HappyHorse 文生视频

Text-to-video without a GPU, a model download, or a single local render. That's the pitch for HappyHorse T2V, and for once it's accurate.

This is the entry point to the Comfyui-Wenyi-demo1 pack (author WY-S, a demo project), and the easiest of its four video nodes to get working. You type a prompt, hit run, and Alibaba's closed HappyHorse 1.0 model does the rendering on its servers - your machine just watches the async job and downloads the result. If you heard the HappyHorse hype and hoped for open weights, stop looking. This is an API wrapper, not a local model. And if anyone offers you "HappyHorse open weights," it's a scam - there are no weights, and Reddit flagged a wave of fake repos the week the model appeared on the arena. The price of that convenience is the usual API-node one: your prompt leaves the machine and every render is metered.

The whole pack - author, install, shared gotchas - lives in the README. T2V is the node that needs the least of it: just an API key, and it works.

How it works

It's a thin async HTTP client. On run, the node POSTs your prompt to Alibaba's DashScope video-synthesis endpoint, gets back a task_id, then polls roughly every 15 seconds until the job finishes, then downloads the mp4 into ComfyUI/output/happyhorse/. Because it's flagged as an output node you don't even need to wire anything downstream - run it, and the clip plays inline in the canvas.

The nice part: T2V is the only HappyHorse node that doesn't need Alibaba Cloud OSS. The sibling nodes upload your images to object storage first; this one has nothing to upload, so it's just key + prompt.

The inputs that matter

• api_key - your DashScope (阿里百炼) key. Fill it once and the node writes it back into config.json, so the next run starts prefilled. This is the one input that will actually block you.
• prompt - multiline. There's no negative prompt, so put the whole scene in the positive.
• resolution - 720P or 1080P, default 1080P.
• ratio - 16:9, 9:16, 1:1, 4:3, 3:4, default 16:9.
• duration - 3 to 15 seconds, default 5.
• seed - for reproducibility.
• watermark - default on; this adds the Happy Horse watermark into the render. Flip it off if you don't want it in the footage.

The model field is fixed to happyhorse-1.0-t2v; it's cosmetic.

Outputs

Three, and they're identical across every node in this pack:

• video - a native ComfyUI VIDEO object. Wire it into anything that consumes video, or just preview it.
• video_url - the Alibaba CDN link. It's public for roughly 24 hours, so treat it as temporary.
• video_path - the local path to the downloaded mp4.

Install

ComfyUI Manager is the easy path: search "Comfyui-Wenyi-demo1" (or "HappyHorse") and install. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/WY-S/Comfyui-Wenyi-demo1
cd Comfyui-Wenyi-demo1
pip install -r requirements.txt

Dependencies are light: dashscope, requests, numpy, Pillow, oss2, and a pinned urllib3==1.26.20. Restart ComfyUI and watch for ===== Comfyui-Wenyi-demo1 HappyHorse 节点已加载 ===== in the console. All nodes land under Add Node → HappyHorse.

Where people get burned

• The key. DashScope is Alibaba's Chinese cloud, and its video API has historically needed a mainland-China account with real-name verification - the friction everyone outside hits first. The README points at the Bailian console (bailian.console.aliyun.com) for a key.
• pip timeouts on the Windows portable build, which inherits your system proxy. Clear ALL_PROXY/HTTP_PROXY/HTTPS_PROXY before installing (exact PowerShell is in the README).
• No GPU at all: launch with python main.py --cpu; the node does no local compute, so it's genuinely fine on CPU.
• Outputs look wrong or missing after an update: the ComfyUI frontend caches /object_info. Hard-refresh (Ctrl+Shift+R) or clear the site's local storage - the README's Q1 walks through all three fixes.