Training policy
Quick mode begins at five accepted photos.
One to four stays blocked, five unlocks quick mode with caution, six turns quick green, and standard mode starts at twelve with a twenty-photo target.
Private staging release
Train portrait models without guessing what the pipeline is doing.
Keep uploads reusable, compare quick and standard training thresholds, import ready-made LoRAs during development, and switch fal on only after the loop feels right.
Upload handling
Large files move straight to R2.
HEIC and HEIF are accepted, large LoRAs use multipart upload, and provider fetches use short-lived signed URLs instead of piping bytes through Workers.
Rollout
Imported LoRAs stay first-class during development.
That keeps costs low while the character flow, prompt presets, and evaluation loop are still taking shape.