Pith. sign in

REVIEW 3 major objections 5 minor 35 references

All fourteen tested text-to-image models still produce recognizable intellectual property; private models refuse some prompts but rates and coverage vary sharply by provider and by IP category.

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 · grok-4.5

2026-07-12 09:36 UTC pith:NWHUDEMR

load-bearing objection Solid comparative snapshot of IP guardrails across 14 T2I systems; the rates are useful even with an imperfect detector and coarse API proxies. the 3 major comments →

arxiv 2607.02582 v1 pith:NWHUDEMR submitted 2026-07-01 cs.CV

Evaluating Intellectual Property Guardrails of Generative Image Models: A Technical Report

Austin T. Hoag , Apostolos Modas , Yunhao Ba , Julienne M. LaChance , Jinru Xue , Wiebke Hutiri , Jan Simson , Tiffany Georgievski
show 4 more authors
Alex Towli Joseph Smith Yuki Mitsufuji Alice Xiang
This is my paper
classification cs.CV
keywords text-to-image modelsintellectual property guardrailsgeneration refusal rateIP occurrence ratefictional characterscelebrity likenesscommercial logosvision-language model detector
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 technical report builds a benchmark of about 1,800 prompts over 201 entities in five IP categories—celebrities, comic and manga characters, superheroes and villains, video-game characters, and commercial logos—then runs an automated pipeline against fourteen widely used text-to-image models. It asks how often inference-time IP guardrails refuse a generation and how often the models still emit images that a vision-language model judges to contain recognizable IP. Open-weight models refuse nothing; every private model refuses something, but refusal rates range from roughly 1 percent to roughly 50 percent, and commercial logos are refused least and generated most. Direct prompts that name the entity trigger more refusals than description-only prompts, yet every model still produces recognizable IP at non-trivial rates as of March 2026. The work therefore supplies a reproducible snapshot of how unevenly current guardrails cover the IP categories that matter to rightsholders and users.

Core claim

As of March 2026, every one of the fourteen text-to-image models tested—three open-weight and eleven private—readily generates images containing recognizable IP; private models activate IP guardrails at some level, but refusal frequency varies substantially by provider and by category, with commercial logos refused least often and successfully generated at the highest average rate.

What carries the argument

The IP Generation Benchmark plus automated pipeline: 1,809 structured direct and indirect prompts over 201 entities, API content-filter error codes as the refusal signal (Generation Refusal Rate), and a Qwen3-VL vision-language model as binary IP detector (IP Occurrence Rate).

Load-bearing premise

The API content-moderation error code is treated as a clean proxy for IP-specific guardrails, and a single vision-language model with modest F1 is treated as a faithful stand-in for human recognition of IP across popularity and region.

What would settle it

Re-run the same 1,809 prompts after a provider updates its IP guardrails or after human annotators re-label a large random sample of the generated images; a large drop in IP Occurrence Rate or a large mismatch with the VLM detector would overturn the March 2026 snapshot.

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

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

3 major / 5 minor

Summary. This technical report introduces an IP Generation Benchmark (201 entities, 1,809 prompts, five prompt types) and an automated pipeline to measure inference-time IP guardrails and IP generation propensity in fourteen text-to-image models (three open-weights, eleven proprietary). Outcomes are summarized by Generation Refusal Rate (GRR), inferred from API content-moderation error codes, and IP Occurrence Rate (IOR), inferred from a VLM (Qwen3-VL-30B-A3B) binary detector selected after a 14-model bake-off on a separately labeled 14,809-image set. Stratified results by prompt directness, IP category, region, and popularity support three main empirical claims: private models refuse at highly variable rates; commercial logos are refused least and generated most on average; and, as of March 2026, all tested models still produce recognizable IP at non-trivial rates.

Significance. If the measurements hold under stronger validation, the work provides a timely, comparative snapshot of IP guardrail coverage across widely used T2I systems and a reusable taxonomy/prompt design that goes beyond prior character-only studies. Strengths include the multi-axis taxonomy, explicit GRR/IOR definitions, a documented VLM detector bake-off with precision/recall reporting, open discussion of limitations, and clear legal/ethical disclaimers. The contribution is primarily empirical and infrastructural rather than theoretical; its value to the community is as a reproducible measurement protocol and a March 2026 baseline against which future guardrail changes can be compared.

major comments (3)
  1. [§5.3, §7, §7.4] Section 5.3 and Section 7: IOR is defined entirely by Qwen3-VL-30B-A3B (F1 71.3%, accuracy 67.7% on the separate labeled set). No human labels are collected on the actual evaluation images produced by the 14 T2I models (explicitly noted in §7.4). Given moderate precision and the paper’s own observation of over-prediction and likely popularity/region bias (§8.2), the absolute IOR levels and the claim that all models “readily” generate recognizable IP are not yet sufficiently grounded. A stratified human audit (even on a few hundred images spanning categories, popularity, and models) is needed to calibrate or bound IOR before the quantitative rates can support the strongest comparative claims.
  2. [§4.5, Appendix B] Section 4.5 and Appendix B: GRR treats CONTENT_FILTERED (and UNKNOWN for Nano Banana) as a proxy for IP guardrail activation. The APIs only expose generic content-moderation codes; non-IP filters could contribute. The manuscript argues that prompts lack overtly violent/sexual content, but this is not demonstrated with a controlled contrast (e.g., matched non-IP prompts, or provider documentation that the code is IP-specific). Without that, cross-model GRR differences and the “IP guardrails” interpretation remain only partially identified. A sensitivity analysis or provider-level clarification is load-bearing for Q1.
  3. [§3.3, §4.2, §8.2] Section 3.3, 4.2, and 8.2: Eastern/Western and more/less-mainstream strata are produced primarily by Llama 3.1-8B, and the same VLM detector is known to be more reliable on mainstream entities. The reported ~1.7× higher IOR for more popular entities may therefore partly reflect detector coverage rather than true generation propensity. The popularity and region claims in Q3 need either human verification on those strata or an explicit bias audit of the detector before they can be treated as model behavior rather than measurement artifact.
minor comments (5)
  1. [Figs. 8–10] Figures 8–10 omit model names “for sensitivity purposes,” which reduces interpretability of the galleries relative to the named models in Figs. 4–7. Consider anonymized but consistent column IDs that map to Table 5 in an appendix.
  2. [§4.5, Eq. (1)] Equation (1) defines IOR over |P| (all prompts) while summing only over generated images; this is intentional but easy to misread as conditioning on non-refusal. A one-sentence clarification next to Eq. (1) would help.
  3. [Table 1] Table 1 comparison to prior work is useful; adding approximate entity counts and whether human vs automatic labeling was used would make the scope claim sharper.
  4. [§4.5] Typo: “IP Occurence rate” should be “Occurrence” in §4.5.
  5. [Abstract, §6.1] The abstract and conclusion state “as of March 2026”; ensure all date stamps (evaluation month, arXiv header) are consistent in the camera-ready version.

Circularity Check

0 steps flagged

No circularity: empirical measurements of external T2I APIs/open-weights under a fixed prompt set; detector selected on a separate labeled set, not on evaluation outcomes.

full rationale

The paper is a technical report of observed rates (GRR via API CONTENT_FILTERED/UNKNOWN codes; IOR via binary VLM ip_present decisions) on 1,809 fixed prompts across 14 external models. There is no claimed first-principles derivation, uniqueness theorem, fitted parameter renamed as prediction, or self-definitional identity. The IP detector (Qwen3-VL-30B-A3B) was chosen by F1/accuracy on an independently constructed and human-QA'd labeled set of 14,809 real+synthetic images (Sec. 5.1–5.3); that selection set is disjoint from the main evaluation images. Metrics G and R are defined by simple proportions over external outcomes (Eqs. 1–2), not by construction from fitted inputs. Self-citations are ordinary background references and do not load-bear any uniqueness or ansatz that forces the reported rates. Limitations of the proxy signals are disclosed (Sec. 8.2, App. B) rather than hidden. The derivation chain is therefore self-contained empirical measurement; score 0 is the correct outcome.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The paper is an empirical systems evaluation, not a derivation. Load-bearing premises are operational definitions (what counts as refusal, what counts as recognizable IP) and tooling choices (which VLM, how popularity is labeled, how API errors map to IP). There are no fitted physical constants; free parameters are engineering thresholds and selection choices that affect measured rates. Invented entities are the benchmark constructs and metrics themselves.

free parameters (5)
  • CLIP token length cap (MAX_TOKENS=75)
    Hard truncation/compression threshold for prompts; can alter semantic content of long indirect prompts and thus measured GRR/IOR.
  • Indirect description length targets (30 vs 60 words) and keyword cap (50)
    Hand-chosen prompt-length settings that define the indirect-prompt conditions and affect refusal comparisons (notably Titan Image G2).
  • VLM detector selection (Qwen3-VL-30B-A3B by max F1)
    Detector chosen by highest F1 (71.3%) among 14 candidates; different selection criteria would change IOR labels.
  • Entity retention thresholds for detector eval set (≥10 IP-positive; ≤50 IP-negative per entity)
    Post-collection filters that shape the detector validation distribution and thus confidence in IOR.
  • Two images per non-refused prompt
    Sampling multiplicity chosen by authors; affects IOR variance but is not justified against alternatives.
axioms (5)
  • domain assumption API CONTENT_FILTERED (or UNKNOWN for Nano Banana) indicates IP-related guardrail refusal rather than unrelated moderation.
    Section 4.5 and Appendix B treat content-moderation error codes as the GRR signal; authors note coarseness but still use it as the IP-guardrail proxy.
  • domain assumption A VLM binary judgment of 'recognizable IP' is a valid operationalization of human-recognizable IP for comparative model ranking.
    Section 5 defines IOR via D(Ii) from the VLM; F1≈0.71 is reported but still used as the sole automatic judge on the main corpus.
  • ad hoc to paper Eastern/Western and more/less mainstream strata produced by Llama 3.1-8B (plus multi-LLM alias QA) adequately represent geographic and popularity diversity.
    Sections 3.3–4.2; authors later admit less-mainstream entities are unlikely near the true long tail because an LLM could name them.
  • domain assumption The five IP categories (celebrities, CCM, superheroes/villains, video-game characters, commercial logos) are sufficient for the stated claims about IP guardrail coverage within scope.
    Section 3.2 explicitly scopes out other IP types; claims are conditioned on this taxonomy.
  • standard math Standard classification metrics (precision/recall/F1/accuracy/name-match) on a mixed web+synthetic labeled set validate detector fitness.
    Section 5.3; ordinary ML evaluation practice.
invented entities (3)
  • IP Generation Benchmark (201 entities / 1,809 prompts with five prompt types) no independent evidence
    purpose: Provide a structured stimulus set spanning categories, regions, popularity, and direct/indirect intent.
    Constructed for this report; not an external standard dataset.
  • Generation Refusal Rate (GRR) and IP Occurrence Rate (IOR) no independent evidence
    purpose: Summarize guardrail activation and successful IP generation under the benchmark.
    Defined in Section 4.5 as paper-specific metrics G and R; useful but not independently standardized.
  • Four-dimension IP taxonomy (user intent, category, geography, popularity) no independent evidence
    purpose: Organize entity selection and stratified analysis for Q1–Q3.
    Section 3 design choice; binary East/West and mainstream splits are simplifications without external validation.

pith-pipeline@v1.1.0-grok45 · 23550 in / 3938 out tokens · 38618 ms · 2026-07-12T09:36:16.875832+00:00 · methodology

0 comments
read the original abstract

Generative image models are capable of producing images that bear a strong resemblance to, or replicate, recognizable intellectual property (IP). In this technical report, we present a benchmark and automated evaluation pipeline to test for evidence of IP guardrails in generative image models along with the propensity for these models to generate images with recognizable IP. The IP categories we tested include fictional characters, celebrity likeness, and commercial logos and do not encompass the full range of IP which may be implicated by image generation models. We evaluated fourteen widely used text-to-image models, including three self-hosted open weights models and eleven private models. While all of the private models were observed to refuse generations at some level due to IP guardrails, the frequency of generation refusals varied substantially among models. The refusal rates also varied considerably across the different IP categories tested. Commercial logos were refused least frequently and were successfully generated at the highest rate, on average. Though the rate varies, all models tested readily generated images containing recognizable IP as of March 2026.

Figures

Figures reproduced from arXiv: 2607.02582 by Alex Towli, Alice Xiang, Apostolos Modas, Austin T. Hoag, Jan Simson, Jinru Xue, Joseph Smith, Julienne M. LaChance, Tiffany Georgievski, Wiebke Hutiri, Yuki Mitsufuji, Yunhao Ba.

Figure 1
Figure 1. Figure 1: Prompt dataset construction pipeline. From a list of IP entities, prompts were generated [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The structured user prompt used for IP detection across all candidate VLMs. [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: IP detection performance of 14 candidate VLMs across five metrics: Precision, Recall, F1, [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Breakdown of prompt outcomes for each T2I model evaluated. O = open weights, P = [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: (Left) Generation refusal rate and (Right) IP occurrence rate, aggregated by prompt type. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: (Left) Generation refusal rate and (Right) IP occurrence rate, aggregated by IP category. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: (Left) Generation refusal rate and (Right) IP occurrence rate, aggregated by the intersection [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Generation results for the Apple logo across all models and prompt types (entity = "Apple", [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Generation results for Wonder Woman across all models and prompt types (entity = [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Generation results for Daniel Craig across all models and prompt types (entity = "Daniel [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Generation Refusal rate for each T2I model aggregated by the intersection of IP category [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: IP occurrence rate for each T2I model aggregated by the intersection of IP category [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Number of IP positive (left) and IP negative (right) images per IP category. [PITH_FULL_IMAGE:figures/full_fig_p026_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Images with multiple IP present for IP positive (left) and IP negative (right) cases. [PITH_FULL_IMAGE:figures/full_fig_p026_14.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

35 extracted references · 9 linked inside Pith

  1. [1]

    Andrea Bartz, Kirk Wallace Johnson, and Charles Graeber. Bartz v. anthropic pbc. https: //www.courtlistener.com/docket/69058235/bartz-v-anthropic-pbc/ , 2024. Case No. 3:24-cv-05417, U.S. District Court, Northern District of California

    arXiv 2024

  2. [2]

    Getty images and others v stability ai

    Mrs Justice Joanna Smith DBE. Getty images and others v stability ai. https:// www.judiciary.uk/judgments/getty-images-and-others-v-stability-ai/ , 2025

    2025

  3. [3]

    IL-2023-000007, High Court of Justice, Business and Property Courts of England and Wales, Chancery Division

    EWHC 38 (Ch), Case No. IL-2023-000007, High Court of Justice, Business and Property Courts of England and Wales, Chancery Division

    2023

  4. [4]

    The new york times company v

    The New York Times Company. The new york times company v. mi- crosoft corporation. https://www.courtlistener.com/docket/68117049/ the-new-york-times-company-v-microsoft-corporation/ , 2023. Case No. 1:23-cv-11195, U.S. District Court, Southern District of New York

    arXiv 2023

  5. [5]

    and others

    Disney Enterprises, Inc. and others. Disney enterprises, inc. et al. v. midjourney, inc. https://www.courtlistener.com/docket/70513159/ disney-enterprises-inc-v-midjourney-inc/ , 2025. Case No. 2:25-cv-04872, U.S. District Court, Central District of California

    arXiv 2025

  6. [6]

    Disney cease-and-desist to bytedance over seedance 2.0.https: //www.aol.com/articles/lawsuits-billion-dollar-deals-disney-183659024

    The Walt Disney Company. Disney cease-and-desist to bytedance over seedance 2.0.https: //www.aol.com/articles/lawsuits-billion-dollar-deals-disney-183659024. html, 2025. Cease-and-desist letter alleging unauthorized use of Disney, Star Wars, and Marvel characters in Seedance 2.0 AI video generator

    2025

  7. [7]

    code ann

    Tenn. code ann. § 47-25-1101 – ensuring likeness, voice, and image security act. https: //www.capitol.tn.gov/Bills/113/Bill/HB2091.pdf, 2024

    2024

  8. [8]

    Extracting training data from diffusion models

    Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In32nd USENIX security symposium (USENIX Security 23), pages 5253–5270, 2023

    2023

  9. [9]

    Detecting, explaining, and mitigat- ing memorization in diffusion models.The Twelfth International Conference on Learning Representations, 2024

    Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigat- ing memorization in diffusion models.The Twelfth International Conference on Learning Representations, 2024

    2024

  10. [10]

    PromptShield: A hybrid framework for copyright-safe text-to-image generation

    Shreya Garg. PromptShield: A hybrid framework for copyright-safe text-to-image generation. InWorkshop on Generative AI for Recommender Systems and Personalization (GenAIRecP) at KDD’25, Toronto, ON, Canada, 2025. August 4, 2025

    2025

  11. [11]

    How to evaluate and mitigate IP infringement in visual generative AI? InForty-second International Conference on Machine Learning, 2025

    Zhenting Wang, Chen Chen, Vikash Sehwag, Minzhou Pan, and Lingjuan Lyu. How to evaluate and mitigate IP infringement in visual generative AI? InForty-second International Conference on Machine Learning, 2025

    2025

  12. [12]

    Fantastic copyrighted beasts and how (not) to generate them

    Luxi He, Yangsibo Huang, et al. Fantastic copyrighted beasts and how (not) to generate them. InThe Thirteenth International Conference on Learning Representations, 2025. 21

    2025

  13. [13]

    Responsible ai and usage guidelines for imagen

    Google Cloud. Responsible ai and usage guidelines for imagen. https://docs.cloud. google.com/vertex-ai/generative-ai/docs/image/responsible-ai-imagen/ ,

  14. [14]

    Accessed: 2026-05-14

    2026

  15. [15]

    Dall·e 3 system card

    OpenAI. Dall·e 3 system card. https://openai.com/index/dall-e-3-system-card/ ,

  16. [16]

    Huggingface: Building the ai community that builds ai

    HuggingFace Team. Huggingface: Building the ai community that builds ai. https:// huggingface.co/, 2025

    2025

  17. [17]

    Diffu- sion art or digital forgery? investigating data replication in diffusion models

    Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffu- sion art or digital forgery? investigating data replication in diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6048–6058, 2023

    2023

  18. [18]

    Red-teaming the stable diffusion safety filter.arXiv preprint arXiv:2210.04610, 2022

    Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter.arXiv preprint arXiv:2210.04610, 2022. ML Safety Workshop, NeurIPS 2022

    Pith/arXiv arXiv 2022

  19. [19]

    Can large vision-language models detect images copyright infringement from genai?arXiv preprint arXiv:2502.16618, 2025

    Qipan Xu, Zhenting Wang, Xiaoxiao He, Ligong Han, and Ruixiang Tang. Can large vision-language models detect images copyright infringement from genai?arXiv preprint arXiv:2502.16618, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    T2i-riskyprompt: A benchmark for safety evaluation, attack, and defense on text-to-image model.The Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), 2026

    Zhang C., Zhang T., Wang L., Chen R., Li W., and Liu A. T2i-riskyprompt: A benchmark for safety evaluation, attack, and defense on text-to-image model.The Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), 2026

    2026

  21. [21]

    On copyright risks of text-to-image diffusion models.arXiv preprint arXiv:2311.12803, 2024

    Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, Haonan Wang, and Kenji Kawaguchi. On copyright risks of text-to-image diffusion models.arXiv preprint arXiv:2311.12803, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Measuring style similarity in diffusion models.arXiv preprint arxiv:2404.01292

    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

    Pith/arXiv arXiv

  23. [23]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Aaron Grattafiori, Abhimanyu Dubey, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Jiang, Alexandre Sablayrolles, et al

    Albert Q. Jiang, Alexandre Sablayrolles, et al. Mixtral of experts.arXiv preprint arXiv:2401.04088, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    DeepSeek-R1 incentivizes reasoning in LLMs through reinforce- ment learning.Nature, 645(8081):633–638, sep 2025

    Daya Guo, Dejian Yang, et al. DeepSeek-R1 incentivizes reasoning in LLMs through reinforce- ment learning.Nature, 645(8081):633–638, sep 2025

    2025

  26. [26]

    Gemma 3 technical report, 2025

    Gemma Team, Aishwarya Kamath, et al. Gemma 3 technical report, 2025

    2025

  27. [27]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors,Proceedings of the 38th International Conference on Machin...

    2021

  28. [28]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Shuai Bai, Yuxuan Cai, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    High-resolution image synthesis with latent diffusion models.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  30. [30]

    Stable diffusion 3 medium

    Stability AI. Stable diffusion 3 medium. https://huggingface.co/stabilityai/ stable-diffusion-3-medium, 2024

    2024

  31. [31]

    Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    Shuai Bai, Keqin Chen, et al. Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models.arXiv preprint arXiv:2504.10479, 2025

    Jinguo Zhu, Weiyun Wang, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models.arXiv preprint arXiv:2504.10479, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    Flux.1-dev

    Black Forest Labs. Flux.1-dev. https://huggingface.co/black-forest-labs/flux. 1-dev, 2025

    2025

  34. [34]

    Pixart-sigma-xl-2

    PixArt-alpha. Pixart-sigma-xl-2. https://huggingface.co/PixArt-alpha/ PixArt-Sigma-XL-2-1024-MS, 2025. 22 Figure 11: Generation Refusal rate for each T2I model aggregated by the intersection of IP category and prompt type. The plot is split into two rows for readability. Prompt type abbreviations: BD = basic direct, DD = direct dilution, ID = indirect des...

    2025

  35. [35]

    out of frame, 24 low quality, bad anatomy, watermark

    Amazon Web Services. Overview - responsible AI - Amazon Titan Image Generator. https://docs.aws.amazon.com/ai/responsible-ai/titan-image-generator/overview.html, 2024. Ac- cessed: 2026-05-08. A Per-Model Detailed Results Here we show generation refusal rate (GRR) and IP occurrence rate (IOR) plots for the intersections of attributes in our prompt dataset....

    2024