Prompt to Prompt Editing with Stable Diffusion

This repository contains our experimental implementation of Prompt-to-Prompt (P2P) editing and Null-Text Inversion (NTI) applied to image generation using Stable Diffusion and to audio generation using Riffusion, for the MsC course in Statistical Machine Learning, 2024/2025.

---

Authors

Valeria Avino, Leonardo Rocci, Riccardo Soleo.

📂 Repository Structure

.
├── riffusion/               # Custom implementation of Riffusion-based audio generation
├── nti_utils.py             # Adapted utilities for NTI (modified from existing repos)
├── ptp_utils.py             # Adapted utilities for P2P (modified from existing repos)
├── seq_aligner.py           # Utility functions for sequence alignment (adapted)
├── NTI.ipynb                # Notebook reproducing NTI experiments with custom prompts
├── P2P.ipynb                # Notebook reproducing Prompt-to-Prompt editing
├── P2P_riffusion.ipynb      # Notebook applying P2P to audio via Riffusion
├── example_images/          # Sample visualizations and reference examples
├── generated_audio.wav      # Example generated audio file
├── p2p_source_12345.wav     # Source audio before editing
├── p2p_edited_12345.wav     # Example of audio after P2P editing
└── __pycache__/             # Cached files

---

🚀 What We Implemented

• From scratch:

  • riffusion/ – adaptation of Riffusion for audio editing tasks.
  • P2P and NTI notebooks with our own pipelines, prompts, and experiments.

• Reused with modifications:

  • nti_utils.py, ptp_utils.py, seq_aligner.py → originally from the public Google Prompt-to-Prompt repo, adapted for compatibility with our hardware and newer model versions.

---

📓 Notebooks

• P2P.ipynb – Reproduction of Prompt-to-Prompt editing with our own experimental prompts.
• NTI.ipynb – Implementation of Null-Text Inversion with adapted utilities.
• P2P_riffusion.ipynb – Audio-focused editing pipeline combining P2P and Riffusion.

---