@sogni-ai/sogni-creative-agent-skill/skills/video_generation.md

Sogni Creative Agent Skill: agent skill and CLI for Sogni AI image, 3D, speech, video, and music generation.

---
name: video_generation
description: Text-to-video synthesis with LTX-2.3, WAN 2.2, Wan 3, Wan 3.0 Enhanced, Seedance, HappyHorse, and MiniMax H3.
always_loaded: false
tool_names:
  - generate_video
---

# Video generation

Text-to-video synthesis with LTX-2.3, WAN 2.2, Wan 3, Wan 3.0 Enhanced, Seedance, HappyHorse, and MiniMax H3. Use when the user wants a new video clip generated from a prompt or a loose multimodal reference set.

## Tools

- `generate_video` — produce a video clip from text, Wan 3/Seedance/H3 multimodal references, or HappyHorse image references.

## Constraints

- Persona-driven video requests must always go through `image_editing` first to produce a conditioned image; never go straight to text-to-video for personas.
- For prompt-only variants with the same model, duration, dimensions, and references, use one Dynamic Prompt branch with `numberOfVariations`/`-n` instead of serial video calls.
- HappyHorse 1.1 is a Premium Spark vendor path: t2v is prompt-only, i2v uses one first-frame image, and r2v uses 1-9 image references. It accepts no reference audio or video.
- Wan 3 (`wan3.0-video`) is a Premium Spark unified vendor path: 2-30 s at fixed 30 fps with native audio and 480P/720P/1080P output. Use either first/last-frame anchors or a loose set capped at 10 images, 5 videos, 5 audios, and 20 files total; never combine the two reference modes.
- Seedance 2.5 Uncensored (`seedance2-5-uncensored`, also "Seedance Uncensored"/"Seedance Spicy") has every Seedance 2.5 mode and limit under its own model id; use it only when the user asks for it. Each account first accepts a one-time likeness and consent agreement in the Sogni app; on error 4103, show the returned `consentUrl` and `consentInstructions`, use the same account as the agent/API key, then return and retry after manual acceptance.
- Seedance 2.0 Mini Uncensored (`seedance2-mini-uncensored`, also "Seedance Mini Uncensored"/"Uncensored Seedance Mini" in any word order) has every Seedance 2.0 Mini mode and limit under its own model id; use it only when the user asks for it. It shares the Seedance 2.5 Uncensored likeness and consent agreement; on error 4103, show the returned `consentUrl` and `consentInstructions`, including the account and return guidance. Never accept for the user or retry while acceptance is pending.
- Wan 3.0 Enhanced (`wan3.0-spicy-video`, MuleRouter provider ID `w3.0-video`, aliases `wan3-enhanced`/`wan3-spicy`) supports fixed or smart 2-30 s at 30 fps, native audio, prompt expansion, 480P/720P/1080P, and adaptive/fixed ratios. Use either first/last-frame anchors or up to 10 image, 5 video, and 5 audio loose references; never mix the modes. It has no document/web context, watermark, negative prompt, source-video edit, or extend mode.
- MiniMax H3 t2v uses `minimax-h3-t2v`; H3 r2v defaults to Balanced R2V, `minimax-h3-r2v-balanced` (use `minimax-h3-r2v-balanced-2stage` for two-stage, 1080p or 2K output; select other tiers only when explicitly requested) and accepts up to 9 images, 3 videos, and 3 audio clips, capped at 12 files with at least one image or video. Address references with per-type `<Picture 1>` / `<Video 1>` / `<Audio 1>` tags and assign each one a role.
- MiniMax H3 keyframes on `generate_video` are Reference to Video only (`minimax-h3-r2v` and its turbo, balanced, FastH3 Reference, and two-stage forms): `keyframes=[{imageIndex, atSeconds}]` (up to 8, strictly inside the clip, `imageIndex` indexed like `endImageIndex`) pins extra images at chosen times. Keyframes are not references: never put them in `referenceImageIndices` or count them toward the reference limits, and R2V still needs its own image or video reference. The prompt names each keyframe `<Picture N>`, numbered in time order after the reference pictures, and describes it at its time. Text-to-video never takes keyframes.