PRYX ComfyUI Higgsfield Video Extend

Every video model has a ceiling. Kling gives you five or ten seconds, Happy Horse tops out around fifteen, and the local open models are the worst of the bunch - Wan 2.2 clips land around five seconds and then you're into frame-interpolation tricks that look like it. Extending a finished clip used to mean chopping the last frame, feeding it back in as a start image, and hoping the seam isn't obvious.

Video Extend is the node for the other approach: hand the provider the whole clip and let it continue it. Be warned, the catalog has exactly one endpoint for this capability - Seedance 2.5 · Video Extend - so there's no model shopping to do here. That scarcity isn't a bug in the node; extends are just rare in the market.

How it works

Connect the source clip to video (required - nothing renders without it), write your prompt, and the pack runs the same path every generator here uses: validate the request against the model's schema, upload connected local media to the provider through its presigned upload flow, call POST /estimate/<endpoint>, and - only when request_mode is generate - submit the paid request and poll it to completion, then download the result into ComfyUI's output directory (or temp, if auto_save is off).

The node is a thin, well-behaved wrapper, which is exactly what you want for a one-model capability. It doesn't pretend to have options the endpoint lacks.

Inputs

model is required and there's realistically one answer. video is the source clip. prompt is a connected input rather than a widget, so wire a text node or a Config UI Prompt node into it - which is also how you'd describe what should happen next, since the model is continuing motion, not starting from scratch.

Then:

• duration - an integer from 4 to 30, default 5. The selected model defines the valid range; this is the number most likely to cost you money, so it's the one to check after an estimate.
• resolution - 480p or 720p.
• bitrate_mode - standard or high; affects the encoded output, not the render.
• generate_audio - whether audio is produced or carried through, when the model supports it.
• image, audio and references - direct media inputs and an ordered reference collection, shown only when the selected model accepts them.
• request_mode, max_usd, auto_save, timeout - the shared safety controls. request_mode defaults to estimate_only; max_usd of 0 disables the limit, a positive value is a hard ceiling and also blocks generation if the provider returns no USD estimate at all; timeout defaults to 1800 seconds and a timed-out generation is never re-submitted.

Outputs

video is the native result, with local_file carrying the path and remote_url carrying the provider URL as JSON text (logging only - it isn't a media input). request_id is empty on an estimate. credits and usd are the provider's estimate for this exact request, -1 when it can't supply one. status is JSON diagnostics: status, request ID, elapsed time, URLs, provider status.

If your video output is empty, you're in estimate_only. That's the default and it's intentional.

Install

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

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. Install into the same interpreter that runs ComfyUI. ComfyUI 0.35.0+, frontend 1.53.6+, Python 3.10–3.13, and three dependencies (higgsfield-client, httpx, jsonschema) - no model downloads, since the model lives on someone else's GPU.

Credentials: Settings → PRYX ComfyUI Higgsfield → Credentials, one key-id:key-secret string, or HF_KEY / HF_API_KEY+HF_API_SECRET for headless installs. Stored locally; never a workflow value.

Two things to know going in

Check the seam, not just the duration. Extension is a continuation, so whether it looks seamless depends on how well your prompt describes the ongoing motion. Run the estimate, then compare a 4–6 second extension before committing to 30.

A missing model here is normal. Capability filtering means this node only lists endpoints documented for extension. Don't go hunting for Kling in the dropdown - the pack won't offer an endpoint the provider doesn't document for the task, and an item absent here isn't a statement about what Higgsfield offers elsewhere.