I am happy to announce the world's first integration of Seedance 2.5 within the context of an agentic framework has been achieved, and is now live in production as well as available on Github- Fully Open-source! Simply select it as the video model setting in Vidgenie or in the API and the ai-video stage will use this model for the render on any genre of topic upto 3 minutes duration.
Seedance 2.5 is configurable in Samsar monorepo via both GMI Cloud and Fal adapters. You can also use samsar-js API key as universal fallback to enable this model. In production simply choose this as your model setting and you're good to go.
Adding integrations to Samsar - the open, modular, agent framework is easy as breeze so for this integration we decided to give ourselves a bit of a challenge, we used GPT 5.6 Luna (Sol can few-shot any adapter or model integration when used within our framework)
GPT 5.6 Luna did a pretty good job with the complex integration spanning across deployed production and standalone environments with only a few missed steps (Mostly because the prompt wasn't detailed enough and Luna doesn't hold enough of the context of the code-base in memory like Sol does)
In the end of an hour long live integration and debugging session, as with every other model integration, Seedance 2.5 is now successfully integrated in production agentic workflow.
Seedance 2.5 incurs 60 credits / second in Agent (cumulative) standalone usage in Studio or via external API calls incurs 50 credits / second.
Hands down the best video-model inference we've seen so far, available in both Samsar T2V agent and open-source distribution. It is a premium model setting and can be used for 60 credits / second for prompt to video (Same price as VEO3.1)
The classifiers for Seedance 2.5 are quite strong and will often fail at close resemblance with human or even fictional characters, for best results with this setting create anime nature or ed-tech media with this model. If you were to use prompts which would result in cinematic renders consisting of human characters, use Seedream 5.0 image model settings and add keywords like anime (to prevent real human likeness) to your prompt for the most predictable results.
Seedance 2.5 / 2.0 when used with GPT 5.6 Sol inference and image models like Seedream and GPT Image 2 in the Samsar Framework, renders premium quality text-to-video in a single attempt with no post-processing required.
Now I'll be showcasing some demo videos, 1-shot with Samsar T2V agent. Prompt to render anything you like to create your own video feed or channels based on your chosen niche.
For the first demo, we travel back to witness an epic battle between the legendary hunters of late jurassic era - The allosaurus and the Torvosaurus. See the exact prompt for each render in the video description.
For the next showcase, we travel to a hidden sanctuary nestled deep within the Himalayan mountains. We asked the assistant for Dan Brown style thriller prompt and then used it to create text to video render via the Samsar engine.
GPT Image 2+Seedance 2.5. GPT 5.6 Sol inference
For the third demo, a live-action from an episode from the novel Malazan Book of the Fallen. You can create live-action from fantasy books or novels or even your own ideas. (Add keywords like anime to create more anime like renditions and prevent trigerring the classifiers)
For the final demo, we will be showcasing a live-action based on a popular book / tv-series. We will be rendering an episode from Mentats of Dune, notice that we used anime render instead of live-action render to avoid triggering the video model classifiers.
Seedream + Seedance 2.5
As you can see, with Seedance 2.5 image to video, SoTA video model you can now create prompt to video on any topic or theme with Samsar - The Visualize anything framework. Use keywords like anime when creating renders that might include resemblance to real humans or go ham with nature / animal / ed-tech creations.
Samsar is a modular general agent that will create a video on any genre, any topic or any use-case.
Some topics as are more popular and gain tons of views when posted on public feeds, others are more niche and could be of personal interest to you only. You can create videos on any topics upto 3 minutes in duration using any model setting of choice.
Try out all the models and all the configurations and the settings..The app is here.
I have public discord for answering general enquiries, enterprise clients get their own private discord and personalized support.
Check out the full-featured source-code (Same source as runs in production) here.