HappyHorse 1.1 (BudgetPixel)

HappyHorse 1.1 is Alibaba's video model, and its whole pitch is that it keeps your subject consistent. Most text-to-video models are perfectly happy to morph a character into a different person halfway through a clip; HappyHorse leans on reference images to hold identity - up to nine of them - which puts it in the character-consistency tier of the video shelf rather than the "anyone, anything" tier. It does text-to-video, image-to-video from a start frame, and reference-images-to-video where up to nine images anchor who the subject is.

That reference count is the headline. Nine reference images is a lot of anchoring: feed it a character sheet, a wardrobe, a location, and the model has a real chance of keeping all of them straight across 4–15 seconds. It's 720p or 1080p, and pricing is per second of output (170 credits/second on the sticker), with a fixed resolution_pricing per resolution tier.

How it works

The usual BudgetPixel video flow: prompt plus optional start image and reference batch go into a POST /videos/happyhorse-1.1 style call, the node polls the status endpoint, and the finished MP4 downloads back as a core VIDEO. The nine reference slots are the interesting part - image batches you wire in map onto reference positions automatically, so a batch of nine becomes nine anchors without extra wiring.

A few constraints the brief is explicit about: 4–15 second clips, audio is determined by the model (there's no toggle - whatever it decides, you get), and there's no video-edit mode. This is a generation model, not an edit tool.

The inputs that matter

• prompt - required, multiline. Describe the action, not just the subject.
• reference_images - up to 9, the identity anchors. This is the node's superpower; use it.
• image - optional start frame for image-to-video.
• resolution - 720p or 1080p; the per-second price differs per tier.
• length_seconds - 4–15. Remember the per-second billing: a 15-second clip is the expensive end of the range.
• aspect_ratio - framing.

Output is a single VIDEO - wire it into Save Video or a frame extractor. You'll want a recent ComfyUI for the core video type.

Install

Manager (search "BudgetPixel") or:

cd ComfyUI/custom_nodes
git clone https://github.com/budgetpixel-ai/comfyui-budgetpixel.git
pip install -r comfyui-budgetpixel/requirements.txt

Restart, set BUDGETPIXEL_API_KEY or budgetpixel.json ({"api_key": "bpx_live_..."}) in the pack folder. Video inputs auto-upload via POST /v1/uploads; requests is the only dependency.

Where people get burned

Three things. First, per-second billing with a 15-second ceiling means the most expensive single clip is also the one that tempts you - a full-length 1080p render at 170 credits/second is a big number, and the Cost Estimate node (with length_seconds and resolution in the params) is the responsible move before queueing. Second, no audio toggle: if your workflow assumes silent video and then needs a voiceover, the model's fixed audio choice will need a strip/silence step rather than a parameter. Third, reference consistency has a ceiling - nine references help a lot, but they can't defeat a prompt that demands the character do something wildly outside what the references depict. Keep the prompt compatible with the anchors and HappyHorse does the rest.