← all repositories
backblaze-labs/genblaze

Swap Sora for Runway in one line, and prove where the pixels came from

Genblaze orchestrates generative video, image, and audio pipelines across providers like OpenAI, Runway, and NVIDIA NIM while generating a SHA-256-backed provenance manifest for every run.

505 stars Python LLMOps · EvalML Frameworks
genblaze
Collecting fresh signals — velocity needs a few days of history.
collecting data…
star history

What it does

Backblaze’s Genblaze is a Python orchestration layer that chains generative media steps—text-to-image, image-to-video, audio synthesis—across providers like OpenAI, Runway, Luma, ElevenLabs, and NVIDIA NIM. A fluent Pipeline API lets you swap one video generator for another by changing a single line, while every execution produces a canonical manifest recording the model, prompt, parameters, and a SHA-256 hash. The manifest can be embedded directly into output files or parked in Backblaze B2 (or any S3-compatible store) alongside the asset.

The interesting bit

Most AI SDKs stop at the API call; Genblaze treats the provenance trail as a first-class output. It essentially notarizes the supply chain of a synthetic media file—provider, model version, prompt, timestamp—so a run can be replayed or audited later. That obsession with lineage feels more like a storage company’s take on generative AI than a typical model wrapper.

Key highlights

  • Adapters for major generative platforms behind one Pipeline/Step API, with fallback chains that retry on MODEL_ERROR.
  • First-class S3-compatible storage sink with content-addressable layouts; designed for B2 but works with AWS S3, Cloudflare R2, and MinIO.
  • Manifests embed into .mp4, .png, .mp3, and other formats, or sit as sidecar JSON with a verifiable canonical hash.
  • probe_models and provider-contract tests aim to catch upstream API drift before production jobs break.
  • Includes a CLI for verify, replay, and extract operations.

Caveats

  • URL-only outputs (when a provider returns a link instead of bytes) do not pass Manifest.verify() unless you separately fetch and hash the bytes.
  • Chat/LLM features are standalone conveniences, not native Pipeline citizens—reach for LangChain if you just need a text loop.
  • The project is explicitly not for JavaScript UI apps or simple single-provider calls.

Verdict

Worth a look if you’re running multi-step generative media in production and need to track lineage, swap vendors, or archive outputs to object storage. Skip it if you’re building a chatbot or a one-off DALL-E script.

Frequently asked

What is backblaze-labs/genblaze?
Genblaze orchestrates generative video, image, and audio pipelines across providers like OpenAI, Runway, and NVIDIA NIM while generating a SHA-256-backed provenance manifest for every run.
Is genblaze open source?
Yes — backblaze-labs/genblaze is open source, released under the MIT license.
What language is genblaze written in?
backblaze-labs/genblaze is primarily written in Python.
How popular is genblaze?
backblaze-labs/genblaze has 505 stars on GitHub.
Where can I find genblaze?
backblaze-labs/genblaze is on GitHub at https://github.com/backblaze-labs/genblaze.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.