PRYX ComfyUI Higgsfield Video Edit

Video editing is the capability local ComfyUI handles worst and the cloud handles best. You can't download Kling's editing model or ByteDance's Seedance 2.5 edit endpoint - they're API products, full stop. What this node gives you is a ComfyUI socket for them, with your source clip coming straight out of whatever local workflow produced it.

The catalog lists five video-edit endpoints, and the dropdown also covers the catalog's motion-control entries:

• Genjutsu · Motion Transfer and Genjutsu · Object Swap - Higgsfield's own, and the two that behave differently (see below).
• Seedance 2.5 · Video Edit - ByteDance's edit endpoint; ByteDance's product models stay API-only while its components (Depth Anything, PuLID, SDXL-Lightning) ship free, and this is on the paid side of that line.
• Kling O3 · Video Edit and Kling Omni · Video Edit - the Kling family's edit routes.

How it works

Connect your source clip to video. Depending on the selected model you'll also get image, audio and the ordered references input; the frontend only shows sockets the model actually accepts, and the backend re-validates everything before uploading. Then the usual path: validate → upload local media to the provider through presigned URLs → call the estimate route → and, only in generate mode, submit the paid request and poll it down, with bounded retries and no automatic re-submission after a timeout.

The Genjutsu pair is the part worth understanding before you wire anything. Both require a video and an image - the clip supplies the motion or the scene, the image supplies the subject or the object. A prompt is optional there, which is unusual: for motion transfer the reference clip is doing most of the describing.

Inputs you'll set

prompt is a connected input rather than a text box - wire a text or Config UI Prompt node into it. Then:

• video - required by the edit models. Nothing renders without it.
• image - the source/target image for motion and object-swap jobs.
• resolution - 480p or 720p on the models that expose it.
• keep_original_sound (yes / no) and generate_audio - whether the source audio survives and whether new audio is produced. Both are model-dependent; a model that doesn't support them won't show the field or will ignore it.
• character_orientation (image or video) - which of your two inputs decides how the character faces in the output. Handy when your reference image and your clip disagree about left and right.
• mode (std / pro / 4k) - the provider's quality tier. Not the billing switch; that's request_mode, which stays estimate_only until you say otherwise.
• elements - up to 4 JSON items on models that document them.
• bitrate_mode (standard / high), plus the shared request_mode, max_usd, auto_save and timeout.

Outputs

video (native VIDEO, with a local_file path alongside it), remote_url as JSON text of the provider URL, request_id - empty in estimate-only mode - credits and usd for the estimate, -1 when the provider supplies none, and status, the JSON diagnostics blob carrying status, elapsed time, URLs and the reference mapping.

Estimate-only runs return an empty video. That's deliberate: the node's whole safety story is that you see the price before the render.

Install

Manager: search PRYX ComfyUI Higgsfield, install, restart. Or 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: ..\python_embeded\python.exe -m pip install -r .\pryx-comfyui-higgsfield\requirements.txt. It pulls higgsfield-client, httpx and jsonschema into ComfyUI's own Python - install with the interpreter that runs ComfyUI or you'll get an import error on restart, which is the classic custom-node foot-gun. Needs ComfyUI 0.35.0+, frontend 1.53.6+, Python 3.10–3.13.

Key in Settings → PRYX ComfyUI Higgsfield → Credentials as a single key-id:key-secret value, or HF_KEY / HF_API_KEY+HF_API_SECRET for headless.

When it goes sideways

• "Missing required parameter" on a Genjutsu model almost always means you connected the clip and skipped the image.
• A weird-looking model ID is not a typo you made. The Genjutsu entries are spelled higgsfiled-genjutsu-motion-transfer-v1-0 and ...object-swap-v1-0 in the catalog - one L missing from "higgsfield". If you're pasting a model ID into an Advanced Request, copy it exactly as the dropdown spells it.
• Video too long or too big gets rejected at validation, before upload, which is the good kind of failure.
• estimate_only and an empty output is the mode working as intended, not a broken node.