Fizgig 8.9.6 adds masked training: auto-generate masks for LoRA training images
shootthesound · reddit · 2026-10-25
Fizgig v8.9.6 ships a long-requested masked training feature. The tool auto-generates a black-and-white mask (photo-masklabel.png) beside each training image, with white marking what the LoRA should learn.
Per the screenshot, mask strategies include: the person (hair to clothes), the person but not their clothes, just face and hair, body and clothes without the face, skin only without face/hair/clothes, and everything but faces — covering likeness, outfit, body and style LoRA cases. Masks run on CPU in a fraction of a second per image, downloading Google's person finder (16MB) on first run; wrong masks can be repainted in any editor or deleted.
More from Multimodal
- Ask ChatGPT to fill a photo of your empty bookshelf with books it thinks you'd own — venturetwins · 2026-10-25
- $1M 1 Billion AI Film Award: 5-7 Min Films Powered by Seedance — lmoroney · 2026-10-25
- AI Recreates Infamous Vintage Cigarette Ad With Eerie Realism — RookeryFilms · 2026-10-25
- $1M AI Film Award Reopens: 5-7 Minute Films, 70% Must Use Seedance — lmoroney · 2026-10-25
- Minecraft Rendered as a 90s Anime: Striking AI Style Transfer Demo — yumaeriel · 2026-10-25
- Top Krea 2 checkpoints on Civitai all contain no-derivative LoRAs — jib_reddit · 2026-10-25