-
-
Graph-node based workflow of attaching multiple I2V, T2V and T2I models
-
The main landing page of the application.
Inspiration
I wanted a simple, self-hosted desktop app for AI media generation that is cheaper, faster, and easier to manage than scattered online tools.
What it does
Zangetsu lets users generate images/videos using prompts, adapters, models, LoRAs, and graph-based workflows.
How we built it
Built with Tauri, Go backend, Go orchestrator, Python Diffusers workers, PostgreSQL, RustFS, Zitadel, Runpod, Prometheus, and Grafana.
Challenges we ran into
Managing GPU pods, cold starts, model loading, worker reliability, and different diffusion/video pipelines was the hardest part.
Accomplishments that we're proud of
We built a full self-hosted generation system with desktop UI, job queue, orchestrator, workers, auth, storage, and graph workflows.
What we learned
We learned how hard real AI infrastructure is beyond model inference: scheduling, storage, reliability, observability, and cost control.
What's next for Zangetsu
Improve pod cleanup, graph workflows, testing, deployment, and explore faster inference runtimes via tensorrt.
Log in or sign up for Devpost to join the conversation.