Both ends of Krea 2 are Qwen and that explains more than the sample grids do
retained text
​ Everyone has been posting grids for two weeks so I went and read the spec instead. The backbone is trained from scratch but the two parts doing the work either side of it are Qwen. Text encoder is Qwen3-VL 4B, and it does not just read the last hidden state like most setups, it takes twelve decoder layers per token and fuses them inside the transformer. Then the Qwen-Image VAE turns the latent back into an image at the other end. If you already know how Qwen-Image treats text and composition you are most of the way there, because that is where the behaviour is coming from. Raw I had wrong for about a week. Assumed it meant unfiltered, the way that word gets used about other checkpoints. It is the undistilled mid-training checkpoint, for fine-tuning and LoRA work. Turbo is what you generate with, 8 steps, guidance off. Raw wants 28 steps with CFG at 4.5 and takes its time, which is the whole point of it. So the complaints about it being heavy are people trying to make portraits with a training base. Around 26GB for the diffusers install. What it is good at is LoRAs off small sets, fifty or sixty images, then you move to Turbo to actually generate, which is what Krea recommend anyway. Quantized Turbo if you just want pictures.
document fetched 2026-09-25 00:34 · canonical https://reddit.com/r/StableDiffusion/comments/1wp2rxo/both_ends_of_krea_2_are_qwen_and_that_explains