-=< AHTUNG >=- This is a very dark, cinematic sci-fi horror themed Krea2 model. It probably requires to apply some technique to increase brightness if you want to get the normal exposure. Easy in ComfyUI.
This model is focused on moody, atmospheric, creative portraits and picturesque landscapes.
Drop my samples to ComfyUI to see the recommeded high quality and high realism workflow. Text-toImage is rendered with this model (low-noise draft). Image-to-Image HiRes Fix finishing is done with CinEro-NG-XL (with low-to-moderate denoise).
PS: I plan to release the workflow as a separate resource and explain why I use the hybrid workflow. In short, SDXL model still give better realism and more natural noisy / grainy textures.
Versions info
v1
It has Qwen3VL Heretic Text Encoder baked in. So, by using this model you take full responsibility on the resulting safety of the rendered images. Qwen Image VAE also embedded. CLIP and Unet parts both quantized to FP8.
Description
initial version
FAQ
Comments (9)
Привет! Нравятся твои модели. Можешь также выложить bf16 версию?
Могу, но она большая и без CLIP. Качество скорее всего эквивалентно.
rdy
о, свои, здравствуйте, братушки )
лайк. качаю.
а в int8 convrot можно пожалуйста?
[Вовка в 3/9 JPG]
Это я пока не умею. Надо книжечку почитать...
love it :)
pretty awesome with Flux.2 as refiner.
kinda soft on it's own, but who cares. its very different from every other krea2 model i tried, and i tried a lot.
Yep.. I saw the images in your showcase with the similar or the same prompt. Your previous images done with deeply TURBO'ed model with a very high contrast and significantly more stuff and details. That is expected - this version is a low noise draft. XL or Flux refiner for high res fix is required for finishing. v1.0 BF16 UnetOnly version is a pure training on Krea2 RAW - no Turbo LORA integrated. You can add this lora to your workflow to shift closer to what results you had with other model. I plan to give more training and probably will make the next version more contrast.