Happy Horse / v1.0 / referenceToVideo

This node runs Happy Horse's reference-to-video mode entirely on Civitai's cloud. You feed it one or more reference images plus a prompt, and it generates a video that borrows the look, character, or style of those images. It's one of a handful of Happy Horse v1.0 nodes in the Civitai category (textToVideo, imageToVideo, referenceToVideo, videoEdit), and the pack itself is Civitai's official bridge to its own Orchestration API. No local GPU, no model download, no weights to babysit - the job happens on Civitai's fleet and you pay in Buzz.

This is the "reference" sibling to plain image-to-video: instead of animating a single source frame, you hand it several images and let the model carry the style or subject across a new scene you describe in the prompt. That's the setup for character consistency or a recurring visual identity without training anything, and it's the pattern the KB's external-API-nodes doc calls the "model isn't on your machine" case - you want it inside your graph, so you call it.

How it works

Happy Horse's engine is opaque - it's a hosted recipe, not open weights - but the node's plumbing is fully visible. When you hit Run, the node builds a videoGen workflow payload with the engine discriminator set to happyHorse / v1.0 / referenceToVideo, submits it to Civitai's orchestration endpoint, and long-polls the job until it finishes. When it does, it downloads the resulting MP4, writes it to ComfyUI's temp folder, and hands you a native VIDEO output plus the workflow_id and raw_json for debugging and cost inspection. Your reference images get encoded and uploaded as part of the request - which means they leave your machine and go to Civitai's servers. Fine for a character sheet you own; think twice before feeding it anything sensitive.

Inputs and outputs that matter

The required inputs are just two:

• prompt - what happens in the clip. Happy Horse is a cloud model with no community prompt-adherence lore yet, so write it like you would for any video model: subject, action, camera, environment, style.
• images - an IMAGE socket. Wire a Load Image (or several) in; this is what the video is referenced against.

The knobs are few and beginner-friendly:

• resolution - 720p or 1080p. Higher costs more Buzz and takes longer.
• duration - 3–15 seconds, default 5.
• seed - 0 for random; set it to reproduce a take.
• aspect_ratio - 16:9, 9:16, 1:1, 4:3, 3:4.
• api_config - optional Civitai Auth connection; otherwise it falls back to your token or stored login.

Outputs: video (the generated clip), workflow_id, and raw_json. video wires into anything that takes a VIDEO - a Video Preview, a Video Combine, or straight to an output.

Installing it

The pack installs once and gives you all the Civitai nodes. In ComfyUI Manager: Custom Nodes Manager → search "Civitai Comfy Nodes" (publisher civitai) → Install → restart. Or from source:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt

The only Python dependency is requests. There are no model files to download - everything runs server-side.

Common issues

• Auth errors. You need a Civitai account with Buzz. Credentials resolve from the Civitai Auth node, then CIVITAI_API_TOKEN, then a stored key/OAuth. On a headless box, set the env var - browser login won't pop.
• Cost surprises. Video jobs are metered per run, and 1080p/longer durations add up fast. The node's ui shows the workflow id and per-currency cost after each run; get in the habit of reading it.
• Early preview. The whole pack is still under active development - nodes and behavior can change without notice. Pin the version if a workflow depends on it.
• No browse button. Recipe nodes like this one don't let you pick a model inline; model sockets are AIR types you fill with a Civitai Model Selector. This Happy Horse node has no model socket, so just leave it be.