Image Prompt Apply 🖼️ (TensorVizion)

You have a reference image - a painting whose palette you love, a photo whose lighting you want to borrow - and a text prompt. The Image Prompt Apply node is where those two become one conditioning. It's the "apply" half of the pair that starts with Image Prompt Loader: the loader brings in a CLIP vision encoder and a style model, this node actually runs the reference through them and pushes the result onto your conditioning at a strength you control.

How it works

Mechanically it's a wrapper around two core ComfyUI nodes: CLIPVisionEncode turns your reference image into an image embedding, and StyleModelApply injects that embedding into your text conditioning through the style model's own path. Because that path is separate from the text path, your prompt keeps working - the reference steers style and content, the text steers the subject and composition. It's the same decoupled trick IP-Adapter made famous, just using ComfyUI's native nodes, which means no extra custom pack and no InsightFace dependency. There's a compat detail worth knowing: if your ComfyUI core is older than the strength parameter on StyleModelApply, the node falls back to calling it without strength rather than erroring.

The inputs

• conditioning - your text conditioning from CLIP Text Encode.
• clip_vision and style_model - straight from Image Prompt Loader.
• reference_image - any IMAGE. Drop in whatever you're borrowing style from.
• strength - default 1.0. How hard the reference pulls. Crank it and the image dominates; at 1.0 with an empty text prompt you're effectively generating purely from the reference. For "steer me, don't copy me" work, values around 0.5–0.8 are the usual starting point.

Output is the blended conditioning (wire it into your KSampler) plus a summary string reporting the strength.

Install and where it fits

Pack install as usual:

cd ComfyUI/custom_nodes/
git clone https://github.com/TensorVizion/OmniNodes

Restart ComfyUI, or find "OmniNodes" in ComfyUI Manager. No extra dependencies or model downloads - the model files come from the loader.

The practical trap is expecting it to be a clone machine. Style models capture style and composition tendencies, not identity; if you want "this exact person," that's an identity tool (PuLID, InstantID, or a LoRA), not an image prompt. And the mechanism is an SD1.5/SDXL-era one - on Flux and the LLM-encoder models this whole path doesn't apply, and you'd reach for the model's own in-context reference editing instead. Strength is also not linear: past about 1.2–1.5 you're mostly just making the image weirder, not more faithful, so treat the top of that 0–10 range as "exists, don't go there."