Open Generative AI Guide
Self-host an open generative AI studio for image, video, and cinema workflows — local models, API routing, and a repeatable production pipeline.
What Open Generative AI is
Open Generative AI (OGA) is a free, open-source creative studio for image, video, audio, lip sync, and cinema workflows. It supports 200+ models — Flux, Kling, Veo, Wan, LTX, and others — through a single self-hosted interface. You can run inference locally, route to your own API keys, or use hosted endpoints without platform guardrails or per-seat markup.
Use this when you want a unified production desk instead of juggling five separate SaaS tabs.
Choose your setup path
| Path | Best for | Requirements |
|---|---|---|
| Hosted (muapi.ai) | Fastest start, no GPU | Free account, browser only |
| Self-hosted Docker | Full control, your keys | Node.js 18+, Docker optional |
| Local GPU stack | Zero per-gen fees | NVIDIA 12GB+ VRAM, CUDA |
Start hosted to learn the studios. Move self-hosted when volume justifies it.
Self-host install
git clone https://github.com/ncsound919/Generative-Video-AI.git
cd Generative-Video-AI
npm install
cp .env.example .env
npm run dev
Open http://localhost:3000. Add API keys in Settings for the providers you use (Replicate, Fal, OpenRouter, etc.).
Studio map
| Studio | Use case |
|---|---|
| Image | Text-to-image, img2img, multi-reference (up to 14 images) |
| Video | Text-to-video, image-to-video, motion control |
| Cinema | Photoreal shots with lens, focal length, aperture controls |
| Lip Sync | Portrait animation driven by audio |
| Workflows | Chain image → video → upscale in one pipeline |
| Agents | Multi-step creative tasks with tool use |
Cinema workflow (step by step)
- Script the shot — one sentence: subject, action, mood, camera move
- Generate keyframe — Image studio, Flux or Seedream, 16:9 aspect
- Animate — Video studio, image-to-video with Kling, Veo, or LTX
- Extend — Smart-continue or last-frame chaining for longer clips
- Upscale — Export at 720p/1080p, run through Real-ESRGAN or Topaz if you need 4K
- Audio — Lip sync studio or external DAW for VO and music
Prompt template for cinematic shots
[Subject] [action] in [environment].
Lighting: [golden hour / neon / soft box].
Camera: [35mm / 85mm], [wide / medium / close-up].
Movement: [static / slow dolly / orbit].
Mood: [tense / warm / documentary].
Avoid: text overlays, warped hands, flicker.
Model selection cheat sheet
| Need | Start with |
|---|---|
| Fast iteration | Flux Schnell, SDXL Turbo |
| Hero still | Flux Pro, Seedream |
| Human motion | Kling, Runway |
| Physics / lighting | Veo, Luma |
| Long clips | LTX-2, chained segments |
| Lip sync | Infinite Talk, LTX Lipsync |
Production checklist
- Keyframe approved before burning video credits
- Consistent aspect ratio across sequence (16:9 or 9:16)
- Reference images locked for character continuity
- Audio synced in post if not using lip-sync studio
- Exports named:
project_scene_take_v01.mp4
Cost control
- Generate stills cheap; animate only approved frames
- Batch similar prompts — model warm-up wastes less time
- Self-host routing: you pay API providers directly, no middleman markup
- Cache seeds and prompts in a project folder for reproducibility
Complementary local tools
For GPU-heavy local work, pair OGA with:
- ComfyUI — node-based workflows for LTX-2 and custom pipelines
- Milimo Video — AI-native NLE with timeline editing
- FFmpeg — concat, trim, format conversion
Troubleshooting
| Issue | Fix |
|---|---|
| Model timeout | Reduce resolution; switch to faster model tier |
| Character drift | Lock reference images; lower denoise on img2img |
| Jerky motion | Shorter clips; use last-frame continuation |
| API errors | Check .env keys; verify provider quota |
Open Generative AI is MIT-licensed. Repo: github.com/ncsound919/Generative-Video-AI