Pith. sign in

REVIEW 5 major objections 5 minor 40 references

The paper proposes that multi-subject image generation can be made reliable by explicitly steering a diffusion transformer's attention to the correct reference regions, using a vision-language model to locate each subject described in the i

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 ↗

T0 review · deepseek-v4-flash

2026-08-03 04:56 UTC pith:GSEVGERS

load-bearing objection Plausible incremental advance in multi-subject generation, but the SOTA claim needs grounding accuracy and statistical rigor before I'd trust it. the 5 major comments →

arxiv 2602.03448 v2 pith:GSEVGERS submitted 2026-02-03 cs.CV cs.AI

Hierarchical Concept-to-Appearance Guidance for Multi-Subject Image Generation

Yijia Xu , Zihao Wang , Haokun Gui , Jinshi Cui This is my paper
classification cs.CV cs.AI
keywords multi-subject image generationcompositional image synthesisdiffusion transformervision-language model groundingmasked attentionidentity preservationVAE dropoutprompt following
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 paper's claim is that multi-subject image generation fails because existing models leave the correspondence between instruction words and reference subjects implicit, and that this can be fixed by making the correspondence explicit at two levels. At the concept level, a vision-language model (VLM) extracts which words in the instruction refer to subjects and localizes each subject in the reference images as a bounding box. At the appearance level, the paper modifies every attention layer of a diffusion transformer so that each instruction word attends only to its matched bounding-box region, and forces the model to learn from semantic VLM features by randomly dropping the low-level pixel (VAE) features during training. On a private 300-sample benchmark, this combination raises prompt-following (7.3) and subject-consistency (7.9) scores above those of prior methods, and it remains strong even when pixel-level reference features are removed at inference. The care point is that the approach trades the model's implicit reasoning for an explicit grounding step that must itself be correct.

Core claim

The central discovery is that a diffusion transformer can generate images with multiple reference identities reliably if each text token is explicitly bound, at every layer, to the specific region of the reference image that the token names. The paper implements this with a correspondence-aware masked-attention module: for a token like 'Hassan (Fig. 2)', the query attends only to the VAE keys inside the bounding box the VLM assigned to Hassan, plus the instruction and target-image tokens; everything else is masked away. The supporting element is a VAE-dropout training strategy that, with probability 0.5, removes all reference VAE features, forcing the model to lean on the VLM's semantics. To

What carries the argument

The key object is the correspondence-aware masked attention operation: in each Diffusion Transformer (DiT) layer—a transformer-based denoising network for image generation—the query of a text token that refers to a reference subject is only allowed to attend to the keys of the VAE tokens inside the bounding box that a vision-language model (VLM) assigned to that subject, plus the text and target-image tokens; all other reference tokens are masked. This explicitly encodes word-to-region binding so the network does not have to infer correspondences implicitly. The complementary mechanism is VAE dropout: during training, reference VAE features are dropped with probability 0.5, which forces the

Load-bearing premise

The vision-language model's extraction of referential words and their bounding boxes is taken as correct at train and inference time; if that grounding step is wrong or noisy, the masked attention actively hides the true references from the network, and the method has no way to recover.

What would settle it

Generate the same 300 prompts twice, once with the real VLM bounding boxes and once with boxes randomly swapped among subjects; if subject-consistency scores do not drop noticeably under the scrambled condition, the masked attention is not carrying the claimed benefit. Alternatively, if a simple baseline using full attention plus the same VLM features matches CAG, the masked attention is not the cause of the improvement.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If CAG works as reported, multi-subject generation systems can move the hard multimodal grounding task out of the diffusion network and into an external VLM, leaving the diffusion model to concentrate on appearance synthesis.
  • The masked-attention prior should scale to larger numbers of references, because each text token ignores all but its own region instead of attending to a growing token stack.
  • With VAE dropout, generation degrades gracefully when reference pixels are missing or noisy: at inference without reference VAE features, the Overall score drops only from 7.568 to 7.273, whereas a model trained without dropout falls to 6.146.
  • The framework is described as modular: it can be inserted into any VLM+DiT architecture, and the paper demonstrates it by upgrading a single-reference editor into a multi-subject generator.
  • On the paper's private benchmark, CAG leads all metrics, including DINO, CLIP-I, and CLIP-T, though the paper cautions that those three are weak evaluators for complex multi-subject composition.

Where Pith is reading between the lines

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

  • A testable implication the paper leaves implicit: because the whole pipeline is conditioned on VLM grounding accuracy, feeding the system reference images with unusual composition (e.g., a character overlapping another) should expose whether bounding-box precision is the bottleneck; measuring grounding accuracy against a human-annotated set would give a direct upper bound on the method's attainabl
  • The masked attention is a type of structural prior that could be reused in other conditional generation settings, such as object-level editing or layout-constrained text-to-image, where any external region proposal (segmentation mask, keypoint, or user click) could stand in for the VLM box.
  • The paper's 300-sample benchmark is private; releasing it would allow the community to check whether the reported score gaps persist across evaluator models and prompt distributions, and would make the outperformance claim independently reproducible.
  • One could push the idea further by training the VLM grounding jointly with generation rather than freezing it, which might let the system learn to adjust its grounding based on the generator's failures.

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

5 major / 5 minor

Summary. The paper proposes CAG (Hierarchical Concept-to-Appearance Guidance), a method for multi-subject image generation built on a VLM–DiT pipeline (Qwen-Image-Edit base). At the conceptual level, it introduces a VAE-dropout training strategy (Eq. 6) that randomly omits reference-image VAE features, forcing the DiT to rely on VLM semantic features. At the appearance level, it uses a VLM (Qwen3-VL-30B-A3B-Instruct) to extract referential words, their reference-image indices, and bounding boxes (Sec. 3.3), and injects this correspondence as a hard attention mask (Eq. 9) that restricts each referential text token to keys from the target VAE, the text VLM, and the VAE tokens inside the predicted bounding box. The model is fine-tuned on a private 24k-sample dataset and evaluated on a private 300-sample set using GPT-4.1 scores of Prompt Following and Subject Consistency. The paper reports state-of-the-art results (PF 7.308, SC 7.906, Overall 7.568) and ablations showing gains from both proposed components.

Significance. If fully validated, CAG would be a meaningful step: it makes text-to-reference correspondence explicit inside the DiT, rather than forcing the diffusion model to infer it implicitly. The attention-mask visualization in Fig. 6 is compelling, and the VAE-dropout idea is simple and potentially useful beyond this specific pipeline. The paper also provides a clear ablation structure. However, the current evidence is not yet strong enough to support the central SOTA claim: the benchmark is private, the evaluator is a single proprietary model, no error bars or significance tests are given, the VAE-dropout probability is tuned on the same test set, the grounding outputs of the VLM are unvalidated, and the reported Overall scores do not match the stated geometric-mean definition. These issues collectively undermine confidence in the quantitative claims, even though the mechanism is conceptually sound.

major comments (5)
  1. [Sec. 3.3, Eq. (9)] The entire correspondence-aware masked attention mechanism depends on VLM outputs W, image indices id_j, and bounding boxes bbox_j. Eq. (9) hard-masks all reference VLM tokens and all VAE tokens outside the predicted box, so errors in word extraction, image-index assignment, or box localization directly suppress the correct reference binding and cannot be corrected by the DiT. The same VLM provides training labels, so systematic errors may be reinforced. The paper reports no accuracy statistics (e.g., word recall, bbox IoU, image-index accuracy). Without such an analysis, the mechanism's benefit is not established. Please report grounding accuracy on the test benchmark and, ideally, compare predicted masks against oracle masks.
  2. [Sec. 4.2, Tables 1–3] The evaluation is not sufficient to support the SOTA claim. Scores come from a single proprietary GPT-4.1 evaluator, the test set (300 samples) and training set are not released, and no confidence intervals or significance tests are provided. Several margins in Table 1 are small (e.g., SC vs. Gemini is 7.906 vs. 7.819, a +0.087 gap) and could easily be evaluator noise. The paper should provide repeated or multi-seed evaluations, statistical tests, release data/evaluator prompts or a public benchmark, and error bars on all reported scores.
  3. [Sec. 4.4.1, Table 2; Sec. 4.2] The paper states that Overall is the geometric mean of PF and SC, but the reported values are inconsistent with the PF/SC entries. For example, baseline sqrt(6.797*7.653)=7.212, not 7.177; for the full model sqrt(7.308*7.906)=7.602, not 7.568. Similar discrepancies appear in Table 3 (e.g., sqrt(6.591*5.807)=6.187, not 6.146). Either the aggregation is different from the stated definition or rounding errors exist. The authors should correct this and re-verify all claimed improvements, since the numeric gaps are load-bearing for the ablation conclusions.
  4. [Sec. 4.1, Appendix B] The VAE-dropout probability p=0.5 is selected by monitoring PF/SC on the same test set (Fig. 10). This constitutes tuning on the evaluation metric, so the final reported improvement over the baseline is optimistic. The paper should either use a held-out validation set for this hyperparameter, report the model's sensitivity across seeds, or otherwise demonstrate that the choice of p does not inflate the reported test-set gains.
  5. [Sec. 3.4, Eq. (9)] The paper does not specify how the VLM's pixel-space bounding boxes are mapped to VAE latent-space token positions when constructing the mask in Eq. (9). VAE features are downsampled relative to the input image, so this mapping is nontrivial and can introduce significant localization error. Please clarify the exact formula or algorithm used to determine which VAE keys are 'inside' a bounding box, and provide sensitivity to this mapping.
minor comments (5)
  1. [Sec. 4.1] The reference for AdamW is cited as Kingma & Ba (2015), but that is the Adam paper. AdamW should cite Loshchilov & Hutter (2019).
  2. [Table 3] Typo: 'w/o V AE droput' should be 'w/o V AE dropout'.
  3. [Sec. 3.1, Eq. (4)] The notation in Eq. (4) uses superscripts in different orders (e.g., Q^r_VLM vs. K^r_VAE). Please unify the notation for clarity.
  4. [Figure 10] The curves in Figure 10 show only point values without error bars or raw numbers; adding them would help assess the sensitivity of the dropout probability.
  5. [Sec. 4.1] The dataset of 'approximately 24k examples' is described only in one sentence. Please provide details on data collection, filtering, and license status, as reproducibility depends on this.

Circularity Check

1 steps flagged

Dropout probability is tuned on the same PF/SC/Overall benchmark used for the headline SOTA, making part of the reported gain a post-selection result; the core masked-attention mechanism is otherwise not circular.

specific steps
  1. fitted input called prediction [Sec. 4.1 and Appendix B (VAE dropout probability)]
    "Following our design in Sec. 3, we apply a V AE dropout with a probability of 0.5. Additional ablation results on the V AE dropout probability are reported in the appendix. ... As shown in Figure 10, the PF, SC, and Overall scores ... reach its peak around 0.3–0.5. ... Specifically, the PF and Overall scores peak at a dropout probability of 0.5, while the SC score peaks at a dropout probability of 0.3."

    The final model uses p=0.5, which is selected by inspecting the PF/SC/Overall curves in Appendix B. These are the same metrics that later constitute the headline state-of-the-art scores in Table 1. No separate validation split is reported for this selection. The reported gains are therefore not fully independent predictions under a fixed a-priori configuration; part of the margin is a selected maximum on the evaluation metric itself. This is a partial fitted-input-called-prediction issue, not an equation-level equivalence, because the masked-attention and dropout mechanisms are still evaluated against external baselines.

full rationale

The central derivation is not circular by construction. The correspondence-aware masked attention in Eq. (9) is a concrete architectural operation that restricts each referential word's query to keys from the text VLM, target VAE, and the predicted reference bbox. It is not defined in terms of the PF/SC scores, and it is tested against external systems (Gemini, OmniGen2, UNO, Qwen-Image-Edit) and through ablations in Table 2. The VAE dropout strategy in Eq. (6) is similarly an independent training-time intervention. The lack of grounding-accuracy statistics for the VLM (Sec. 3.3, Appendix A) is a correctness and robustness concern, not circularity: the method's predictions are not derived from the measured quantity that is being reported. The only partial circularity is the choice of the VAE dropout probability: Appendix B selects p using the same PF/SC/Overall benchmark that later generates the headline SOTA numbers in Table 1, with p=0.5 chosen because it peaks those metrics. This makes part of the claimed improvement fitted to the evaluation metric rather than a clean out-of-sample prediction. There is no load-bearing self-citation chain and no equation-level reduction of the reported result to the method's inputs, so the score is moderate rather than high.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The approach relies on standard diffusion transformer assumptions plus two domain assumptions: that VLM grounding is reliable and that GPT-4.1 evaluation reflects actual quality. One hyperparameter (VAE dropout p=0.5) is tuned on the same benchmark used for the main results.

free parameters (1)
  • VAE dropout probability = 0.5
    Chosen to maximize the same GPT-4.1 benchmark used for the main results; Appendix B shows PF/Overall peak at 0.5 and SC at 0.3.
axioms (4)
  • domain assumption DiT tends to over-rely on VAE features and ignore VLM semantic features without dropout
    Motivates the dropout strategy; asserted from Figure 2 rather than proven quantitatively.
  • domain assumption Qwen3-VL provides accurate word-region grounding
    The whole masked attention mechanism depends on reliable referent-word extraction and bounding boxes (Sec. 3.3).
  • domain assumption GPT-4.1 scores are a valid proxy for multi-subject generation quality
    Primary evaluation metric; no human study or independent judge is provided.
  • domain assumption The 24k training examples are sufficient and representative
    Dataset construction and quality are not described; the model's ability to generalize depends on this.

pith-pipeline@v1.3.0-alltime-deepseek · 16130 in / 8756 out tokens · 85098 ms · 2026-08-03T04:56:54.812732+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Hierarchical Concept-to-Appearance Guidance for Multi-Subject Image Generation." pith.science (2026). https://pith.science/paper/GSEVGERS

@misc{pith2026260203448,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Concept-to-Appearance Guidance for Multi-Subject Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GSEVGERS}},
  note         = {Machine review of arXiv:2602.03448}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multi-subject image generation aims to synthesize images that faithfully preserve the identities of multiple reference subjects while following textual instructions. However, existing methods often suffer from identity inconsistency and limited compositional control, as they rely on diffusion models to implicitly associate text prompts with reference images. In this work, we propose Hierarchical Concept-to-Appearance Guidance (CAG), a framework that provides explicit, structured supervision from high-level concepts to fine-grained appearances. At the conceptual level, we introduce a VAE dropout training strategy that randomly omits reference VAE features, encouraging the model to rely more on robust semantic signals from a Visual Language Model (VLM) and thereby promoting consistent concept-level generation in the absence of complete appearance cues. At the appearance level, we integrate the VLM-derived correspondences into a correspondence-aware masked attention module within the Diffusion Transformer (DiT). This module restricts each text token to attend only to its matched reference regions, ensuring precise attribute binding and reliable multi-subject composition. Extensive experiments demonstrate that our method achieves state-of-the-art performance on the multi-subject image generation, substantially improving prompt following and subject consistency.

Figures

Figures reproduced from arXiv: 2602.03448 by Haokun Gui, Jinshi Cui, Yijia Xu, Zihao Wang.

Figure 1
Figure 1. Figure 1: Our multi-character, scene-referenced generation examples show that the outputs not only preserve consistency across characters and scenes, but also faithfully follow the specified action prompts (in bold). Abstract Multi-subject image generation aims to synthe￾size images that faithfully preserve the identities of multiple reference subjects while following textual instructions. However, existing methods … view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison of OmniGen2 (Wu et al., 2025b) under inference with and without using VAE features from refer￾ence images. OmniGen2 (Wu et al., 2025b) leads to inconsistencies not only in fine-grained details but also in semantic attributes such as shape and color. This observation suggests that the model lacks sufficient capability to effectively utilize VLM features. Effectively integrating the co… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of CAG. Top: The overall training pipeline of our CAG framework, which integrates the proposed VAE Dropout strategy and the Correspondence-Aware Masked Attention mechanism. Bottom: Illustration of the proposed Correspondence-Aware Masked Attention, where text tokens are explicitly matched to their associated regions in the reference images, enabling fine-grained and identity-consistent multi-subje… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison with different methods on multi-subject driven image generation. containing approximately 24k examples with diverse char￾acter and scene references. We employ the AdamW (Kingma & Ba, 2015) optimizer with a learning rate of 1e-5, training for 9k steps using a total batch size of 8. Following our design in Sec. 3, we apply a VAE dropout with a probability of 0.5. Additional ablation re… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison under inference without using VAE features from reference images. w/o VAE dropout denotes the model trained without the VAE dropout strategy and evaluated without reference-image VAE features, while w/ VAE dropout denotes the model trained with the VAE dropout strategy under the same inference setting. improvements of +0.266, +0.200, and +0.238, respectively. By explicitly guiding ea… view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of attention maps from textual phrases to the VAE features of the reference images. “Full Attn.” denotes using full attention over all tokens, while “Mask Attn.” denotes our proposed correspondence-aware masked attention. The example visualizes the attention map for the phrase “Hassan ( [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Prompt used to extract words associated with the reference images. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Prompt used to obtain the bounding boxes corresponding to each word. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Example of Instruction-Reference Alignment. (a) PF scores. (b) SC scores. (c) Overall scores [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Curves of PF, SC, and Overall scores under varying VAE dropout probabilities. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Generated samples from CAG. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Generated samples from CAG. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Generated samples from CAG. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_13.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

40 extracted references · 19 linked inside Pith

  1. [1]

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., et al. Qwen2. 5-vl technical report, 2025

    2025

  2. [2]

    Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., English, J., English, Z., Esser, P., Kulal, S., et al. Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space, 2025

    2025

  3. [3]

    Surf: Speeded up robust features

    Bay, H., Tuytelaars, T., and Van Gool, L. Surf: Speeded up robust features. In ECCV, pp.\ 404--417, 2006

    2006

  4. [4]

    Brooks, T., Holynski, A., and Efros, A. A. Instructpix2pix: Learning to follow image editing instructions. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 18392--18402, 2023. doi:10.1109/CVPR52729.2023.01764

    arXiv 2023

  5. [5]

    Hunyuanimage 3.0 technical report, 2025

    Cao, S., Chen, H., Chen, P., Cheng, Y., Cui, Y., Deng, X., Dong, Y., Gong, K., Gu, T., Gu, X., et al. Hunyuanimage 3.0 technical report, 2025

    2025

  6. [6]

    Xverse: Consistent multi-subject control of identity and semantic attributes via dit modulation

    Chen, B., Zhao, M., Sun, H., Chen, L., Wang, X., Du, K., and Wu, X. Xverse: Consistent multi-subject control of identity and semantic attributes via dit modulation. arXiv preprint arXiv:2506.21416, 2025 a

    Pith/arXiv arXiv 2025

  7. [7]

    Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset, 2025 b

    Chen, J., Xu, Z., Pan, X., Hu, Y., Qin, C., Goldstein, T., Huang, L., Zhou, T., Xie, S., Savarese, S., Xue, L., Xiong, C., and Xu, R. Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset, 2025 b . URL https://arxiv.org/abs/2505.09568

    Pith/arXiv arXiv 2025

  8. [8]

    Janus-pro: Unified multimodal understanding and generation with data and model scaling, 2025 c

    Chen, X., Wu, Z., Liu, X., Pan, Z., Liu, W., Xie, Z., Yu, X., and Ruan, C. Janus-pro: Unified multimodal understanding and generation with data and model scaling, 2025 c . URL https://arxiv.org/abs/2501.17811

    Pith/arXiv arXiv 2025

  9. [9]

    Umo: Scaling multi-identity consistency for image customization via matching reward, 2025

    Cheng, Y., Wu, W., Wu, S., Huang, M., Ding, F., and He, Q. Umo: Scaling multi-identity consistency for image customization via matching reward, 2025. URL https://arxiv.org/abs/2509.06818

    Pith/arXiv arXiv 2025

  10. [10]

    Emerging properties in unified multimodal pretraining, 2025

    Deng, C., Zhu, D., Li, K., Gou, C., Li, F., Wang, Z., Zhong, S., Yu, W., Nie, X., Song, Z., Shi, G., and Fan, H. Emerging properties in unified multimodal pretraining, 2025. URL https://arxiv.org/abs/2505.14683

    Pith/arXiv arXiv 2025

  11. [11]

    H., Chechik, G., and Cohen-Or, D

    Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A. H., Chechik, G., and Cohen-Or, D. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022

    Pith/arXiv arXiv 2022

  12. [12]

    Seed-data-edit technical report: A hybrid dataset for instructional image editing, 2024

    Ge, Y., Zhao, S., Li, C., Ge, Y., and Shan, Y. Seed-data-edit technical report: A hybrid dataset for instructional image editing, 2024. URL https://arxiv.org/abs/2405.04007

    Pith/arXiv arXiv 2024

  13. [13]

    Gemini 2.5 flash image

    Google. Gemini 2.5 flash image. https://developers.googleblog.com/en/introducing-gemini-2-5-flash-image/, 2025

    2025

  14. [14]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. In ICLR, 2015

    2015

  15. [15]

    Viescore: Towards explainable metrics for conditional image synthesis evaluation, 2023

    Ku, M., Jiang, D., Wei, C., Yue, X., and Chen, W. Viescore: Towards explainable metrics for conditional image synthesis evaluation, 2023

    2023

  16. [16]

    Sfnet: Learning object-aware semantic correspondence, 2019

    Lee, J., Kim, D., Ponce, J., and Ham, B. Sfnet: Learning object-aware semantic correspondence, 2019. URL https://arxiv.org/abs/1904.01810

    Pith/arXiv arXiv 2019

  17. [17]

    Reference-based sketch image colorization using augmented-self reference and dense semantic correspondence, 2020

    Lee, J., Kim, E., Lee, Y., Kim, D., Chang, J., and Choo, J. Reference-based sketch image colorization using augmented-self reference and dense semantic correspondence, 2020. URL https://arxiv.org/abs/2005.05207

    Pith/arXiv arXiv 2020

  18. [18]

    Onecat: Decoder-only auto-regressive model for unified understanding and generation, 2025

    Li, H., Peng, X., Wang, Y., Peng, Z., Chen, X., Weng, R., Wang, J., Cai, X., Dai, W., and Xiong, H. Onecat: Decoder-only auto-regressive model for unified understanding and generation, 2025. URL https://arxiv.org/abs/2509.03498

    arXiv 2025

  19. [19]

    Lowe, D. G. Distinctive image features from scale-invariant keypoints. IJCV, 60: 0 91--110, 2004

    2004

  20. [20]

    Regiondrag: Fast region-based image editing with diffusion models, 2024

    Lu, J., Li, X., and Han, K. Regiondrag: Fast region-based image editing with diffusion models, 2024. URL https://arxiv.org/abs/2407.18247

    Pith/arXiv arXiv 2024

  21. [21]

    Dreamo: A unified framework for image customization

    Mou, C., Wu, Y., Wu, W., Guo, Z., Zhang, P., Cheng, Y., Luo, Y., Ding, F., Zhang, S., Li, X., et al. Dreamo: A unified framework for image customization. arXiv preprint arXiv:2504.16915, 2025

    arXiv 2025

  22. [22]

    OpenAI. Gpt-4-1. https://openai.com/index/gpt-4-1, 2025

    2025

  23. [23]

    and Xie, S

    Peebles, W. and Xie, S. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 4195--4205, 2023

    2023

  24. [24]

    Qwen3-vl

    QwenTeam. Qwen3-vl. https://qwen.ai/blog?id=99f0335c4ad9ff6153e517418d48535ab6d8afef&from=research.latest-advancements-list, 2025

    2025

  25. [25]

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

    Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., and Aberman, K. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 22500--22510, 2023

    2023

  26. [26]

    Mosaic: Multi-subject personalized generation via correspondence-aware alignment and disentanglement, 2025

    She, D., Fu, S., Liu, M., Jin, Q., Wang, H., Liu, M., and Jiang, J. Mosaic: Multi-subject personalized generation via correspondence-aware alignment and disentanglement, 2025. URL https://arxiv.org/abs/2509.01977

    Pith/arXiv arXiv 2025

  27. [27]

    Emu edit: Precise image editing via recognition and generation tasks, 2023

    Sheynin, S., Polyak, A., Singer, U., Kirstain, Y., Zohar, A., Ashual, O., Parikh, D., and Taigman, Y. Emu edit: Precise image editing via recognition and generation tasks, 2023. URL https://arxiv.org/abs/2311.10089

    Pith/arXiv arXiv 2023

  28. [28]

    Ominicontrol: Minimal and universal control for diffusion transformer

    Tan, Z., Liu, S., Yang, X., Xue, Q., and Wang, X. Ominicontrol: Minimal and universal control for diffusion transformer. In ICCV, 2025

    2025

  29. [29]

    P., and Hariharan, B

    Tang, L., Jia, M., Wang, Q., Phoo, C. P., and Hariharan, B. Emergent correspondence from image diffusion, 2023. URL https://arxiv.org/abs/2306.03881

    Pith/arXiv arXiv 2023

  30. [30]

    Chameleon: Mixed-modal early-fusion foundation models, 2025

    Team, C. Chameleon: Mixed-modal early-fusion foundation models, 2025. URL https://arxiv.org/abs/2405.09818

    Pith/arXiv arXiv 2025

  31. [31]

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

    Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution, 2024

    2024

  32. [32]

    Qwen-image technical report, 2025 a

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., Yin, S.-m., Bai, S., Xu, X., Chen, Y., et al. Qwen-image technical report, 2025 a

    2025

  33. [33]

    Omnigen2: Exploration to advanced multimodal generation

    Wu, C., Zheng, P., Yan, R., Xiao, S., Luo, X., Wang, Y., Li, W., Jiang, X., Liu, Y., Zhou, J., et al. Omnigen2: Exploration to advanced multimodal generation. arXiv preprint arXiv:2506.18871, 2025 b

    Pith/arXiv arXiv 2025

  34. [34]

    Less-to-more generalization: Unlocking more controllability by in-context generation, 2025 c

    Wu, S., Huang, M., Wu, W., Cheng, Y., Ding, F., and He, Q. Less-to-more generalization: Unlocking more controllability by in-context generation, 2025 c . URL https://arxiv.org/abs/2504.02160

    Pith/arXiv arXiv 2025

  35. [35]

    Llmga: Multimodal large language model based generation assistant, 2024

    Xia, B., Wang, S., Tao, Y., Wang, Y., and Jia, J. Llmga: Multimodal large language model based generation assistant, 2024. URL https://arxiv.org/abs/2311.16500

    Pith/arXiv arXiv 2024

  36. [36]

    Dreamomni2: Multimodal instruction-based editing and generation, 2025

    Xia, B., Peng, B., Zhang, Y., Huang, J., Liu, J., Li, J., Tan, H., Wu, S., Wang, C., Wang, Y., et al. Dreamomni2: Multimodal instruction-based editing and generation, 2025

    2025

  37. [37]

    P., Jampani, V., Sun, D., and Yang, M.-H

    Zhang, J., Herrmann, C., Hur, J., Cabrera, L. P., Jampani, V., Sun, D., and Yang, M.-H. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence, 2023 a . URL https://arxiv.org/abs/2305.15347

    Pith/arXiv arXiv 2023

  38. [38]

    Telling left from right: Identifying geometry-aware semantic correspondence, 2024

    Zhang, J., Herrmann, C., Hur, J., Chen, E., Jampani, V., Sun, D., and Yang, M.-H. Telling left from right: Identifying geometry-aware semantic correspondence, 2024. URL https://arxiv.org/abs/2311.17034

    Pith/arXiv arXiv 2024

  39. [39]

    Magicbrush: a manually annotated dataset for instruction-guided image editing

    Zhang, K., Mo, L., Chen, W., Sun, H., and Su, Y. Magicbrush: a manually annotated dataset for instruction-guided image editing. In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS '23, Red Hook, NY, USA, 2023 b . Curran Associates Inc

    2023

  40. [40]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...