Pith. sign in

REVIEW 5 major objections 5 minor 24 references

Knowledge-intensive image synthesis is a distinct, largely unmapped capability; ExpertVerse measures it with 58 sub-disciplines, and KnowThinker shows a reasoning engine can begin to close the gap.

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

2026-08-01 12:41 UTC pith:DN6ZAATR

load-bearing objection A real benchmark and training recipe with strong external gains, but the headline ExpertVerse numbers rest on an unreleased proprietary evaluator and withheld data. the 5 major comments →

arxiv 2607.19341 v1 pith:DN6ZAATR submitted 2026-07-21 cs.CV

ExpertVerse: A General-Purpose Benchmark for Expert-Level Reasoning in Knowledge-Intensive Visual Synthesis

Yuan Wang , Yongchao Du , Mengting Chen , Jinsong Lan , Xuetao Feng , Xiaoyong Zhu This is my paper
classification cs.CV
keywords knowledge-intensive image generationreasoning benchmarkimage editingtext-to-image synthesisvision-language modelreinforcement learningchain-of-thoughtmulti-objective optimization
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.

This paper tries to establish that generating images that require expert-level knowledge—nutritional plans, historical scenes, scientific diagrams, character relationship maps—is a distinct capability, separate from visual fidelity, and that current models, including proprietary ones, have deep blind spots there. To make that capability legible, it builds ExpertVerse, a held-out benchmark of 1,611 expert-annotated instances organized by an orthogonal taxonomy of 9 cognitive abilities and 8 expert disciplines (58 sub-disciplines), covering single-image editing, multi-image composition, and text-to-image. It then argues that the bottleneck is reasoning, not pixels, and presents KnowThinker, a vision-language reasoner that externalizes a thinking trace and a refined instruction before a frozen image editor renders the output. KnowThinker is trained with BPPO, which combines bootstrapped reward rectification and Pareto-aware advantage fusion to stabilize multi-reward RL. On five benchmarks it sets open-source state-of-the-art results, approaching and in some editing tasks surpassing proprietary systems, which the paper reads as evidence that decoupled, knowledge-anchored reasoning is the key lever.

Core claim

The paper's central claim is that reasoning is the limiting factor in knowledge-intensive visual synthesis: a vision-language reasoner that decomposes a high-level request into a chain-of-thought and a refined instruction, trained on a large-scale reasoning-trace corpus with BPPO, raises a frozen editor from 19.4% to 40.0% on RISEBench, from 67.14% to 83.02% on KRIS-Bench, and from 60.2 to 82.4 on WiseEdit-Complex, while scoring 69.73% on the new ExpertVerse. The authors take this as evidence that knowledge-intensive generation is a measurable capability—stratified by the benchmark's 58 sub-disciplines—and that current models, both open and proprietary, suffer systematic knowledge blind spot

What carries the argument

ExpertVerse's orthogonal taxonomy (9 cognitive capabilities × 8 expert disciplines, yielding 58 sub-disciplines) is the measurement backbone: it operationalizes 'knowledge-intensive' and forces each instance to require cross-disciplinary, multi-hop reasoning rather than commonsense recall. KnowThinker is the reasoning engine: a vision-language model that ingests a prompt and optional reference image, emits an explicit reasoning trace and a refined text instruction, and then calls a frozen image editor—so all learning happens in the thinker. BPPO is the training algorithm: Bootstrapping Reward Rectification (BRR) recalibrates each modality's advantage with the paired modality's reward as a ba

Load-bearing premise

The benchmark's scores—including the new Knowledge Reasoning and Rationale Alignment metrics—come from a proprietary multimodal evaluator that has not been validated against human expert ratings, so the claim that ExpertVerse measures expert-level knowledge correctness depends on that evaluator's 1–5 scores being faithful.

What would settle it

A study in which domain experts independently rate a random sample of generated images on the same four dimensions (Visual Quality, Input Consistency, Knowledge Reasoning, Rationale Alignment) would settle it: if the proprietary evaluator's scores correlate weakly (e.g., r < 0.5) with expert scores, or if a no-reasoning baseline that simply echoes the input image scores high on Knowledge Reasoning, the benchmark's core measurement claim would be refuted.

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

If this is right

  • If ExpertVerse measures what it claims, then knowledge-intensive generation is a first-class evaluation dimension, and future editing benchmarks should stratify by cognitive ability and expert discipline rather than by task type alone.
  • KnowThinker's gains on a frozen editor (e.g., +20.6 on RISEBench) imply that large reasoning-trace datasets and RL can substitute for retraining or scaling the diffusion backbone—a cheaper path to knowledge-aware models.
  • The zero-shot transfer of refined instructions across four editor backbones implies that a single reasoner can upgrade proprietary and open editors alike without touching their weights.
  • BPPO's ablations show that both reward rectification and conflict-aware fusion are needed; this suggests that any multi-reward RL training for multimodal generation should handle cross-modal credit assignment explicitly.
  • The per-domain results (e.g., a top proprietary model scoring 62.04 in science I2I but 97.58 in fashion) imply that even frontier models have lopsided knowledge, so diagnoses from ExpertVerse could guide targeted data collection for both editors and reasoners.

Where Pith is reading between the lines

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

  • The mechanism behind BPPO is reward-agnostic; the same bootstrapped cross-modal rectification and sign-based Pareto fusion could transfer to other multi-reward settings such as text-to-video generation or multimodal RLHF, though the paper does not test this.
  • Because the benchmark is held-out and uses a proprietary evaluator, a natural audit would be to open the scoring prompts and run a human-correlation study; if the evaluator correlates with expert ratings, ExpertVerse could serve as a reusable qualification exam for image-generation models.
  • The taxonomy's orthogonality suggests a diagnostic use: plotting a model's score by sub-discipline yields a knowledge-profile fingerprint, letting developers see, for instance, that a model is strong in fashion but weak in data visualization—an application the paper only sketches.
  • A controlled test to isolate 'knowledge' from 'verbosity' would be to feed KnowThinker's refined prompts to the editor with the rationale stripped out; the paper's ablations on instruction rewriting hint at this but do not fully separate prompt detail from factual grounding.

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 introduces ExpertVerse, a benchmark of 1,611 expert-annotated instances spanning a 9-capability-by-8-discipline taxonomy (58 sub-disciplines) for knowledge-intensive visual synthesis, covering single-image editing, multi-image composition, and text-to-image generation. It also presents ExpertVerse-100K, a synthetic training corpus with step-by-step reasoning traces and knowledge-anchored rationales. On the modeling side, the paper trains KnowThinker, a VLM reasoner that produces explicit reasoning traces and refined instructions while a frozen image editor executes synthesis, and proposes Bootstrapped Pareto Policy Optimization (BPPO), combining Bootstrapping Reward Rectification (BRR) and Conflict-Aware Pareto Advantage Fusion (CPAF). Experiments report state-of-the-art open-source results on ExpertVerse (69.73%), RISEBench (40.0%), KRIS-Bench (83.02%), UniREditBench (65.34%), and WiseEdit/WiseEdit-Complex (77.3/82.4), with the central claim that current models have severe knowledge blind spots in knowledge-intensive visual synthesis and that KnowThinker narrows the gap to proprietary systems.

Significance. If the ExpertVerse measurements are valid, the benchmark addresses a genuine gap in evaluating knowledge-intensive image generation, and the thinker-centric paradigm with BPPO is a useful contribution. The external benchmark results (RISEBench, KRIS-Bench, UniREditBench, WiseEdit) provide partly independent evidence that KnowThinker is a strong open-source reasoner, and the extensive ablations and cross-backbone transfer experiments are valuable. However, the load-bearing ExpertVerse numbers rest on a proprietary, unvalidated multimodal evaluator, the training set shares its handcrafted seeds/taxonomy with the held-out test set, and the BPPO Pareto-improvement claim is asserted without proof. Since the benchmark and evaluation prompts are not released, the headline claims are not yet independently checkable. The paper's central contributions are defensible in principle, but the current evidence is not sufficient to accept the ExpertVerse-specific conclusions.

major comments (5)
  1. [Evaluation Metrics] The 1-5 scores for Knowledge Reasoning and Rationale Alignment are produced by a proprietary multimodal evaluator, with no human-evaluator correlation, no inter-annotator agreement, and no release of the scoring prompts or the benchmark itself; the Appendix that supposedly contains the prompts is absent from this version. Since Table 2 and Figs. 4-6 all rely on this evaluator, the central claim that ExpertVerse exposes knowledge blind spots and that KnowThinker achieves 69.73% cannot be independently verified. Provide human correlation, IAA, error bars, and release the evaluation set/prompts.
  2. [Data Collection and Annotations] ExpertVerse-100K is synthesized from the same handcrafted seed prompts and the same 9x8 taxonomy that structure the held-out ExpertVerse test set. Even if individual instances are disjoint, training on ExpertVerse-100K and evaluating on ExpertVerse partly measures distribution overlap rather than general expert-level ability. The manuscript states the test set is 'completely held-out' but this is not sufficient; please demonstrate that test instances are not generated from the same seed templates (e.g., by releasing seeds and showing non-overlap) or evaluate on independently sourced out-of-distribution tasks.
  3. [Eq. (4), Bootstrapped Pareto Policy Optimization] The claim that the min-operator in Eq. (4) 'ensures stable and pareto-improving policy refinement' is unsupported and, as stated, false in general. In a conflict with A_img=+1 and A_cot=-5, selecting A=+1 moves toward a response that is worse than baseline for the cot objective; selecting A=-5 harms the image objective. Neither choice is Pareto-improving. Provide a formal proof under explicit assumptions about the reward model and policy parameterization, or substantially weaken the claim to say that the heuristic stabilizes updates.
  4. [Eq. (2), Bootstrapping Reward Rectification] The cross-modal anchor s_i^{\bar m} is never formally defined, and Eq. (2) subtracts (s_i^{\bar m} - \mu^{\bar m})/\sigma^m, which mixes reward units/statistics of the two domains in a dimensionally inconsistent way (the numerator is in the paired domain's units while the denominator is in domain m's standard deviation). Clarify whether s is a raw reward, a normalized score, or an advantage, and derive the rectified advantage with consistent normalization. The free parameters \beta and \alpha also have no sensitivity analysis; report how results vary with these choices.
  5. [Tables 2 and 9] There is an internal inconsistency: Table 2 reports KnowThinker (SFT) at 64.22 on ExpertVerse, but Table 9's SFT-only ablation (row b) reports 66.22 on ExpertVerse. These numbers should agree if they describe the same model and evaluation setup. Please reconcile this discrepancy, as it directly affects the claimed contribution of SFT and the ablation story.
minor comments (5)
  1. [Appendix] The text says 'Full scoring prompts are provided in Appendix,' but no appendix is present in the submitted version. Add the appendix or remove the reference.
  2. [Table 2] In the GPT-Image-2 row, the Fash.&Beau and Life&Consu columns are identical (97.58 for T2I and 75.11 for I2I). This looks like a copy-paste error and should be corrected.
  3. [Tables 3 and 4] Model naming is inconsistent: 'Qwen-Image-Edit-2511' appears in Table 3, while 'Qwen-Image-Edit' is used elsewhere; also 'Know-Thinker' and 'KnowThinker' are mixed. Please unify.
  4. [Availability] The benchmark, ExpertVerse-100K, and evaluation code are not released. Add a URL or an availability statement for reproducibility.
  5. [Figure 6] The caption says 'zero-shot transferability of KnowThinker' when the figure shows instruction refinement applied to four different editors. Clarify what is zero-shot in this setting, since KnowThinker is trained on ExpertVerse-100K and then applied to other editors.

Circularity Check

0 steps flagged

No significant circularity: held-out ExpertVerse evaluation and independent external benchmarks ground the central claims; proprietary-evaluator validity is a measurement concern, not a circular reduction.

full rationale

The paper explicitly separates training from evaluation: "To ensure rigorous evaluation, the benchmark is completely held-out and has never been exposed to the model during SFT, RL, reward modeling, or preliminary data distillation." KnowThinker is trained on ExpertVerse-100K and evaluated on the held-out 1,611-instance ExpertVerse set, so the headline ExpertVerse result is not a self-definitional fit to the training objective by construction. The strongest SOTA claims are also supported on external benchmarks with independent protocols (RISEBench, KRIS-Bench, UniREditBench, WiseEdit), which do not depend on ExpertVerse-100K or the authors' proprietary evaluator. The main validity threat — "we utilize a separate proprietary multimodal evaluator as the evaluation model and design dimension-specific prompts" with no human correlation and scoring prompts deferred to an absent appendix — is a measurement-validity gap, not a circular derivation: the paper's equations (BRR advantage in Eq. 2, CPAF fusion in Eq. 3) are optimization updates with hand-set constants β=0.3 and α=0.5, not predictions reconstructed from the evaluation metric. The thinker-centric paradigm is motivated by cited prior work that may share authors, but the paper also provides direct comparisons against joint-training baselines on external benchmarks, so the design choice is not supported solely by self-citation. Overall, no load-bearing step reduces to its own input by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central results rest on unverifiable data-generation and evaluation assumptions. The train/eval split is asserted but not auditable, and BPPO's key Pareto claim is adopted as a design axiom rather than derived. Hyperparameters β, α, and the distillation thresholds are hand-set, with limited or no sensitivity analysis.

free parameters (4)
  • β (BRR cross-modal anchor strength) = 0.3
    Set in Implementation Details; no sensitivity analysis reported; controls how much the paired-modality reward rectifies each advantage in Eq. 2.
  • α (CPAF fusion weight) = 0.5
    Set in Implementation Details; used in Eq. 3 for sign-agreement fusion; no sensitivity analysis reported.
  • RL difficulty distillation thresholds τ_low, τ_high, τ_var = τ_low=2, τ_high=4, τ_var determined by thresholding rollout reward distribution
    Used to construct D_RL by keeping moderate-mean, high-variance instructions; thresholds are chosen to create the training set and affect all RL results.
  • Conservative conflict-resolution choice = min-operator over advantages
    Selected after comparing Drop/Mean/PCGrad/Image-Priority/Conservative in Tab. 8; model selection on the benchmark can inflate reported gains.
axioms (4)
  • domain assumption Synthetic CoT traces and knowledge-anchored rationales in ExpertVerse-100K are factually correct and knowledge-dense.
    Data Collection states these are generated by a proprietary VLM, filtered by an automated VLM, and manually verified by three PhD specialists; the correctness of 100K traces is not independently audited.
  • domain assumption The proprietary multimodal evaluator's scores are valid and unbiased measures of the four metrics.
    Evaluation Metrics: scoring prompts are not shown (Appendix omitted); no human correlation or inter-rater agreement is reported.
  • domain assumption ExpertVerse evaluation instances were never seen by the model during SFT, RL, or reward modeling.
    Data Collection asserts the set is 'completely held-out', but since ExpertVerse and ExpertVerse-100K are generated from the same handcrafted seeds via the same proprietary VLM, distributional overlap is plausible and cannot be checked without release.
  • ad hoc to paper The CPAF min-operator yields Pareto-improving updates under conflicting advantages.
    Eq. 3-4 states this without theorem or proof; the min-operator bounds updates, but no Pareto guarantee is established.

pith-pipeline@v1.3.0-alltime-deepseek · 22267 in / 11640 out tokens · 113322 ms · 2026-08-01T12:41:21.275854+00:00 · methodology

0 comments
read the original abstract

Recent advances in multimodal generative models have enabled instruction-based image generation to move beyond semantic manipulation to knowledge-driven visual reasoning. However, these methods focus on explicit commonsense reasoning, shallow causal understanding, and direct knowledge recall, failing at knowledge-intensive generation. We develop \textbf{ExpertVerse}, a capability-centric benchmark to evaluate generative models via knowledge-intensive lens. ExpertVerse stratifies reasoning generation across an orthogonal taxonomy of \textit{9 cognitive capabilities} and \textit{8 expert disciplines}, yielding \textit{58 sub-disciplines}. We curate 1,611 expert-annotated instances covering single-image editing, multi-image composition, and text-to-image generation. We further develop an automated workflow to produce \textbf{ExpertVerse-100K}, a large-scale dataset with reasoning traces and knowledge-anchored rationale annotations. Based on this, we train \textbf{KnowThinker} with RL fine-tuning, a VLM reasoning engine with world knowledge that jointly generates thinking processes and refined instructions. Towards the cross-modal credit misalignment and multi-objective gradient conflicts in multi-reward optimization, we propose a tailored Bootstrapped Pareto Policy Optimization (BPPO), which synergizes Bootstrapping Reward Rectification (BRR) and Conflict-Aware Pareto Advantage Fusion (CPAF). Extensive results of both open-source and proprietary models exposes critical reasoning deficits, highlighting imperative for knowledge-intensive benchmarks towards next-generation visual generation.

Figures

Figures reproduced from arXiv: 2607.19341 by Jinsong Lan, Mengting Chen, Xiaoyong Zhu, Xuetao Feng, Yongchao Du, Yuan Wang.

Figure 3
Figure 3. Figure 3: Statistical overview of our ExpertVerse across task [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: Impact of refined instructions across the evaluation [PITH_FULL_IMAGE:figures/full_fig_p010_6.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

24 extracted references · 16 linked inside Pith

  1. [2]

    Chang, D.; Cao, M.; Shi, Y.; Liu, B.; Cai, S.; Zhou, S.; Huang,W.;Wetzstein,G.;Soleymani,M.;andWang,P.2025

    Hidream-i1: A high-efficient image generative foundation model with sparse diffusion transformer.arXiv preprint arXiv:2505.22705. Chang, D.; Cao, M.; Shi, Y.; Liu, B.; Cai, S.; Zhou, S.; Huang,W.;Wetzstein,G.;Soleymani,M.;andWang,P.2025. Bytemorph: Benchmarking instruction-guided image editing with non-rigid motions.arXiv preprint arXiv:2506.03107. Chen, ...

    Pith/arXiv arXiv 2025

  2. [4]

    Deng, C.; Zhu, D.; Li, K.; Gou, C.; Li, F.; Wang, Z.; Zhong, S.; Yu, W.; Nie, X.; Song, Z.; et al

    Gemini 2.5: Pushing the fron- tier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261. Deng, C.; Zhu, D.; Li, K.; Gou, C.; Li, F.; Wang, Z.; Zhong, S.; Yu, W.; Nie, X.; Song, Z.; et al

    Pith/arXiv arXiv

  3. [5]

    Emerging prop- erties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683. Google

    Pith/arXiv arXiv

  4. [6]

    Han, F.; Wang, Y.; Li, C.; Liang, Z.; Wang, D.; Jiao, Y.; Wei, Z.; Gong, C.; Jin, C.; Chen, J.; et al

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948. Han, F.; Wang, Y.; Li, C.; Liang, Z.; Wang, D.; Jiao, Y.; Wei, Z.; Gong, C.; Jin, C.; Chen, J.; et al

    Pith/arXiv arXiv

  5. [7]

    Hurst, A.; Lerer, A.; Goucher, A

    UniREditBench: AUnifiedReasoning-basedImageEditingBenchmark.arXiv preprint arXiv:2511.01295. Hurst, A.; Lerer, A.; Goucher, A. P.; Perelman, A.; Ramesh, A.; Clark, A.; Ostrow, A.; Welihinda, A.; Hayes, A.; Rad- ford, A.; et al

    arXiv

  6. [9]

    arXiv preprint arXiv:2601.03467

    ThinkRL-Edit: Thinking in ReinforcementLearningforReasoning-CentricImageEditing. arXiv preprint arXiv:2601.03467. Li, H.; Zhang, M.; Zheng, D.; Guo, Z.; Jia, Y.; Feng, K.; Yu, H.; Liu, Y.; Feng, Y.; Pei, P.; et al

    arXiv

  7. [10]

    Lin, B.; Li, Z.; Cheng, X.; Niu, Y.; Ye, Y.; He, X.; Yuan, S.; Yu, W.; Wang, S.; Ge, Y.; et al

    Editthinker: Unlocking iterative reasoning for any image editor.arXiv preprint arXiv:2512.05965. Lin, B.; Li, Z.; Cheng, X.; Niu, Y.; Ye, Y.; He, X.; Yuan, S.; Yu, W.; Wang, S.; Ge, Y.; et al

    arXiv

  8. [11]

    Liu, C.; Chen, M.; Huang, Y.; Wu, H.; Ju, C.; Xiao, S.; Lan, J.; and Wang, Y

    Uniworld-v1: High- resolutionsemanticencodersforunifiedvisualunderstanding and generation.arXiv preprint arXiv:2506.03147. Liu, C.; Chen, M.; Huang, Y.; Wu, H.; Ju, C.; Xiao, S.; Lan, J.; and Wang, Y. 2026a. Improving Human Image Animation via Semantic Representation Alignment. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognitio...

    Pith/arXiv arXiv

  9. [12]

    Liu, T.; Yan, H.; Chen, M.; Hu, T.; Yue, Z.; Pan, Z.; Lan, J.; Zhu, X.; Cheng, M.-M.; Zheng, B.; et al

    Step1x-edit: A practicalframeworkforgeneralimageediting.arXivpreprint arXiv:2504.17761. Liu, T.; Yan, H.; Chen, M.; Hu, T.; Yue, Z.; Pan, Z.; Lan, J.; Zhu, X.; Cheng, M.-M.; Zheng, B.; et al. 2026b. Continuous- Time Distribution Matching for Few-Step Diffusion Distilla- tion.arXiv preprint arXiv:2605.06376. Niu, Y.; Ning, M.; Zheng, M.; Jin, W.; Lin, B.; ...

    Pith/arXiv arXiv

  10. [13]

    Pan, K.; Chen, W.; Qiu, H.; Yu, Q.; Bu, W.; Wang, Z.; Zhu, Y.; Li, J.; and Tang, S

    Wise: A world knowledge-informed semantic evaluation for text-to-image generation.arXiv preprint arXiv:2503.07265. Pan, K.; Chen, W.; Qiu, H.; Yu, Q.; Bu, W.; Wang, Z.; Zhu, Y.; Li, J.; and Tang, S

    Pith/arXiv arXiv

  11. [14]

    Seedream, T

    Uni-cot: Towards unified chain- of-thought reasoning across text and vision.arXiv preprint arXiv:2508.05606. Seedream, T

    arXiv

  12. [15]

    Song,Q.;Shen,Y.;Chen,M.;Sun,H.;Lan,J.;Zhu,X.;Zheng, B.;andCao,L.2026

    Seedream 4.0: Toward next- generation multimodal image generation.arXiv preprint arXiv:2509.20427. Song,Q.;Shen,Y.;Chen,M.;Sun,H.;Lan,J.;Zhu,X.;Zheng, B.;andCao,L.2026. FashionChameleon:TowardsReal-Time andInteractiveHuman-GarmentVideoCustomization.arXiv preprint arXiv:2605.15824. Sun, H.; Yan, H.; Chen, M.; Song, Q.; Li, Y.; Cao, J.; Lan, J.; Zhu, X.; Zh...

    Pith/arXiv arXiv 2026

  13. [16]

    Sun, Q.; Cui, Y.; Zhang, X.; Zhang, F.; Yu, Q.; Wang, Y.; Rao, Y.; Liu, J.; Huang, T.; and Wang, X

    TryOnCrafter: Unleashing Camera Trajectories for Realistic Video Virtual Try-on via a Renderable 4D Try-on Proxy.arXiv preprint arXiv:2606.26092. Sun, Q.; Cui, Y.; Zhang, X.; Zhang, F.; Yu, Q.; Wang, Y.; Rao, Y.; Liu, J.; Huang, T.; and Wang, X

    Pith/arXiv arXiv

  14. [17]

    Wang, G.-H.; Zhao, S.; Zhang, X.; Cao, L.; Zhan, P.; Duan, L.; Lu, S.; Fu, M.; Chen, X.; Zhao, J.; et al

    Deepgen 1.0: A lightweight unified multimodal model for advancing image generation and editing.arXiv preprint arXiv:2602.12205. Wang, G.-H.; Zhao, S.; Zhang, X.; Cao, L.; Zhan, P.; Duan, L.; Lu, S.; Fu, M.; Chen, X.; Zhao, J.; et al. 2025a. Ovis-u1 technical report.arXiv preprint arXiv:2506.23044. Wang, Y.; Li, Z.; Zang, Y.; Bu, J.; Zhou, Y.; Xin, Y.; He,...

    arXiv

  15. [18]

    Wu, C.; Li, J.; Zhou, J.; Lin, J.; Gao, K.; Yan, K.; Yin, S.-m.; Bai,S.;Xu,X.;Chen,Y.;etal.2025a

    Holistic- motion2d: Scalable whole-body human motion generation in 2d space.arXiv preprint arXiv:2406.11253. Wu, C.; Li, J.; Zhou, J.; Lin, J.; Gao, K.; Yan, K.; Yin, S.-m.; Bai,S.;Xu,X.;Chen,Y.;etal.2025a. Qwen-imagetechnical report.arXiv preprint arXiv:2508.02324. Wu, C.; Zheng, P.; Yan, R.; Xiao, S.; Luo, X.; Wang, Y.; Li, W.; Jiang, X.; Liu, Y.; Zhou,...

    Pith/arXiv arXiv

  16. [19]

    InProceedings of the IEEE International Conference on Computer Vision, 1926–1936

    DepthDance: Complex-pose Human ImageAnimationwithAppearance-agnosticDepthGuidance. InProceedings of the IEEE International Conference on Computer Vision, 1926–1936. Xiao, S.; Wang, Y.; Zhou, J.; Yuan, H.; Xing, X.; Yan, R.; Li, C.; Wang, S.; Huang, T.; and Liu, Z

    1926

  17. [20]

    Ye, J.; Jiang, D.; Wang, Z.; Zhu, L.; Hu, Z.; Huang, Z.; He, J.; Yan, Z.; Yu, J.; Li, H.; et al

    DDA-Thinker: Decoupled Dual-Atomic Reinforcement Learning for Reasoning-Driven Image Editing.arXiv preprint arXiv:2604.25477. Ye, J.; Jiang, D.; Wang, Z.; Zhu, L.; Hu, Z.; Huang, Z.; He, J.; Yan, Z.; Yu, J.; Li, H.; et al

    Pith/arXiv arXiv

  18. [21]

    Ye, Y.; He, X.; Li, Z.; Yuan, S.; Yan, Z.; Hou, B.; Yuan, L.; et al

    Echo-4o: Harnessing the power of gpt-4o synthetic images for improved image generation.arXiv preprint arXiv:2508.09987. Ye, Y.; He, X.; Li, Z.; Yuan, S.; Yan, Z.; Hou, B.; Yuan, L.; et al

    Pith/arXiv arXiv

  19. [24]

    InAdvances in Neural Information Processing Systems (NeurIPS)

    In-Context Edit: Enabling Instructional Image Editing with In-Context Generation in Large-Scale Diffusion Transform- ers. InAdvances in Neural Information Processing Systems (NeurIPS). ArXiv:2504.20690. Zhao, H.; Ma, X.; Chen, L.; Si, S.; Wu, R.; An, K.; Yu, P.; Zhang,M.;Li,Q.;andChang,B.2024. Ultraedit:Instruction- basedfine-grainedimageeditingatscale.Ad...

    Pith/arXiv arXiv 2024

  20. [38]

    Reasonedit:Towards reasoning-enhanced image editing models

    Yin,F.;Liu,S.;Han,Y.;Wang,Z.;Xing,P.;Wang,R.;Cheng, W.;Wang,Y.;Li,A.;Yin,Z.;etal.2026. Reasonedit:Towards reasoning-enhanced image editing models. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 23283–23293. Yu, T.; Kumar, S.; Gupta, A.; Levine, S.; Hausman, K.; and Finn, C

    2026

  21. [2020]

    Advances in Neural Information Processing Systems, 33: 5824–5836

    Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33: 5824–5836. Zhang,K.;Mo,L.;Chen,W.;Sun,H.;andSu,Y.2023. Mag- icbrush: A manually annotated dataset for instruction-guided image editing.Advances in Neural Information Processing Systems, 36: 31428–31449. Zhang, Z.; Xie, J.; Lu, Y.; Yang, Z.; and Yang, Y

    2023

  22. [2024]

    Jiang, H.; Fang, J.; Zhang, N.; Ma, G.; Wan, M.; Wang, X.; He,X.;andChua,T.-s.2025

    Gpt-4o system card.arXiv preprint arXiv:2410.21276. Jiang, H.; Fang, J.; Zhang, N.; Ma, G.; Wan, M.; Wang, X.; He,X.;andChua,T.-s.2025. Anyedit:Editanyknowledgeen- coded in language models.arXiv preprint arXiv:2502.05628. Jiao, S.; Lin, Y.; Zhong, Y.; She, Q.; Zhou, W.; Lan, X.; Huang, Z.; Yu, F.; Yu, Y.; Zhao, Y.; et al

    Pith/arXiv arXiv 2025

  23. [2025]

    Batifol, S.; Blattmann, A.; Boesel, F.; Consul, S.; Diagne, C.; Dockhorn, T.; English, J.; English, Z.; Esser, P.; Kulal, S.; et al

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631. Batifol, S.; Blattmann, A.; Boesel, F.; Consul, S.; Diagne, C.; Dockhorn, T.; English, J.; English, Z.; Esser, P.; Kulal, S.; et al

    Pith/arXiv arXiv

  24. [2026]

    Chen, X.; Feng, Y.; Chen, M.; Wang, Y.; Zhang, S.; Liu, Y.; Shen, Y.; and Zhao, H

    Tstars-Tryon 1.0: Robust and Realistic Virtual Try-On for Diverse Fashion Items.arXiv preprint arXiv:2604.19748. Chen, X.; Feng, Y.; Chen, M.; Wang, Y.; Zhang, S.; Liu, Y.; Shen, Y.; and Zhao, H. 2024a. Zero-shot image editing with reference imitation.Advances in Neural Information Processing Systems, 37: 84010–84032. Chen, X.; Liu, Z.; Chen, M.; Feng, Y....

    Pith/arXiv arXiv