Asian Character LoRA for Krea 2 Released With Full AI-Toolkit Training Settings
Fideltampan12345 · reddit · 2026-09-25
A Reddit user released an Asian character LoRA trained on Krea 2 Raw/Turbo, sharing the full AI-Toolkit recipe: 20-30 hand-picked photobook images captioned with qwen/gemini/grok, balanced close-up/medium/full-body framing, plus a fix for Krea's sensitivity to moles (remove them first). Training setup: Linear Rank 32, qfloat8 quantization, AdamW8Bit, LR 0.0001, 2500 steps, multi-aspect buckets 512/768/1024, 1h15m on an RTX 5090. Showcase and download available on a Hugging Face Space.
More from Multimodal
- Clockwork Masquerade: a steampunk image prompt template with color variables — LudovicCreator · 2026-09-25
- Opus 5.5 can now drive Blender to make claymations from a single prompt, and the skill is open-sourced — sidahuj · 2026-09-25
- Tested every learning rate: a working photorealistic LoRA recipe for Qwen-Image-2.1 in AI-Toolkit — No-Bad-4838 · 2026-09-25
- New Midjourney SREF code delivers retro 80s comic-style images — tisch_eins · 2026-09-25
- Redditor shares AI sci-fi action short film "Resistance" — JL_Films · 2026-09-25
- Children's Story Video Made Entirely With MiniMax H3 in ComfyUI on an AMD R9700 — discardthemold · 2026-09-25