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

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.