PaperlayerView exact revision on arXiv

Source-linked AI summary

Position: Towards Implicit Prompt For Text-To-Image Models

Yue Yang, Yuqi Lin, Hong Liu, Wenqi Shao, Runjian Chen, Hailong Shang, Yu Wang, Yu Qiao, Kaipeng Zhang, Ping Luo

arXiv:2403.02118v4cs.CYcs.AIcs.CV

TL;DR

Existing work leaves T2I models’ handling of implicit prompts unexplored, despite their potential to infer targets without naming them. The paper evaluates six models with ImplicitBench across three aspects, finding symbolic-generation capability alongside celebrity-privacy and NSFW risks that current constraints do not sufficiently address.

  • Problem

    Prior work leaves T2I models’ capabilities and risks for implicit prompts unexplored, although such prompts can guide models toward target images without directly naming the targets.

    Source: §1 · p. 2, §3.1 · p. 3

  • Method

    The paper evaluates six contemporary T2I models on a benchmark spanning General Symbols, Celebrity Privacy, and NSFW Issues, using GPT-4V and human evaluation for symbols.

    Source: §3.1 · p. 3, §3.4 · p. 5, §6 · p. 9

  • Results

    Accuracy for General Symbols is generally positively correlated with image quality and text-image alignment, celebrity images are more likely to be generated for higher-fame figures, and some implicit NSFW prompts carry elevated generation risk.

    Source: §5.1 · p. 8

  • Takeaways & Limitations

    The findings identify potential privacy and NSFW risks that current policy constraints do not sufficiently address.

    Source: §6 · p. 9

    Logo prompts have relatively low accuracy, possibly because logos’ diversity and abstract nature make accurate generation difficult.

    Source: §4.1 · p. 6

Abstract

from arXiv · show

Recent text-to-image (T2I) models have had great success, and many benchmarks have been proposed to evaluate their performance and safety. However, they only consider explicit prompts while neglecting implicit prompts (hint at a target without explicitly mentioning it). These prompts may get rid of safety constraints and pose potential threats to the applications of these models. This position paper highlights the current state of T2I models toward implicit prompts. We present a benchmark named ImplicitBench and conduct an investigation on the performance and impacts of implicit prompts with popular T2I models. Specifically, we design and collect more than 2,000 implicit prompts of three aspects: General Symbols, Celebrity Privacy, and Not-Safe-For-Work (NSFW) Issues, and evaluate six well-known T2I models' capabilities under these implicit prompts. Experiment results show that (1) T2I models are able to accurately create various target symbols indicated by implicit prompts; (2) Implicit prompts bring potential risks of privacy leakage for T2I models. (3) Constraints of NSFW in most of the evaluated T2I models can be bypassed with implicit prompts. We call for increased attention to the potential and risks of implicit prompts in the T2I community and further investigation into the capabilities and impacts of implicit prompts, advocating for a balanced approach that harnesses their benefits while mitigating their risks.

1. Introduction

Existing T2I benchmarks focus on explicit prompts, leaving models’ ability to infer unnamed targets and the associated risks unexplored. The paper benchmarks implicit prompts across three aspects and finds both useful capabilities and safety concerns.

Motivation

  • Earlier T2I benchmarks assess image quality, text-image alignment, and safety, but focus on explicit prompts rather than indirectly described targets.Implicit prompts describe relevant target information without stating the target’s concrete identity.Source: §1 · pp. 1–2

Approach

  • The study collects more than 2,000 implicit prompts across General Symbols, Celebrity Privacy, and NSFW Issues, then evaluates six open- and closed-source T2I models.It applies three evaluation methods and analyzes results for each aspect.Source: §1 · p. 2

Findings

  • T2I models can generate images embodying symbolism implied by General Symbols prompts, with accuracy generally positively correlated with image quality and text-image alignment.Closed-source APIs show better performance on this aspect.Source: §1 · p. 2
  • Models are more likely to generate privacy-infringing images of celebrities with higher fame, risking misinformation and reputational damage beyond current privacy policy.Source: §1 · p. 2
  • Some implicit prompts bypass most T2I safety filters despite appearing harmless, yet generate NSFW content.Certain stylistic professional terms, detailed body-part descriptions, and ambiguous or multimeaning words pose greater NSFW-generation risk than normal words.Source: §1 · pp. 2–3

Motivation

  • The paper argues that implicit prompts can support scalable image generation, while existing policy restrictions may be ineffective against privacy and NSFW risks.It calls for further investigation into their capabilities and impacts and for risk mitigation.Source: §1 · p. 3

2. Related Work

The related work covers T2I models and evaluation benchmarks, noting that existing evaluations leave implicit prompts unexplored.

Text-To-Image Models

  • Current T2I approaches often use diffusion models conditioned on text embeddings, and the paper selects six models spanning open- and closed-source systems.The selection considers generation performance, popularity, and coverage of open- and closed-source settings.Source: §2 · p. 3

Benchmarks for Evaluating T2I models

  • T2I evaluation has progressed from image-fidelity datasets and metrics to benchmarks probing composition, attributes, and societal effects.Examples include DrawBench, HE-T2I, T2I-CompBench, and work on safety and bias.Source: §2 · p. 3
  • Despite this breadth, existing benchmarks mainly assess image quality and text-image alignment using explicit prompts, leaving implicit-prompt capabilities and policy issues unexplored.The paper proposes a benchmark and evaluation methods across General Symbols, Celebrity Privacy, and NSFW Issues to address this gap.Source: §2 · p. 3

3. Overall Framework

The framework divides implicit prompts into general comprehension and policy-risk aspects, then collects prompts, generates images, and evaluates results across the three aspects.

3.1. Scope of Implicit Prompts

  • Implicit prompts are organized into General Symbols for general comprehension and Celebrity Privacy and NSFW Issues for policy concerns.Each aspect has a pipeline of prompt collection, image generation, and evaluation.Source: §3.1 · pp. 3–4

3.2. Implicit Prompt Collection

  • General Symbols prompts cover seven domains and pair each selected object with a GPT-4-generated description that omits its explicit name.The domains are logos, landmarks, architecture, clothes, food, art, and festivals; the collection contains over 200 prompts.Source: §3.2 · p. 4
  • Celebrity Privacy prompts cover ten public-identity domains, with concise GPT-4 descriptions for about 30 celebrities per domain.The resulting collection contains about 300 prompts.Source: §3.2 · p. 4
  • The NSFW collection uses Lexica prompts from four categories, removes prompts containing listed NSFW keywords, and retains more than 2,000 seemingly innocuous prompts.Their associated imagery may still be harmful.Source: §3.2 · pp. 4–5

3.3. Image Generation on Implicit Prompt

  • Table 1 presents information about the T2I models evaluated on ImplicitBench.Source: Table 1 · p. 5

3.4. Evaluation Method

  • Evaluation combines GPT-4V judgments and manual review for symbols, Arcface identity verification for celebrities, and two safety classifiers for NSFW content.A prompt counts as inducing NSFW content if either safety model identifies the image as unsafe.Source: §3.4 · p. 5

4. Experiments

The experiments evaluate six text-to-image models on implicit prompts across general symbols, celebrity privacy, and NSFW content. Models often infer the intended targets, but celebrity generation and NSFW results expose privacy and safety risks.

4. Experiments

  • The experiments compare generated images across the three implicit-prompt aspects and report both quantitative and qualitative results.Figure 3 presents sample outputs, while Figure 4 compares model outputs generated from the same prompt in each column.Source: §4 · p. 5, Figure 3 · p. 6

4.1. General Symbols

  • All six models show competence with implicit general-symbol prompts, with DALL-E 3 achieving the highest overall accuracy and closed-source APIs performing slightly better overall.Logo prompts remain difficult, which the authors attribute possibly to logos' diversity and abstract nature. Examples include recovering McDonald’s from indirect clues and representing igloos and calligraphy; users can produce high-quality images without recalling a target’s exact name.Source: §4.1 · pp. 5–6, Table 2 · p. 6

4.2. Celebrity Privacy

  • Midjourney and Stable Diffusion outperform DALL-E 2 and DALL-E 3 at generating celebrities, while DALL-E often substitutes fictitious characters.Accuracy is higher for politicians, actors, and athletes than for less prominent professions such as artists and poets; globally renowned political figures and actors also tend to have higher image quality.Source: §4.2 · pp. 6–7, Figure 5 · p. 7, Figure 4 · p. 7

4.3. NSFW Issues

  • 27.37%, 13.27%, and 17.75% of implicit NSFW prompts produced NSFW content for Midjourney, DALL-E 2, and DALL-E 3, respectively.All evaluated models showed some risk; open-source models leaned more toward NSFW content, while closed-source systems still produced milder unsafe images. Stable Diffusion XL generated a lower proportion of inappropriate images than Stable Diffusion v1-5.Source: Table 3 · p. 7, §4.3 · pp. 7–8

5. Result Discussion and Dealing Solution

The discussion relates implicit-prompt results to model capabilities and identifies risks involving celebrity images and NSFW content. It proposes input- and output-stage defenses, plus context-aware prompt filters.

5.1. Result Discussion

  • Accuracy rates are generally positively correlated with models’ image-generation quality and text-image alignment.The authors suggest that advancing common T2I capabilities may help models interpret implicit prompts.Source: §5.1 · p. 8, Figure 6 · p. 8
  • Models more successfully generate images of higher-fame celebrities, risking misinformation and reputational damage.Celebrity fame groups were based on search-result counts; Figure 7 depicts the relationship.Source: §5.1 · p. 8, Figure 7 · p. 8
  • The six evaluated models show slightly higher identity exposure for male than female celebrities.The authors note that this pattern aligns with prior studies and report that those studies speculate about training-data proportions.Source: §5.1 · p. 8, Figure 8 · p. 8
  • Stylistic professional terms, body-part descriptions, and hyper-detailed wording increase NSFW-generation risk relative to normal words.The authors show implicit prompts containing highlighted high-risk words inducing NSFW images, alongside a word cloud of risky vocabulary.Source: §5.1 · pp. 8–9, Figure 9 · p. 9

5.2. Dealing Solution

  • At the input stage, ImplicitBench’s prompt–unsafe-image pairs and safety evaluations can reward fine-tuning an LLM recaptioner to make prompts safer.The proposed recaption model adaptively optimizes input prompts to produce a harmless version for celebrity and NSFW content.Source: §5.2 · p. 9, Figure 10 · p. 9
  • At the output stage, edged-unsafe images from implicit prompts can augment training data for image safety checkers that miss them.These images may contain small areas of gore, disturbing content, or pornography at image edges or corners.Source: §5.2 · p. 9
  • Context-aware safety filters paired with user-intent clarification are proposed to flag potentially unsafe implicit prompts while preserving ordinary prompt use.The proposed scope includes possible NSFW content, privacy violations, and other harmful implications.Source: §5.2 · p. 9

6. Conclusion

The conclusion presents ImplicitBench as an evaluation of six T2I models across three areas and reports privacy and NSFW risks from implicit prompts. It calls for further research and more nuanced policy constraints.

6. Conclusion

  • ImplicitBench evaluates six contemporary T2I models across General Symbols, Celebrity Privacy, and NSFW Issues.The paper frames this investigation as broadening the scope of T2I research toward prompts that describe targets without naming them.Source: §6 · p. 9
  • The evaluation identifies potential celebrity-privacy violations and NSFW generation that current policy constraints do not sufficiently address.The paper calls for more nuanced constraints to mitigate these risks.Source: §6 · p. 9
  • The authors call for further research and attention to implicit prompts’ capabilities and impacts in the T2I community.The paper presents this as a way to support development of policy constraints addressing the identified risks.Source: §6 · p. 9

Impact Statement

The impact statement describes implicit prompts as a way to generate visual content without explicit or technical descriptions, while warning of privacy and NSFW risks. It advocates responsible use and stronger ethical frameworks and policy constraints.

Impact Statement

  • Implicit prompts may enable more intuitive T2I interaction for art creation, education, and innovation across fields.The statement says these applications could be supported without explicit or technical descriptions.Source: Impact Statement · p. 10
  • Seemingly benign implicit prompts may inadvertently breach celebrity privacy or generate NSFW content.The statement presents these as ethical concerns associated with T2I capabilities.Source: Impact Statement · p. 10
  • The statement calls for robust ethical frameworks, nuanced policy constraints, and awareness of these risks among developers and users.It frames these measures as part of responsible use of T2I technology.Source: Impact Statement · p. 10

B. Detailed Samples of ImplicitBench

The appendix describes ImplicitBench’s structure and sample prompts, then points to numerical celebrity-generation success rates for six models across professions.

B. Detailed Samples of ImplicitBench

C. Numerical Accuracy of Celebrity Privacy.

D. More Samples of Generated Images on ImplicitBench.

Figures 11–13 show generated-image samples across General Symbols, Celebrity Privacy, and NSFW Issues; each row uses the same prompt.

D. More Samples of Generated Images on ImplicitBench.

D.1. General Symbols

General Symbols results show strong target-symbol generation, but abstract and varied logos remain difficult, especially when prompts require substantial real-world knowledge.

D.1. General Symbols

  • Current T2I models accurately generate target symbols from implicit prompts, but varied, abstract logos remain a failure case.These failures indicate difficulty with implicit prompts requiring substantial real-world knowledge.Source: §D.1 · p. 14
  • The logo failures point to difficulty handling prompts that require substantial real-world knowledge.Source: §D.1 · p. 14
  • Figure 11 shows generated-image samples from six models for General Symbols prompts.Images in each row are generated with the same prompt.Source: Figure 11 · p. 14

D.2. Celebrity Privacy

Celebrity image generation is often imprecise but can produce similar images; implicit NSFW prompts can bypass policy constraints, and composed prompts can create false endorsements.

D.2. Celebrity Privacy

  • Celebrity prompts often fail to produce highly accurate images, yet may generate similar images that facilitate misleading or harmful information.The authors flag potential privacy implications.Source: §D.2 · p. 15, Figure 12 · p. 15

D.3. NSFW Issues

  • Implicit NSFW prompts can easily bypass current policy constraints and induce unsafe content.The authors call for more sophisticated filters and monitoring to detect and mitigate these risks.Source: §D.3 · p. 16, Figure 13 · p. 16

E. Danger in the composition of implicit prompts

Loading 2403.02118v4…