PRYX ComfyUI Higgsfield Reference Preview

Reference ordering is the least visible and most expensive thing to get wrong in a cloud-video workflow. You connect four images, the provider numbers them by type, your prompt says "Image 1 is the character", and you find out you were wrong after the paid request came back with somebody else's face.

Reference Preview is the seatbelt. It maps your media to the exact fields the request will carry, resolves your prompt aliases, prints the whole thing as readable text, and does no media upload and no API call at all. There is no cost to running it, which is the entire reason it's worth having.

What it does

You give it a model and your media - either a Collector's references output, direct image / video / audio / end_image sockets, or both - plus a prompt. It runs the pack's real payload mapper against the selected model's parameter schema, so the mapping it reports is the same mapping the generator will build. It isn't a separate guess that can drift.

The output reference_info reads like this:

Wan 3.0 · Reference to Video - reference mapping
Numbering is separate for images, videos and audio.
Direct inputs come first; collector inputs follow. Collector slots use numeric order; batches keep their order.
Image 1 ← person (Collector image_1)
Image 2 ← outfit[1] (Collector image_2)
Resolved prompt: Use Image 1 as the main character, Image 2 as the clothing.

Where the endpoint doesn't document prompt tokens at all, that text says so plainly, and the numbered entries mean payload order only - they are not tokens you can type. The Preview is deliberately conservative about this: it will not invent @ mentions or angle-bracket syntax for an endpoint that doesn't publish them.

Aliases work through the same path. Write {{ref:person}} in the prompt, label image_1=person in the Collector, and the Preview shows you the resolved sentence. Unknown labels and duplicate labels are errors - you find out here, in a preview, rather than mid-generation.

Inputs and outputs

model is required and is a dropdown of the catalog's active models that accept connected media (55 at revision 2026-09-23.5). Optional inputs are prompt, references, image, video, audio, end_image.

Outputs:

• prompt - your prompt passed through, with aliases intact, so the generator downstream can resolve them itself against its own final validation.
• references - the ordered collection, passed through, ready to wire into a generator.
• model - the selected model ID as a string, useful for wiring into a generator's model input so the preview and the generator can't disagree.
• reference_info - the readable mapping plus the resolved prompt text.

The node is an output node, so it renders its text in the ComfyUI UI and will execute even if nothing is wired downstream of it. That's how you use it in practice: park it next to your collectors, run the workflow, read the panel, fix labels, run again.

Install

Manager: search PRYX ComfyUI Higgsfield, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/PRYX-STUDIO/pryx-comfyui-higgsfield.git pryx-comfyui-higgsfield
python -m pip install -r pryx-comfyui-higgsfield/requirements.txt

Windows portable: ..\python_embeded\python.exe -m pip install -r .\pryx-comfyui-higgsfield\requirements.txt. Needs ComfyUI 0.35.0+ and frontend 1.53.6+, Python 3.10–3.13. No checkpoints, no model files - higgsfield-client, httpx, jsonschema, that's the lot. Credentials (Settings → PRYX ComfyUI Higgsfield → Credentials, or HF_KEY) are only needed when something downstream actually calls the provider; the Preview itself won't complain about them.

When to run it

After any change to media, labels, slot order, or the selected model. Those four things are what move the mapping, and all four are cheap to change and easy to change accidentally - a Collector batch splitting differently because you loaded four images instead of one, for instance. Thirty seconds of preview beats re-running a paid render because outfit landed in slot 3.

If the Preview disagrees with what you expected, trust the Preview. It's reading the same validation the request will.