MuAPI handles the 420+ model catalog; local inference is narrower
The README advertises more than 420 models across image, video, audio, lip sync, clipping, workflows, and creative agents. Open Generative AI puts them into one Electron or Next.js interface. The shared studio changes controls according to the selected model and input type. Uploaded references and generation history remain available in browser storage. For someone comparing many media models, that is less friction than learning a separate interface for each provider.
The architecture matters more than the feature count. Most generation requests go to MuAPI: the app submits a job, polls for completion, and uses hosted URLs for uploaded media and results. The checked-out repository has 329 files and about 85,421 lines of source, yet running that interface yourself does not move the main cloud models onto your server. You still need a MuAPI access key for that catalog.
Local inference is a separate desktop feature. The bundled sd.cpp engine can run selected image models on CPU or supported GPU back ends. Wan2GP adds Flux, Qwen-Image, and video families through a server you operate elsewhere. The README says that server needs CUDA or ROCm, so an Apple Silicon Mac can be the client but cannot host the described Wan2GP runtime.
What happened when we ran it
Our sandbox installed commit c90e908 in 47 seconds. npm added 1,239 packages and the checkout occupied 1,375 MB after installation. The run used a fresh unprivileged Debian container with 3 CPUs, 8 GB of RAM, Node 22, and no secrets. Installation succeeded, so the failure came at the next documented production step.
The build stopped with exit code 1 after 18 seconds. Next.js reported that AgentChatClient.js, AgentCreateClient.js, and AgentEditClient.js could not resolve ai-agent/dist/tailwind.css. The log shows the missing module import and does not establish why it was absent. We therefore cannot pin the result on submodules, workspace packaging, or any other unshown cause.
There was no test script or target, so tests were skipped. A tests directory exists, but our run had no project-level command to execute it. We also measured 0 CI workflow files. npm audit reported 32 known vulnerabilities: 1 critical, 25 high, 4 moderate, and 2 low. The repository includes a Dockerfile and a compose file, though neither changes the failed build result.
Source setup needs submodules before either app can start
The README requires Node.js 18 or newer and a recursive clone because workflow and agent packages arrive as submodules. Contributors are told to run npm run setup, which installs dependencies and builds the workspace packages before starting Electron or Next.js. A plain npm install is explicitly described as insufficient. That is a fair warning, but our documented build still failed at commit c90e908.
Prebuilt desktop apps remove the Node setup. They introduce platform trust prompts instead: the macOS build is not notarized, and the Windows installer is not code-signed. Ubuntu users are steered toward the .deb, which includes an AppArmor profile. AppImage users may need libfuse2 or a user-namespace policy change. Those are manageable steps, but they belong in any deployment or support estimate alongside the 1,375 MB npm footprint.
Local model weights run from 2.1 GB to 6.9 GB
The README lists local model files from roughly 2.1 GB to 6.9 GB and warns that one Z-Image setup can hang an 8 GB Apple Silicon Mac. The sd.cpp path can generate images without a MuAPI key once the user downloads an engine and weights. An environment variable can redirect the local model directory to another drive, which is useful when several files would fill the default application-data location.
Wan2GP is a larger commitment. It brings a separate Python and PyTorch service, GPU drivers, model storage, and a network endpoint for the desktop app. That route can keep generation on hardware you control, but it is a second system to patch and monitor. Our 3-CPU sandbox did not download model weights or measure generation quality, latency, or GPU memory, so this review makes no performance claim about either engine.
The 32 advisories and open XSS report block production use
Our run found 32 npm advisories, while open issue #309 gives a separate reason to inspect the desktop boundary. Its reporter says image, video, cinema, and lip-sync history values from localStorage are inserted into innerHTML without escaping. The report describes possible script execution and MuAPI-key exposure in Electron. We did not reproduce the issue, so it remains a public report rather than a confirmed sandbox result.
Of those 32 advisories, 1 was critical and 25 were high severity, while no test target or CI workflow was available to catch regressions. Open issue #337 separately asks for DOM coverage of the standalone video model picker. A team can investigate and patch these items, but the current evidence does not support placing the web app on a public production route.
August commits show activity; they do not repair our checkout
GitHub recorded the last push on August 24, 2026, and an issue received an update on August 25. The repository had 27,195 stars and 31 open issues and pull requests when fetched. The latest release is v2.0.0 from May 23, 2026. Recent pushes and issue activity show ongoing work, so the older release alone is not evidence that maintenance stopped.
Open Generative AI is easiest to justify as a desktop evaluation for someone who already accepts MuAPI or wants to test sd.cpp through a friendlier interface. The 47-second install makes inspection easy enough, but the failed 18-second build changes the production answer. Fully local media teams should compare ComfyUI or InvokeAI, while MuAPI users should still verify a newer checkout before distributing or hosting this code.