Home › Custom LoRAs
Custom LoRA training
Tell me the character you want and the model you work with, and I train the LoRA for it. From €29.99.
Which models
Krea 2, Flux, SDXL, Illustrious, Pony, Qwen-Image and WAN — and others on request. The base model matters more than people expect: a LoRA trained for one will not work properly on another, so it is the first thing to decide.
How it works
- You send the reference material. As little as two or three images works for a character; a folder of forty is better.
- I build the dataset. Captioning, framing spread and trigger word, which is where most of the quality comes from and where most LoRAs go wrong.
- I train and test it. Twenty preview renders across different framings before anything is delivered.
- You get the .safetensors plus a sheet with the trigger word, recommended weight and the settings used.
Why the settings are not guesses
Rank 32, 2000 steps, resolutions 512 and 1024 — each of those was compared against its alternative on the same dataset and seed before being fixed. Krea 2 was never pretrained at 768, so training at that resolution wastes time for a worse result, however many example configs suggest it.
What I do not take on
Likenesses of real people, and styles lifted from a living artist. Not a moral position so much as a practical one: both get taken down, and neither is something you can build on.