The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides explicit instructions on how to bypass safety filters built into the Diffusers library.
Evidence: The skill suggests setting pipe.safety_checker = None in SKILL.md and references/troubleshooting.md to resolve issues with black or noisy images.
[INDIRECT_PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by processing untrusted text prompts to generate images and reading local files for model training.
Ingestion points: The prompt parameter in DiffusionPipeline calls and the instance_images_path in the DreamBoothDataset class (in references/advanced-usage.md).
Boundary markers: None. Prompts are interpolated directly into the model pipelines without delimiters or instructions to ignore embedded commands.
Capability inventory: The skill can write files to the local system (image.save), save trained models (pipe.save_pretrained), and perform network operations to download models from the Hugging Face Hub.
Sanitization: No sanitization or validation of the input prompts or image metadata is demonstrated.
[EXTERNAL_DOWNLOADS]: The skill references several external sources for library installation and model weights.
Evidence: Instructions to install xformers from a specific PyTorch index URL and a recommendation to install the package directly from Meta's (facebookresearch) official GitHub repository in references/troubleshooting.md. It also fetches models and configuration from the Hugging Face Hub (a well-known service).

stable-diffusion-image-generation