StyleRef Reference Images

The prompt from StyleRef Apply tells the sampler what a style reads like. This node hands you what it looks like - the style's actual inspiration images, downloaded and returned as one IMAGE batch you can wire into image conditioning, a reference input, or just a Preview Image node so you can eyeball what you're about to generate with.

It's a nice complement to the rest of the pack. The prompt from Apply is already doing the heavy lifting; the images here are for the workflows where words aren't enough. StyleRef deliberately keeps them out of Apply's positive prompt - a sampler can't fetch a URL out of prompt text, so in there they'd just be junk tokens. Here they're real pixels.

How it works

Wire style in from StyleRef Load and the node downloads each inspiration image stored with the style, caches the bytes for the session, resizes everything to the first image's size, and hands it back as a single batch. Two details worth knowing:

• The image requests are deliberately anonymous. Inspiration images live on public storage URLs, and the node never sends your token to a host other than the API. That's a good privacy default, not an accident.
• The node needs an internet connection to fetch them, and it errors cleanly if a style has no inspiration images at all - check the style on styleref.io before you wire this up and assume the batch will appear.

The IP-Adapter caveat

The README points at IPAdapter, and on SD 1.5 or SDXL that's the obvious destination - it's the standard no-training image-prompt tool there, injecting a reference's embedding alongside the text path. But here's the wider truth you'll hit quickly: IP-Adapter doesn't run on FLUX or Z-Image - Tencent never shipped weights for those, and the pack's own templates target FLUX.1 dev and Z-Image Turbo. So on the models these templates actually use, this node's output belongs in whatever image conditioning your model does support - BFL's Redux, InstantX's Flux adapter, PuLID, or just a Preview Image node. The node gives you plain IMAGEs; choosing the right consumer is on you. That's not a flaw in the pack, it's the honest shape of the ecosystem right now.

Inputs and outputs

One required input, style (from StyleRef Load). One optional, max_images - an integer from 1 to 12, default 12, capping how many inspiration images come through.

Outputs are images, the full batch as IMAGE, and count, an INT telling you how many actually loaded. Wire images into an IP-Adapter node, a reference input, or Preview Image; count is handy when you're building logic around whether the style had enough references.

Install

ComfyUI Manager - search StyleRef - or:

cd ComfyUI/custom_nodes
git clone https://github.com/StyleRef/styleref-comfyui

Restart and confirm [StyleRef] v1.0.6 - 5 nodes registered in the console. The pack declares no Python dependencies; this node reaches for Pillow and numpy lazily at run time, and every ComfyUI install already ships those. Like the rest of the pack, it's a thin client over styleref.io - first fetch of a style is a network call, and offline it errors rather than degrading. Nothing spends credits.