zhenzhen-krea-v2-fal

Krea v2 is the model that made the "style fusion" idea mainstream: give it two reference images and it merges their aesthetics into a single new image - one image's color grading plus another's composition, or one's character design with another's material feel. That's this node's entire reason to exist, and it's the clearest "I need a specific look" tool in the Comfyui-zhenzhen pack. Two style_image inputs, a style_strength slider, a creativity dial, and the model does the fusion.

Worth knowing what you're paying for: Krea started as a closed subscription product and only opened its weights in 2026 - but through this node you're calling the closed API, not running the open 12B locally. Per-call credits via Zhenzhen's fal proxy, key in the widget, no GPU on your side.

How it works

Text-to-image with optional style references. prompt describes the subject; style_image1 and style_image2 (or their URL twins) carry the reference aesthetics; style_strength (0–2, default 1) decides how hard the style hits; creativity (raw/low/medium/high) decides how free the model is. model_size (medium / medium_turbo / large) trades quality against speed/cost. The node submits, polls, downloads, and returns images (an IMAGE tensor), plus response and image_urls strings.

It's a non-output node, so images can flow onward into upscaling or compositing steps.

Inputs that matter

• prompt - the subject. The style comes from the reference images, not the prompt.
• style_image1 / style_image2 - the stars of the show. One reference gives you a flavor; two give you the actual fusion trick. Both are optional in schema, but a style-fusion node with no styles is just a text-to-image node.
• style_strength - 0 = barely there, 2 = heavy. If the output ignores your references, this is the first dial to check.
• creativity - raw sticks closest to the reference composition; high wanders. For preserving a brand look, stay low.
• model_size - medium_turbo to iterate fast and cheap, large for the final.
• seed - max 65535, 0 = random.

Installing it

Manager → search Comfyui-zhenzhen → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen

Restart, paste your api_key into the widget. No model files.

Common issues

• Styles not showing up - check style_strength isn't at 0 and that you've actually connected style_image* (not left them empty). Also image_way matters: base64 vs image_url changes how the refs travel, and hosted URLs can expire mid-job.
• Too literal a copy - that's creativity: raw doing its job. If you want a fusion rather than a photocopy, bump creativity to medium.
• The pack standard gotchas - empty key, 443/VPN-in-TUN-mode for the overseas API, vendor 500s needing a re-run, fal's pre-deducted billing, and a real seed cap of 65535.

One aside: Krea v2's open weights exist if you want to run it locally - but the open tier is capped below $1M revenue, and the API route here is the "no 12B card in my rig" answer. For a quick brand-style test, this node beats a weekend of local setup every time.