Krea 2:
reduced Dataset to 50 Images, captioning with joycaptioner and medium length output. Training wit ai-toolkit standard settings for Krea 2.
ANIMA:
Nothing special here. Trained with ~250 images, mostly of landscapes. Resolution was 1536. Captioning with a custom Qwen 3 VL setup and pure natural language to describe the artstyle of the images.
Description
FAQ
Comments (3)
Is that Simon Stålenhag style?
I try to use only synthetic or open source images - often by describing a specific style I like. Sometimes this can come close, especially when promting for the painterly artstyle and the specific @artist tag.
I test a lot of combinations and if I find something I like, I'll generate hundreds or thousands of images and pick the best ones for training.
I like the result but have no clue if this is a way to go - but most of the MachineLearning people seem to just stir the datasoup so long till it tastes right ;-)
In Anima it is fun to play around with comfyui-prompt-control (start the first 20% with a specific tag and after that switch to another tag is great to mix styles and get someting "creative"...
in otherwords yes it is, there are barely any other artists known to me that have a similar style. Everytime I see a similar aesthetic outside of AI and seek the artist it's Stålenhag - love that guy