Bobs Studio is the generation tool I wanted to use, so I built it. It is a frameless desktop client with a panel-docking workspace. My own GPU does local inference on one side, and cloud APIs for frontier models sit on the other.
The native inference engine
The native engine supports more than twenty model architectures across image and video, in every checkpoint format the ecosystem uses, including quantized weights for cards without datacenter VRAM. GPU work runs in its own process and streams live previews back while it renders. The app sets up its own Python environment on first launch, because nobody should have to install CUDA by hand, me included.
Built for daily production work
Distilled models are detected automatically and their constraints enforced, so one wrong slider cannot quietly ruin a generation. A combination that will not fit is refused before it loads, because the memory budget is checked against the VRAM that is actually free and not the VRAM the card advertises. Long jobs can be interrupted cleanly in the middle of a step. There is a hires-fix pipeline that does tiled upscaling and refinement without seams, and the same workspace drives a Discord bot relay, because the community I built still likes generating together.
Training moved into the app as well, so a run is set up, costed, and watched in the same place the results land. The language models the app talks to can now be local ones running beside the image engine. Both of those features exist because I kept leaving the app to do them and coming back with files that no longer matched what the app thought it had.
This is also the tool behind the Studio on this site. Everything there was generated and upscaled in this app, on checkpoints I trained.