Now you can train your own LoRAs. The undistilled base of COLOSS: train on Raw, run on Turbo. SFW and NSFW, adults 18+ only. Made by COLOSS Labs, an independent AI lab building open generative models. Together with Krea 2 Coloss Turbo it makes COLOSS v1. Train on Raw, run on Turbo Train your LoRA on this model. Load it on Krea 2 Coloss Turbo with LoraLoaderModelOnly, strength about 1.0. Generate at 8 steps, cfg 1. Raw and Turbo share the same COLOSS weights in the same blocks, so a LoRA trained here lands on Turbo as is, without any conversion. Why train on COLOSS Bodies are already learned. Male and female anatomy, from portraits to fully explicit scenes. Your LoRA teaches your character, style or concept, not anatomy from scratch. Quick to learn. Our own character LoRA locked in the face after about 1000 steps. Same quality on the way out. Images come from Turbo: living skin, legible text, 8 steps. Ideas to try Your own character, a style, a concept, a product. A full fine-tune on your own dataset. Merge back into Turbo, the way we made COLOSS: your Turbo = Krea 2 Turbo + (your Raw − Krea 2 Raw). Preference tuning (DPO) on Raw, then carry it over to Turbo the same way. LoRA training settings What worked for our character LoRA with musubi-tuner: rank 32, learning rate 1e-4, modules blocks.* and txtfusion.*, 1000 to 3500 steps for a character (about 100 images), plain sentence captions with a trigger word. musubi-tuner can render samples straight on Turbo (--turbo_dit). For ai-toolkit and diffusers, use the transformer/ folder from huggingface.co/colosslabs/Krea-2-Coloss-Raw. Made by - colosslabs https://civitai.red/models/3024721/krea-2-coloss-raw