backblaze-labs/genblaze

What it solves

Genblaze is a pipeline SDK designed to orchestrate generative AI media workflows across video, image, and audio. It solves the problem of provider lock-in by providing a unified API to swap between different AI models (e.g., Sora, Runway, ElevenLabs) without rewriting orchestration logic. Additionally, it addresses the issue of media provenance by automatically generating a canonical, hash-verified manifest for every run, ensuring that the origin and parameters of generated assets are trackable and tamper-evident.

How it works

The SDK uses a Pipeline API to define multi-step generation workflows. Each Step specifies a provider, model, and prompt. The system integrates with various provider adapters (such as OpenAI, Google, and NVIDIA NIM) to handle the API lifecycle. Generated assets are handled by Sinks, which can persist media and their corresponding provenance manifests to S3-compatible storage (like Backblaze B2) or local Parquet files. The provenance manifest captures the full chain of provider, model, prompt, and parameters, and can be embedded directly into media files like .mp4 or .png.

Who it’s for

It is intended for developers building production-grade generative media applications who need to manage multiple AI providers, maintain a strict audit trail of generated content (provenance), and integrate durable storage for their assets.

Highlights

• Unified Provider API: Supports a wide array of providers including OpenAI, Google, Runway, Luma, and ElevenLabs across multiple modalities.
• Built-in Provenance: Automatically generates SHA-256 covered manifests that can be embedded in media files or stored as sidecars.
• S3-First Storage: Native integration with Backblaze B2 and other S3-compatible stores for durable, content-addressable asset management.
• Replayable Runs: Every run is captured in a manifest, allowing users to reconstruct a generation via the CLI using genblaze replay.
• Fallback Chains: Supports fallback_models to automatically retry generation on model errors.
• Multi-step Orchestration: Ability to chain steps (e.g., text-to-image then image-to-video) within a single pipeline.