Pith. sign in

REVIEW 4 major objections 5 minor 14 cited by

The paper's central claim: one model can match both a reference style and a reference subject at once.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

USO trains one DiT model for subject-driven, style-driven, and joint generation by disentangling content and style from triplet data and adding a style-reward objective, claiming SOTA on USO-Bench.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A useful subject-consistency model with a credible DreamBench result, but the style-similarity SOTA claim is circular because the CSD model serves as both training reward and evaluation metric. the 4 major comments →

arxiv 2508.18966 v1 pith:4S255XSS submitted 2025-08-26 cs.CV cs.LG

USO: Unified Style and Subject-Driven Generation via Disentangled and Reward Learning

Shaojin Wu , Mengqi Huang , Yufeng Cheng , Wenxu Wu , Jiahe Tian , Yiming Luo , Fei Ding , Qian He This is my paper
classification cs.CV cs.LG
keywords style-driven generationsubject-driven generationcontent-style disentanglementcross-task co-disentanglementdiffusion transformersstyle reward learningimage customizationunified generation benchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

USO argues that style-driven and subject-driven image generation are not opposite tasks but two sides of the same coin: each one requires the model to include one kind of visual feature and exclude the other. To prove this, the authors build a large triplet dataset of content images, style images, and stylized content images, then train a single diffusion-transformer model with two complementary objectives—style alignment and content–style disentanglement—and a style-reward learning loop. On their new USO-Bench and on DreamBench, the resulting model reports the highest subject-consistency and style-similarity scores among open-source customization models, including in the joint setting where a subject and a style are given together. If the claim holds, users no longer need separate customization models for style transfer and identity preservation.

Core claim

The central discovery is that jointly modeling the two tasks produces better disentanglement than training either task alone. USO processes style references through a semantic encoder with a hierarchical projector and content references through a frozen autoencoder, feeding both token streams into one in-context diffusion transformer. Training proceeds in two stages: first align style tokens, then add content tokens and train on triplets so the model learns to recombine arbitrary subjects with arbitrary styles, including layout shifts. A style-reward loss then pushes the generated image's style-similarity score toward the reference style, which also improves subject consistency and reduces p

What carries the argument

Cross-task co-disentanglement: a training scheme in which each task (style-driven versus subject-driven) teaches the model which features to exclude by learning to include them in the complementary task. It is implemented with a triplet dataset (content image, style image, stylized target), separate style and content encoders whose token streams are concatenated in the transformer, and two training objectives—style-alignment and content–style disentanglement—plus a style-reward learning loop that scores generated images against the style reference.

Load-bearing premise

The whole scheme rests on the triplet data being truly disentangled: the de-stylized content reference must carry no residual style, and the style reference must carry no residual subject.

What would settle it

Take one subject, stylize it into several different styles, then de-stylize each into a content reference; feed these content references through the model under the same style prompt and compare outputs. If the outputs systematically differ with the source style, content-style disentanglement is incomplete and part of the reported performance comes from correlated references rather than clean separation.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single open model can handle subject-driven, style-driven, and combined style-subject-driven prompts without task-specific fine-tuning.
  • The model can re-render a subject in a new style and reposition it in a new layout, not just recolor the original composition.
  • Adding a style-similarity reward improves subject consistency and image quality, an effect the paper observes even for tasks the reward never targets.
  • The released USO-Bench gives the field a common yardstick for evaluating subject fidelity and style similarity jointly.
  • If the approach generalizes, customization systems can be trained on cross-task data instead of separate per-task datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • It should be possible to reuse the same co-disentanglement recipe for other feature axes—lighting, viewpoint, object category—by building analogous include/exclude reference pairs.
  • A stricter test of the core idea would swap the style-similarity reward for a subject-consistency reward (or apply both) and check whether the mutual-benefit effect persists, which the paper does not report.
  • The triplet curation step is the latent bottleneck: if the experts used to produce content/style references are imperfect, later gains may partly reflect correlated noise rather than true disentanglement; a human-validated subset would directly test this.
  • The observation that style reward reduces 'plastic' faces suggests reward learning might replace identity-specific data augmentation; a targeted user study on identity tasks could verify it.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper presents USO, a unified customization model that jointly handles subject-driven, style-driven, and joint style-subject-driven generation. The method constructs a triplet dataset using UNO-based stylization/de-stylization experts, trains in two stages (style-alignment and content–style disentanglement), and adds a style-reward learning (SRL) term. The authors release USO-Bench, a new benchmark, and report state-of-the-art results on subject consistency, style similarity, and text alignment relative to existing open-source baselines.

Significance. If the claims hold, this is a useful contribution: a single model that supports both subject and style conditioning, a systematic triplet-curation pipeline, a new benchmark, and released code/models. The subject-consistency half is supported by externally validated DreamBench results (DINO 0.777, CLIP-I 0.838, CLIP-T 0.317) that do not use a subject reward during training. However, the style-similarity half of the headline claim currently rests on a single metric (CSD) that also appears as the SRL reward, so independent validation is needed. The assumed disentanglement of the curated references is also not verified. The paper is clearly written and the method is reproducible in principle.

major comments (4)
  1. [§3.3.3, Eq. (6); §4.1; Tables 1–3] The style-similarity SOTA claim is circular as reported. Section 4.1 evaluates style similarity via the CSD score, while Section 3.3.3 defines the SRL reward as style similarity 'measured by either a VLM-based filter or the CSD model MRM(·)'. The paper does not state which reward was used to train or select the final model. Table 3's ablation (w/o SRL: CSD drops from 0.495 to 0.413 on the joint task) is exactly what would be expected if the model were optimized against the evaluation metric itself. I request an independent style metric (or a numerical user study with reported N), and disclosure of which reward model was used for the released checkpoint. The subject-consistency claims from DreamBench (Table 5) are not affected by this concern.
  2. [§3.2, Figure 3] The triplet construction assumes that the stylization and de-stylization experts produce genuinely disentangled references: I_s_ref carries style but no subject content, and I_c_ref carries content but no residual style. The VLM filter validates style similarity and subject consistency relative to the target image, but not mutual independence of the two references. No human validation or statistical leakage check is provided. Because the central cross-task co-disentanglement mechanism and the ability to freely recombine arbitrary subjects and styles depend on this assumption, please add diagnostics such as CSD(I_c_ref, target), DINO(I_s_ref, target), or human ratings of residual style/content leakage.
  3. [§4.2, Figure 9; Tables 1 and 5] The text states that USO 'significantly outperforms' baselines, but no error bars, significance tests, or run-to-run variance are reported. Several differences in Table 1 are small (e.g., CLIP-I 0.623 vs 0.605, DINO 0.793 vs 0.789), and the user study in Figure 9 reports no sample size, confidence intervals, or statistical comparison. Please provide these details to support the 'significant' wording and the SOTA claims.
  4. [Table 4 vs Table 2] Table 4 reports the hierarchical projector with CSD=0.402, whereas Table 2 reports full USO with CSD=0.495 on the style-subject task. If these are the same evaluation, the numbers are inconsistent; if Table 4 uses a different task split (e.g., style-driven only), the table header must say so. As written, the projector ablation cannot be interpreted.
minor comments (5)
  1. [§3.3.3] Duplicate word: 'conditioned on both an image reference and and its corresponding text prompt' should be 'and its'.
  2. [Algorithm 1, Eq. (6)] The sign convention is inconsistent: Eq. (6) defines LSRL as a positive expectation over reward scores, while Algorithm 1 line 11 sets LSRL = -MRM(...). Clarify that phi maps rewards to losses or remove the negative sign.
  3. [§3.3.2] 'UnoPE' should be consistently written as 'UNO-PE' or 'UnoPE' and explained once.
  4. [Related Work] In Section 2.1, 'InstanceStyle' appears to be a typo for 'InstantStyle'.
  5. [Appendix F.1.2, Figure 12] The label 'Layout-preserved prompts' includes both an empty prompt and an example prompt. Clarify the intended prompt structure so readers know when the empty prompt is used.

Circularity Check

1 steps flagged

Style-similarity SOTA is partially circular: the SRL reward and the evaluation metric are the same CSD model MRM(·), so the headline style gains partly measure reward optimization rather than independent style fidelity.

specific steps
  1. fitted input called prediction [Section 3.3.3, Eq. 6; Section 4.1 Evaluation Metrics; Tables 1-3]
    "we define the reward score as the style similarity between the reference style image I_s_ref and the generated stylized image I_theta, measured by either a VLM-based filter or the CSD model MRM(·) [26, 37]. The reward loss is defined as: LSRL = E_yi∼Y [ϕ(MRM(yi, Iθ(yi)))] (6) ... style similarity, reported via the CSD score [26] for both style-driven and style-subject-driven generation, following [37]."

    The SRL loss in Eq. 6 uses MRM(·) as the style-similarity reward, and Section 4.1 evaluates style similarity 'via the CSD score [26]', i.e., the same MRM model. Algorithm 1 line 11 sets LSRL ← −MRM(Î0, I_s_ref), so the model is explicitly trained to maximize this exact score, and the headline style-similarity results in Tables 1 and 2 are reported as CSD scores. The paper never states that the final model's reward used a different metric (e.g., the VLM-based filter instead of CSD). Therefore the CSD gains (0.557 vs 0.540 InstantStyle; 0.495 vs 0.407 StyleID) partly reflect how well the model optimizes its own evaluator, not an externally validated style-fidelity improvement. The w/o SRL ablation (CSD 0.495→0.413) is consistent with metric-specific overfitting, and no independent style metr

full rationale

The paper's subject-consistency claims are not circular: DreamBench DINO/CLIP-I and USO-Bench CLIP-I/DINO are external metrics that were not used as rewards. The triplet-curation disentanglement assumption is an unvalidated assumption, not a circular step. The self-citations to UNO and VMix are prior-work foundations and do not by themselves make the central contribution circular. USO-Bench is author-constructed, but that is standard practice; the concern is confined to the style metric. The style-similarity half of the SOTA claim is partially circular because the evaluation metric (CSD score) is identical to the reward model (MRM) in Eq. 6, and the paper does not rule out that the final model was trained with that same metric. A partial user study and qualitative comparisons provide some independent support, but no numerical user-study results or alternative style benchmarks are reported. Score 6 reflects partial circularity: one of the two headline dimensions is not independently measured.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The method does not introduce new physical or formal entities. It relies on several domain assumptions about data curation, representation, and reward/evaluation validity, plus a set of manually chosen hyperparameters that are not ablated. The most consequential gap is that the quality of the UNO-generated triplets, which is the foundation of the claimed disentanglement, is not independently validated.

free parameters (7)
  • reward balancing coefficient lambda = 0 before step S, 1 after
    Set by hand in Eq. 7; controls when SRL loss starts to dominate.
  • reward step S = 16,000 (Stage 1) / 18,000 (Stage 2)
    Chosen thresholds in Appendix F.1.1 that switch on the style-reward loss.
  • training steps = 23,000 (Stage 1), 21,000 (Stage 2)
    Manual stopping points; no convergence or early-stopping analysis.
  • batch size = 16 (Stage 1), 64 (Stage 2)
    Manual hyperparameter; affects gradient noise and training dynamics.
  • learning rate = 8e-5
    Manual hyperparameter for both stages.
  • LoRA rank = 128
    Manual choice for parameter-efficient tuning; not ablated.
  • resolution = 768 (Stage 1), 1024 (Stage 2)
    Manual training resolutions; affect detail and content/style trade-off.
axioms (5)
  • ad hoc to paper UNO-based stylization and de-stylization experts produce valid, disentangled style and content references for triplet construction.
    Section 3.2 builds the entire USO dataset on these expert outputs; no independent verification that the generated triplets are truly disentangled.
  • domain assumption SigLIP embeddings and the hierarchical projector capture style information needed for large geometric and material transformations.
    Section 3.3.1 states style is an abstract cue requiring richer semantic information and replaces the VAE with SigLIP, but the paper does not prove this representation is sufficient.
  • domain assumption The CSD model is a valid, unbiased measure of human-perceived style similarity.
    Used both as the training reward in Eq. 6 and as the evaluation metric in Section 4.1; the paper treats CSD as ground truth.
  • domain assumption The VLM-based filter correctly enforces style similarity between target and style reference and subject consistency between target and content reference.
    Used in Section 3.2 to filter triplets, but no accuracy or human agreement metrics are reported.
  • standard math Standard flow-matching pre-training objective (Eq. 5) is a sufficient base for the customization model.
    Adopted from prior diffusion literature; this is a background assumption, not a new contribution.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of USO: Unified Style and Subject-Driven Generation via Disentangled and Reward Learning." pith.science (2026). https://pith.science/paper/4S255XSS

@misc{pith2026250818966,
  author       = {Pith},
  title        = {Pith review of: USO: Unified Style and Subject-Driven Generation via Disentangled and Reward Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4S255XSS}},
  note         = {Machine review of arXiv:2508.18966}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing literature typically treats style-driven and subject-driven generation as two disjoint tasks: the former prioritizes stylistic similarity, whereas the latter insists on subject consistency, resulting in an apparent antagonism. We argue that both objectives can be unified under a single framework because they ultimately concern the disentanglement and re-composition of content and style, a long-standing theme in style-driven research. To this end, we present USO, a Unified Style-Subject Optimized customization model. First, we construct a large-scale triplet dataset consisting of content images, style images, and their corresponding stylized content images. Second, we introduce a disentangled learning scheme that simultaneously aligns style features and disentangles content from style through two complementary objectives, style-alignment training and content-style disentanglement training. Third, we incorporate a style reward-learning paradigm denoted as SRL to further enhance the model's performance. Finally, we release USO-Bench, the first benchmark that jointly evaluates style similarity and subject fidelity across multiple metrics. Extensive experiments demonstrate that USO achieves state-of-the-art performance among open-source models along both dimensions of subject consistency and style similarity. Code and model: https://github.com/bytedance/USO

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 14 Pith papers

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

  1. UniCustom: Unified Visual Conditioning for Multi-Reference Image Generation

    cs.CV 2026-05 unverdicted novelty 7.0

    UniCustom fuses ViT and VAE features before VLM encoding and uses two-stage training plus slot-wise regularization to improve subject consistency in multi-reference diffusion-based image generation.

  2. DreamStyle3D: Efficient 3D Stylized Asset Generation via Dual-Attention Disentanglement

    cs.CV 2026-07 conditional novelty 6.0

    A new model generates stylized 3D objects in about 10 seconds by separating style from geometry inside a native 3D diffusion model.

  3. Scaling Multi-Reference Image Generation with Dynamic Reward Optimization

    cs.CV 2026-06 unverdicted novelty 6.0

    Introduces OmniRef-Bench benchmark and DyRef two-stage framework using Difficulty-aware Advantage Reweighting and Discriminative Reward Scaling to improve open-source models on complex multi-reference image generation.

  4. Lance: Unified Multimodal Modeling by Multi-Task Synergy

    cs.CV 2026-05 unverdicted novelty 6.0

    Lance presents a dual-stream mixture-of-experts model with modality-aware positional encoding and staged multi-task training that outperforms prior open-source unified models on image and video generation while keepin...

  5. UniCustom: Unified Visual Conditioning for Multi-Reference Image Generation

    cs.CV 2026-05 unverdicted novelty 6.0

    A unified visual conditioning approach fuses semantic and appearance features before VLM processing, with two-stage training and slot-wise regularization, to improve consistency in multi-reference image generation.

  6. Fashion130K: An E-commerce Fashion Dataset for Outfit Generation with Unified Multi-modal Condition

    cs.CV 2026-05 unverdicted novelty 6.0

    Fashion130K dataset and UMC framework align text and visual prompts to generate more consistent fashion outfits than prior state-of-the-art methods.

  7. Fashion130K: An E-commerce Fashion Dataset for Outfit Generation with Unified Multi-modal Condition

    cs.CV 2026-05 unverdicted novelty 6.0

    Fashion130K dataset and UMC framework align text and visual prompts with embedding refiner, Fusion Transformer, and redesigned attention to generate more consistent outfits than prior methods.

  8. Scone: Bridging Composition and Distinction in Subject-Driven Image Generation via Unified Understanding-Generation Modeling

    cs.CV 2025-12 conditional novelty 6.0

    Scone adds a semantic-bridge attention-masking step to a unified understanding-generation model, improving subject distinction in multi-candidate reference images, and introduces the SconeEval benchmark.

  9. Scone: Bridging Composition and Distinction in Subject-Driven Image Generation via Unified Understanding-Generation Modeling

    cs.CV 2025-12 conditional novelty 6.0

    Scone unifies subject understanding and generation in a two-stage trained model to improve both composition and distinction in multi-subject image generation, outperforming prior open-source models on new benchmarks.

  10. iMontage: Unified, Versatile, Highly Dynamic Many-to-many Image Generation

    cs.CV 2025-11 conditional novelty 6.0

    iMontage repurposes a pretrained video diffusion model to generate coherent yet highly dynamic image sets from arbitrary numbers of input images.

  11. DreamStyle3D: Efficient 3D Stylized Asset Generation via Dual-Attention Disentanglement

    cs.CV 2026-07 conditional novelty 5.0

    Decoupled dual cross-attention plus style/content augmentations let a TRELLIS-based model inject image style into 3D assets in ~10s while better preserving geometry than prior 2D-to-3D pipelines.

  12. FreeStyle: Free Control of Style-Content Dual-Reference Generation from Community LoRA Mining

    cs.CV 2026-06 unverdicted novelty 5.0

    FreeStyle proposes community LoRA mining plus attention and frequency disentanglement to enable scalable style-content dual-reference generation with reduced leakage.

  13. Lance: Unified Multimodal Modeling by Multi-Task Synergy

    cs.CV 2026-05 unverdicted novelty 5.0

    Lance introduces a dual-stream MoE model with modality-aware rotary positional encoding and staged multi-task training that outperforms open-source unified models on image and video generation while retaining understa...

  14. EditIDv2: Editable ID Customization with Data-Lubricated ID Feature Integration for Text-to-Image Generation

    cs.CV 2025-09 reject novelty 3.0

    EditIDv2 fine-tunes only PerceiverAttention cross-attention weights on 3K images to inject editability into Flux-based ID customization, reporting selective gains on the self-proposed IBench benchmark.

Reference graph

Works this paper leans on

41 extracted references · 8 canonical work pages · cited by 9 Pith papers · 2 internal anchors

  1. [1]

    Re-imagen: Retrieval-augmented text-to- image generator

    Wenhu Chen, Hexiang Hu, Chitwan Saharia, and William W Cohen. Re-imagen: Retrieval-augmented text-to- image generator. arXiv preprint arXiv:2209.14491, 2022

    Pith/arXiv arXiv 2022

  2. [2]

    Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer

    Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8795–8805, 2024

    2024

  3. [3]

    Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683, 2025

    Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, et al. Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024

    2024

  5. [5]

    Implicit style-content separation using b-lora

    Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or. Implicit style-content separation using b-lora. In European Conference on Computer Vision, pages 181–198. Springer, 2024

    work page 2024

  6. [6]

    An image is worth one word: Personalizing text-to-image generation using textual inversion.arXiv preprint arXiv:2208.01618, 2022

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion.arXiv preprint arXiv:2208.01618, 2022

    Pith/arXiv arXiv 2022

  7. [7]

    Styleshot: A snapshot on any style.arXiv preprint arXiv:2407.01414, 2024

    Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. Styleshot: A snapshot on any style.arXiv preprint arXiv:2407.01414, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021

    Pith/arXiv arXiv 2021

  9. [9]

    In-context lora for diffusion transformers.arXiv preprint arXiv:2410.23775, 2024

    Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers.arXiv preprint arXiv:2410.23775, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Realcustom: narrowing real text word for real-time open-domain text-to-image customization

    Mengqi Huang, Zhendong Mao, Mingcong Liu, Qian He, and Yongdong Zhang. Realcustom: narrowing real text word for real-time open-domain text-to-image customization. InCVPR, pages 7476–7485, 2024

    work page 2024

  11. [11]

    Arbitrary style transfer in real-time with adaptive instance normalization

    Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017

    2017

  12. [12]

    Visual style prompting with swapping self-attention

    Jaeseok Jeong, Junho Kim, Yunjey Choi, Gayoung Lee, and Youngjung Uh. Visual style prompting with swapping self-attention. arXiv preprint arXiv:2402.12974, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Infiniteyou: Flexible photo recrafting while preserving your identity.arXiv preprint arXiv:2503.16418, 2025

    Liming Jiang, Qing Yan, Yumin Jia, Zichuan Liu, Hao Kang, and Xin Lu. Infiniteyou: Flexible photo recrafting while preserving your identity.arXiv preprint arXiv:2503.16418, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Flux: Official inference repository for flux.1 models, 2024

    Black Forest Labs. Flux: Official inference repository for flux.1 models, 2024. URL https://github.com/ black-forest-labs/flux. Accessed: 2025-02-07

    2024

  15. [15]

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, et al. Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space.arXiv preprint arXiv:2506.15742, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Stylestudio: Text-driven style transfer with selective control of style elements

    Mingkun Lei, Xue Song, Beier Zhu, Hao Wang, and Chi Zhang. Stylestudio: Text-driven style transfer with selective control of style elements. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 23443–23452, 2025

    2025

  17. [17]

    Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing.Advancesin Neural Information Processing Systems, 36:30146–30166, 2023

    Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing.Advancesin Neural Information Processing Systems, 36:30146–30166, 2023

    work page 2023

  18. [18]

    Realgeneral: Unifying visual generation via temporal in-context learning with video models.arXiv preprint arXiv:2503.10406, 2025

    Yijing Lin, Mengqi Huang, Shuhan Zhuang, and Zhendong Mao. Realgeneral: Unifying visual generation via temporal in-context learning with video models.arXiv preprint arXiv:2503.10406, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    Realcustom++: Representing images as real-word for real-time customization.arXiv preprint arXiv:2408.09744, 2024

    Zhendong Mao, Mengqi Huang, Fei Ding, Mingcong Liu, Qian He, and Yongdong Zhang. Realcustom++: Representing images as real-word for real-time customization.arXiv preprint arXiv:2408.09744, 2024. 15

    arXiv 2024

  20. [20]

    Dreamo: A unified framework for image customization.arXiv preprint arXiv:2504.16915, 2025

    Chong Mou, Yanze Wu, Wenxu Wu, Zinan Guo, Pengze Zhang, Yufeng Cheng, Yiming Luo, Fei Ding, Shiwen Zhang, Xinghui Li, et al. Dreamo: A unified framework for image customization.arXiv preprint arXiv:2504.16915, 2025

    arXiv 2025

  21. [21]

    SDXL: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. InICLR, 2024. URL https://openreview.net/forum?id=di52zR8xgf

    work page 2024

  22. [22]

    BootPIG: Bootstrapping Zero-shot Personalized Image Generation Capabilities in Pretrained Diffusion Models

    Senthil Purushwalkam, Akash Gokul, Shafiq Joty, and Nikhil Naik. Bootpig: Bootstrapping zero-shot personalized image generation capabilities in pretrained diffusion models.arXiv preprint arXiv:2401.13974, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  23. [23]

    Deadiff: An efficient stylization diffusion model with disentangled representations

    Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentangled representations. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8693–8702, 2024

    2024

  24. [24]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InCVPR, pages 10684–10695, 2022

    2022

  25. [25]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. InCVPR, pages 22500–22510, 2023

    2023

  26. [26]

    Measuring style similarity in diffusion models.arXiv preprint arXiv:2404.01292, 2024

    Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models.arXiv preprint arXiv:2404.01292, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Ominicontrol: Minimal and universal control for diffusion transformer.arXiv preprint arXiv:2411.15098, 3, 2024

    Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. Ominicontrol: Minimal and universal control for diffusion transformer.arXiv preprint arXiv:2411.15098, 3, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    Instantstyle: Free lunch towards style-preserving in text-to-image generation.arXiv preprint arXiv:2404.02733, 2024

    Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation.arXiv preprint arXiv:2404.02733, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Omnistyle: Filtering high quality style transfer data at scale

    Ye Wang, Ruiqi Liu, Jiang Lin, Fei Liu, Zili Yi, Yilin Wang, and Rui Ma. Omnistyle: Filtering high quality style transfer data at scale. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 7847–7856, 2025

    2025

  30. [30]

    Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation

    Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. InCVPR, pages 15943–15953, 2023

    work page 2023

  31. [31]

    Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Omnigen2: Exploration to advanced multimodal generation.arXiv preprint arXiv:2506.18871, 2025

    Chenyuan Wu, Pengfei Zheng, Ruiran Yan, Shitao Xiao, Xin Luo, Yueze Wang, Wanli Li, Xiyan Jiang, Yexin Liu, Junjie Zhou, et al. Omnigen2: Exploration to advanced multimodal generation.arXiv preprint arXiv:2506.18871, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    Vmix: Improving text-to-image diffusion model with cross-attention mixing control.arXiv preprint arXiv:2412.20800, 2024

    Shaojin Wu, Fei Ding, Mengqi Huang, Wei Liu, and Qian He. Vmix: Improving text-to-image diffusion model with cross-attention mixing control.arXiv preprint arXiv:2412.20800, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    Less-to-more generalization: Unlocking more controllability by in-context generation.arXiv preprint arXiv:2504.02160, 2025

    Shaojin Wu, Mengqi Huang, Wenxu Wu, Yufeng Cheng, Fei Ding, and Qian He. Less-to-more generalization: Unlocking more controllability by in-context generation.arXiv preprint arXiv:2504.02160, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    StyleAlign: Analysis and Applications of Aligned StyleGAN Models

    Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. Stylealign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  36. [36]

    Omnigen: Unified image generation.arXiv preprint arXiv:2409.11340, 2024

    Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation.arXiv preprint arXiv:2409.11340, 2024

    Pith/arXiv arXiv 2024

  37. [37]

    Csgo: Content-style composition in text-to-image generation.arXiv preprint arXiv:2408.16766, 2024

    Peng Xing, Haofan Wang, Yanpeng Sun, Qixun Wang, Xu Bai, Hao Ai, Renyuan Huang, and Zechao Li. Csgo: Content-style composition in text-to-image generation.arXiv preprint arXiv:2408.16766, 2024

    Pith/arXiv arXiv 2024

  38. [38]

    Imagereward: 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. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903–15935, 2023

    2023

  39. [39]

    Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.arXiv preprint arXiv:2308.06721, 2023

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.arXiv preprint arXiv:2308.06721, 2023. 16

    Pith/arXiv arXiv 2023

  40. [40]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975–11986, 2023

    2023

  41. [41]

    The girl is riding a bike in the street

    Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. InCVPR, pages 8069–8078, 2024. 17 USO: Unified Style and Subject-Driven Generation via Disentangled and Reward Learning Appendix F.1 Experiments Setting F.1...

    work page 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.