Genblaze Studio QC

Agentic multimodal studio with self-healing QC loops that evaluate AI generations, refine prompts on failure, and archive media, metadata, and C2PA provenance directly into Backblaze B2.

Comment

Inspiration

Generative AI models often produce visual artifacts, text hallucinations, or out-of-frame errors, forcing creators into frustrating manual retry loops. Furthermore, as creative pipelines scale, managing asset storage and verifying AI provenance becomes a massive data challenge.

We built Genblaze Studio QC to automate quality assurance in generative media. By pairing an agentic self-healing orchestration engine with immutable object storage, creators get production-ready media alongside a full cryptographic trail of how every frame was generated.

What it does

Genblaze Studio QC is an end-to-end multimodal production suite:

  • Multimodal Generation Studio: Generates high-fidelity Images, Audio, and Video via GMI Cloud models (stable-diffusion-xl-base-1.0 and seedance-2-0-260128), with resilient fallback mechanisms (pollinations.ai / Google TTS).
  • Agentic Self-Healing QC Loop: Evaluates generated media against automated quality thresholds (e.g., artifact density, resolution, prompt compliance). If a generation scores below 85%, the engine automatically refines parameters (like CFG scale) and retries up to 3 times.
  • Durable Backblaze B2 Storage Vault: Every approved asset is archived directly into a private Backblaze B2 bucket (genblaze-ai-media-vault).
  • C2PA Provenance & Sidecar Metadata: Simultaneously generates JSON sidecar files containing model provenance, generation latency, cost metrics, and full QC retry history under a metadata/ prefix.
  • Stateless B2 Gallery: The Next.js UI queries the Backblaze B2 bucket in real-time via S3 SDK APIs (ListObjectsV2Command), fetching JSON metadata directly from B2 without relying on a traditional SQL database.

How we built it

  • Frontend: Next.js 14 (App Router), React, Tailwind CSS, Lucide Icons.
  • Backend Orchestrator: Python FastAPI server natively running the genblaze SDK and b2sdk.
  • Storage Layer: Backblaze B2 Cloud Storage (using AWS S3-compatible endpoints for real-time reads and native SDK for uploads).
  • AI Model Pipeline:
    • Image: GMI Cloud (stable-diffusion-xl-base-1.0)
    • Video: GMI Cloud (seedance-2-0-260128 ByteDance model)
    • Audio: Google TTS & Fallbacks

Challenges we ran into

  1. Managing Real-Time B2 Metadata Synchronization: Rather than using a database, we architected the system so the Next.js frontend fetches JSON sidecars directly from Backblaze B2. Optimizing ListObjectsV2Command requests to keep the UI snappy required careful prefix structuring (/assets vs /metadata).
  2. Implementing Automated QC Retry Logic: Building a reliable feedback loop that adjusts parameters gracefully without hitting rate limits required tuning prompt-adjustment heuristics in our FastAPI orchestration layer.

Accomplishments that we're proud of

  • Successfully integrated Backblaze B2 as the exclusive source of truth for both heavy media binaries and lightweight JSON state sidecars.
  • Built a functional self-healing generative workflow that turns raw AI outputs into verified, production-grade media assets.
  • Designed a clean, secure UI (backblaze2026) that feels ready for production agency deployment.

What we learned

  • Deep integration patterns using genblaze and b2sdk in Python.
  • How S3-compatible cloud storage like Backblaze B2 can replace traditional databases for asset-heavy workflows when paired with JSON sidecar metadata patterns.

What's next for Genblaze Studio QC

  • Adding automated video frame rate and audio sync evaluation into the self-healing pipeline.
  • Integrating C2PA hardware signature signing directly on B2 bucket upload triggers using Backblaze Event Notifications.

Built With

  • b2sdk
  • backblaze-b2
  • c2pa
  • fastapi
  • genblaze
  • gmi-cloud
  • json
  • next.js
  • python
  • react
  • s3-api
  • stable-diffusion-xl
  • tailwind-css
Share this project:

Updates