Loading
Have a ComfyUI workflow? Get a readable Workflow Doc in your browser.Make a workflow readable.Try it now ↗
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.
Runtime profile
Source description
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.
AI-generated commentary
AI-generated explanation based on source and configuration details. Suggestions are clearly labeled.
This workflow is for realistic style transfer: give it a creative stylized image, optionally with the prompt used to create it on the manual path, and it can produce a realistic-looking image while generally retaining the image’s creative features.
The Automatic path uses a vision-language model to transfer information from the source image.
The manual path can first use a prompt for realistic style transfer, then feed the resulting image and optionally its prompt into Text Encode (Krea2) to denoise the original image.
Input: a creative stylized image; the manual path can also take the prompt used to generate it.
Output: a realistic-looking image that generally retains the input image’s creative features.
This is the Automatic version of a Krea 2 workflow published by Bvga.
Open the workflow at https://civitai.com/models/2790688?modelVersionId=3176355.
The workflow requires these model files: krea2Raw_v10.safetensors, qwen_image_vae.safetensors, and qwen3vl_4b_bf16.safetensors.
The workflow requires these node packs: comfyui-easy-use (pin 1.3.6), comfyui-kjnodes (pin bc8e4ce4254bcd0050383386ee2f9d753dbf1fa5), comfyui-krea2-text-encoder (pin 38da10b0d4655098d867c14af10093baa76a85c4), RES4LYF (pin 0dc91c00c4c3fb38e7874fcd7a2a327765e8882c), and rgthree-com
Use Automatic for the vision-language-model path; manual prompting may provide stronger realistic transfer but may be less clean.
Suggestion · not verified
For the manual path, use a realistic image as the Text Encode (Krea2) input to reduce stylized characteristics carried over from the original image.
Suggestion · not verified
Verify the exact file and source for krea2Raw_v10.safetensors before setup because no source is identified.
Suggestion · not verified
License terms are not specified; verify usage terms before use. Text-encoder-generated prompts may not be ideal, so review them before relying on them.
Suggestion · not verified
Does this need to be edited?
Sign in to send an edit request.
Sources
1 sourceSource excerpts
1 excerptSource context: 1127 downloads · Type Workflows · Base model Krea 2
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.
Remaining hardware or download-size estimates are not available yet.
Requirements
krea2Raw_v10.safetensors
Not resolvedCheckpoint · Unknown
VAE · 242 MB · Civitai · 1842123
Text encoder · 8.27 GB · SAFETENSORS · bf16 · Civitai · 2740758
comfyui-easy-use
FoundNode pack · Registry
Node pack · Registry
comfyui-krea2-text-encoder
FoundNode pack · Registry
RES4LYF
FoundNode pack · Registry
rgthree-comfy
FoundNode pack · Registry
