MOSS Character LoRas Only 70-80% Identity-Accurate, Community Seeks Fixes
Bostonparis · reddit · 2026-10-19
Reddit user Bostonparis trained two character LoRAs for MOSS using datasets that worked well on other models like Krea 2 and Qwen 2.1, but identity fidelity only reaches about 70-80%. They used AI toolkit with both the base and sft models and asked whether others hit the same issue.
The thread highlights that training recipes don't transfer cleanly across media-generation models — switching base models usually means reworking datasets and hyperparameters.
More from Multimodal
- MiniMax H3 video model recreates Totally Spies scene with striking fidelity — Certain_Potato_4509 · 2026-10-19
- Local video model LoRAs for creative control, then Wan 3.0 re-lights the scene — TinfoilTricorn · 2026-10-19
- PhotoQuilt accepted at WACV 2027: training-free arbitrary-resolution photomosaics — CSProfKGD · 2026-10-19
- University AI Film Course Generates Shorts Shot-by-Shot in Flow, With a Peer-Review Exercise — lmoroney · 2026-10-19
- freavy: open-source local-first node canvas for AI image/video on ComfyUI — framedeptco · 2026-10-19
- Fixing broken AI hands: crop-then-edit with Qwen-Image-Edit-2511 beat every alternative — aoi_secret_jp · 2026-10-19