Samsar the Text to Visualize anything framework now has first-class support for BackBlaze B2 cloud and GenBlaze generative adapter via GMI Cloud.
As we continue with the improvement of prompt-to-visualize anything framework, we yesterday added model and adapter config support for BackBlaze B2 and GenBlaze - Premium generative cloud repository with a plethora of generative models and scalable S3 compatible file storage. This update adds even more adapters to the model and adapter composable Text-to-video agent as we now support Backblaze B2 - Your trusted generative Cloud.
This was a day and half long integration which required understanding the nuances of B2 cloud and GenBlaze, the model offerings , the adapters and the naming scheme / payload structure of the GMI Cloud repository and completed in response to Devpost BackBlaze hackathon.
Users can now setup GenBlaze as the generative adapter by adding a GMICloud configuration via the Setup wizard from the community edition.
BackBlaze B2 cloud, an S3 compatible storage adapter is now available in addition to the existing Minio and AWS adapters as well. This is an affordable and reliable data storage for users who want high volume generative cloud storage w/o breaking the bank.
What is GenBlaze-
Genblaze is a universal adapter for generative models including LLMs, image and video models. When enabled it minimally provides almost the entire suite of models required for text-to-video and image-list-to-video run except for lip-sync and sound-effect models.
You can enable the full-suite of functionality for example by setting up GenBlaze + Samsar-js or GenBlaze + Fal config during setup.
GenBlaze adapter can also be added along with native adapters or your custom adapters and will act as a fallback for when other adapters are missing or upon failures in primary adapter calls, thus adding redundancy and resilience to the agent pipeline.
GenBlaze is adapter composable and can be configured with native provider keys as well, however to keep things simple you can use GMI Cloud to enable all providers via GenBlaze and if you'd like to, you can directly add supported native providers in the setup wizard.
BackBlaze B2 Cloud-
B2 is an affordable remote storage alternative for individuals who would like to run the Docker container and don't want to lose data if the container or volume is deleted or needs restart.
To create a B2 cloud account simply register in the dashboard, then purchase a minimal amount of credits and create a bucket, note down the bucket endpoint, the bucket name and the master or application key and keyId. The setup wizard can configure B2 cloud, with both Master and application key (recommended)
Current implementation requires the B2 cloud to be set as public access for sending requests to external API providers, future iterations will explore the possibility of configuring private secure B2 bucket for user installation while creating public urls only when required for communicating with external adapters.
Once setup, all generations and final renders will be persisted to the Backblaze B2 cloud while intermediate frames still remain in the local volume and are cleanup up with the task cleanup process every day.
Integration challenges-
Whenever we are adding a new adapter or model, once we've defined the specs thoroughly, it just seems like In and out, 20 minute adventure.. However the actual integration and testing begins, that's when the real challenges surface. Here are some of the challenges I faced during this integration and how I overcame them.
GenBlaze model keys are different
The payload structure is a bit different from the schema we were using, also the model names and keys are not the same as what we were using (For example NonoBanana Pro is referred to as gemini-3.1-pro-image and other models have custom keys as well)
Mapping all the models available (Almost all models are available in GMICloud) to the model definitions in our system was a bit of a challenge and took multiple iterations to get this done.
Access control and CDN for B2 buckets
It wasn't obvious how to setup publicly accessible links for private B2 buckets, so we left the B2 buckets are configured as public accessible for now. Also B2 currently does not natively support CDN, so current implementation of B2 resources will return S3 compatible urls instead of CDN urls. We will explore the possibility of configurable external CDN in the future.
Limited Concurrency
The pipeline generally runs for 5 concurrent image to video requests at a time, however we are seeing some requests getting silently dropped when requesting 5 items at a time with GenBlaze with GMI Cloud integration. (This could also be because we just created our account yesterday so potentially in a lower concurrency tier) To handle this we reduced the concurrency to 2 for GMI cloud while still maintaining the concurrency of 5 for other providers and native providers
Another failed attempt at integrating Seedance 2.0
We have been trying to integrate Seedance 2.0 for few months now, the quality looks great from promos. However it does not yet seem suitable for agentic usage.
Previous attempts failed because the classifier would too often label ai generated images as real pictures and refuse to animate.
The current attempt with GMI Cloud worked better, however now seeing a new issue where some requests just get stalled for a long time intermittently with no errors until a timeout after 60 minutes and refusal rate still remains ridiculously high.
After several attempts to get Seedance 2.0 working reliably, we gave up. The failure rates are too high for agentic usage. We still have video models like Happy-Horse 1.1 , VEO 3.1 and Cosmos 3 which provide equally good animations and work with ~100% reliability in our pipeline but Seedance 2.0 integration will have to wait for another day.
Now, more fallbacks and redundancy, more choices and better pricing
GenBlaze is a super adapter which supports LLM, Image and Vision models so can enable all the models required for the Samsar T2V pipeline with only minimal external adapters needed for auxiliary tasks such as lip-sync and sound-effects.
This adapter can also add as a fallback with native or other external adapters. You can choose the order of preference of adapters for models in the case where more than one adapter is enabled for the same model for the user installation, The system also automatically fallbacks to lower priority adapters when higher priority adapter requests fail or time-out.
This latest addition, makes Samsar - The 1-shot T2V agent, supporting full-stack of generative models got even more robust, decentralized and gives more options to consumers as to how they would like to configure their installation.
The project is fully open-source under MIT license and free to use for local installations. You only pay for the generative models / adapters you configure, and you pay directly to the provider you configure.
Checkout the code here (And leave a Star if you find it useful 😸 )
samsarone