What Just Happened
Higgsfield charges $30/month. Freepik charges $24/month. Krea, OpenArt — same story. Stack them all and you're spending $80-100/mo to generate AI images and videos.
A developer named Anil Matcha pushed a project called Open-Generative-AI to GitHub. Within 24 hours it had over 8,000 stars. The repo description doesn't pull punches — it calls itself an "uncensored, open-source alternative to Higgsfield AI, Freepik, Krea, OpenArt."
What's actually inside: 200+ models (Flux, Midjourney, Kling, Sora, Veo, SDXL, Ideogram), image + video + lip sync + cinema studios, no content filters, MIT licensed, and the option to plug in your own API keys or run local models with no keys at all.
Why This Matters
The AI creative tooling market has been a subscription tax. You pay $20-30/month for one tool with one set of models, hit your monthly cap, and pay again for the next one. Each tool also nerfs outputs with content filters that block half the prompts you actually want to run.
Open-Generative-AI flips this. You pay only for the API credits you actually use, and you decide what you generate. If you want to run something locally with no API at all, models like Z-Image Turbo and SDXL work via stable-diffusion.cpp.
Translation:the same workflow that costs $80/mo across four subscriptions now costs whatever you put on a Replicate or Muapi account. For most people that's $5-10/mo of actual usage.
Two Ways To Run It
Option 1 — Use the hosted version (no setup): Go to dev.muapi.ai/open-generative-ai. Plug in your API keys in the settings panel. Start generating. Keys live in your browser's localStorage — they don't get shipped to anyone's server. This is the path if you don't want to touch a terminal.
Option 2 — Self-host (full control): Clone the repo and run it yourself. Tell Claude Code:
Claude reads the repo, installs the dependencies, runs the dev server, and tells you when it's live. Total time: about 3 minutes.
The repo:
MIT licensed · 8,000+ stars in 24 hours · actively maintained
What API Keys You Actually Need
You don't need all of them. Pick based on what you're generating:
Replicate — runs Flux, SDXL, most open models. Best general-purpose key. Pennies per image.
Muapi — unified API for premium models (Midjourney, Kling, Sora, Veo). Drop-in for the paid tools.
OpenAI — only if you want DALL-E or Sora 2.
Nothing — for local-only models (Z-Image, SDXL via stable-diffusion.cpp). Slower, but $0.
Start with Replicate. Add Muapi when you want premium models. Skip the rest until you need them.
What To Generate First
Three workflows to test it on:
1. Replace your Higgsfield Spark videos— try the Kling or Veo integration with the same prompts you'd use in Higgsfield. Same outputs, no monthly cap.
2. Lip sync your own UGC clips — drop a face video + audio file, get back a synced output. Replaces Captions and Synthesia for short-form content.
3. Run a "no filter" prompt that gets blocked elsewhere — anything legal but flagged by content moderation usually runs fine here. Useful for edgy ad creative, satire, or stylized work that mainstream tools refuse.
If you do creative work, this should be a one-time setup that saves you hundreds per year.
What To Watch Out For
Speed varies by API. Local models are slow. Replicate is fast. Muapi premium is fastest.
No filter ≠ no responsibility.You're still responsible for what you generate. Don't be weird with it.
It's MIT, not magic.If a model upstream goes paywalled, the integration breaks until somebody patches it. The repo is active, but it's still a community project.
Keys live in localStorage. Clearing your browser data wipes your keys — back them up somewhere.