happyhorse

The happyhorse skill generates and edits physically realistic videos using Alibaba's HappyHorse 1.0 models through the inference.sh CLI, invoked as belt. It solves the problem of producing character-consistent, physically realistic video and editing existing footage with natural-language instructions, without a custom integration — everything runs through belt app run commands. After belt login, users call one of four models by app ID and pass a JSON --input payload.

Four models are available: T2V (alibaba/happyhorse-1-0-t2v) for text-to-video with realistic motion; I2V (alibaba/happyhorse-1-0-i2v) to animate a single image via a first_frame; R2V (alibaba/happyhorse-1-0-r2v) to preserve characters from up to 9 reference_images; and Video Edit (alibaba/happyhorse-1-0-video-edit) to edit existing videos, optionally with reference_images (up to 5) and an audio_setting of auto, generate, or keep_original. All models support 720P/1080P resolution and up to 15 seconds duration. Documented parameters include prompt, duration (3-15, default 5), resolution, ratio (16:9, 9:16, 1:1, 4:3, 3:4, 21:9), seed, and watermark, with model-specific inputs like first_frame, reference_images, and video. Pricing is $0.14 per second at 720P and $0.24 per second at 1080P, with Video Edit billed on input plus output duration. Users can discover apps with belt app store search "happyhorse".

The skill targets creators building product demos, social media content, character-consistent videos, and AI-driven video edits. It requires the inference.sh CLI (belt), installable via npx skills add belt-sh/cli, and its allowed tool is limited to Bash(belt *). Related skills cover the full inference.sh platform and other video/image models such as Seedance and Google Veo.