Video models
Video models turn keyframe stills into motion. A shot on CreatorStudio is typically: image model generates a keyframe, video model animates that keyframe into a clip, audio models score and voice it, the scene editor composes and renders. The creator briefs once. Ra routes every stage, including which video model runs which shot.
All video models are accessed through Ra. Model choice is never the creator’s problem. Ship intent, not model IDs.
What video generation is used for
Section titled “What video generation is used for”- Motion from a keyframe. The most common path. Creator approves a keyframe; Ra picks the right motion model for the shot’s action class (dialogue, cinematic wide, object manipulation, ambient background).
- Multi-shot sequences. Scene-level pipelines where Ra chains multiple shots through possibly different models, staying consistent via the locked character and location references.
- Ambient and B-roll. Non-dialogue shots where pacing and atmosphere matter more than character fidelity.
Supported video models
Section titled “Supported video models”| Model | Provider | Strengths | Typical use |
|---|---|---|---|
| Veo 3 | Photoreal, dialogue-safe, strong prompt adherence | Character dialogue, documentary shots | |
| Sora 2 | OpenAI | Complex camera moves, cinematic composition | Cinematic wides, establishing shots |
| Kling 2.5 | Kuaishou | Strong body motion, natural gestures | Action, movement, body language |
| Runway | Runway | Versatile, fast iteration, video-to-video | Quick drafts, style transfers |
| Pika | Pika Labs | Short-form energy, stylized motion | Shorts, stylized sequences |
| Luma Dream Machine | Luma | Camera motion, fluid transitions | Dolly, pan, orbit shots |
| Seedance | ByteDance | Cinematic grain, ambient quality | Ambient, atmospheric, mood |
| Hailuo | MiniMax | Stylized performance, expressive faces | Stylized character work |
More models are routed as they ship. The app’s story-config picker exposes current and legacy variants (Kling v1 through v2.5, Runway Gen-2 through Gen-3, and others) for power users who want explicit control.
How Ra picks a video model
Section titled “How Ra picks a video model”For each shot, Ra considers:
- Action class. Dialogue lands better on Veo. Ambient lands better on Seedance. Complex body motion lands better on Kling.
- Continuity constraints. If the character was locked on Hailuo, staying on Hailuo protects face consistency across cuts.
- Length and aspect. Platform-specific constraints (9:16 vertical for Shorts, 16:9 for YouTube main). Some models handle vertical better than others.
- Credit budget. Higher-fidelity models cost more. Ra biases toward budget-fit models for B-roll and saves the premium budget for hero shots.
- Outcome signal. If the creator’s Memory shows Kling shots outperform Runway shots on their audience, Ra routes more Kling.
See How Ra picks a model for the full heuristic including the latent-vs-deterministic split.
Working with video models in the app
Section titled “Working with video models in the app”Video generation runs through the Movie Maker pipeline. Each scene’s Generate tab surfaces the pipeline (Keyframe → Video → Dialogue → Audio → Effects → Render). Creators can regenerate any stage without re-running earlier stages. Model choice per stage is available as an explicit override, but the default is always “let Ra route.”
Every generated clip is written to Assets with full provenance. Branch, remix, or re-render from any prior clip at any time.
Related
Section titled “Related”- Orchestration explains the thin-harness routing layer.
- The storyteller workflow shows where video generation sits in the end-to-end arc.
- Writing briefs is the single most important read for getting good video on the first render.