GOAT UGC AI

Type a topic. Ship a product video.

GOAT UGC AI is an open-source, self-hostable AI video studio for app founders, creators, and agencies. You give it a brief; it plans the shot list, writes every prompt, generates real motion clips on fal.ai (Seedance 2.0 / Kling v2.1), and stitches them in a Remotion player you can preview in the browser and render to MP4 locally.

Live demo · goat-ugc-ai.vercel.app

Built by Burhan Kocabıyık. Inspired by and adapted from goatstarter/GOAT-youtube (Remotion patterns, scene component, prompting guide).

---

What it does

Stage	What happens	Powered by
1. Director (/api/video/plan)	Topic → scene-by-scene shot list with image prompts, motion prompts, voice lines, duration, and animation.	fal-ai/any-llm (default openai/gpt-4o-mini)
2. Visuals (/api/generate/image) — image-anchored mode	Anchor frame for each scene at 1280×720.	Flux Schnell · Flux Dev · Flux 1.1 Pro · Nano Banana Pro · Nano Banana 2 · Seedream v4 · SDXL
3. Motion (/api/generate/video)	Real AI motion clips per scene — image-to-video or text-to-video.	Seedance 2.0 · Seedance 2.0 Fast · Kling v2.1 Master / Pro / Standard · Seedance 2.0 t2v
4. Stitch (client)	Remotion Player composes scenes with fades, subtitles, optional Ken Burns. Edit / reorder / regenerate scenes inline.	@remotion/player, remotion, jszip

Up to 5 minutes total / 30 scenes per video. All scenes render in parallel.

---

Two creation modes

• Image-anchored (default) — generate a still per scene, then animate it. More controllable, lower variance, great for product UGC.
• Text-to-video — skip stills entirely; send the scene prompt straight to Seedance 2.0 t2v. Best for abstract/dynamic shots where you don't need pixel-level control of the first frame.

A toggle in the brief panel switches between them; the available video models update accordingly.

---

Inline scene editor

Right under the live Remotion player you get a Remotion-style timeline editor. For every scene:

• ↑ / ↓ to reorder
• ✕ to delete
• "+ Add scene" to insert a new one
• Inline edit subtitle / voice line, image prompt, motion prompt
• 5 s / 10 s clip length toggle
• Camera animation picker (zoom in/out, pan L/R, breathing, static) for the still
• Per-scene Regenerate button (re-runs only that one)
• Top toolbar: "Render missing scenes" and "Download assets (.zip)"

Edits to subtitle / animation / duration update the player live. Prompt edits only kick in on the next regeneration so you don't burn credits.

---

Self-host

1. Local dev

git clone <this-repo> goat-ugc-ai
cd goat-ugc-ai
cp .env.example .env.local
# Get a key at https://fal.ai/dashboard/keys, paste into FAL_KEY
npm install
npm run dev   # http://localhost:3000

2. Deploy to Vercel

vercel link
vercel env add FAL_KEY production       # paste your fal.ai key when prompted
vercel env add AI_PROVIDER production    # value: fal
vercel deploy --prod

That's it — no databases or auth required for the core flow. The dashboard's "Projects" view persists workspaces in localStorage until you wire a real DB.

3. Run on your own GPU (no fal.ai)

AI_PROVIDER=local LOCAL_INFERENCE_URL=http://127.0.0.1:7860 npm run dev

The local adapter is ~30 lines (lib/providers/local.js) — point it at your Ollama / ComfyUI / sd.cpp endpoint and adapt to your runtime's request shape.

---

Environment variables

Var	Required	Purpose
FAL_KEY	yes (cloud)	fal.ai API key
AI_PROVIDER	no	fal (default) or local
FAL_LLM_MODEL	no	Override scene-planner LLM (default openai/gpt-4o-mini)
LOCAL_INFERENCE_URL	yes (local mode)	Base URL of your local inference server

See .env.example.

---

Architecture at a glance

app/
├── page.js                       Marketing landing
├── video/                        Primary surface — brief, planner, editor, player
├── create/                       Single-image quick-fire
├── dashboard/                    SaaS shell + projects
└── api/
    ├── config/                   Active provider + capability summary
    ├── generate/image/           Provider-abstracted image generation
    ├── generate/video/           Provider-abstracted video generation
    └── video/plan/               Scene planner via fal-ai/any-llm

components/
├── SaasNav.js                    Top nav across SaaS pages
└── remotion/
    ├── Scene.jsx                 Ken Burns + fade + subtitles
    ├── VideoComposition.jsx      <Series> of scenes, 1280×720 @30fps
    └── Player.jsx                Client-side <Player> wrapper

lib/providers/
├── config.js                     AI_PROVIDER resolution + capability map
├── fal.js                        fal.ai image + video adapter (Seedance 2.0, Kling v2.1, Flux, Nano Banana 2/Pro)
├── fal-llm.js                    fal any-llm scene planner
├── local.js                      Local HTTP inference adapter
└── index.js                      Unified façade

---

Contributing

PRs welcome. The code is intentionally small and reads top-to-bottom — no framework gymnastics. Open an issue first for anything bigger than a 20-line change so we can align on direction.

---

Credits

• Remotion — the renderer that makes in-browser composition possible
• fal.ai — every generation in this pipeline
• goatstarter/GOAT-youtube — the Remotion + Scene component blueprint we adapted
• Built by Burhan Kocabıyık

---

License

MIT.