One repo that bundles image generation, video generation, lip sync, and cinematic camera controls into a single app. MIT licensed, 10.9K stars on GitHub, built by Anil Matcha. No subscription. You bring your own API keys and pay per call. Runs as a desktop app (Mac, Windows, Linux) or a web interface.
Repo: github.com/Anil-matcha/Open-Generative-AI
The problem nobody talks about
Higgsfield charges $100/month for a wrapper around the same AI models you can access directly. You're paying for the interface, not the AI. Most of the models underneath (Flux, Kling, Veo, Sora) are available through API providers at per-call pricing. If you generate 20 images a month, the subscription is a bad deal. If you generate 2,000, it might make sense. But you never get to see that math because the pricing is a flat monthly fee.
Open Generative AI strips the wrapper off. Same models, same quality, same workflow. The difference is you see exactly what each generation costs because you're paying the API directly.
Why this matters
200+ models in one UI. 50+ for text-to-image, 55+ for image-to-image, 40+ text-to-video, 60+ image-to-video, 9 lip sync models. Switch between Flux, Midjourney, Kling, Sora, and Veo without leaving the app.
Cinema Studio is the headline feature. Pick your camera (8K Digital, 70mm Film, S35 Studio), choose from 11 lens types, set focal length from 8mm ultra-wide to 85mm portrait, dial in aperture from f/1.4 to f/11. The app builds the cinematic prompt for you. This is what Higgsfield was charging for.
Local inference if you want zero API cost. The desktop app bundles sd.cpp for running Stable Diffusion models on your machine. Apple Silicon gets Metal GPU acceleration. You can also connect a Wan2GP server on a CUDA machine for Flux, Wan 2.2, and Hunyuan Video locally.
Workflow Studio for chaining operations. Node-based visual pipeline builder. Chain image gen into video gen into lip sync in one flow.
Step 1: install the desktop app
Download the installer for your platform from the GitHub releases page.
Mac (Apple Silicon): Download the .dmg arm64 file. After installing, you may need to bypass Gatekeeper:
xattr -cr "/Applications/Open Generative AI.app"Then right-click the app and select Open.
Mac (Intel): Same process, grab the x64 .dmg instead.
Windows: Download and run the x64 .exe installer. Standard install.
Linux: AppImage or .deb package. On Ubuntu 24.04+, if the AppImage fails:
sudo sysctl -w kernel.apparmor_restrict_unprivileged_userns=0Step 2: get your API key
Open Generative AI routes cloud model requests through an API gateway called Muapi. You need a key from their site to use the cloud models.
Create an account, generate a key, and the app prompts you for it on first launch. It stores the key locally. You only pay for what you generate. No monthly minimum.
Step 3: pick a studio and generate
Launch the app. You get five tabs across the top.
Image Studio. Type a prompt, pick a model (Flux Dev, Seedream 5.0, Ideogram, Midjourney, or 50+ others), set your aspect ratio and resolution. Hit generate. Some models support multi-image input where you upload up to 14 reference images.
Video Studio. Text-to-video or image-to-video. Models include Kling v3, Sora 2, Veo 3, Wan 2.6, Seedance 2.0, Hailuo, Runway. Duration and quality options depend on which model you pick.
Lip Sync Studio. Upload a portrait image or video, drop in your audio file. 9 models including Infinite Talk, Wan 2.2 Speech-to-Video, LTX 2.3. Resolution options up to 1080p.
Cinema Studio. This is the Higgsfield replacement. Select camera type, lens, focal length, aperture. The app constructs the full cinematic prompt from your selections. You still pick an underlying model for the actual generation.
Workflow Studio. Visual node editor. Drag and connect studios into a pipeline. Output of one feeds into the next.
Step 4: set up local inference (optional)
If you want to generate without hitting any API, the desktop app supports two local engines.
sd.cpp (bundled with the app): Go to Settings, then Local Models. Download the model you want (SD 1.5, SDXL, Z-Image Turbo, Dreamshaper, Realistic Vision). Metal GPU on Apple Silicon, CUDA/Vulkan/ROCm on Windows/Linux, CPU fallback everywhere. Z-Image models need 16GB RAM. SD 1.5 works on 8GB.
Wan2GP (self-hosted on a separate GPU machine): For local Flux, Wan 2.2, and Hunyuan Video. Requires a CUDA GPU. No Apple Silicon support for this engine.
git clone https://github.com/deepbeepmeep/Wan2GP
cd Wan2GP
./install.sh
python wgp.py --listen --server-name 0.0.0.0Then in the app: Settings, Local Models, enter your Wan2GP server URL.
Step 5: run from source (alternative)
If you want to self-host the web version or hack on the code:
git clone --recurse-submodules https://github.com/Anil-matcha/Open-Generative-AI.git
cd Open-Generative-AI
npm run setup
npm run devOpens at http://localhost:3000. Needs Node.js v18+.
For production:
npm run build && npm run startTo compile the desktop app yourself:
npm run electron:build # macOS
npm run electron:build:win # Windows
npm run electron:build:linux # LinuxThe cost math
This is the part most people skip. "Free" means no subscription. It does not mean zero cost.
Cloud API pricing through Muapi depends on the model and output size. Light use (a few dozen images, a handful of short videos per month) will cost well under $20. Heavy production use (hundreds of video generations) can outpace Higgsfield's $100/month.
Before you switch: estimate your monthly generation volume, check the pricing for the models you actually use, and compare against your current bill. For most people making content a few times a week, the per-call model wins by a lot.
Local inference with sd.cpp or Wan2GP costs nothing after hardware. But the model selection is much smaller and quality varies compared to the cloud options.
Honest limitations
Muapi dependency.Cloud models all route through one API gateway. If it goes down or changes pricing, you're affected. No direct API key passthrough for other providers.
Local inference is limited.sd.cpp handles Stable Diffusion variants but nothing close to Flux or Kling quality. Wan2GP needs a separate CUDA machine. Apple Silicon users can't run Wan2GP locally.
Z-Image can hang 8GB machines. The README warns about this. If you have 8GB RAM, stick to SD 1.5 or Dreamshaper for local work.
"Uncensored" has a caveat. The app itself has no content filter, but the underlying model providers still enforce their own terms.
Desktop-only features. Local inference only works in the Electron app. The web version is cloud-only.