AI Video Generator with Consistent Characters
FTL Studio turns one sentence into a directed, multi-shot cinematic film — the same character and the same set in every shot. Open source, built on Google Veo 3.1, Gemini and Claude.
node ftl-studio.mjs "an old lighthouse keeper repairing the great lamp as a storm closes in" --shots 4
The problem: AI video characters change every clip
Everyone who has generated AI video knows the pattern — the first clip is magic, the second stars a different actor in a different room, and by the third your "film" is a slideshow of strangers. Text prompts cannot lock a face, because every generation re-imagines every visual fact.
The fix: compile prompts, then animate approved images
- Compile, don't write. A sentence is scattered into typed nodes, expanded through six lenses (form, material, motion, light, sound, physics), and converged under a precedence law so contradictions die before the model sees them.
- Lock identity in stills. A canon character still and set still are approved once, then every shot's first frame is composited from them and chained to the previous frame.
- Animate only approved frames. Image-to-video preserves the still's identity; motion prompts stay short and describe only what moves, the camera, and the audio.
- Edit like film. Hard cuts on action, unified grade, fine grain, loudness-normalised audio — one film, not four clips.
What it produces
| Stage | Output |
|---|---|
| Compile | plan.json — the reproducible film recipe: character, set, per-shot framing, motion prompts, audio |
| Canon | character.png, set.png — approved once, reused forever |
| Frames | shotN_frame.png — vision-audited before any video spend |
| Video | shotN.mp4 via Veo 3.1 or Gemini Omni Flash (image-to-video) |
| Film | film.mp4 — graded, hard-cut, mixed |
Frequently asked questions
How do I keep the same character across multiple AI video clips?
Generate a canon character still once, then composite every shot's first frame from that image and animate only approved frames. FTL Studio automates this and runs a vision judge on every frame before spending on video.
Can I generate a multi-shot AI film from a single prompt?
Yes — one command compiles the plan, generates the stills, animates each shot, and assembles the finished film.
What is the best prompt format for Veo 3?
Google's official formula: Cinematography + Subject + Action + Context + Style & Ambiance, with audio as separate SFX: and Ambient noise: sentences and dialogue in quotes. FTL Studio emits this automatically.
Which AI video models are supported?
Google Veo 3.1 and Gemini Omni Flash today, selected by one config line. Adapters for Sora, Kling, Runway and Luma are open contribution issues.
Is it free?
The code is MIT licensed and free. Model usage is billed by Google and Anthropic; a --skip-video mode lets you approve plans and stills for cents first.