Breathing Color Back into Memories with DeOldifyRevisioned
How I extended Jason Antic’s original DeOldify with FastAPI, modern deployment patterns, and ergonomic defaults for creative teams.
When the original DeOldify project was archived in 2024, I forked and rewired it into DeOldifyRevisioned—a colorization service that fits modern creative pipelines. The goal was simple: make it effortless for product teams to drop vintage media into an API and get consistent, vibrant results back.
WORKING!
From notebook to production
The archival repo shipped with notebooks and a CLI. Useful for tinkering, not for teams. I introduced a FastAPI layer that exposes `/colorize`, `/test`, and `/results/{id}` endpoints, all documented with Swagger. Combined with render-factor presets, the API lets editors choose between artistic and stable models without touching Python.
- **Artistic** model for marketing stills.
- **Stable** model for portraits and historic archives.
- Video mode with frame-aware batching to keep performance predictable.
Infrastructure choices
Colorization is GPU hungry, so the project ships with reproducible conda environments, CUDA checks, and simple `docker-compose` files. For production, a Render or RunPod deploy pulls model weights from Hugging Face at boot, caches them in a shared volume, and reports health via the FastAPI root endpoint.
I also exposed environment variables—`MODEL_TYPE`, `RENDER_FACTOR`, `RESULTS_DIR`—so creative ops can tune fidelity without redeploying. Those knobs proved invaluable for agencies that bounce between 4 GB and 16 GB VRAM machines.
Human workflows first
The most satisfying feedback came from archivists who batch hundreds of family photos. They script against the API using curl or Python, dump results to S3, and review them in Notion galleries. DeOldifyRevisioned made the jump from a hacker’s weekend project to a dependable service these teams can trust.
Great ML tooling isn’t defined by how many models it offers—it’s defined by how calm it makes people feel when they put irreplaceable media in its hands.
If you want to explore the code, the repo lives at https://github.com/sarfarazh/DeOldifyRevisioned. Contributions are welcome, especially around multi-tenant auth and GPU orchestration patterns.