{"id":"315f97bd-dde4-4efd-83d6-bb34a0e42e29","arxiv_id":"2603.12506","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"A lightweight predictor ranks initial noises by expected human-preference score for a prompt, selecting the best few for diffusion generation and reporting prompt difficulty.","lead":"Naïve PAINE predicts a diffusion model's image preference score from the prompt embedding and initial Gaussian noise alone, then keeps only the top-scoring noises for full generation. It also estimates how hard a prompt is for a given model, cutting wasted sampling while remaining plug-and-play.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Top-K selection may not systematically land in the true upper tail once preference-label noise and offline training distribution are taken into account.","rationale":"The reader already isolated the correct weakest assumption: that offline preference regression yields sufficiently accurate ranking of unseen noises. The manuscript supplies SRCC/MAE numbers and competitive tables, but never quantifies how much of the true upper tail is actually recovered, nor reports uncertainty on the metric deltas. The concrete test above directly measures that recovery gap; a large gap would keep the verdict CONDITIONAL (or push it toward REJECT), while a small gap would justify moving toward ACCEPT. No stronger internal inconsistency is present, so the reader’s CONDITIONAL verdict is left unchanged.","tokens_in":23837,"tokens_out":510,"duration_ms":4857,"concrete_test":"On the held-out Pick-a-Pic test prompts for each of the four DMs, generate the full set of N=100 candidate noises, compute both the PAINE-predicted ranking and the true post-generation PickScore ranking, then measure (i) the fraction of true top-K that appear in PAINE’s top-K and (ii) the mean true-score gap between PAINE top-K and an oracle top-K. If either quantity is statistically indistinguishable from a random-K baseline (bootstrap 95 % CI), the load-bearing ranking assumption fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on the premise (Sec. 3.1, Eq. 4, training protocol 4.1) that a scalar regressor trained offline on PickScore (or similar) labels of fully-generated images will rank unseen initial noises accurately enough that the top-K selected noises systematically occupy the upper tail of the true preference distribution for new prompts. Validation SRCC values of 0.74–0.87 (Table 6) and MAE < 0.84 are reported, yet these are still imperfect rank correlations; human-preference labels themselves are known to be noisy (cited by the authors). Consequently the selected top-K may be only modestly better than random draws from the same prompt-conditioned distribution, and the reported metric gains (Table 1) could be inflated by the absence of error bars, multiple-comparison correction, or an explicit “oracle top-K vs. PAINE top-K” gap measurement. If that gap is large, the practical value of the lightweight predictor collapses.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Naïve PAINE, a lightweight plug-and-play predictor that estimates human-preference scores (primarily PickScore) of a would-be T2I image from the prompt embedding and the initial Gaussian noise latent, before reverse diffusion. Candidate noises are ranked and the top-N_img are forwarded for full generation; zeroing the noise encoder yields a prompt-only mean-score estimate framed as a naïve-Bayes prior. The method is trained offline on 5k Pick-a-Pic prompts × 20 noises per model (Hunyuan-DiT, PixArt-α, DreamShaper-XL, SDXL) with MAE + differentiable SRCC loss, and is evaluated against Golden Noise, NoiseAR and HyperNoise on HPSv2/v3, ImageReward, PickScore and GenEval, with latency and ablation studies.","tokens_in":24159,"tokens_out":981,"duration_ms":7900,"significance":"If the ranking premise holds, the work supplies a practical, model-agnostic, fine-tuning-free alternative to noise mutation or RL-based optimizers that is easy to drop into Diffusers/ComfyUI pipelines and that additionally returns an interpretable prompt-difficulty signal. Multi-model, multi-benchmark tables (Tables 1–2, 7–9), hardware measurements (Tables 3–4), and ablations on target metric, loss, K and text-encoder design constitute a solid empirical package; code is released. The contribution is incremental but useful for the large community that still runs multi-sample generation on open DMs.","major_comments":[{"comment":"§3.1 Eq. (4) and §4.1 training protocol rest on the claim that offline PickScore (or similar) labels of fully generated images produce rankings accurate enough for top-K noises to systematically occupy the upper tail of the true preference distribution on new prompts. Validation SRCC of 0.74–0.87 (Table 6) is only moderate; human-preference labels are themselves noisy (authors’ own citations). The manuscript never reports an oracle gap (true top-K by post-generation score vs. PAINE-selected top-K) or error bars / significance tests on the metric deltas in Table 1. Without that measurement it is hard to know how much of the reported gains survive label noise and distribution shift.","section":null},{"comment":"§4.2 / Table 1: all preference-metric gains are point estimates. Given the known variance of HPSv2/v3, ImageReward and PickScore across seeds and the multiple-comparison setting (4 models × 4 prompt sets × 4 metrics), the absence of standard errors or a simple paired test leaves open the possibility that several “best” entries are not reliably better than Golden Noise or the standard baseline. Adding these would make the central claim falsifiable rather than merely directionally consistent.","section":null}],"minor_comments":[{"comment":"§2.3 / Figs. 1–3: the preliminary distribution study is informative but uses only 50 prompts and 20 seeds; a short note on how sensitive the PCC matrices are to prompt sampling would strengthen the motivation.","section":null},{"comment":"§3.2: the “naïve Bayes” framing is heuristic (zeroing the noise encoder is not a true likelihood). Soften the language or move the analogy to discussion so it is not read as a formal derivation.","section":null},{"comment":"Table 3 vs. Table 4: parameter counts favor Golden Noise while latency favors PAINE; a single combined table of wall-clock cost for N_img = 1 and N_img = 4 would make the efficiency claim clearer.","section":null},{"comment":"Supplementary Table 6 reports validation SRCC/MAE; moving a one-line summary into the main text (near §4.1) would help readers assess predictor quality without leaving the paper.","section":null},{"comment":"Scattered OCR/encoding artifacts (e.g., “Na\\\"ive”, “PixArt-�”, “�����”) should be cleaned for the camera-ready version.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The work is a solid systems/engineering contribution rather than a conceptual breakthrough; it is appropriate for a CV venue that values practical T2I tooling. The two major points (oracle gap + statistical significance) are addressable with additional experiments already within the authors’ pipeline and do not require new theory. I would not block acceptance once those are supplied."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful bit is simple and well-executed. They treat initial-noise selection as scalar regression: feed the DM’s own prompt embedding plus a batch of Gaussian latents into a small predictor (token-pooled transformer + ResNet noise encoder + MLP), rank by predicted PickScore, and only run the expensive reverse process on the top-K. Masking the noise encoder also gives a prompt-only mean-score estimate they frame as a naïve-Bayes prior. That dual use is the actual novelty relative to Golden Noise’s mutation network and the heavier RL fine-tunes (NoiseAR, HyperNoise).\n\nWhat they do well: the Section 2.3 diagnostics are honest and informative—prompt dominates the preference distribution, optimal noise is prompt-specific, and different DMs excel on different prompt subsets. Tables 1–2 and the supplements show consistent gains over Golden Noise across four backbones (and competitive numbers vs NoiseAR) on HPSv2/v3, ImageReward, PickScore, and GenEval; the SANA-Sprint comparison is a nice extra. Latency and checkpoint numbers favor them even at K=100. Ablations on target metric, loss (MAE+SRCC), K, and text-encoder design are present. Qualitative examples (hands, character fidelity) match the numbers. Code is promised.\n\nSoft spots, in proportion: preference labels are noisy (they cite this themselves) and validation SRCC of 0.74–0.87 is only decent, so the stress-test worry that top-K may not systematically hit the true upper tail is real. They never report an oracle-top-K gap or error bars, which would have tightened the claim. Gains are metric-driven rather than human-study-driven. None of that sinks the central result; the multi-model consistency still holds.\n\nThis is for people who ship or evaluate T2I pipelines and care about sampling efficiency without fine-tuning. Not a new scientific paradigm, but a practical, reproducible plug-in that deserves a serious referee. I would engage with it and expect it to clear peer review with ordinary requests for stats and a clearer oracle comparison.","headline":"Clean engineering win: regress preference from noise+prompt, rank top-K, and get free prompt-difficulty feedback without touching the DM.","tokens_in":24709,"tokens_out":529,"would_cite":true,"duration_ms":10708,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A lightweight predictor ranks initial noises from the prompt alone so diffusion models generate preferred images without re-rolling the dice.","keywords":["Diffusion Models","Text-to-Image","Initial Noise Optimization","Human Preference Prediction","PickScore","Prompt Evaluation","Lightweight Inference"],"falsifier":"Hold out a fresh prompt set and a new diffusion model; generate many images from random noises, score them with the same preference metric used in training, then check whether the noises PAINE ranks highest actually produce higher measured scores than random or Golden-Noise baselines at the same compute budget.","tokens_in":24775,"feed_emoji":"🎨","tokens_out":821,"duration_ms":11095,"temperature":0.7,"pith_summary":"Diffusion text-to-image models start from random Gaussian noise, so the same prompt yields different quality each run and users must regenerate until they like the result. This paper argues that image quality is not pure chance: the prompt largely sets the score distribution, while the initial noise only picks where on that distribution a sample lands. Naïve PAINE therefore trains a small network to predict a human-preference score directly from the prompt embedding and the candidate noise tensor, before any reverse diffusion runs. It samples many noises, keeps the top few, and only those go through the expensive generator. The same network, with the noise branch masked, also estimates the average quality a model can achieve for that prompt, giving the user feedback before generation. Across four diffusion models and several preference benchmarks, the method beats or matches other lightweight noise-selection baselines while adding only milliseconds of latency.","feed_headline":"Rank the noise first, then generate the image","feed_subtitle":"A small predictor picks better starting seeds so diffusion models waste fewer rolls","key_machinery":"Naïve PAINE (Prompt-Aware Initial Noise Evaluator): a three-module predictor (prompt encoder with a learnable summary token, ResNet noise encoder, MLP score head) that estimates preference from (prompt embedding, initial noise). Masking the noise branch yields a Naïve-Bayes-style prior on the prompt-conditioned mean score.","core_discovery":"The authors show that a scalar human-preference score for a would-be image can be predicted well enough from only the prompt encoding and the initial noise tensor that ranking many candidate noises and forwarding the top-K systematically raises HPSv2/v3, ImageReward, and PickScore relative to unoptimized sampling and to Golden Noise, without fine-tuning the diffusion model itself.","pith_inferences":["If the ranking signal generalizes across preference metrics, practitioners could train once on the cheapest metric and still improve others, lowering annotation cost.","The same noise-ranking idea should transfer to text-to-video or autoregressive image models whose quality also depends on a stochastic seed.","When the predictor’s mean-score feedback is low, interactive prompt rewriting guided by that score may raise final quality more than noise selection alone."],"forward_implications":["Users can reduce the number of full generation cycles needed to obtain a satisfactory image for a given prompt.","Existing Diffusers or ComfyUI pipelines can insert the predictor as a pre-generation filter with only milliseconds of added latency.","Prompt-only mode supplies an interpretable difficulty signal so a user can rewrite a prompt before spending GPU time.","The same selection idea applies without architecture changes to both older U-Net models and newer Diffusion Transformers."],"fun_headline_variants":["Predict image quality from noise before you generate","Rank noises by predicted score then run diffusion","Lightweight scorer picks better seeds for T2I models","Score initial noise first to raise preference metrics","Select top-K noises via prompt to cut bad rolls"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"A predictor trained offline on preference scores of fully generated images will rank never-before-seen noises accurately enough that the top-K choices land in the upper tail of the true preference distribution for new prompts.","fun_headline_variants_meta":{"raw":{"variants":["Predict image quality from noise before you generate","Rank noises by predicted score then run diffusion","Lightweight scorer picks better seeds for T2I models","Score initial noise first to raise preference metrics","Select top-K noises via prompt to cut bad rolls"]},"model":"grok-4.5","effort":"low","cost_usd":0.003204,"raw_usage":{"total_tokens":1090,"prompt_tokens":738,"num_sources_used":0,"completion_tokens":55,"cost_in_usd_ticks":32040000,"prompt_tokens_details":{"text_tokens":738,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":297,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":738,"tokens_out":55,"duration_ms":3303,"temperature":1.0,"reasoning_tokens":297,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T22:22:14.531743+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Hold out a fresh prompt set and a new diffusion model; generate many images from random noises, score them with the same preference metric used in training, then check whether the noises PAINE ranks highest actually produce higher measured scores than random or Golden-Noise baselines at the same compute budget.","supporting_citations":[],"review_version":1}