Happy Horse Text to Video (Ace Data Cloud)

Here's the thing that confuses everyone in the first five minutes: this node does not make a video. It makes a task ID. Press Queue and it fires a paid job at api.acedata.cloud, prints a status line, and hands you four strings. The video shows up later, through a different node, using that ID. Understand that split and this pack is easy. Miss it and you will pay for the same clip three times.

What Happy Horse actually is

Happy Horse appeared in the video-arena leaderboards under a horse-shaped alias in early 2026, and for most of April nobody knew whose model it was - Alibaba's other team, a Seedance 3.0, a disguised Veo. The theory-crafting got loud enough that fake HuggingFace repos and "open source" download pages started circulating - exactly the shape of scam this community warns about. It never went open. What shipped was a closed, API-only model: 720p/1080p, 3–15 second clips, native multi-language audio, sold per second through partner platforms. Ace Data Cloud is one of those partners, and this node pack is its front door.

That puts it squarely in the category of API-wrapper nodes: no weights, so no local door at any VRAM. The trade is data out (your prompt leaves your machine) and money out (metered per call). In exchange you get a video model you were never going to run on a 4070.

The two-graph pattern

The pack keeps submission and retrieval in separate nodes, and so should you - its examples ship as two files, 01-submit-text-to-video.json for the ID and 02-retrieve-and-save.json for the file.

The inputs that matter

prompt is the obvious one, multiline, and the author's own default is a white horse on a moonlit beach. model gives you two choices, happyhorse-1.1-t2v (default) and happyhorse-1.0-t2v - 1.1 is the newer of the pair, so start there. resolution is 720P or 1080P and nothing else. ratio is your framing: 16:9, 9:16, 1:1, 4:3, 3:4. duration is an integer from 3 to 15 seconds, default 5.

seed deserves a sentence. The default is -1, and that isn't "random seed 0" - the node only sends a seed when the value is 0 or higher, so -1 means "you pick" and the server decides. Set a real number if you ever want to re-run the same prompt with the same intent.

watermark is off by default; leave it off unless a downstream platform wants the mark.

The outputs, and where they go

Four strings: status, task_id, video_url, trace_id. On a submit, video_url is empty - that's not a failure, that's the async design. The one you care about is task_id. Copy it out of the status panel, or wire it into Happy Horse Retrieve Task, which polls until the clip is done and hands you a video_url you can feed to Happy Horse Load Video. The trace_id is the support handle: if a submission times out, quote it in the Ace console's request history.

Install

ComfyUI Manager → Custom Nodes → search Happy Horse → confirm publisher acedatacloud and package Happy Horse by Ace Data Cloud → install → restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/AceDataCloud/HappyHorseComfyUI.git
cd HappyHorseComfyUI
python -m pip install -r requirements.txt

That requirements.txt is one line - requests. No torch, no model files, no ffmpeg, because nothing runs locally. The pack wants ComfyUI 0.39.0 or newer, but only Load Video actually needs that (it imports ComfyUI's native VIDEO type), so a stale install will still submit tasks fine and then fall over at the last step.

Then give the process a key and restart it:

export ACEDATACLOUD_API_KEY='paste-your-key-here'
python main.py --cpu

ACEDATACLOUD_API_TOKEN works too. The key has to be in the environment of the ComfyUI process itself - launch from a desktop shortcut and your shell export isn't there, giving you an empty-key error that looks like a broken install.

Where people get burned

The submit node is an output node. ComfyUI runs output nodes on every Queue press, which means every Queue press is a new paid submission. The README's own instruction is "run once." So: submit, copy the ID, then switch to your retrieve graph. The node's own status text says it plainly - "Use Retrieve with this ID; do not resubmit to poll."

A submitted task is also not a completed video. If it's still pending, querying the same ID again is free; resubmitting is not. And before you queue twenty seed variants of a 10-second clip, check the price page and do the multiplication - this class of model runs around ten to fifteen cents per second on the reseller market, which adds up faster than people expect. Cost is the honest objection to every API node, and it's the one nobody argues with.