ONE-PEACE

Why we included this project

ONE-PEACE is a single representation model that maps text, images, and audio into one shared embedding space, trained from scratch without reusing an existing vision or language backbone. The payoff is zero-shot cross-modal retrieval: you can query an image library with an audio clip, or combine sound and words to narrow a search, and the model aligns modalities it never saw paired during training. The repo backs this up with fine-tuning scripts and checkpoints for vision, vision-language, and audio-language tasks, plus a visual grounding API that locates objects inside a picture. Teams building multimodal search or retrieval systems will find the pretraining and evaluation code a useful reference, though this is research-grade software with a workflow that expects some familiarity, not a turnkey serving layer.