I built Image MetaHub because my local library hit 60k images and managing them was painful. Its a local desktop app (Electron/React) designed specifically for privacy and scale. It parses generation metadata from Stable Diffusion (A1111), ComfyUI, InvokeAi, Fooocus, Midjourney, and others, allowing for deep filtering by prompt, seed, model hash, and parameters.
Instead of running heavy ML models that eat up the VRAM needed for generation, I implemented a clustering engine that runs entirely on Web Workers. It uses Jaccard Similarity and Levenshtein Distance with prompt normalization to automatically stack 50+ variations of the same prompt into a single view.
Parsing ComfyUI workflows is tricky because of the spaghetti node graph. The app includes a declarative traversal engine to trace parameters backwards from SINK nodes, but I also built a specific Custom Save Node that forces a clean metadata dump at generation time to bypass the complexity entirely.
I just released v0.13.0 with a focus on performance and organization logic
In this update, I rewrote the file indexer (Phase B). I switched from asynchronous header reads to synchronous reads with controlled concurrency. This eliminated disk contention on my SSD and reduced read times from ~800ms to ~10ms per file, making indexing a lot quicker.
I maintain automated builds for Windows, macOS, and Linux (AppImage), so you can run it directly without needing to set up a Node environment.
I built Image MetaHub because my local library hit 60k images and managing them was painful. Its a local desktop app (Electron/React) designed specifically for privacy and scale. It parses generation metadata from Stable Diffusion (A1111), ComfyUI, InvokeAi, Fooocus, Midjourney, and others, allowing for deep filtering by prompt, seed, model hash, and parameters.
Instead of running heavy ML models that eat up the VRAM needed for generation, I implemented a clustering engine that runs entirely on Web Workers. It uses Jaccard Similarity and Levenshtein Distance with prompt normalization to automatically stack 50+ variations of the same prompt into a single view.
Parsing ComfyUI workflows is tricky because of the spaghetti node graph. The app includes a declarative traversal engine to trace parameters backwards from SINK nodes, but I also built a specific Custom Save Node that forces a clean metadata dump at generation time to bypass the complexity entirely.
I just released v0.13.0 with a focus on performance and organization logic
In this update, I rewrote the file indexer (Phase B). I switched from asynchronous header reads to synchronous reads with controlled concurrency. This eliminated disk contention on my SSD and reduced read times from ~800ms to ~10ms per file, making indexing a lot quicker.
I maintain automated builds for Windows, macOS, and Linux (AppImage), so you can run it directly without needing to set up a Node environment.