RL Set Reference Image

RL Set Reference Image is where you hand the AI Render pipeline extra images to lean on. The README is blunt about the use case: it "works well for overlay styles such as IPAdapter," and it's how you stack multiple reference images in one workflow. One node per reference, each carrying a path plus a weight curve, and the plugin knows which is which by the name field.

It's a config node, not an engine - it doesn't apply IPAdapter or blend anything itself. It resolves the image path, validates the weight type, and passes the settings onward for whatever node in your graph does the actual style or composition work.

The inputs that matter

• name - a label, mostly for the plugin to tell references apart. Default additional_image; if you're stacking several, give them real names or you'll lose track of which is which.
• image_path - the reference file, relative to ComfyUI's input folder.
• weight_type - the interesting one. Fifteen options, and they're the classic IPAdapter weight curves: the easing family (linear, ease in, ease out, ease in-out, reverse in-out), the input/output emphasis family (weak input, weak output, weak middle, strong middle), the style/composition family (style transfer, composition, strong style transfer, style and composition), and two precise variants (style transfer precise, composition precise) that pair with IPAdapter Plus-style models.
• weight - 0 to 1, the overall influence of this reference.
• active - a boolean on/off. Toggle a reference off without deleting the node; the node then forces its output weight to 0.

If you're new to this, the mental model from the IPAdapter playbook: style-oriented types steer how it's drawn, composition types steer what's where, and the easing curves mostly exist for video, where a reference fades in or out across frames. For a still render, style transfer at a moderate weight is the safe starting point.

The outputs are just image_path (resolved absolute), weight_type, and weight. Wire them into your IPAdapter-style apply node and let that do the actual work.

Install

Same zero-dependency pack as the rest:

cd ComfyUI/custom_nodes
git clone https://github.com/reallusion/ComfyUI-Reallusion

Restart ComfyUI; the nodes live under the Reallusion category (or install via ComfyUI Manager by searching "Reallusion"). No pip step, no models to fetch - the IPAdapter models you'd actually use are a separate download on your side.

Gotchas

• The node doesn't apply anything. Path resolves, weight passes through. Forget the downstream apply node and this looks like it does nothing - because it does nothing, by design.
• Missing file, instant error. If the image isn't in ComfyUI's input folder, you get FileNotFoundError: cannot find input file. The AI Render plugin normally drops files where they're expected; hand-wired workflows hit this constantly.
• active off = weight 0. That's the whole behavior; the path still resolves, so a muted node won't fail on a missing file the way ControlNet does.
• Weight type names are lowercase-with-spaces (ease in-out, style and composition). Pick from the dropdown rather than typing; a typo gets rejected with an invalid weight_type error.