Happy Horse Video Edit (Ace Data Cloud)

Maskless instruction editing is where video went in 2026: you stop masking and start describing. "Replace the background with a sunny meadow" is the author's own default prompt on this node, and it tells you exactly what the interaction is - no mask, no ControlNet, no SAM pipeline, just a sentence and a source clip. Locally, Bernini on a Wan 2.2 base is the open-weight answer to the same job. This node is the hosted answer, and it's the most expensive thing in the pack per second of output.

What you feed it

video_url is required and must be a public HTTPS URL - the API downloads the source itself, so a local path is out, as are http:// links, 127.0.0.1, .local hostnames and private IPs. The flip side: if the clip came out of Happy Horse in the first place, you already have exactly the URL it wants, sitting in a Retrieve node's video_url. Chaining a generation into an edit is the one place you don't have to host anything yourself.

prompt is the edit. Be specific about what changes and stay quiet about what doesn't - a short, surgical instruction beats a paragraph, because everything you describe is something the model will try to alter.

reference_image_urls is optional here, and it's a multiline string taking the same formats as the reference-to-video node: a JSON array, or one URL per line, up to five. Use it when the change needs a source of truth - the new background, the replacement wardrobe. Leave it empty and you're doing a pure instruction edit.

The widgets that aren't there

There's no duration and no ratio on this node, and that's not an oversight: the edit follows the input clip. Its length is your clip's length, its framing is your clip's framing. What you do get is resolution (720P or 1080P, the render target), watermark (off by default), and seed (default -1, meaning the server picks; anything below zero simply isn't sent).

audio_setting is the interesting one, with two values: auto and origin. Set it to origin and the source audio is preserved through the edit; auto lets the service decide. If your source clip has dialogue or a music bed you actually want to keep, this is the widget that saves you from re-syncing audio afterwards, and it's the kind of thing hosted editors have historically fumbled.

The version problem

model offers exactly one choice: happyhorse-1.0-video-edit. No 1.1. Every other action in this pack gives you a 1.0 and a 1.1 - text-to-video, image-to-video and reference-to-video all have both - so editing is the action running a version behind. Fine if you need the capability. Worth knowing before you assume the quality matches the 1.1 generation you liked.

Wiring and install

Outputs are the usual four strings: status, task_id, video_url, trace_id. It's a paid async submit, so video_url is empty on the way out - send task_id to Happy Horse Retrieve Task, its video_url to Happy Horse Load Video, and the resulting native video to Save Video.

cd ComfyUI/custom_nodes
git clone https://github.com/AceDataCloud/HappyHorseComfyUI.git
cd HappyHorseComfyUI && python -m pip install -r requirements.txt
export ACEDATACLOUD_API_KEY='paste-your-key-here'

Or through Manager: search Happy Horse, publisher acedatacloud, package Happy Horse by Ace Data Cloud. Dependency is requests and nothing else. Key goes in the environment of the ComfyUI process, then restart.

Before you use it

Two honest calculations. First, cost: a video edit is priced by the second of video you push through it, so a one-minute source clip is not a cheap experiment no matter how short your prompt is. Trim first if you only care about a few seconds. Second, alternatives: if you already run Wan 2.2, Bernini does instruction-driven edits and object removal locally with no per-call meter, and it's genuinely good at it - heavy on compute, but the cost curve is flat instead of linear. Reach for this node when you don't have the hardware, when you want somebody else's quality rather than your own, or when a clip is already coming out of Happy Horse and the edit is one more hop.

And the standing rule for every submit node in this pack: press Queue once. Output nodes re-run every time you do, and this one bills you for the privilege.