Character LoRA trained on shadowless backgrounds learns to "float" 0.5-3 inches off the ground
TraditionalCity2444 · reddit · 2026-09-30
A Reddit user recounts a character LoRA training pitfall: after learning that isolated character images don't work (use a small set with varied backgrounds) and dropping Photoshopped shadows to avoid confusing the model, the resulting Z-Image Turbo LoRA (AI-Toolkit + ComfyUI Realtime LoRA Trainer, Joy Caption .txt captions) works well — but the model concluded the character "can float and is often seen 0.5-3 inches off the ground," likely because every training image had a floating, ungrounded figure. They're asking for a prompt-level fix short of retraining.
More from Multimodal
- Four imaginary tokens for Midjourney v8.2 produce memory ghosts and bone echoes — LudovicCreator · 2026-09-30
- Hyper-personalized music is BS: music is culture and inherently social, argues developer — jordiponsdotme · 2026-09-30
- NUS Proposes StoryEngine: A State-Grounded Agentic Framework for Coherent Long-Form Video Storytelling — NationalUniversityofSingapore · 2026-09-30
- One Year of Local Image Generation: Why Civitai and ComfyUI Both Fall Short — BenDLH · 2026-09-30
- Opus made a launch video for Violetto 1B in 50 minutes amid zero media coverage — tensorqt · 2026-09-30
- Meshy hits $100M ARR in under two years as GPT-6 Astra stirs the AI 3D debate — 量子位 · 2026-09-30