genblaze-core

Python SDK for building generative AI pipelines across video, image, and audio — with built-in SHA-256 provenance.

genblaze-core is the core of genblaze, an open-source orchestration framework by Backblaze for composing multi-step AI media generation workflows. It gives you a single, provider-agnostic Pipeline API for text-to-video, text-to-image, text-to-speech, image-to-video, and audio generation — so you can swap models (Sora, Veo, Runway, Luma, Flux, DALL·E, ElevenLabs, Stable Audio, LMNT, GMICloud) without rewriting pipeline logic.

Every pipeline run emits a canonical, hash-verified provenance manifest — a tamper-evident JSON document capturing the provider, model, prompt, parameters, timestamps, and output asset metadata. Outputs are sha256-covered once sha256 is populated, typically by a storage sink or byte-returning provider path. Manifests can be embedded directly into PNG, JPEG, WebP, MP4, MP3, and WAV files, uploaded alongside assets to S3-compatible storage, or exported to Parquet for analytics.

Why genblaze-core

• One API for every generative AI provider — Pipelines, not per-vendor SDK glue. Fluent, composable, chainable.
• Built-in provenance — Every asset gets a SHA-256–verified manifest. Prove how media was made; detect tampering.
• Production-ready — Retries, timeouts, progress streaming, moderation hooks, OpenTelemetry tracing, step caching.
• Storage-agnostic sinks — Drop into Backblaze B2, AWS S3, Cloudflare R2, MinIO, Parquet, or local disk.
• Policy + privacy controls — Redact prompts, strip params, pointer-mode for sensitive content.
• Agent loops + templates — Evaluator-driven iteration, reusable pipeline and step templates.
• Zero lock-in — MIT licensed, typed, lazy imports, provider adapters are separate packages.

Features

Install

pip install genblaze-core

Optional extras:

pip install "genblaze-core[parquet]"   # ParquetSink for analytics
pip install "genblaze-core[audio]"     # Audio metadata embedding (mutagen)

Add provider adapters separately:

pip install genblaze-openai genblaze-google genblaze-runway genblaze-luma \
            genblaze-decart genblaze-replicate genblaze-elevenlabs \
            genblaze-stability-audio genblaze-lmnt genblaze-gmicloud

pip install genblaze-s3    # Storage backend for Backblaze B2 / AWS S3 / R2 / MinIO
pip install genblaze-cli   # Extract / verify / replay / index manifests

Quickstart — local, zero API keys

from genblaze_core import Modality, Pipeline
from genblaze_core.testing import MockVideoProvider

run, manifest = (
    Pipeline("hello-genblaze")
    .step(MockVideoProvider(), model="mock-v1",
          prompt="A drone shot over a coastal city at golden hour",
          modality=Modality.VIDEO)
    .run()
)

print(manifest.canonical_hash)   # deterministic SHA-256 of the run
print(manifest.verify_hash())    # True: canonical payload hash matches
print(manifest.verify())         # True when every output asset declares sha256

Quickstart — Sora + Backblaze B2 storage

Generate a video, upload it + its manifest to Backblaze B2, verify the hash:

pip install genblaze-core genblaze-openai genblaze-s3
export OPENAI_API_KEY="sk-..."
export B2_KEY_ID="..."
export B2_APP_KEY="..."

from genblaze_core import KeyStrategy, Modality, ObjectStorageSink, Pipeline
from genblaze_openai import SoraProvider
from genblaze_s3 import S3StorageBackend

storage = ObjectStorageSink(
    S3StorageBackend.for_backblaze("my-bucket"),
    key_strategy=KeyStrategy.HIERARCHICAL,
)

result = (
    Pipeline("hero-reel")
    .step(SoraProvider(), model="sora-2",
          prompt="Aerial flyover of a mountain lake at sunrise",
          modality=Modality.VIDEO, seconds=4, size="1280x720")
    .run(sink=storage, timeout=300)
)

print(result.run.steps[0].assets[0].url)   # durable B2 URL
print(result.manifest.canonical_hash)      # SHA-256 of the full run
assert result.manifest.verify()

Manifest.verify() does not fetch remote asset URLs. It verifies the manifest hash and requires output assets to declare valid sha256 values. If your code fetches asset.url, hash those bytes separately and compare them to asset.sha256.

Storage — Backblaze B2 recommended

Backblaze B2 is the recommended default sink for genblaze — purpose-built for large AI-generated media with S3-compatible APIs, resilient multipart uploads, Object Lock for immutable manifests, and strong cost economics at scale. One-liner credentials from B2_KEY_ID / B2_APP_KEY. See the genblaze-s3 backend for the full recipe plus AWS S3, Cloudflare R2, and MinIO variants.

Documentation

• Main repo: https://github.com/backblaze-labs/genblaze
• Architecture: https://github.com/backblaze-labs/genblaze/blob/main/ARCHITECTURE.md
• Feature docs: https://github.com/backblaze-labs/genblaze/tree/main/docs/features
• Runnable examples: https://github.com/backblaze-labs/genblaze/tree/main/examples

Related packages

Provider adapters: genblaze-openai · genblaze-google · genblaze-runway · genblaze-luma · genblaze-decart · genblaze-replicate · genblaze-elevenlabs · genblaze-stability-audio · genblaze-lmnt · genblaze-gmicloud

Storage + tooling: genblaze-s3 · genblaze-cli · genblaze-langsmith

License

MIT