Pith. sign in

REVIEW 5 major objections 5 minor 7 cited by

HPSv3: Towards Wide-Spectrum Human Preference Score

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims HPSv3 is the most reliable automated proxy for human preference across a wide spectrum of generated and real images, reporting a Spearman correlation of 0.94 with human rankings and state-of-the-art prediction accuracy on…

desk verdict HPDv3 is a genuinely useful dataset and HPSv3 is a plausible step forward, but the headline correlation is not yet supported because the human ground truth and test splits are under-described. read the letter →

arxiv 2508.03789 v2 pith:WSTCFYEN submitted 2025-08-05 cs.CV

classification cs.CV
keywords humanpreferencescoretext-to-imageevaluationpairwisecomparisondatasetuncertainty-awarerankinglossvision-languagemodelchain-of-human-preferenceimagegenerationrefinementHPDv3
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Human preference is the yardstick for text-to-image generation, but automated metrics have been trained on a narrow slice of data: mostly diffusion-model outputs without high-quality real photographs. This paper sets out to close that gap with HPSv3, a scoring model trained on HPDv3, a newly released dataset of 1.08 million text-image pairs and 1.17 million pairwise human comparisons spanning 16 generative models and low-to-high-quality real images. The model uses a vision-language backbone and an uncertainty-aware ranking loss that treats each image's score as a distribution rather than a fixed number. The paper reports that HPSv3 aligns with human ranking better than prior metrics (Spearman $r = 0.94$) and sets state-of-the-art accuracy on the PickScore, HPDv2, and HPDv3 test sets. If correct, this gives the field a reliable automated proxy for human taste and a reward signal that can refine generation quality without retraining, via the proposed Chain-of-Human-Preference (CoHP) loop.

What carries the argument

Three components carry the argument. First, the HPDv3 dataset: a wide-spectrum collection whose 12-category prompt distribution follows JourneyDB, which pairs outputs of 16 generative models with high-quality real photographs and annotates each pair with 9 to 19 expert judgments, training on pairs whose agreement exceeds 95%. Second, the uncertainty-aware ranking loss: instead of a point score, the model outputs $\mu$ and $\sigma$ for each image, treats the score as $r \sim \mathcal{N}(\mu, \sigma)$, and optimizes the negative log-likelihood of the integrated pairwise preference $P(x_1 \succ x_2 \mid c) = \int \int \mathrm{sigmoid}(r_1 - r_2)\,\mathcal{N}(r_1\mid\mu_1,\sigma_1)\,\mathcal{N}(r_2\mid\mu_2,\sigma_2)\,dr_1\,dr_2$. Third, Chain-of-Human-Preference (CoHP): a two-stage iterative loop that first selects the best base model for a prompt by average HPSv3 score over several rounds, then repeatedly adds noise to the current best image and re-samples with HPSv3 picking the best candidate, refining quality without any additional training data.

What would settle it

On the released HPDv3 dataset, compute the fraction of the 1.17 million comparisons with inter-annotator agreement above 95%, then evaluate HPSv3 on the held-out low-agreement pairs (or retrain without the filter). If accuracy on those ambiguous, hard pairs falls well below the reported 76.9% on the HPDv3 benchmark, the wide-spectrum claim is an artifact of filtering to easy, high-agreement pairs.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a preference model becomes wide-spectrum when its training data spans the full quality range and its loss respects annotation uncertainty. HPDv3 adds high-fidelity real photographs as an upper quality bound and outputs from state-of-the-art diffusion, DiT, GAN, and autoregressive models, with every pair judged by 9 to 19 annotators and pairs kept for training only when inter-annotator agreement exceeds 95%. HPSv3 then predicts a Gaussian score $r \sim \mathcal{N}(\mu, \sigma)$ for each image and defines the probability that image 1 beats image 2 as the double integral of $\mathrm{sigmoid}(r_1 - r_2)$ over both score distributions. The paper reports this yields Spearman $r = 0.94$ and Kendall $\tau = 0.8222$ against human model rankings and accuracy of 72.8%, 85.4%, and 76.9% on the PickScore, HPDv2, and HPDv3 test sets, with a much smaller performance drop than previous metrics when moving to the harder HPDv3 test set. That is the load-bearing claim: a wide-spectrum dataset plus uncertainty-aware ranking makes HPSv3 both a better evaluator and a better reward model.

Load-bearing premise

The load-bearing premise is that the pairs kept for training, those on which annotators agreed more than 95% of the time, represent the same preferences as the full pool of 1.17 million comparisons, so a model trained only on easy-to-agree pairs still measures wide-spectrum preference.

Editorial extensions

If this is right

  • HPSv3 can be used as a drop-in automated evaluator that ranks current and future text-to-image models across the full quality spectrum, including comparisons against real photographs.
  • CoHP improves generation quality without extra training data or fine-tuning, by selecting the best model per prompt and then iteratively refining the best image with HPSv3 as reward.
  • Because high-quality real photos anchor one end of HPDv3, model rankings produced by HPSv3 have an explicit upper-quality reference point that earlier datasets lacked.
  • When used as a reward model in reinforcement-learning fine-tuning, HPSv3 produces fewer reward-hacking artifacts than ImageReward, PickScore, or HPSv2, per the paper's DanceGRPO experiments.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper never reports what fraction of the 1.17 million comparisons survives the 95% agreement filter; if the surviving training set is small, HPSv3's high scores may reflect easy, high-consensus pairs rather than the full preference distribution.
  • The model's learned uncertainty parameter $\sigma$ could be repurposed as a confidence signal to flag ambiguous image pairs for human review or active learning, a use the paper does not explore.
  • If HPSv3 generalizes as claimed, it should transfer to quality assessment outside text-to-image, such as ranking real photographs or selecting images for curation, which the paper does not test.
  • The CoHP model-wise stage chooses among a fixed pool of base models; the same mechanism could in principle select among models released after HPSv3's training if its ranking remains valid, which would require an out-of-distribution test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper introduces HPSv3, a VLM-based human preference score for text-to-image evaluation, together with HPDv3, a large pairwise preference dataset (claimed 1.08M text-image pairs and 1.17M annotated comparisons) covering real high-quality images, old and recent generative models, and Midjourney user choices. The model uses a Qwen2-VL backbone and an uncertainty-aware ranking loss. The paper also proposes CoHP, a training-free iterative image refinement method that selects the best model and the best sample at each step using HPSv3 scores. Claims include state-of-the-art preference accuracy on PickScore/HPDv2/HPDv3 test sets and the highest model-level correlation with human preference rankings (Spearman r = 0.94, Kendall tau = 0.8222).

Significance. If the central claims hold, HPSv3 would be a valuable wide-spectrum preference metric and HPDv3 a substantial community resource: it is the largest dataset in the paper's comparison, it uniquely includes high- and low-quality real images alongside outputs of GAN, diffusion, DiT, and autoregressive models, and the CoHP refinement scheme is an interesting zero-training alternative to RL-based alignment. However, the evaluation pieces that carry the headline claims—Table 3's human preference rankings, Table 4's HPDv3 test accuracy, and Table 6's CoHP round ablation—are currently insufficiently documented or internally circular. These are load-bearing issues, not cosmetic ones, because the paper's main conclusions ('most reliable metric', 'state-of-the-art accuracy', 'CoHP improves generation quality') rest directly on them. The dataset and model design are plausible and worth publishing, but the manuscript needs substantial revision to substantiate the evaluation.

major comments (5)
  1. [§5.1.3, Table 3, Figure 4] The provenance of the 'human preference rankings' used to compute Spearman r = 0.94 and Kendall tau = 0.8222 is never stated. The HPDv3 Benchmark described in §5.1.2 is a set of 132,000 generated images; the paper does not say whether these images were annotated by humans, how the model-level human ranking was aggregated, or how many prompts/pairs underlie each of the 11 model scores. Without this information, the headline correlation cannot be interpreted, and if the rankings were derived from HPDv3 training annotations or from the same prompts used to train HPSv3, the number is in-distribution fit rather than evidence of human alignment. Please specify the exact human-annotation procedure for the benchmark, the aggregation rule, and the overlap (at the prompt and image level) between the human-ranked data and HPSv3's training data.
  2. [§5.1.2, Table 4, Table S1] The HPDv3 testset is not defined as a held-out split. Table S1 labels several image sources as 'Train & Test' without giving a split criterion, and the main text reports 76.9% accuracy on 'the HPDv3 testset' without stating whether the test pairs, prompts, or images are disjoint from the pairs used to train HPSv3. Because HPSv3 is trained on HPDv3 pairs from the same prompts and models, any prompt- or image-level overlap will inflate the reported accuracy. The same issue applies to the PickScore and ImageReward columns: the training mix in Table S4 randomly selects 350k and 120k pairs, and the paper never states that these subsets are disjoint from the official test sets. Please provide an explicit pair-level/prompt-level split, a statement of exclusion, and re-reported numbers on truly held-out partitions.
  3. [§3.2, §3.3, Supplementary Limitation] The paper trains only on image pairs with inter-annotator confidence exceeding 95%, but it reports an average annotation convergence of 76.5% for HPDv3 and later admits in the supplementary Limitation that annotators 'often struggled to converge on a consistent score' for difficult cases. The fraction of the 1.17M comparisons that pass the 95% threshold is never reported, nor is any analysis showing that the high-confidence subset is representative of the full preference distribution. If the filtered subset is small or biased toward easy-to-agree aesthetic differences, HPSv3's claimed 'wide-spectrum' coverage and its robustness on hard cases are unsupported. Please report the confidence-score distribution, the retained fraction, and a comparison of model performance on high-confidence versus low-confidence held-out pairs.
  4. [§5.2.5, Table 6] The quantitative CoHP round ablation measures improvement using HPSv3 scores while HPSv3 is also the model used to select the best sample at each round (Eqs. (5)–(6)). The reported monotonic gains in Table 6 are therefore partly self-confirming: a reward model that selects the highest-scoring image in each round will trivially increase its own score, and this does not by itself demonstrate alignment with human quality judgments. The 100-pair user study in §5.2.4 compares CoHP-HPSv3 against other preference models but does not test round count. Please add a human evaluation of the round-number curves, or at least an evaluation with an independent reward model, and report whether the CoHP improvements translate to human-preferred outputs.
  5. [§3.2, §5.1.1, Table S4] There is a quantitative inconsistency in the reported training data size: §3.2 states 'we gathered 1.17 million high-confidence pairwise comparisons for training,' §5.1.1 says 'We train HPSv3 using 1.5 million well-annotated pairwise samples,' and Table S4 sums to 1,522k pairs that includes 652k HPDv3 pairs plus a 250k golden set, 350k Pick-A-Pic, 120k ImageReward, and 150k Midjourney pairs. These numbers cannot all be correct under the same definition of 'training pairs.' Please reconcile them and specify which pairs constitute the 1.17M HPDv3 comparisons versus the full 1.5M/1.522M training mix.
minor comments (5)
  1. [Equation (3)] The integral defining P(x1 ≻ x2|c) is not evaluated or approximated; please state whether it is computed with a closed-form expression, numerical integration, or Monte Carlo sampling, since this affects both training efficiency and reproducibility.
  2. [Table 3] 'Normalized MSE' is not defined in the main text or the caption; please specify the normalization and the units used in its calculation.
  3. [§5.2.1] The denoising strength schedule (0.8 for the first two rounds and 0.5 for the remaining rounds) is stated but not motivated; a sentence explaining why the schedule changes would help readers assess the sensitivity of CoHP to this hyperparameter.
  4. [Throughout] There are several typographical issues, including 'Howerver' in §4.1, 'as shown in Table 5' with 'shwon' in §5.1.4, and 'candicate' in the Figure 5 caption; these should be corrected in a final pass.
  5. [Table 2] The benchmark scores in Table 2 are presented as point estimates without confidence intervals or significance tests; given that they are averaged over 1,000 prompts per category, reporting standard errors would help readers judge whether the model ordering is stable.

Circularity Check

2 steps flagged · score 6.0 of 10

CoHP's quantitative improvement is self-confirming because the same HPSv3 score is used both as the selection objective and as the reported outcome; additionally, the headline HPDv3 testset accuracy is reported on the same corpus admitted into training, with no held-out split demonstrated.

  1. self definitional [Sec. 4.2 (Eqs. 5-6) and Sec. 5.2.5 (Table 6)]
    "In each stage, we employ HPSv3 as the reward model to evaluate generated images and guide the iterative selection of the best candidates, improving generation quality over multiple rounds. ... We calculate the HPSv3 score with different rounds of the preference stages, opting to run each stage for 4 rounds."

    CoHP selects the golden model and the final image by argmax over HPSv3 scores (Eqs. 5 and 6), and the quantitative evidence in Table 6 is an increase in that same HPSv3 score across rounds. By construction, the maximum of a batch of scored candidates cannot decrease relative to the average of the batch, so iterative selection by the reward model guarantees non-decreasing selected scores even if no true human preference is captured. Table 6 therefore demonstrates that CoHP optimizes HPSv3, not that it independently improves quality; the only non-circular evidence for human-aligned improvement is the separate user study in Sec. 5.2.4.

  2. fitted input called prediction [Sec. 5.1.3 (Table 4) vs. Sec. 5.1 and Table S4]
    "As presented in Table 4, HPSv3 demonstrates outstanding performance, achieving state-of-the-art accuracy scores of 72.8%, 85.4%, and 76.9% on the PickScore, HPDv2, and HPDv3 testsets, respectively."

    The paper's own Table S4 lists 'HPDv3 Real images and comparisons 652k' as part of the HPSv3 training dataset, and Table S1 marks the same HPDv3 sources as 'Train & Test' without defining any held-out partition. Reporting accuracy on an 'HPDv3 testset' drawn from the same corpus used for training, with no demonstrated exclusion, presents in-distribution fit as cross-dataset prediction. The HPDv2 result is similarly burdened because the HPDv2 training subset was curated by a filtering model that 'shares the same architecture and training methodology as HPSv3,' so the HPDv2 evaluation is not clearly independent of the training-data selection process.

full rationale

Two concrete reductions are visible from the paper's own text. First, CoHP's round-ablation evidence is circular: the same HPSv3 score is both the selection objective (Eqs. 5-6) and the reported outcome (Table 6), so the reported increase is a mathematical consequence of max-selection rather than independent evidence of quality improvement. Second, the headline HPDv3 testset numbers are reported on the same corpus admitted into training (Table S4), with no split described, so the 76.9% accuracy figure is at least partly an in-distribution fit. The r=0.94 correlation in Table 3 is a related transparency gap because the provenance of the 'human preference rankings' is never stated; I did not count it as a separately proven circular step, but it should be resolved before the central metric claim is accepted. The qualitative user study (Sec. 5.2.4) and the evaluations on PickScore and ImageReward are external and non-circular, providing some independent support for CoHP and HPSv3; the paper is therefore not wholly self-referential. Score 6 reflects the demonstrated partial self-confirmation in the main quantitative evidence, not complete equivalence of the derivation.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the dataset pipeline's filtering choices (95% confidence, aesthetic top 10%, category alignment) and on the compatibility of preference labels from several collection protocols. The main free parameters are thresholds and round counts selected by the authors rather than derived. The axioms listed capture assumptions that, if violated, would weaken the wide-spectrum generalization claim.

free parameters (5)
  • training pair confidence threshold = 0.95
    Pairs with inter-annotator confidence above 0.95 are used for HPSv3 training; threshold chosen without sensitivity analysis, while average convergence of the full dataset is 76.5%.
  • aesthetic filter percentile = top 10% per category
    Real images are filtered by a retrained aesthetic predictor, keeping the top 10% within each category; this choice controls the real-image subset size and may bias categories differently.
  • model-wise round count = 4
    The number of rounds for model selection in CoHP is chosen from the Table 6 ablation, where gains plateau after round 4.
  • sample-wise round count = 4
    The number of refinement rounds is chosen from the Table 6 ablation, where HPSv3 scores improve until round 4 and then plateau.
  • denoising strength schedule = 0.8 then 0.5
    Sample-wise refinement uses denoising strength 0.8 for the first two rounds and 0.5 for the remaining rounds, set manually in §5.2.1 without a reported sweep.
assumptions (6)
  • domain assumption Human preference can be represented as a single scalar and aggregated from pairwise choices across annotators.
    Annotation guidelines combine prompt alignment, aesthetics, and coherence into one choice (§4.3); no evidence is given that these dimensions are commensurable.
  • domain assumption The 95% confidence-filtered subset of HPDv3 is representative of the full preference distribution.
    Training uses only pairs with confidence above 0.95 (§3.2), while average convergence is 76.5% (§3.3); no coverage analysis of the filtered subset is provided.
  • domain assumption Preferences from different sources (professional annotators, Midjourney Discord users, HPDv2, Pick-a-Pic, ImageReward) are compatible and can be pooled.
    Training mixes 1.5M pairs from these sources (§5.1, Table S4) without modeling source-specific biases.
  • domain assumption Qwen2-VL-7B, after full fine-tuning on the pairwise loss, encodes preference-relevant information from text and image.
    Backbone choice is justified by general VLM capability (§4.1), not by a preference-specific analysis beyond the reported ablations.
  • ad hoc to paper In CoHP, re-noising the selected best image and re-generating with the same prompt samples from a distribution where HPSv3 score improvements track human quality improvements.
    The sample-wise stage uses denoising strengths 0.8 and 0.5 (§5.2.1); this is a specific procedure with no formal guarantee beyond the reported experiments.
  • domain assumption Per-image scores are independent Gaussians in the uncertainty-aware ranking objective.
    Equation 3 integrates a product of independent Gaussian densities; no justification is given for ignoring correlation within a pair.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HPSv3: Towards Wide-Spectrum Human Preference Score." pith.science (2026). https://pith.science/paper/WSTCFYEN

@misc{pith2026250803789,
  author       = {Pith},
  title        = {Pith review of: HPSv3: Towards Wide-Spectrum Human Preference Score},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WSTCFYEN}},
  note         = {Machine review of arXiv:2508.03789}
}
read the original abstract

Evaluating text-to-image generation models requires alignment with human perception, yet existing human-centric metrics are constrained by limited data coverage, suboptimal feature extraction, and inefficient loss functions. To address these challenges, we introduce Human Preference Score v3 (HPSv3). (1) We release HPDv3, the first wide-spectrum human preference dataset integrating 1.08M text-image pairs and 1.17M annotated pairwise comparisons from state-of-the-art generative models and low to high-quality real-world images. (2) We introduce a VLM-based preference model trained using an uncertainty-aware ranking loss for fine-grained ranking. Besides, we propose Chain-of-Human-Preference (CoHP), an iterative image refinement method that enhances quality without extra data, using HPSv3 to select the best image at each step. Extensive experiments demonstrate that HPSv3 serves as a robust metric for wide-spectrum image evaluation, and CoHP offers an efficient and human-aligned approach to improve image generation quality. The code and dataset are available at the HPSv3 Homepage.

Discussion (0). Sign in to comment.

Forward citations

Cited by 7 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Latent Reward Registers for Diffusion Preference Alignment

    cs.LG 2026-08 conditional novelty 6.0 of 10

    Prepending learnable register tokens to a frozen DiT yields dense latent reward estimates, enabling faster on-policy distillation and training-free guided sampling that improve preference alignment.

  2. Learning Sampling Parameters for Diffusion Models

    cs.LG 2026-07 conditional novelty 6.0 of 10

    An LLM policy trained with GRPO can emit prompt-conditioned, timestep-varying diffusion sampling parameters that beat fixed defaults and prior LLM schedulers on preference metrics.

  3. Z-Reward: Beyond Scalar Rewards by Internalizing Reasoning into Score Distributions

    cs.CV 2026-06 unverdicted novelty 6.0 of 10

    Z-Reward trains a 27B reasoning teacher VLM on score distributions via GDSO and distills it via RISD into a 9B student, reaching 89.6% and 88.6% human preference accuracy with 41.3% optimization gain over SFT baseline.

  4. Directly Aligning the Full Diffusion Trajectory with Fine-Grained Human Preference

    cs.AI 2025-09 conditional novelty 6.0 of 10

    Direct-Align and SRPO fine-tune FLUX using ground-truth-noise recovery and text-conditional relative rewards, improving human-evaluated realism and aesthetics roughly 3x.

  5. Visko Orbis 1.0: A Live Model for Real-Time Interactive Long Video Generation

    cs.CV 2026-07 conditional novelty 5.0 of 10

    A chunk-wise streaming video model with bounded multi-scale memory and streaming 4K upscaling reports real-time interactive long-form generation and top Arena preference/stability scores.

  6. FAIL: Flow Matching Adversarial Imitation Learning for Image Generation

    cs.CV 2026-02 conditional novelty 5.0 of 10

    Post-training of flow matching can be framed as adversarial imitation learning, and the proposed FAIL methods improve FLUX's generation quality using 13K expert images without preference pairs.

  7. RewardDance: Reward Scaling in Visual Generation

    cs.CV 2025-09 conditional novelty 5.0 of 10

    RewardDance reframes visual reward modeling as a yes/no judgment task in a VLM and reports consistent gains in text-to-image, text-to-video, and image-to-video generation as the reward model scales from 1B to 26B.

Reference graph

Works this paper leans on

60 extracted references · 45 canonical work pages · cited by 7 Pith papers

  1. [1]

    Flux.1 [dev]: 12b-parameter open-source text-to-image diffusion model, 2024

    Black Forest Lab. Flux.1 [dev]: 12b-parameter open-source text-to-image diffusion model, 2024. Architecture: 12B- parameter diffusion model with T5-XXL text encoder; Fea- tures: Anatomically precise human generation, native mul- tilingual support, commercial-grade text rendering; License: Non-commercial research use (Apache 2.0 with additional restriction...

  2. [2]

    Rank analysis of incomplete block designs: I

    Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired compar- isons. Biometrika, 39(3/4):324–345, 1952. 4, 5

  3. [3]

    Pixart- α: Fast training of diffusion transformer for photorealistic text-to-image synthesis

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart- α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2

  4. [4]

    Pixart- σ: Weak-to-strong training of dif- fusion transformer for 4k text-to-image generation, 2024

    Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- σ: Weak-to-strong training of dif- fusion transformer for 4k text-to-image generation, 2024. 6, 2

  5. [5]

    Lawrence Zitnick

    Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedan- tam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO Captions: Data Collection and Evaluation Server, 2015. 4

  6. [6]

    Cogview2: Faster and better text-to-image generation via hi- erarchical transformers

    Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hi- erarchical transformers. NeurIPS, 35:16890–16902, 2022. 2

  7. [7]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873–12883, 2021. 2

  8. [8]

    Scaling rectified flow trans- formers for high-resolution image synthesis, 2024

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yan- nik Marek, and Robin Rombach. Scaling rectified flow trans- formers for high-resolution image synthesis, 2024. 2, 3, 6, 7

Show all 60 references
  1. [9]

    Fluid: Scaling autoregressive text-to-image generative models with continuous tokens

    Lijie Fan, Tianhong Li, Siyang Qin, Yuanzhen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024. 2

  2. [10]

    Infinity: Scaling bit- wise autoregressive modeling for high-resolution image syn- thesis

    Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bit- wise autoregressive modeling for high-resolution image syn- thesis. arXiv preprint arXiv:2412.04431, 2024. 2, 3, 6

  3. [11]

    Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Ma- tiana, Joe Penna, and Omer Levy. Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation. arXiv preprint arXiv:2305.01569, 2023. 1, 2, 3, 7, 4

  4. [12]

    Instantportrait: One-step portrait editing via diffusion multi-objective distillation

    Zhixin Lai, Keqiang Sun, Fu-Yun Wang, Dhritiman Sagar, and Erli Ding. Instantportrait: One-step portrait editing via diffusion multi-objective distillation. In The Thirteenth In- ternational Conference on Learning Representations , 2025. 2

  5. [13]

    Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image genera- tion, 2024

    Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Lin- miao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image genera- tion, 2024. 6, 7

  6. [14]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888–12900. PMLR, 2022. 2

  7. [15]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML,

  8. [16]

    Ecnet: Effective controllable text-to-image diffusion models

    Sicheng Li, Keqiang Sun, Zhixin Lai, Xiaoshi Wu, Feng Qiu, Haoran Xie, Kazunori Miyata, and Hongsheng Li. Ecnet: Effective controllable text-to-image diffusion models. arXiv preprint arXiv:2403.18417, 2024. 2

  9. [17]

    Hunyuan-dit: A powerful multi-resolution diffusion trans- former with fine-grained chinese understanding, 2024

    Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, Dayou Chen, Jiajun He, Jiahao Li, Wenyue Li, Chen Zhang, Rongwei Quan, Jianxiang Lu, Jiabin Huang, Xiaoyan Yuan, Xiaoxiao Zheng, Yixuan Li, Ji...

  10. [18]

    Llm4gen: Leveraging semantic representation of llms for text-to-image generation, 2024

    Mushui Liu, Yuhang Ma, Yang Zhen, Jun Dan, Yunlong Yu, Zeng Zhao, Zhipeng Hu, Bai Liu, and Changjie Fan. Llm4gen: Leveraging semantic representation of llms for text-to-image generation, 2024. 2

  11. [19]

    Fusedream: Training-free text-to-image generation with improved clip+ gan space op- timization

    Xingchao Liu, Chengyue Gong, Lemeng Wu, Shujian Zhang, Hao Su, and Qiang Liu. Fusedream: Training-free text-to-image generation with improved clip+ gan space op- timization. arXiv preprint arXiv:2112.01573, 2021. 2

  12. [20]

    New theory about light and colours

    Isaac Newton. New theory about light and colours. Philo- sophical Transactions of the Royal Society , 6:3075–3087,

  13. [21]

    GLIDE: Towards Photorealistic Image Gener- ation and Editing with Text-Guided Diffusion Models

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and 9 Mark Chen. GLIDE: Towards Photorealistic Image Gener- ation and Editing with Text-Guided Diffusion Models. In ICML, 2021. 2

  14. [22]

    Inverting generative adversarial renderer for face reconstruction

    Jingtan Piao, Keqiang Sun, Quan Wang, Kwan-Yee Lin, and Hongsheng Li. Inverting generative adversarial renderer for face reconstruction. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 15619–15628, 2021. 2

  15. [23]

    Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 6, 2

  16. [24]

    Learning Transferable Visual Models From Natural Language Supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML,

  17. [25]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

  18. [26]

    Hierarchical Text-Conditional Image Gen- eration with CLIP Latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Gen- eration with CLIP Latents. ArXiv, abs/2204.06125, 2022. 2

  19. [27]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, 2022. 6

  20. [28]

    Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer

    Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. CVPR, pages 10674–10685,

  21. [29]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479–36494, 2022. 2

  22. [30]

    StyleGAN-T: Unlocking the power of gans for fast large-scale text-to-image synthesis

    Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. StyleGAN-T: Unlocking the power of gans for fast large-scale text-to-image synthesis. arXiv preprint arXiv:2301.09515, 2023. 2

  23. [31]

    Clip+mlp aesthetic score pre- dictor

    Christoph Schuhmann. Clip+mlp aesthetic score pre- dictor. https://github.com/christophschuhmann/improved- aesthetic-predictor, 2022. 1

  24. [32]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Worts- man, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:221...

  25. [33]

    Controllable 3d face syn- thesis with conditional generative occupancy fields

    Keqiang Sun, Shangzhe Wu, Zhaoyang Huang, Ning Zhang, Quan Wang, and HongSheng Li. Controllable 3d face syn- thesis with conditional generative occupancy fields. Ad- vances in Neural Information Processing Systems , 35: 16331–16343, 2022. 2

  26. [34]

    Journeydb: A benchmark for generative image under- standing, 2023

    Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, Jifeng Dai, Yu Qiao, Limin Wang, and Hongsheng Li. Journeydb: A benchmark for generative image under- standing, 2023. 3, 1

  27. [35]

    Cgof++: Controllable 3d face synthesis with conditional generative occupancy fields

    Keqiang Sun, Shangzhe Wu, Ning Zhang, Zhaoyang Huang, Quan Wang, and Hongsheng Li. Cgof++: Controllable 3d face synthesis with conditional generative occupancy fields. IEEE transactions on pattern analysis and machine intelli- gence, 46(2):913–926, 2023. 2

  28. [36]

    Genca: A text- conditioned generative model for realistic and drivable codec avatars

    Keqiang Sun, Amin Jourabloo, Riddhish Bhalodia, Moustafa Meshry, Yu Rong, Zhengyu Yang, Thu Nguyen-Phuoc, Christian Haene, Jiu Xu, Sam Johnson, et al. Genca: A text- conditioned generative model for realistic and drivable codec avatars. arXiv preprint arXiv:2408.13674, 2024. 2

  29. [37]

    Probabilistic uncertain reward model, 2025

    Wangtao Sun, Xiang Cheng, Xing Yu, Haotian Xu, Zhao Yang, Shizhu He, Jun Zhao, and Kang Liu. Probabilistic uncertain reward model, 2025. 5

  30. [38]

    Kolors: Effective training of diffusion model for photorealistic text-to-image synthesis

    Kolors Team. Kolors: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint,

  31. [39]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Li- wei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in neural in- formation processing systems, 37:84839–84865, 2024. 2

  32. [40]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution, 2024

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Jun- yang Lin. Qwen2-vl: Enhancing vision-language model’s ...

  33. [41]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis, 2023

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis, 2023. 1, 2, 3, 7, 4

  34. [42]

    Better Aligning Text-to-Image Models with Hu- man Preference, 2023

    Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hong- sheng Li. Better Aligning Text-to-Image Models with Hu- man Preference, 2023. 3, 7

  35. [43]

    Human preference score: Better aligning text- to-image models with human preference

    Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hong- sheng Li. Human preference score: Better aligning text- to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 2096–2105, 2023. 1, 2

  36. [44]

    Deep reward supervisions for tuning text-to-image diffusion models

    Xiaoshi Wu, Yiming Hao, Manyuan Zhang, Keqiang Sun, Zhaoyang Huang, Guanglu Song, Yu Liu, and Hongsheng Li. Deep reward supervisions for tuning text-to-image diffusion models. In European Conference on Computer Vision, pages 108–124. Springer, 2024. 2

  37. [45]

    Dc-ar: Efficient masked autoregressive image generation with deep compression hybrid tokenizer

    Yecheng Wu, Junyu Chen, Zhuoyang Zhang, Enze Xie, Jincheng Yu, Junsong Chen, Jinyi Hu, Yao Lu, Song Han, and Han Cai. Dc-ar: Efficient masked autoregressive image generation with deep compression hybrid tokenizer. arXiv preprint arXiv:2507.04947, 2025. 2

  38. [46]

    ImageReward: 10 Learning and Evaluating Human Preferences for Text-to- Image Generation, 2023

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. ImageReward: 10 Learning and Evaluating Human Preferences for Text-to- Image Generation, 2023. 1, 2, 3, 7, 4, 5

  39. [47]

    Imagere- ward: Learning and evaluating human preferences for text- to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere- ward: Learning and evaluating human preferences for text- to-image generation. Advances in Neural Information Pro- cessing Systems, 36, 2024. 2, 1

  40. [48]

    Dancegrpo: Unleashing grpo on visual generation

    Zeyue Xue, Jie Wu, Yu Gao, Fangyuan Kong, Lingting Zhu, Mengzhao Chen, Zhiheng Liu, Wei Liu, Qiushan Guo, Weilin Huang, et al. Dancegrpo: Unleashing grpo on visual generation. arXiv preprint arXiv:2505.07818, 2025. 6

  41. [49]

    Learning multi- dimensional human preference for text-to-image generation,

    Sixian Zhang, Bohan Wang, Junqiang Wu, Yan Li, Tingt- ing Gao, Di Zhang, and Zhongyuan Wang. Learning multi- dimensional human preference for text-to-image generation,

  42. [50]

    Cogview3: Finer and faster text-to-image generation via relay diffusion, 2024

    Wendi Zheng, Jiayan Teng, Zhuoyi Yang, Weihan Wang, Ji- dong Chen, Xiaotao Gu, Yuxiao Dong, Ming Ding, and Jie Tang. Cogview3: Finer and faster text-to-image generation via relay diffusion, 2024. 6, 2 11 HPSv3: Towards Wide-Spectrum Human Preference Score Supplementary Materia...

  43. [51]

    Image Sources of HPDv3 Table S1 summarizes the source models and images in HPDv3. Our dataset includes outputs from recent state- of-the-art image generation models, high-quality real-world images, and images generated by Midjourney, resulting in a total of 1.08M text-image pa...

  44. [52]

    As shown in Figure S2, we compare the category distri- butions of HPDv3, HPDv2, ImageReward, and Pick-a-Pic datasets

    Category distribution of HPDv3 To better reflect user preferences for prompt categories, we categorize user prompts in JourneyDB [34] into 12 dis- tinct classes, ensuring that the class proportions in HPDv3 closely match those in JourneyDB. As shown in Figure S2, we compare th...

  45. [53]

    HPDv3 Dataset Construction 3.1. Real Image Collection We collect aound 5M high-quality real images from the In- ternet, covering a wide range of categories such as architec- ture, people, objects, animals, plants, landscapes, products, Figure S2. Distribution of real images ac...

  46. [54]

    Image Annotation

    Annotation Details 4.1. Image Annotation. To build a reliable training dataset, we subject the pairwise image data to thorough human annotation, following a stan- dardized evaluation protocol. As shown in Table S2, human evaluators are provided with detailed guidelines that de...

  47. [55]

    Training Dataset For training our final model, we use data from four sources: HPDv3, subsets of Pick-A-Pic and ImageReward, and real user preference data collected from Midjourney

    HPSv3 Training Details 5.1. Training Dataset For training our final model, we use data from four sources: HPDv3, subsets of Pick-A-Pic and ImageReward, and real user preference data collected from Midjourney. In total, the training dataset comprises 1.5 million pairwise sample...

  48. [56]

    Dataset Visualization Figure S10 showcases examples from the HPDv3 dataset

    HPDv3 Dataset Visualization 6.1. Dataset Visualization Figure S10 showcases examples from the HPDv3 dataset. Each image pair consists of different images generated from the same prompt, with the images sourced from various im- age generation models as well as real-world photog...

  49. [57]

    We showcase diverse outputs pro- 5 duced across multiple iterations

    More Result of CoHP In this section, we present an extensive collection of gener- ation results from CoHP. We showcase diverse outputs pro- 5 duced across multiple iterations. The first row of Figure S11 and S12 shows the best result of each model (Flux, Kolors and Playground ...

  50. [58]

    HPSv3 as Reward Model When using reinforcement learning (RL) to improve the quality of generated images, the design of the reward model is critically important. A well-designed reward model can significantly improve outputs by boosting realism, aesthetic quality, and text-imag...

  51. [59]

    The HPDv3 dataset con- tains some parts of images obtained from the Internet, which are not the property of MizzenAI

    Term of Use of HPDv3 Ownership and Responsibility. The HPDv3 dataset con- tains some parts of images obtained from the Internet, which are not the property of MizzenAI. MizzenAI is not responsible for the content or the meaning of these images. Noncommercial Usage. Our funding...

  52. [60]

    Limitation While HPDv3 contains 1.08M text-image pairs and 1.17M pairwise data, aiming to reflect real-world user prefer- ences, it is important to acknowledge its inherent limita- tions, which may affect its generalizability and applicability in certain contexts. Prompt Distr...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.