":"9107dbd3-c949-4f59-93d3-8f26be61ab4e","arxiv_id":"2504.13392","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"POET is a tool that finds hidden sameness in AI-generated images, widens those dimensions for more variety, and personalizes future outputs from user feedback.","lead":"POET is a tool that finds hidden patterns of sameness in AI-generated images, expands those patterns to produce more varied pictures, and learns from user likes and dislikes. A study with 28 people found it increased perceived variety and helped users reach satisfying images in fewer prompts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ICAD diversity gains may be inflated because the CLIP representation used in the filtering objective also defines the evaluation metric; the human study is independent but does not validate the automatic numbers.","rationale":"The reader's weakest assumption identifies CLIP as the load-bearing proxy, and I agree that this is the core risk. My stress-test sharpens the concern: the problem is not merely that CLIP might miss creative attributes, but that the automatic evaluation metric appears to share the same representation as the selection objective, creating a circularity that can inflate the central quantitative result. The human Phase 1 study provides some independent evidence for perceived pairwise difference, which is a real mitigating factor, but it does not cover the automatic ICAD numbers or the fidelity half of the diversity-fidelity trade-off. The paper does not report human fidelity ratings, and the ICAD computation is not specified in enough detail to rule out CLIP-based evaluation. I therefore recommend keeping the reader's CONDITIONAL verdict: the central claim is plausible and partially supported by the user study, but it should not be fully accepted until the automatic evaluation is shown to be robust under a representation-independent metric. I do not move the verdict because the existing conditional assessment already captures the need for this additional evidence.","tokens_in":25767,"tokens_out":4110,"duration_ms":42398,"concrete_test":"Recompute the ICAD values in Figure 4 on the same generated image sets using a CLIP-independent perceptual metric such as LPIPS or DreamSim instead of CLIP embeddings, and collect human fidelity ratings for a random subset of Base and POET outputs. If the Base-to-POET increase (0.24 to 0.48) is not reproduced, or if human fidelity ratings show POET images deviate more from the original prompt, then the automatic diversity and fidelity claims are artifacts of self-similar evaluation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim's automatic evidence is the ICAD comparison in Figure 4: Base 0.24, POET w/o HDI 0.41, and POET 0.48. However, the pipeline's selection criterion in Equations 3 and 4 uses CLIP text-image similarity both for the diversity term (Sim_image(t1, I_hat)) and the fidelity term (Sim_image(t0, I_hat)), and the paper does not specify a CLIP-independent metric for ICAD. If ICAD is computed in the same CLIP embedding space, then the evaluation metric is not independent of the selection objective: prompts are chosen to produce images that CLIP regards as distant from t1, and the same CLIP model then reports those sets as diverse. This creates an inflation loop that can overstate diversity even when pixel-level or human-perceived diversity is unchanged. The Phase 1 user study is a genuine independent check for pairwise perceived difference and supports the perceived-diversity claim, but it does not validate the specific automatic ICAD numbers or the fidelity component of 'maintaining fidelity and consistency with user intent' because no human fidelity rating is reported. The weakest load-bearing assumption is therefore that CLIP similarity is a sufficient proxy for both homogeneity and diversity in the automatic evaluation; this is a validity risk, not a claim that CLIP is worthless.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents POET, a real-time text-to-image generation tool that automatically discovers homogeneous dimensions in generated image sets via a novel prompt-inversion method, expands prompts along those dimensions using an LLM under a CLIP-based diversity-fidelity filter, and personalizes future expansions from user feedback. The authors evaluate the system with automatic metrics (ICAD, inference cost) on 1,000 Flickr30K prompts and with a within-subjects user study (N=28) across four creative scenarios. They report that POET produces image sets users perceive as significantly more diverse (Phase 1), helps users reach satisfaction in fewer prompt iterations (marginal in Phase 2, significant in Phase 3), and that POET-Personalize outperforms Base-Personalize on rounds and satisfaction. The paper also contributes a project page with code and discusses implications for pluralistic, user-centered generative tools.","tokens_in":26033,"tokens_out":5734,"duration_ms":49037,"significance":"If the findings hold, POET is a meaningful step toward creativity support tools that push back against normative and homogeneous outputs of text-to-image models. The prompt-inversion method for discovering interpretable homogeneous dimensions is a novel and potentially reusable contribution, and the user study is generally well designed with counterbalanced conditions, mixed-methods analysis, and a clear focus on ideation-stage needs. The Phase 1 perceived-diversity result is strong (coefficient 2.28, p<0.001) and provides independent human evidence supporting the diversity claim. The paper also ships code and a project page, aiding reproducibility. However, the automatic evaluation's reliance on CLIP without an explicitly independent metric, the marginal Phase 2 rounds result, and the confounded personalization comparison mean the strongest claims require careful revision before the contribution is fully convincing.","major_comments":[{"comment":"The automatic ICAD evaluation is potentially circular with the filtering objective. The paper does not specify the feature space in which ICAD is computed, while the filter in Eqs. (3)-(4) uses CLIP text-image similarity for both the diversity term Div(t_hat) = Sim_image(t0, I_hat) - Sim_image(t1, I_hat) and the fidelity term Sim_text(t_hat, t0). If ICAD is also computed in a CLIP embedding space, the reported gains (Base 0.24, POET w/o HDI 0.41, POET 0.48) partially reflect the selection procedure's own objective rather than an independent measure of visual diversity. Please specify the embeddings used for ICAD; if they are CLIP-based, add a complementary CLIP-free diversity metric (e.g., LPIPS or pixel-level distance) and report error bars or significance tests for the ICAD values, which are currently single numbers without uncertainty.","section":"§4.2.2, Figure 4, Table 1, and Eqs. (3)-(4)"},{"comment":"The Phase 2 rounds coefficient for POET vs Base is -0.857 with p<0.1 (indicated by the dot), which is not significant at the conventional 0.05 level. The text states that \"POET was able to help users achieve satisfiable results in fewer rounds than Base in Phase 2\" without qualifying this as a marginal trend. Please report the exact p-value and explicitly characterize this as a marginal result, or temper the claim in the abstract and conclusions, which currently state that POET helps users \"reach satisfaction in fewer prompts\" as a general finding.","section":"§5.3.2, Table 4"},{"comment":"The Phase 3 comparison of POET-Personalize vs Base-Personalize does not isolate the contribution of personalization, because Base-Personalize already includes a personalization mechanism on the base model. The observed reduction in rounds (1.35, p<0.01) and higher satisfaction may be driven entirely by POET's homogeneous-dimension expansion, not by the personalization step. To support the personalization contribution, add a POET vs POET-Personalize comparison (or a fully factorial design), or explicitly reframe the Phase 3 result as evaluating the combined POET-plus-personalization system rather than the personalization component alone.","section":"§5.1.3, Table 5"},{"comment":"The claim that POET maintains \"fidelity and consistency with user intent\" is not evaluated by users. The only consistency evidence is qualitative (Figure 7), and the filter relies on CLIP similarity as a proxy. The user study measures perceived diversity and satisfaction but not how well the generated images match the original prompt. Please include a user-rated fidelity measure (e.g., a match or preservation rating for each generated image against the original prompt) or revise the claim to specify that fidelity is defined in terms of CLIP similarity rather than human judgment.","section":"§4.2.3 and §5.3"}],"minor_comments":[{"comment":"The description of prompt inversion says the original prompt is padded with random tokens, but Algorithm 1 initializes Z from the original tokens without explicitly showing the padding step; please clarify how the padded tokens enter the initialization and how m relates to the padded length.","section":"§3.1, Algorithm 1"},{"comment":"The projection function has inconsistent notation: the denominator uses ||e_i|| while the argmax is over e; the index in the denominator should match the variable being projected.","section":"Eq. (2)"},{"comment":"The y-axis is labeled only \"ICAD\"; please add the units or note that it is unitless, and include error bars or shaded intervals for the reported means.","section":"Figure 4"},{"comment":"The ICAD values are reported as point estimates without variability; please report the standard deviation or confidence interval across the 1,000 prompts.","section":"Table 1"},{"comment":"The repeated use of \"satisfiable\" is incorrect in this context; the intended word is \"satisfying\" or \"satisfactory\" (e.g., \"more satisfying results,\" \"satisfactory results\").","section":"§5.3.2 and elsewhere"},{"comment":"The maximum of five re-prompts may create a ceiling effect that compresses differences in the number of rounds; please discuss this limitation explicitly.","section":"§5.1.2"}],"recommendation":"major_revision","confidential_remarks":"The automatic evaluation is the main validity risk: without specifying the ICAD feature space and adding a CLIP-independent diversity metric, the diversity numbers may be inflated by the selection objective. The user study is the strongest evidence, but the Phase 2 rounds result is marginal and the personalization claim is confounded. The paper would be substantially strengthened by a factorial design isolating personalization and by human-rated fidelity measurements. Overall, the core idea is promising and the human results are largely supportive, so major revision seems the right call rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take: the POET paper is a solid HCI systems contribution with a genuine user study. The central perceived-diversity claim—users find POET's output more diverse—holds up, because it is grounded in the Phase 1 human ratings, which are statistically significant across four scenarios. The Phase 3 personalization results (fewer rounds to satisfaction, higher satisfaction) are also significant. The paper deserves a serious referee.\n\nWhat is actually new is the pipeline: inverting a set of generated images into shared discrete token embeddings to surface homogeneous dimensions, then using an LLM to expand those dimensions, and a user-preference profile for personalization. Textual inversion, hard prompt optimization, and prompt-suggestion systems exist separately, but this combination—multi-image set inversion for homogeneity discovery plus expansion—is a real contribution. The code link and the honest qualitative reporting (e.g., negative prompting failures, realism tradeoffs) are both positive signals.\n\nThe main soft spot is the automatic evaluation. The filter in Equations 3–4 is defined in CLIP space, and the ICAD metric in Figure 4 likely uses the same embedding family. So the automatic diversity gains (0.24→0.48) are partly built into the selection procedure. The human study is an independent check and does reduce the circularity burden, but it does not validate the specific automatic numbers. The paper should either report a CLIP-independent diversity metric or clearly mark the automatic metrics as descriptive and not causal. I'd also flag the Phase 2 rounds effect as marginal (p<0.1), so the 'fewer prompts' claim is best supported by the personalization phase. And the personalization comparison inherits the POET expansion advantage, so it is not a clean test of personalization alone. No error bars or significance tests appear on the ICAD numbers, and the free parameters (m, λ, n, optimizer settings) are unablated.\n\nMinor point: code is shared but there's no commit hash or user-study data, which limits reproducibility.\n\nWho this is for: HCI researchers working on creativity support tools, and CV researchers interested in diversity in text-to-image generation. It should go to peer review, with major revision: tighten the automatic evaluation, rebalance the claims toward the human study, ablate the hyperparameters, and report full statistics on the automatic metrics.","headline":"A well-run HCI study on diversifying text-to-image output, but the automatic ICAD gains partly reflect the CLIP filter/evaluation loop.","tokens_in":26548,"tokens_out":3629,"would_cite":true,"duration_ms":32879,"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":"This paper claims that a text-to-image tool can find the repetitive visual patterns a model secretly favors, deliberately vary those patterns, and learn from user picks, so people exploring ideas see more diverse images and reach…","keywords":["text-to-image generation","prompt inversion","homogeneous dimensions","creative ideation","output diversification","personalization","CLIP similarity","user study"],"falsifier":"Have independent human raters score pairwise diversity on images that POET selects as diverse and images it filters out as redundant; if human scores do not separate the two groups, the filtering function is optimizing the wrong signal, and replacing the CLIP-based filter with human diversity judgments should make the iteration savings disappear.","tokens_in":25575,"feed_emoji":"🎨","tokens_out":3950,"duration_ms":34820,"temperature":0.7,"pith_summary":"The paper introduces POET, a tool for the ideation stage of creative work with text-to-image models. POET automatically finds the shared, homogeneous visual dimensions that a pretrained model tends to reproduce (for example, always drawing a bearded elderly man for \"ancient artist\"), then expands those dimensions into alternatives while filtering to stay close to the user's original intent. A user study with 28 participants across four creative scenarios found that people rated POET image sets as more different from each other, reported higher satisfaction, and needed about 0.85 fewer rounds (and 1.35 fewer with personalization) to reach satisfaction. The value, if the claim holds, is that creative users get more varied starting points without doing the tedious work of rewriting prompts, and the system can later tailor expansions to a user's expressed likes and dislikes.","feed_headline":"Auto-expanded prompts give users more diverse AI images","feed_subtitle":"A new tool finds the repetitive patterns a model secretly favors, varies them, and learns from your picks to cut prompt rounds.","key_machinery":"The load-bearing mechanism is prompt inversion: initializing a sequence of learnable token embeddings from the user's prompt, optimizing them against the CLIP text–image similarity objective over the generated image set, and projecting the continuous embeddings back to nearest neighbor discrete tokens at every step so the result is human-readable. This inverted prompt reveals the model's tacit homogeneous dimensions. A second machinery is the filtering function $F(\\hat{t}) = \\mathrm{Div}(\\hat{t}) + \\lambda \\cdot \\mathrm{Sim}_{\\text{text}}(\\hat{t}, t_0)$, where $\\mathrm{Div}$ rewards images that differ from the homogeneous pattern while still resembling the original prompt, and $\\mathrm{Sim}_{\\text{text}}$ keeps the candidate prompt close to the user's original wording. Together they decide which expanded prompts enter the displayed image set.","core_discovery":"The central discovery is that the overly similar images a text-to-image model produces for one prompt can be traced to a small set of latent \"dimensions of homogeneity,\" and those dimensions can be read out automatically by inverting the prompt: optimize a sequence of token embeddings, projected onto the model's discrete vocabulary, to maximize CLIP similarity to the generated image set. Once the inverted prompt is read as text (e.g., \"considering experienced beard apostle writing\"), an LLM expands each semantic category into alternative attributes, and a filtering function trades diversity against semantic fidelity to the original prompt. Adding a feedback loop, where the user rates images and marks favorites and least-liked outputs, conditions future expansions on a stored preference profile. In the authors' evaluations, this pipeline outperforms both the base model and the base model with personalization but without dimension discovery on perceived diversity, satisfaction, and rounds to satisfaction.","pith_inferences":["The same inversion-plus-expansion loop could be applied to other generative backbones or modalities, so long as a CLIP-like text–image embedding space exists; the paper fixes one backbone, so cross-model generality is not tested.","Because the system's notion of diversity is defined by CLIP distances, its gains may be strongest for attributes CLIP can see (demographics, setting, activity) and weakest for subtle stylistic or conceptual differences human creatives care about; a human study swapping in different diversity criteria would settle this.","If CLIP's bias toward Western or stereotypical attributes is what drives the homogeneous dimensions, POET could in principle amplify or counteract specific stereotypes depending on how the LLM samples alternatives; the paper documents resistance to stereotypes but does not directly aim the expansion at specific axes of bias.","The 28-user study is modest in scale, so the size of the iteration savings (0.85–1.35 rounds) should be treated as an early estimate rather than a stable effect size."],"forward_implications":["Creative users in the studied domains (advertising, tourism, character design, interior design) can start from one prompt and see a deliberately broadened set of options, rather than several near-duplicates of the same concept.","The number of prompt rewrites needed to reach satisfaction drops measurably, which reduces the burden of prompt engineering for novices.","Personalization can be achieved without fine-tuning model weights, by conditioning a language model on a user preference profile built from satisfaction ratings and favorite and least-favorite images.","Automatic homogeneous-dimension identification beats both open and closed vision-language models at the same task in this setup and is cheaper, because the inversion runs once over the image set instead of querying a multimodal model repeatedly."],"supporting_citations":[{"why":"Supplies the textual-inversion idea of representing a concept as learned embedding tokens, which prompt inversion adapts to discover shared dimensions.","marker":"[28]"},{"why":"Provides the frozen CLIP text–image similarity that both the inversion objective and the diversity filtering rely on.","marker":"[84]"},{"why":"Provides the discrete-token projection technique that keeps the inverted prompts human-readable and semantically grounded.","marker":"[108]"},{"why":"Is the LLM used to categorize homogeneous dimensions and sample expanded attributes.","marker":"[72]"},{"why":"Supplies the ICAD metric used to measure image-set diversity in automatic evaluations.","marker":"[48]"},{"why":"Provides the 1,000 text prompts used as the automatic evaluation testbed.","marker":"[81]"},{"why":"Motivates the preference-conditioning approach for personalizing prompt expansion from user history.","marker":"[14]"}],"fun_headline_variants":["POET tool auto-expands prompts for more diverse AI images","Prompt inversion exposes hidden style biases in text-to-image models","Automated prompt expansion personalizes image generation","Tool learns your favorites to diversify AI image outputs","Fewer prompts to satisfied users via automated expansion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system treats CLIP cosine similarity as the measure of both sameness and diversity, so if CLIP's notion of visual similarity does not track what creative users mean by variety, the discovered dimensions and measured diversity gains could reflect the embedding space rather than the images themselves.","fun_headline_variants_meta":{"raw":{"variants":["POET tool auto-expands prompts for more diverse AI images","Prompt inversion exposes hidden style biases in text-to-image models","Automated prompt expansion personalizes image generation","Tool learns your favorites to diversify AI image outputs","Fewer prompts to satisfied users via automated expansion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000466,"raw_usage":{"total_tokens":2341,"prompt_tokens":976,"completion_tokens":1365,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":1290}},"tokens_in":592,"tokens_out":1365,"duration_ms":9860,"temperature":1.0,"reasoning_tokens":1290,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:09:35.778279+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have independent human raters score pairwise diversity on images that POET selects as diverse and images it filters out as redundant; if human scores do not separate the two groups, the filtering function is optimizing the wrong signal, and replacing the CLIP-based filter with human diversity judgments should make the iteration savings disappear.","supporting_citations":[],"review_version":1}