TraceFrame

Inspiration

Generative AI can create media in seconds, but creative teams often cannot prove which prompt, model, parameters, or source produced an asset. TraceFrame treats provenance as a product feature so every generation can be inspected, verified, and reproduced.

What it does

TraceFrame is a provenance-first AI image pipeline. It generates an image, stores the output and canonical manifest, records prompts, model parameters, timestamps, SHA-256 hashes, and storage URLs, displays a searchable provenance history, and can replay any previous run while preserving parent-child lineage.

How we built it

The Python FastAPI backend uses Genblaze 0.4.3 to orchestrate OpenAI gpt-image-1 through its official provider. Genblaze's S3-compatible ObjectStorageSink writes generated assets to Backblaze B2 using a hierarchical traceframe key layout. TraceFrame adds a SQLite provenance index, canonical manifest hashing, replay endpoints, and a responsive web interface. Docker, GitHub Actions, tests, sample data, and deployment documentation provide a production path.

The public judge-accessible deployment includes a credential-free demonstration mode. Supplying the documented OpenAI and Backblaze B2 environment variables activates the live generation-and-storage path without code changes.

Challenges we ran into

The hardest part was keeping the asset, canonical manifest, application record, and replay lineage consistent across external generation and storage boundaries. We also designed a safe demonstration path that exposes no credentials while retaining the same provenance model.

Accomplishments that we're proud of

• Meaningful Genblaze orchestration instead of a wrapper-only integration
• Durable B2-ready asset and manifest storage
• SHA-256 verification for both media and canonical provenance manifests
• Replayable runs with explicit parent lineage
• A complete Python, Docker, CI, and web deployment workflow

What we learned

Provenance must be captured during generation, not added after the fact. Canonical metadata and content hashes make verification portable, while Genblaze adapters keep the workflow independent from any single provider.

What's next for TraceFrame

Next steps include signed provenance attestations, automated integrity checks against B2 objects, additional image, video, and audio providers, collaborative review, and policy-driven retention.

Built With

• backblaze-b2
• genblaze
• openai-api
• python