Shift or sigmoid? LoRA trainers hit broken results on new model architectures
Inevitable_Pen9043 · reddit · 2026-10-25
A Redditor reports a completely broken LoRA when training on MOSS, while sigmoid works fine on Krea 2 and Qwen 2.1 — reigniting the shift-vs-sigma scheduling debate and whether 'shift balanced' then 'shift low noise' late in training is the better recipe for the new generation of image models.
More from Multimodal
- Five days of AI-generated visuals blended with ASCII dither techniques — Teknium · 2026-10-25
- One frame, one prompt: Seedance 2.5 vs Kling 4 vs MiniMax H3 vs Wan 3.0 — LudovicCreator · 2026-10-25
- Midjourney prompt for cinematic eyewear reflections, with three sample renders — azed_ai · 2026-10-25
- Turning mouse sketches into clean drawings via gestalt-like priors and staged interpretation — debreuil · 2026-10-25
- Finger sync was AI video's #1 tell — hands are now scary good — techhalla · 2026-10-25
- Art direction first: a Firefly Boards workflow from palette to poster — LudovicCreator · 2026-10-25