PRYX ComfyUI Higgsfield Model Catalog

Every generator in this pack already has a model dropdown. So why does a separate model node exist? Because the moment your workflow has three generators in it - a text-to-video pass, a video-edit pass, a reference pass - you now have three dropdowns to keep in sync, and changing one of them is a five-minute hunt through the graph. The Model Catalog is a plumbing node for exactly that: pick the model in one place, wire the model_id output into every generator that should use it.

This is the value-holder pattern the ecosystem has always needed and rarely names. A node that touches no pixels, holds one choice, and carries it down a single wire is genuinely useful - the same reason rgthree's context nodes exist for wires as a whole. This one carries a string.

What it actually does

It reads the pack's bundled catalog - 81 documented Higgsfield image and video endpoints in catalog revision 2026-09-23.5, spanning Alibaba, ByteDance, Kling, MiniMax, PixVerse, Recraft, Ideogram, Lightricks, xAI and others - and lets you narrow that list before you choose. The three filters are the point:

• provider - one label, or all. (The list genuinely contains both Xai and xAI. Yes, two spellings of one company. That is how the catalog spells them; pick either and check the results.)
• capability - the task, e.g. image_to_video, reference_to_video, text_to_video, video_edit, video_extend, image_generate, image_edit, video_motion, or all. This is the same capability gate the generator nodes use, so a model you can't use in a video node won't tempt you here either.
• status - active, deprecated, unavailable, or all. Defaults to active, which is what you want.

Then model_id picks the entry. One catch worth knowing: if you select a model and then change provider or capability so your pick is no longer in the filtered set, the node falls back to the first model that still matches instead of erroring. Change your filters, then re-check the selection.

Inputs and outputs

Inputs are model_id, provider, capability, status - all four required, all four widgets. There are no optional inputs.

Outputs are two strings:

• model_id - the catalog ID, e.g. seedance-2-5-image-to-video. Wire this into any generator's model input and the generator uses it instead of its own dropdown.
• model_info - JSON with the catalog revision, the full selected model record, supported_media, max_references, and a matches array listing everything your filters allowed. Handy when you want to know why a model vanished from a filtered list.

One caveat: model_id is a plain string, and the generator's model input is validated against the catalog. If you point a catalog at an image_edit model and wire it into the Text to Video node, you get a validation error - the capability gate still applies downstream.

Install and credentials

ComfyUI Manager: search PRYX ComfyUI Higgsfield, install, restart. The README notes Manager availability is post-Registry, so if Manager can't find it, install by hand:

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 users need the bundled interpreter: ..\python_embeded\python.exe -m pip install -r .\pryx-comfyui-higgsfield\requirements.txt. The dependencies are light - higgsfield-client, httpx, jsonschema - and there are no model files to download. What you need instead is ComfyUI 0.35.0+, frontend 1.53.6+, Python 3.10–3.13, and a Higgsfield API key.

Key goes in ComfyUI Settings → PRYX ComfyUI Higgsfield → Credentials as one key-id:key-secret string, or as HF_KEY / HF_API_KEY+HF_API_SECRET for headless installs. It is stored locally in your ComfyUI user directory and never becomes a workflow value.

Also: model dropdowns only list endpoints the catalog documents, and account access is ultimately Higgsfield's call. The catalog node doesn't change that. If a model is missing, refresh the catalog (Settings → Model catalog → Refresh catalog), restart ComfyUI, and hard-reload the browser with Ctrl+F5 - the Python side and the frontend script both need to agree on the revision.