Skip to content

Home › Krea 2 LoRAs

Krea 2 LoRAs

48 original synthetic characters, trained for Krea 2 and sold only here. Not on Civitai, not on any other marketplace.

What a Krea 2 LoRA is

Krea 2 is a 12-billion-parameter diffusion transformer with a Qwen3-VL text encoder. A LoRA is a small file — around 220 MB — that teaches it one specific thing without retraining the whole model. These teach it a person: a consistent face, body and presence you can put in any scene you can describe.

Each one here is an original synthetic character. Not a real person, not a celebrity likeness, not a face scraped from anywhere. They exist because they were trained into existence.

How these are trained

Every setting here was measured against its alternative before being fixed, not copied from an example config:

  • Rank 32 / alpha 32. Rank 16 at 1000 steps gave a visibly weaker likeness on the same dataset and seed.
  • 2000 steps. Compared against 2500 across two characters and five framings with the same seed: identity was already resolved at 2000.
  • Resolutions 512 and 1024. Krea 2 was pretrained at 256, 512 and 1024 — never at 768. Training at 0.25 MP alone cuts the time almost in half but degrades likeness visibly.
  • 40 to 80 images per dataset. Below forty the model binds incidental detail to the character; past a hundred the extra images mostly repeat.

Need a different character or a different model?

These are trained for Krea 2. If you need a specific character, or the same character for Flux, SDXL, Illustrious, Pony, Qwen-Image or WAN, I train custom LoRAs to order from €29.99.