This repository contains the code to implement methods and experiments presented in the paper: Fantastic Copyrighted Beasts and How (Not) to Generate Them, by Luxi He^*1, Yangsibo Huang^*1, Weijia Shi^*2, Tinghao Xie¹, Haotian Liu³, Yue Wang⁴, Luke Zettlemoyer², Chiyuan Zhang, Danqi Chen¹, Peter Henderson¹.

^*Equal contribution

¹Princeton University, ²University of Washington, ³University of Wisconsin-Madison, ⁴University of Southern California

Website | Paper

• Environment Setup
• Image Generation
• Input Prompts
• Character Detection
• Consistency Evaluation
• Citation

We provide the environment for our experiments in environment.yml. You can reproduce the environment using conda env create -f environment.yml . Note that you need to specify environment path within the environment.yml file.

You can generate images using one of our five image-generation models and one video-generation model by calling the main.py function. For example:

python main.py \
    --model playground-v2.5-1024px-aesthetic \
    --output_dir 'results' \
    --data prompts/targets_dalle_prompt.csv \
    --neg_prompt keyword-5-and-laion-5 \
    --run_id 1

Specification of arguments:

• --model: Currently support these following models for generation:
  • playground-v2.5-1024px-aesthetic, Playground v2.5
  • stable-diffusion-xl-base-1.0, SDXL v1.0
  • PixArt-XL-2-512x512, PixArt
  • IF-I-XL-v1.0, DeepFloyd-IF
  • damo-vilab/text-to-video-ms-1.7b, VideoFusion
• --output_dir: Directory for storing model output
• --neg_prompt: Optional negative prompts to be included in generation. The default is "none". Our code currently supports the following negative prompts
  • "target": target character's name
  • "keyword-5-greedy":top 5 keywords selected through LM's greedy decoding
  • "keyword-5-embedding": top 5 keywords selected through embedding similarity comparison
  • "keyword-5-LAION": top 5 keywords selected through co-occurrence analysis with LAION.
  • You can also custimize your negative prompts by adding/modifying options in main.py to load relevant files.
• --data: Directory to input prompt for model generation.
• --run_id: Identifier for different runs. Will be reflected in output directory name.

In addition to the generated image, we store a log.json file containing a dictionary for each generation with the following key information:

{
  "target": "Mario",
  "prompt": "Mario",
  "negative_prompt": "",
  "image_path": "23.png",
  "has_error": false
}

The prompts folder contains different categories of input prompts. The full list of 50 characters in our evaluation suite $\mathsf{CopyCat}$ are listed in targets.csv. targets_dalle_prompt.csv is a collection of DALL·E rewritten prompts using target name as input to the rewriting pipeline.

• Keywords-based prompts: The keywords folder contains LM generated keywords with/ without greedy decoding. The keywords_co-occurrence folder contains prompts consisting of keywords having top-occurrence frequency with popular training corpora like LAION. The keywords_embedding folder contains keywords with high embedding space similarity to target characters.

• Description-based prompts: targets_60words.csv contains character descriptions around 60 words in length (to keep under CLIP's 77 tokens input length restriction). targets_60words_dallerewrite.csv are the rewritten description-based prompts.

The generation scripts for related keywords and rewritten prompts are located in the indirect_anchors_utils directory. Run python indirect_anchors_utils/collect_keywords.py or python indirect_anchors_utils/get_dalle_rewrite.py to generate these two sets of indirect anchor prompts respectively.

Generate binary score for each generated image: Example:

python detect_score/gpt4v_eval.py --dir OUTPUT_DIR

Replace OUTPUT_DIR with the folder containing output files.

The GPT-4v evaluator stores each evaluation in binary_result.json under OUTPUT_DIR in the following format:

{
  "id": "23.png",
  "character": "Mario",
  "score": 1,
  "has_error": false,
  "error_msg": ""
}

For video evaluations, which are conducted on the first, middle, and last frames of the video, run:

python gpt4_eval_video.py --dir OUTPUT_DIR

1. Collect key characteristics

To generate key characteristics for the characters:

python consistency_score/collect_entities.py --num_words N

Replace N with the number of desired key characteristics.

2. Run consistency check

Install t2v-metrics

pip install t2v-metrics

Run evaluation

python consistency_score/calculate_consistency.py --entity_file ENTITY_FILE --dir DIR

Replace ENTITY_FILE with the path to the file containing the key characteristics (generated in step 1). Replace DIR with the path to the directory containing the generated images you want to evaluate.

If you find our code and paper helpful, please consider citing our work:

@article{he2024fantastic,
  title={Fantastic Copyrighted Beasts and How (Not) to Generate Them},
  author={Luxi He and Yangsibo Huang and Weijia Shi and Tinghao Xie and Haotian Liu and Yue Wang and Luke Zettlemoyer and Chiyuan Zhang and Danqi Chen and Peter Henderson},
  journal={arXiv preprint arXiv:2406.14526},
  year={2024}
}