• 

  Enable GenBlaze Super Adapter via GMI Cloud

• 

  Backblade B2 storage via application or master key config

• 

  Genblaze containers are built only when config is added

• 

  Model composable inference - Hover to see adapter preference

• 

  Preview available models and adapters in agent while selecting

Inspiration

With Samsar, I have always envisioned it as being a decentralized networked router- the backbone agent, connecting the various adapters and orchestrating them. BlazeConnect is the next step in this direction.

The Devpost hackathon gave me the opportunity to achieve another step in making the agent more robust and decentralized while also ensuring predictable outputs across adapters and models. The Text-to-visualize anything framework has been to allow multiple adapters and model settings. Previous version of the community edition supported only S3 and Minio based local file-system. I wanted to add additional remote S3 compatible storage adapter which was affordable to end users, B2 exactly matched with the requirements and a super adapter capable of handling LLM, Image and Video calls. By combining GMI Cloud’s GenBlaze infrastructure with reliable, scalable Backblaze B2 storage, I was able to achieve all of my goals while also making the platform more distributed and affordable as well as enable fallback mechanisms and more options for the end-user.

What it does

The core product allows user to create prompt to video using any combination of inference, video and image model settings up to 3 minutes in duration. BlazeConnect allows user to setup GenBlaze via GMI cloud for the Samsar pipeline. This setting when enabled allows GenBlaze to be used as the primary provider for all parts for the text to video and image list to video agent. This integration also allows user to setup Backblaze B2- S3 compatible storage server, for user to store their renders and images in a remote file-system to protect against accidental deletion etc. and is more affordable option with fast access and easy setup.

How I built it

I deep connected GenBlaze and GMI cloud adapters in all service projects, then I connected the GenBlaze generation pipeline to Backblaze B2 using its S3-compatible API, enabling generated files to be uploaded and retrieved automatically. This integration was built with heavy assistant of Codex with GPT 5.6 Sol over multiple long prompt, verification and debugging sessions.

I added adapters for GMI Cloud for all services, and added listeners in the services as well, in the generators and listeners for GenBlaze, the system pre-processes the payload to make it compatible with Genblade API for each supported model before sending the generate or polling request.

Challenges I ran into

Normalization of model keys in GMI Cloud and mapping against the Samsar Model keys.

Limited concurrency (perhaps due our account to usage tier) However user can use their own GMI Cloud API key, the GenBlaze container shipped as part of the bundle is fully open-source.

Seedance 2 model integration faced some challenges, such as high refusale rate, output intermittently never returning and getting stuck for ever.

Accomplishments that I am proud of

I built the text to video and image list to video agent for over 3 years and iterating over each model release which is now a premier model composable framework to visualize anything from text prompts.

The latest Genblaze and B2 cloud integration makes the platform even more distributed and robust than before.

Users gets the most choices for adapters and models as well as the best in class technology for 1-shot Text to video and image list to video.

The app runs in secure sandboxed environment and the user only enables the credentials they need to minimally enable all the services for T2V agent.

What I learned

I learnt more about the GenBlaze platform and GMI Cloud, this was my first introduction to the BackBlade generative media cloud and I was pleasantly surprised to find that there is a generative adapter available with such comprehensive tooling, which seem to be exactly what a stateful long running agent like Samsar needs. I learnt how to handle models adapter wise and also that behavior of same model across adapters differs substantially.

I learnt that payload input needs to be normalized across adapters even for the same models.

I have also blogged about my learnings here https://samsar.one/blog/backblaze-b2-cloud-and-genblaze-adapter-are-now-configurable-in-samsar/

What's next for BlazeConnect

BlazeConect is the GenBlaze and B2 integration within Samsar monorepo. I will continue maintaining model and version upgrades for GenBlaze and parity with the latest features in Samsar - the Text to visualize anything agent and generative video cloud.

Built With

• javascript
• mongodb
• nextjs
• node.js
• react