TizenryA/colab-gpu-studio

Overview

Colab GPU Studio is a self-hosted AI creation workbench aimed at people who want to drive ComfyUI on Google Colab from a phone. The React frontend is built around touch input, while a FastAPI backend holds the web UI, task queue and finished works on a user's own VPS. GPU time is requested only when a generation starts, and the Colab runtime is destroyed right after the result is saved, so compute units are not spent idling README lines 5–10.

Model weights and the ComfyUI environment are cached in the user's Google Drive: a cheap CPU runtime prepares the environment and downloads weights, while the GPU runtime only restores the environment, samples and shuts down. Four built-in workflows cover text-to-image (Krea+MOSS dual sampling, OpenMOSS) and MiniMax H3 first/last-frame and multi-reference image-to-video, with T4 or L4 recommended depending on the workflow built-in workflows.

Quick start

A read-only UI demo is available with the public password demo-readonly; it uses mock data and cannot generate, upload or authorize anything demo notes. For real use, deploy from the 0.1.31-dev.8 release or the source tree.

The Docker path is the recommended VPS route. Prerequisites are Docker, Docker Compose and Python (used to generate the initial random secrets):

uv run python scripts/init_env.py --site-url https://your-domain
docker compose up -d --build

The stack starts api, worker and watchdog containers and listens on 127.0.0.1:8788 only; put Caddy or Nginx with HTTPS in front for public access. Stable data lives in the studio-data volume, so avoid docker compose down -v Docker deployment.

For a local run you need Python 3.11+, uv, Node.js and npm:

uv sync --extra test
uv run python scripts/init_env.py --site-url http://127.0.0.1:8788
npm --prefix web ci
npm --prefix web run build

Then start three processes — Web API, queue worker and timeout watchdog local run steps. Open http://127.0.0.1:8788 and log in with STUDIO_PASSWORD from the generated .env.

Enabling real GPU

Real GPU use is off by default (GPU_REAL_ENABLED=0), and the built-in workflows stay as drafts until a Google OAuth client and Drive are connected. Import the Google Cloud Web client JSON in settings, connect Colab and Drive, then enable workflows and set a budget limit before turning GPU on, following the OAuth setup and workflow review documents. CU consumption shown in the UI is an estimate; Google Colab's own billing is authoritative.

Constraints

• Secrets matter: .env, the database and the OAuth client JSON must not be published. ENCRYPTION_KEY decrypts stored OAuth tokens, so losing it means re-authorizing.
• The project is designed for single-person use; there is no multi-tenant isolation, and it should not be exposed as a public multi-user SaaS.
• Code is AGPL-3.0, meaning network services built on modified code must publish the full modified source. ComfyUI, third-party nodes, models and LoRAs keep their own licenses, and no model weights are included in the repository.
• scripts/verify_offline.py runs backend tests, Ruff, frontend contract tests and builds against local mock services without touching real Colab, OAuth, Drive or GPU, so it does not prove real cloud or model availability; the workflow file exists but has not run on GitHub Actions yet offline verification.

HotRepo's take

There is limited site data to go on: the project appeared once in HotRepo's new-project list for Python on 2026-11-06, gaining 10 stars and no forks in about an hour, with 117 stars and 18 forks in total. With no daily all-language rank recorded and only a two-point star curve, that is not enough to call a trend. The notable substance is the design choice itself: decoupling cheap CPU environment preparation from short GPU runs and reusing the Colab runtime only within a generation, which is a concrete cost control rather than a generic wrapper around ComfyUI.