Krea 2 OSS runs locally in ComfyUI today. Both open-weight checkpoints — Krea 2 RAW and Krea 2 Turbo — load through the standard diffusion model nodes, so you can train and generate locally on your own hardware. Krea 2 OSS excels at generating diverse aesthetic outputs and is definitely an image model we recommend to add to your stack. For information directly from Krea on the new checkpoints, check out the github repository and read the technical report.
Two models, designed to work together
Krea 2 OSS ships as two checkpoints with distinct jobs.
Krea 2 RAW is the base model: a pretrained checkpoint with no distillation and no post-training. It produces diverse, malleable outputs and is the checkpoint to use for fine-tuning, post-training research, and LoRA training. RAW is not tuned for final-quality generation on its own.
Krea 2 Turbo is an 8-step distilled checkpoint built for fast, high-quality text-to-image. It’s distilled from a fully post-trained RL checkpoint, so it produces higher quality than RAW but with lower variation between outputs.
The two are meant to be paired. Train a LoRA on RAW, then apply it on Turbo. LoRAs trained on RAW express strongly on Turbo, which lets you train on the malleable base and run fast inference on the distilled model.
Model Highlights
Speed. Turbo generates in 8 steps. Generation is quick and allows for fast iteration.
Stylistic range. Krea 2 OSS covers a wide aesthetic and stylistic range, from photographic portraits to anime to high-contrast editorial macro work.
Malleability. RAW is undistilled, which makes it well suited to fine-tuning and research. Diverse output and a tunable base give you room to push the model toward specific styles or subjects.
Prompting
The model was trained on short, medium, and long natural-language prompts. Detailed, long prompts give the best quality. Our templates ship with a Prompt Enhancer system instruction directly from the Krea team.
Architecture
Krea 2 OSS is a 12B dense DiT. It uses the Qwen Image VAE and a Qwen3-VL text encoder with multi-layer feature aggregation.
Get started
Update ComfyUI to the latest version 0.26.0.
Download the workflow below, or find it in the template library
Follow the note in the workflow to download models and save them in the correct model directory
For the people who wants to download the models, here is the link:
https://docs.comfy.org/tutorials/image/krea/krea-2#krea-2-comfyui-workflow-example
Running the image checkpoints in the same ComfyUI workflow would make a controlled style comparison easier to organize. I would retain the prompt and reference, change only the checkpoint, and note which output details matter for the next scene. When a still is intended for video, a usable composition and consistent subject can matter more than novelty alone. Disclosure: I represent the independent WAN30 website (https://wan30.io/), an AI video creation platform.