HappyHorse-1.0

HappyHorse-1.0 (alibaba:[email protected]) is the video node in the Runware pack that behaves least like a generator and most like an editor. Its most interesting input is video: a source clip you feed in, and instead of generating from nothing, the model re-renders your footage according to a prompt. Prompt-based video editing - describe what changes, keep the motion and structure, change the style or content.

It's part of Alibaba's Tongyi video line (the Wan team's neighborhood), and like every video model in this pack it runs on Runware's cloud because that's the only way most people get it: no open weights, no local run, just an API key. Install the pack once - ComfyUI Manager → search Runware, or git clone https://github.com/Runware/ComfyUI-Runware into custom_nodes/ - add your Runware API key, and it's a node like any other.

How it works

The editing path: wire a video into the video field (URL or UUID - the pack routes video inputs as media references, not tensors) and prompt the transformation. Dimensions are inherited from the input clip, so what you feed is roughly what you get back, size-wise. Without a source video it falls back to generation, using referenceImages and frameImages for conditioning. The settings.audio dropdown (auto by default) decides how the model handles the source audio through the edit.

Inputs that matter

• video - the source clip (URL/UUID) for editing. This is the feature that makes it worth installing.
• positivePrompt - what to change. "Turn the car into a horse-drawn carriage" is the vibe of it; the model is about semantic re-rendering of footage.
• referenceImages / frameImages - anchors for generation/consistency when you're not editing a source clip.
• size - 13 presets, default 1080p (16:9).
• duration - 3–15 seconds. Editing inherits dimensions from the source but you still set length.
• seed, numberResults (1–4), safety gate - the standard Runware set.

The video output (VIDEO) goes into SaveVideo, same as the Seedance nodes.

The honest notes

• It's an editor, not a magician. Editing models keep the structure of the source and change what the prompt asks for. Feed it a clip where the transformation is plausible and it shines; feed it a clip and ask for something it can't reconcile with the footage and you get mush. Source clips with clean motion and clear subject matter edit dramatically better.
• Video input is a URL/UUID, not a socket. This catches people - you can't drag a file onto it. Host the clip or use Runware's upload flow to get a UUID, then paste it.
• Audio handling is a choice, not a given. settings.audio set to auto means the model decides; if the edit should be silent or should keep the original track, that's the field to control.
• Cloud latency applies. A full-res edit is a real round-trip to Alibaba's hardware; budget for it, draft small.

HappyHorse is the pack's sleeper: most people queue the generation nodes and never notice the one that rewrites footage. If prompt-based video editing is a workflow you've been wanting to play with, this is the cheapest way into it.