How joycenerd/p4d works
Project type and purpose
• This is an open-source research tool (Prompting4Debugging, P4D) for red-teaming text-to-image diffusion models by finding problematic prompts.
• It provides two main optimization routines (P4D-N and P4D-K) to attack deployed safe diffusion pipelines (ESD, SLD, SD-NEGP).
• It is a CLI-driven Python project with data processing, prompt optimization, safety model pipelines, and classification modules.
File structure
Top-level scripts and CLI entry points
• run_p4dn.py, run_p4dk.py: orchestration for P4D-N and P4D-K runs
• process_data.py: preprocessing and formatting raw prompt datasets
Configuration and data files
• configs/*.json: model-specific hyperparameters and pipeline settings
• data/*.csv: original and processed prompt datasets
Core modules
• prompt_optimization/: optimization algorithms (optimize_n.py for P4D-N, optimize_k.py for P4D-K)
• model/: custom pipelines built on top of Stable Diffusion (modified_sld_pipeline.py, modified_stable_diffusion_pipeline.py) and UNet (unet.py)
• nudenet/: integrates NudeNet for NSFW classification and detection
• open_clip/: OpenAI CLIP implementation for prompt encoding, model configs, and utility functions
Utilities and global scripts
• main_utils.py: shared helper functions
• figure/: visualization assets for the README
Additional insights from the README
• Describes two optimization workflows: P4D-N (needle optimization) and P4D-K (key optimization)
• Highlights required components:
– Pretrained safe pipelines (ESD UNet checkpoints, SLD, SD-NEGP)
– NudeNet ONNX classifier model
– Torch and CUDA dependencies
• Emphasizes outputs: optimized prompt CSVs, generated images, logs