LoRA training help: white-background datasets leak shading into scenes
sonsuka · reddit · 2026-11-13
A user training a style LoRA on a danbooru artist (with the @anima tag for comparison) found the mostly white/simple-background training set causes outfit shading to bleed onto backgrounds like kitchen tiles when prompting scenes. They suspect inpainting might fix it but are unsure and ask for advice on compensating for missing background detail.
More from Multimodal
- AI-generated 'Stranger Beans (1975)' clip passes as a vintage B-movie — Mother-Pear-2392 · 2026-11-13
- AI reimagines Zelda: Breath of the Wild as a 1970s retro cartoon — techhalla · 2026-11-13
- This Midjourney --sref code nails motion blur: 3238681836 in action — umesh_ai · 2026-11-13
- Riverside launches Tonality, first voice model that edits single words in existing recordings — nikola_mr64990 · 2026-11-13
- NVIDIA's TTB bakes scene info into an MLP, speeding view synthesis 23x — nvidia · 2026-11-13
- Different reasoning depths drastically change image model outputs — vincent_koc · 2026-11-13