In my time experimenting with realism models I've always been disappointed, as I could never get the sort of creativity, variance, and quality I wanted.But with Krea 2 I've found something interesting. I'm able to make creative stylized images with a model like Illustrious or Anima, and then run a denoised KSampler over these with Krea 2 to get good realistic-looking results while generally retaining the creative features. Once you get everything configured right I've found the outputs look even better than what Krea 2 is able to generate through prompting alone._________After a lot of frustration trying to optimize the manual prompting workflows, I've found the "automatic" version of the workflow generally works best. The way the vision language model injects the information just seems to transfer the understanding better than manual prompting can. If you want to try out manual prompting (which provides stronger realistic transfer, but may not be as clean), the v1.1 workflow is set up for that._________Further experimenting, it seems the best quality results come from using a manual prompt to do an initial realistic style transfer, and then feeding this image (and optionally the prompt that was used to generate it) into the "Text Encode (Krea2)" node to denoise the original input image with. What this does is get around the issue of the VLM injecting stylized characteristics of the original input image, as you're now feeding it a realistic image instead of a stylized one. However, this is an absolute pain to prompt for. Using the text encoder to help generate a prompt does work, but the quality isn't ideal, and I'm doubtful Qwen3-4b is capable enough to generate the prompts needed for this. You might instead need to use models like ChatGPT and Grok.

Files with SHA256: 96dd40e879675fa9ff6d85519bfb880840239a94ce01b2baf2cc84a2cace59e9