Pro artist's style-LoRA failure: 66 illustrations yield barely any change or more hallucinations
Shady-Dragon · reddit · 2026-09-23
A professional illustrator, pushed by an industry shift toward volume, is trying to reproduce his own art style — strokes, color, shading, proportions, not just characters — in AI. Training a LoRA on 66 of his illustrations in Drawthings on iPad, with captions auto-generated by a small Qwen model then hand-corrected, produced LoRAs that either barely changed output or increased hallucinations; most tutorials only cover character LoRAs. He asks whether there's a better route than LoRAs for style transfer.
More from Multimodal
- GPT-6 Astra color graded, captioned and stitched a full video in one shot from raw phone LOG footage — OpenAIDevs · 2026-09-23
- Qwen Image 2.1 remakes look great with CFG>1, beta_57 and AuraFlow Shift 16 — GTManiK · 2026-09-23
- Seedance 2.5 Draft Mode lands on Pika at as low as 10 cents per second — Kyrannio · 2026-09-23
- Qwen-Image-2.1 tested: 'insane' prompt adherence on candid phone-photo looks, full workflow embedded — kosovack · 2026-09-23
- Is AI art just polished repetition? Debate asks where the Neo-Pop of AI art is — PAstynome · 2026-09-23
- WanAnimate 2.2 exact inpainting struggles: mask expansion wrecks faces, dev seeks fix — Reasonable-Aide1259 · 2026-09-23