{"id":"8dca147a-7356-499e-a395-c0c824d691cc","arxiv_id":"2508.17718","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An MLLM-driven, training-free pipeline extracts preference keywords from a reference image and modulates diffusion cross-attention at global and regional levels for instant, multi-round preference-aligned image generation.","lead":"This paper proposes a training-free framework that uses a multimodal large language model to read a user's preferred image, extract preference keywords, enrich the text prompt, and steer a diffusion model's attention to generate matching images. It targets the practical problem of adapting image generation to individual taste on the fly, without retraining.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table S5 reports ~180s per image for the proposed method, directly contradicting the central 'real-time/instant' claim; the core novelty is unsupported by the paper's own measurements.","rationale":"The reader's verdict is CONDITIONAL, focusing on evaluation weaknesses (forced-choice taxonomy validation, MLLM-as-judge circularity, missing code/data). I agree those are real concerns, but the single most load-bearing issue is the internal contradiction between the paper's central 'real-time/instant' claim and its own reported execution time in Table S5. The paper's stated contribution is 'pioneer the instant preference alignment framework for real-time T2I generation.' If generation takes ~180s per image, that claim is false by the authors' own measurement, not by an outside standard. This is more decisive than the taxonomy concern because it directly falsifies the primary novelty. The concrete test is straightforward: reproduce the timing. The paper already provides the number, so unless 'real-time' is redefined in an unconventional way, the concern lands. I recommend REJECT because the central claim is contradicted by the paper's own data; a revised manuscript that removes the real-time claim and reframes the contribution would be a different paper. I mark agreement as 'partial' because the reader identified the timing conflict in the rationale but selected the taxonomy as the weakest assumption; the timing issue is stronger and deserves the central position.","tokens_in":17930,"tokens_out":5346,"duration_ms":61991,"concrete_test":"Reproduce the timing experiment in Supp. Table S5 on the same GPU (e.g., A100) with the released configuration (SDXL, Qwen-VL-72B, DPM-Solver 30 steps, α=0.7, λ=0.2), measuring end-to-end wall-clock from reference-image+prompt input to final image. If the measured latency is ≈180s (or >10s for a 1024×1024 image), the 'real-time' claim is falsified. Additionally, request a component-level breakdown (MLLM calls vs. diffusion sampling) and a screen recording of an interaction round; if total latency exceeds 2s, the paper must be revised to remove 'real-time' from its claims.","verdict_should_be":"REJECT","load_bearing_attack":"The abstract and contributions repeatedly claim 'real-time' and 'instant' preference-aligned generation (Sec. 1, Contribution 1). Yet Supp. Table S5 lists the method's execution time as ~180s, roughly 25x slower than the NP baseline (~7s) and on par with Viper and Fabric (>180s), the very methods the paper criticizes for time overhead. The supplementary text even claims the method avoids 'the substantial time overhead associated with methods like Viper and Fabric,' which is inconsistent with its own table. This is not an external disagreement about what 'real-time' means; it is an internal contradiction between the headline claim and the paper's own reported measurement. If 180s is accurate, the central contribution—instant, real-time preference alignment—is falsified. The human-study and metric comparisons may still show quality gains, but the key differentiator over prior work collapses, making the concern load-bearing: it targets the primary novelty, not a peripheral detail.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a training-free framework for preference-aligned text-to-image generation. Given a reference image and a base prompt, an MLLM extracts preference keywords in four fixed categories, enriches the prompt with preference-informed entities, plans bounding boxes, and then performs inference with global keyword-embedding projection (Eq. 6) and local region-wise cross-attention modulation (Eqs. 7-10). The authors report qualitative and quantitative comparisons against NP, Fabric, MGIE, Prompt Rewrite, Viper, and InstantStyle, along with ablations, multi-round interaction examples, and generalization across MLLMs and diffusion backbones. The central claim is that this is an 'instant/real-time' preference alignment method that outperforms prior approaches in both metrics and user studies.","tokens_in":18147,"tokens_out":7995,"duration_ms":89709,"significance":"If the central claims are upheld, the framework would offer a flexible training-free alternative to preference fine-tuning and to Viper-style personalization, with a clean global/local decomposition and a single-image preference interface. The paper is well motivated and provides extensive qualitative evidence, module ablations, and multi-round interaction examples. However, the strongest differentiators—'instant', 'real-time', and quantitative superiority—are currently not supported by the paper's own measurements and reporting. The execution-time table (Supp. Table S5) directly contradicts the headline claim, and several metrics are reported in ways that prevent independent verification. The underlying architecture remains plausible and potentially valuable, but the present manuscript needs substantial revision before the central claims can be accepted.","major_comments":[{"comment":"Table S5 reports our method's execution time as ~180 s per image, compared with ~7 s for NP and ~10 s for InstantStyle, and on par with Viper/Fabric (>180 s). This directly contradicts the abstract's 'real-time and training-free' claim and Contribution 1 ('instant preference alignment framework for real-time T2I generation'). The supplementary text even claims the method avoids 'the substantial time overhead associated with methods like Viper and Fabric,' which its own table disproves. Please provide a wall-clock breakdown (MLLM inference vs. diffusion steps vs. attention composition) and either revise the real-time/instant claims or reduce the actual latency; as written, the main differentiator over prior work is falsified by the paper's own measurement.","section":"Supp. Sec. B.2 / Table S5 vs. Abstract, Sec. 1"},{"comment":"ImageReward is reported as dimensionless percentages (e.g., 25.4% in Table 1 and 80.2% in Table 2) without any stated normalization or win-rate protocol; ImageReward is normally a scalar reward, so the units and the text 'text corresponds to the base prompt and the extracted keywords' do not explain these numbers. In the same table, 'Style Loss' is placed under 'Metrics↑', yet a loss is conventionally lower-is-better, and Sec. 4.2 says InstantStyle (0.712, the maximum in that column) 'excels in Style Loss' while Ours is 0.663. The direction of this metric must be clarified, or the quantitative superiority claim cannot be evaluated.","section":"Sec. 4.1, Table 1 and Table 2"},{"comment":"The completeness of the four-category preference taxonomy is validated by asking users to choose reasons from the same four categories (plus 'Others'); with 4.3% choosing 'Others', the 95.7% coverage is partly a construction. The method and its evaluation rely on MLLMs (Qwen-VL-72B for extraction; GPT-4o as judge), so shared MLLM bias is a real risk, and the human study (20 participants, no significance testing) is not sufficient to rule it out. The paper's own Limitations paragraph concedes that 'certain minor preferences may not be fully captured,' which further tempers the 'broader and more fine-grained coverage' claim. Please add an open-ended preference-elicitation study and independent human judgments with error bars.","section":"Sec. 4.1, Table 3; Sec. 4.3; Limitations paragraph"}],"minor_comments":[{"comment":"Typo: 'generate generate corresponding' should be 'generate corresponding'.","section":"Eq. (7)"},{"comment":"The text says 'This distribution of these selections (Fig. 3, first row)' but the referred result is in Table 3, not Fig. 3.","section":"Sec. 4.2"},{"comment":"ImageReward is cited as (Gal et al. 2022); the reference list attributes ImageReward to Xu et al. (2023). Please correct the citation.","section":"Sec. 4.1"},{"comment":"The text says 'as shown in failure cases in Supp. C,' but Supp. C does not contain a failure-case section. Either provide the promised failure cases or correct the reference.","section":"Limitations and Future Work"}],"recommendation":"major_revision","confidential_remarks":"The internal contradiction between the 'real-time/instant' headline and Supp. Table S5 is serious enough to block acceptance. I would ask the editor to require a revised timing analysis, metric definitions, and a more cautious framing of the central contribution before any further decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things. First, the method itself is a sensible combination of existing ideas: MLLM keyword extraction with a fixed taxonomy, orthogonal-projected global embedding addition, and mask-blended regional cross-attention. The qualitative results are genuinely convincing, and the user study, though small, shows a real preference advantage over Viper and Prompt Rewrite. Second, the paper's central \"real-time/instant\" claim is contradicted by its own Table S5, which reports ~180s per image—about the same as Viper and Fabric, the very methods it criticizes for time overhead. The supplementary text even claims the method avoids that overhead. That is not a semantic quibble; it is an internal contradiction in the headline contribution.\n\nCredit where due: the global projection in Eq. 6 is a clean way to preserve base-prompt semantics while injecting preference keywords, and the regional cross-attention blending avoids RPG's resizing artifacts. The prompt templates are detailed and the framework is genuinely extensible, as shown in the appendix. No code or data are released, which hurts reproducibility, but the pipeline description is concrete enough to re-implement.\n\nThe soft spots are real but mostly fixable. The timing contradiction is load-bearing: if 180s is accurate, the \"instant\" differentiator collapses, leaving \"training-free\" as the contribution—still useful, but not what the abstract promises. The evaluation also leans on GPT-4o both as the MLLM prior inside the method and as judge, so shared-bias is a live concern; the 20-participant user study is too small to fully offset it. The preference-coverage validation in Table 3 forces users to choose among the paper's own categories, making the high coverage partly circular. Hyperparameters α and λ appear tuned on the test benchmarks, with no held-out validation. And ImageReward is reported in percentage units without explanation—odd for a scalar reward.\n\nOverall, this is a solid training-free method wrapped in an inflated real-time narrative. It deserves a serious referee, but the referee should demand measured runtime, a clarified definition of \"real-time,\" an independent judge or larger human study, and code release. I'd send it to review with the expectation of major revision, not desk reject.","headline":"A plausible training-free preference-alignment framework, undermined by its own supplementary time table showing ~180s/image—the 'real-time' claim is internally contradicted.","tokens_in":18640,"tokens_out":2386,"would_cite":false,"duration_ms":28192,"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 training-free pipeline aligns text-to-image generation with a user's preferences in real time, using MLLM-extracted keywords and global-local attention guidance from a single reference image.","keywords":["text-to-image generation","preference alignment","training-free control","multimodal large language models","prompt enrichment","cross-attention modulation","diffusion models","interactive refinement"],"falsifier":"Run a preference-elicitation study with fully open-ended user responses (no category menu) and check whether the MLLM keywords and generated images align with the stated reasons; if a substantial fraction of preferences fall outside the four categories or the keywords miss them, the coverage claim is falsified.","tokens_in":17826,"feed_emoji":"🎨","tokens_out":6649,"duration_ms":76787,"temperature":0.7,"pith_summary":"This paper tries to establish that preference-aligned text-to-image generation can be done instantly and without fine-tuning: a user hands over a reference image they like, and the model generates new images that carry over the same artistic style, emotional tone, themes, and visual elements. The key move is to separate the problem into understanding preferences and using them to guide generation. Understanding is handled by a multimodal large language model that turns a reference image into structured keywords and an enriched prompt; guidance is handled at inference time by adding a projected preference embedding to the prompt embeddings and by modulating cross-attention region-wise. If this works as claimed, users and dialog systems can steer image generation in real time and iteratively, without retraining per preference or per user.","feed_headline":"One preferred image becomes the brief for new generations","feed_subtitle":"MLLM keywords plus global-local attention guidance align images to style, mood, and theme on the fly—no retraining.","key_machinery":"The machine is a two-stage inference-time pipeline. Preference understanding is carried by a fixed four-category keyword taxonomy (artistic style, emotional/atmospheric, thematic, visual elements) plus chain-of-thought MLLM prompting that extracts keywords from a reference image and enriches the base prompt. Preference-guided generation is carried by two mechanisms: global preference guidance, P_yi = T_text(y_i) + α·Proj⊥(T_text(y_key)), which injects preference keywords into text embedding space while subtracting the component already present in the prompt; and local cross-attention modulation, which computes region-specific cross-attention latents from per-entity sub-prompts and writes the","core_discovery":"The central claim is that the full preference-alignment loop can be moved out of training. Given a reference image, an MLLM extracts keywords in four preference categories—artistic style, emotion/atmosphere, theme, visual elements—and rewrites the base prompt into a complex prompt, per-entity sub-prompts, and a background prompt. Generation then uses two training-free controls: global preference guidance, which orthogonally projects preference keywords into each prompt embedding so the output is shifted toward reference-image attributes without losing the original prompt's meaning; and local cross-attention modulation, which renders each entity within its planned bounding box by replacing ba","pith_inferences":["Reading beyond the paper, the ceiling of this framework likely tracks the MLLM's ability to see and verbalize a preference; replacing the MLLM with a stronger one should raise alignment without changing the guidance equations.","The four-category taxonomy is probably a pragmatic simplification; a natural stress test is free-form preference elicitation, since the paper's own coverage table gives users only these categories to choose from, making 'Others' at 4.3% partly an artifact of the question format.","The orthogonal-projection trick could be reused as a generic semantic-preservation module for any conditioning signal that lives in text space, independent of image preferences.","A cleaner controlled comparison—feeding the same enriched prompt to a plain rewrite baseline versus the full pipeline—would isolate how much of the gain comes from global keyword guidance versus local cross-attention modulation."],"forward_implications":["A single reference image, not a per-user dataset, becomes the preference specification, so preference changes take effect in one generation step.","The framework can serve as an interactive image editor: users refine keywords, prompts, entities, or boxes across rounds, and each round amounts to inference-time control.","Components are swappable: the paper shows the same pipeline with different MLLMs and different diffusion backbones, so gains transfer rather than being locked to one model.","Because guidance lives in text-embedding and attention space, it sidesteps RLHF/DPO-style retraining and coarse binary feedback.","Prompt fidelity and preference alignment are treated as separable axes: the orthogonal projection preserves base-prompt meaning while preference strength is controlled by α."],"supporting_citations":[{"why":"Supplies the main prior baseline and the Viper dataset used for generalization evaluation; the paper positions itself against Viper's static, expert-collected preference definitions.","marker":"Salehi et al. 2024"},{"why":"Provides the regional planning and complementary regional diffusion idea that the local cross-attention modulation extends; the paper compares against it on T2I-CompBench.","marker":"Yang et al. 2024"},{"why":"Defines the Stable Diffusion latent diffusion formulation that the method builds on.","marker":"Rombach et al. 2022"},{"why":"The default diffusion backbone used in all main experiments.","marker":"Podell et al. 2023"},{"why":"Provides the CLIP text encoder used in the embedding-space preference guidance equations.","marker":"Radford et al. 2021"},{"why":"An LLM-based prompt-rewrite baseline; the paper's prompt enrichment extends this to structured multimodal preference categories.","marker":"Chen et al. 2024"},{"why":"A training-based style-transfer baseline; comparison shows that style-only control is insufficient for multifaceted preference alignment.","marker":"Wang et al. 2024"},{"why":"An iterative-feedback baseline; comparison highlights the need for a richer preference signal than binary liked-or-disliked images.","marker":"V on Rütte et al. 2023"},{"why":"Used as the MLLM judge for evaluating thematic and visual element alignment.","marker":"Hurst et al. 2024"}],"fun_headline_variants":["No retraining: one reference image instantly steers T2I output","One image, instant style: MLLM guides diffusion without training","Real-time preference alignment for T2I, no fine-tuning needed","From one reference image to aligned generations: zero training"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole pipeline rests on the premise that a user's preference is adequately captured by MLLM-generated keywords organized into four fixed categories extracted from a single reference image—and that the evaluation's category menu does not conceal preferences outside that taxonomy.","fun_headline_variants_meta":{"raw":{"variants":["No retraining: one reference image instantly steers T2I output","One image, instant style: MLLM guides diffusion without training","Real-time preference alignment for T2I, no fine-tuning needed","From one reference image to aligned generations: zero training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000671,"raw_usage":{"total_tokens":2896,"prompt_tokens":745,"completion_tokens":2151,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":2078}},"tokens_in":489,"tokens_out":2151,"duration_ms":16288,"temperature":1.0,"reasoning_tokens":2078,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:44:55.160571+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a preference-elicitation study with fully open-ended user responses (no category menu) and check whether the MLLM keywords and generated images align with the stated reasons; if a substantial fraction of preferences fall outside the four categories or the keywords miss them, the coverage claim is falsified.","supporting_citations":[],"review_version":1}