Identity LoRA trained on 66 images gives inconsistent faces in ComfyUI with Krea 2 Raw
DragonfruitNo9822 · reddit · 2026-10-12
The author trained an identity LoRA on 66 images of a character, but results in ComfyUI with Krea 2 Raw fluctuate: sometimes the face matches, sometimes it looks like a different person. Unsure of the original training setup, they ask whether to change prompting/trigger words, adjust LoRA strength, or retrain.
More from Multimodal
- Recreating Fleet Week in 3D from live flight and vessel tracking data — bilawalsidhu · 2026-10-12
- AuroraIMG-6M: a 6M-parameter text-to-image model that renders recognizable 64x64 images — Tall_Abrocoma_3533 · 2026-10-12
- Father-Daughter AI Podcast 'Maya and the Fog' Episode 6 Turns Fog Into the Monster — jacalulu · 2026-10-12
- MiniMax H3 ComfyUI configs are a mess — user asks for a centralized benchmark — Lightningstormz · 2026-10-12
- MiniMax video model's uncanny 'waving girl' is pure nightmare fuel — hi_im_fatter · 2026-10-12
- 'Echoes of Autumn': a 2-minute AI period romance short film — Abject-Conflict-2673 · 2026-10-12