Research map / Video / Video generation
Fantastic Copyrighted Beasts and How (Not) to Generate Them
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds an evaluation framework (DETECT and CONS metrics over 50 copyrighted characters) showing that image and video generators reproduce copyrighted characters via indirect keywords, and that prompt rewriting combined with negative prompts substantially reduces character generation.
Problem
Image and video generators can reproduce copyrighted characters even when the character name is not in the prompt, and it is unclear how effective existing runtime mitigations such as prompt rewriting are at preventing this while preserving user intent.
Approach
Introduces a two-metric evaluation (GPT-4V-based copyright detection count and VQAScore-based consistency with user intent), a two-stage method for finding indirect anchor keywords/descriptions (LLM candidate generation plus embedding-similarity, LAION co-occurrence, or LM-ranked reranking), and tests negative prompts (including the character name and anchor keywords) combined with simulated DALL·
Key results
- Combining prompt rewriting with negative prompts (target name plus 5 EmbeddingSim and 5 LAION co-occurrence keywords) reduces DETECT on Playground v2.5 from 30.33 to 4.33 while CONS stays about 0.81.
- Prompt rewriting alone reduces DETECT on Playground v2.5 from 30.33 to 14.33, showing it is not fully effective.
- Keywords ranked by LAION co-occurrence with character names can generate more characters than 60-word descriptions when only 20 keywords are used; GPT-4V detector reaches 82.5% accuracy against majority human labels (Cohen's kappa 0.648).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine prompt rewriting with target name and 5 EmbeddingSim plus 5 Co-Occurrence-LAION negative keywords | Playground v2.5 · Rewritten character prompts | DETECT: 14.33_{\pm 2.62} -> 4.33_{\pm 0.47} | Helped |
| Combine prompt rewriting with target name and 5 EmbeddingSim plus 5 Co-Occurrence-LAION negative keywords | Playground v2.5 · Rewritten character prompts | CONS: 0.80_{\pm 0.01} -> 0.81_{\pm 0.00} | No clear effect |
| Add negative prompt set to prompt rewriting | Stable Diffusion XL, PixArt-alpha, DeepFloyd IF, and VideoFusion · Character generation | DETECT: 33.00_{\pm1.00} -> 1.67_{\pm 0.94}; 24.67_{\pm0.58} -> 4.67_{\pm 0.47}; 33.67_{\pm 1.53} -> 2.00_{\pm 1.00}; 28.33_{\pm | Helped |
| Add negative prompt set to prompt rewriting | Stable Diffusion XL, PixArt-alpha, DeepFloyd IF, and VideoFusion · Character generation | CONS: 0.73_{\pm 0.01} -> 0.77_{\pm 0.03}; 0.79_{\pm 0.01} -> 0.79_{\pm 0.01}; 0.71_{\pm 0.01} -> 0.72_{\pm 0.01}; 0.68_{\pm 0. | No clear effect |
| Add target name and 5 EmbeddingSim plus 5 Co-Occurrence-LAION keywords as negative prompts | Playground v2.5 · Character-name prompts | DETECT: 30.33_{\pm 1.89} -> 20.67_{\pm 0.47} | Helped |
| Add 5 EmbeddingSim plus 5 Co-Occurrence-LAION keywords, without target name, as negative prompts | Playground v2.5 · Character-name prompts | DETECT: 30.33_{\pm 1.89} -> 23.33_{\pm 3.30} | Helped |
| Use target name plus Co-Occurrence-LAION keywords rather than target name alone as negative prompts (ablation) | Playground v2.5 · Rewritten character prompts | DETECT: 7.67_{\pm 0.47} -> 5.00_{\pm 0.82} | Helped |
| Add 5 Co-Occurrence-LAION keywords to the target name as negative prompts (ablation) | Playground v2.5 · Character-name prompts | DETECT: 23.67_{\pm 2.62} -> 20.67_{\pm 2.05} | Helped |
| Use the top-ranked rather than bottom-ranked 60-word description by embedding similarity | Playground v2.5 · Indirect anchoring | DETECT: 16 -> 24 | Helped |
| Use 5 Co-Occurrence-LAION keywords as negative prompts rather than 5 LM-Ranked keywords (ablation) | Playground v2.5 · Character-name prompts | DETECT: 30.33_{\pm 1.89} -> 27.33_{\pm 0.00} | Helped |
| Use generic copyrighted-character wording rather than the specific character name in the evaluator prompt (ablation) | GPT-4V evaluator · 200 annotated images | Accuracy: 69\% -> 82.5\% | Helped |
| Use generic copyrighted-character wording rather than the specific character name in the evaluator prompt (ablation) | GPT-4V evaluator · 200 annotated images | Cohen-Kappa agreement: 0.39 -> 0.65 | Helped |
Surprising findings
- Two generic keywords (e.g., 'videogame, plumber') can consistently generate Mario without the character name.
- Character name anchoring is blocked by DALL·E 3 filters, but indirect anchoring descriptions still produce copyrighted characters on DALL·E 3.
Limitations
- The LLM-as-judge detector makes mistakes in judging character similarity, with imperfect accuracy (82.5%).
- Evaluation covers only runtime mitigations; models are assumed not to be modified, and the character-intent detector is assumed to be an oracle.
- Consistency metric captures only some aspects of user intent; legal sufficiency is jurisdiction-dependent.
Experimental setup
Datasets: LAION-2B C4 OpenWebText The Pile
Benchmarks: 50 copyrighted characters (custom list) 50 unseen characters (generalization check)
Metrics: DETECT CONS accuracy Cohen's kappa AUROC
Baselines: LM-Ranked keywords character name prompting DALL·E 3 prompt rewriting
Builds on: classifier-free guidance VQAScore LAION-2B Playground v2.5
Closest papers by meaning
- How to Evaluate and Mitigate IP Infringement in Visual Generative AI? (ICML 2025 · similarity 0.88)
- Evaluating Copyright Takedown Methods for Language Models (NeurIPS 2024 · similarity 0.81)
- Rethinking Artistic Copyright Infringements In the Era Of Text-to-Image Generative Models (ICLR 2025 · similarity 0.81)
- Bridging the Copyright Gap: Do Large Vision-Language Models Recognize and Respect Copyrighted Content? (AAAI 2026 · similarity 0.80)
- VA3: Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models (CVPR 2024 · similarity 0.80)
- Can Watermarking Large Language Models Prevent Copyrighted Text Generation and Hide Training Data? (AAAI 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.