How apple/pico-banana-400k works
Project type and purpose
• Type: This is an open-source dataset repository (“Pico-Banana-400K”) for text-guided image editing.
• Purpose: Provide ~400K text–image–edit triplets plus associated manifests and a small helper script to map Open Images URLs to local files.
File structure
• Top-level files only; no nested code modules or services.
• LICENSE_DATA, CODE_OF_CONDUCT.md, CONTRIBUTING.md, README.md – documentation.
• map_openimage_url_to_local.py – a single Python script for URL→local path mapping.
• No “frontend” or “backend” directories; this is a data-centric repository with one utility script.
README for additional insights
• Two-stage pipeline:
b) Editing + Self-Evaluation (Nano-Banana for edits; Gemini-2.5-Pro for quality scoring)
• Data staging: Open Images tar files are downloaded from AWS S3, unpacked, and mapped via the script.
• Distribution: Final dataset manifests and JSONL files published on Apple’s public CDN.