krea2-character-lora — "Masafee"

A character LoRA for Krea 2 (Qwen-Image–family single-stream MMDiT, ~12B params), trained on a single NVIDIA RTX 3060 (12 GB VRAM) via FP8 quantization + block swapping. Trigger token: masafee.

The subject is an original red-panda character created by the author. Train on Krea 2 RAW, infer on Krea 2 Turbo (8-step). See the full technical report (JA/EN) and reproduction materials on GitHub:

https://github.com/masafykun/krea2-character-lora

Checkpoints

File	Epoch	Notes
masafee_krea2_e06.safetensors	6	retains eye catchlights, flexible backgrounds
masafee_krea2_e08.safetensors	8	balanced
masafee_krea2_e10.safetensors	10	strongest form (recommended at strength 0.8)

Recommended: e10 at LoRA strength 0.8, with a prompt containing big sparkling eyes with bright highlights.

Usage (ComfyUI)

Place a checkpoint in ComfyUI/models/loras/, apply it to the Krea 2 Turbo (GGUF Q5_K_M) UNet with LoraLoaderModelOnly, and prompt with the trigger masafee, e.g. masafee, red panda, full body, playing an electric guitar, big sparkling eyes with bright highlights, simple background. Sampler er_sde, scheduler simple, 8 steps, CFG 1.0.

License & attribution

This LoRA is a derivative of Krea 2 and is subject to the Krea 2 Community License. The base weights are not redistributed here. The character "Masafee" is the intellectual property of the author (Masato Suzuki / Masafy).