Voxta Krea 2
Local image generation with Krea AI's flow-matching diffusion transformer.
Krea 2 (K2) is a from-scratch flow-matching diffusion transformer from Krea AI, running locally. Voxta pairs it with a Qwen3-VL text encoder and the Qwen-Image VAE, and installs everything on first use.
Setup
Add the service
Manage Services → + Add Services → Krea 2 → Add.
Pick a checkpoint to match your card
| Checkpoint | Size | For |
|---|---|---|
| Turbo fp8 | ~13 GB | The recommended default. Eight-step distilled. |
| Turbo INT8 ConvRot | ~13 GB | Best quantized quality, RTX 20-series and up. |
| Turbo NVFP4 | ~7.7 GB | 8–12 GB cards, at some quality. |
| Turbo bf16 | ~26 GB | 32 GB+ VRAM. |
| RAW fp8 | ~13 GB | Undistilled. The base to train LoRAs against. |
Pick a text encoder
Qwen3-VL-4B fp8 (~4 GB) is the recommended one. A bf16 build and an abliterated build are also offered.
LoRAs load from the configured LoRAs directory.