Maker Shares 100 Krea 2 Asian Character LoRAs Plus Full Training Recipe
Fideltampan12345 · reddit · 2026-10-18
A Reddit maker shares 100 Asian-character LoRAs trained on Krea 2 (raw), freely available on a Hugging Face Space. Datasets come mostly from gravure idols and Japanese actresses, hand-picked at 20-30 images each (quality over quantity), captioned with Qwen/Gemini/Grok.
Training recipe: AI-Toolkit, Linear Rank 32, qfloat8 quantization, AdamW8Bit, LR 0.0001, 2500 steps, multi-aspect buckets (512/768/1024), caption dropout 0.05, and deliberately balanced close-up/medium/full-body framing. The author plans no new characters for now (Krea 3 open weights are near) but may retrain for Krea3, Flux, or Qwen later.
More from Multimodal
- Fizgig v8.5 adds MOSS-Image training: full LoRA suite on a 12GB card, tuned ComfyUI settings included — shootthesound · 2026-10-18
- Stunning fuzzy iris renders dubbed "fuzzy little agent" go viral — RileyRalmuto · 2026-10-18
- Behind the MV: made with Claude, every visual based on real data or simulations — anderssandberg · 2026-10-18
- Oxford researcher releases AI-generated transhumanist Eurodance MV "Almost" — anderssandberg · 2026-10-18
- Moss vs Flux: Blurry Faces vs Artifacts, and It's Not Just About Censorship — Friendly-Fig-6015 · 2026-10-18
- Leak: Google's Nano Banana 3 Pro May Launch November 17 — DeliciousEgg1781 · 2026-10-18