Skip to content

Redrob Studio

Available on GitHub

Local studio,
on your own GPU.

A web app that puts image and audio tools on one screen. Generation happens on a GPU you run, and nothing leaves for a hosted inference API.

Clone on GitHub

Every tool in one shell

Make it, upscale it, compress it, convert it, read it back. Image and audio tools share one surface and the same gates, so you are not hopping between apps to finish one piece of work.

  • Image: create, upscale, compress, convert, vision
  • Audio: speak, clone, design, music, transcribe, extract, convert
Every tool in one shell

The browser never touches the GPU

Workflows are assembled on the server from allowlisted templates, and progress streams back from there. The GPU address stays in server-side config and never reaches the browser.

  • Browser → your server → your GPU
  • GPU address lives in server-only config
  • Progress relayed live from the server
The browser never touches the GPU

Only work you can actually use

A local model will happily make something you cannot put in front of a client. So a prompt that asks for it gets stopped before generation, and anything that slips through gets caught once the file exists. Two passes, because a bad prompt and a bad image are not the same problem.

  • Before: an LLM reads the prompt and can refuse it
  • After: a VLM looks at the finished file
  • Zero false negatives on 6,000 labeled prompts
  • Still a help, not a substitute for your own review
Only work you can actually use

Bring your own model and check it

The shipped setup is a starting point, not a lock-in. Drop in your own checkpoint or LoRA, run it against the same prompt as the current one, and look at both results next to each other before you commit to it.

  • Swap in your own checkpoint or LoRA
  • Same prompt on both, results side by side
  • Compared in the tool, not in a spreadsheet
Bring your own model and check it

Without a GPU it only opens

With no GPU host attached, the UI loads but no tool produces output. There is no cloud fallback to quietly pick up the work, which is a condition of the design rather than a setup problem.

  • Generation needs a GPU host
  • No cloud fallback
  • Image generation inherits the Hangul and typography gap
Without a GPU it only opens
Every tool in one shell

And much more

Every module, and the rules we hold ourselves to, in one list.

Local by design

Browser → your app → your GPU process

Image generation

Runs Redrob Image inside the studio

Music generation

Runs Redrob Tune, released only inside Studio

Dual safety gates

LLM on the prompt, VLM on the output

No hosted API

No request leaves for hosted inference

Apache-2.0

Clone it, change it, ship it again

What we measured

The safety gates were measured on 6,000 labeled prompts across image, text, and audio: 99.92% accuracy, zero false negatives, five false positives. How the set was built and what the confusion matrix looks like are written up outside this page.

Read the measurement writeup

Get started

Clone the public repo. Run it locally. Inspect the method.

01

Clone Studio

git clone the public repo and install with yarn.

02

Attach your GPU

Point the app at the local inference process running on your GPU.

03

Run tools

Prompts and outputs stay on your GPU host.

Clone on GitHub
git clone https://github.com/redrob-labs/redrob-studio.git && cd redrob-studio && yarn install && yarn dev

Common questions

License, limits, and how the path actually runs. Claims travel with method and caveat.