One of the cooler use-cases for the prompt-to-visualize anything framework is creating generative nature documentaries on extinct paleofauna and prehistoric animals.
We did generative documentaries on prehistoric subjects with previous generation of models which while pretty good, had minor inaccuracies across the entire stack such as out-of-place fauna, lip-sync on animals, incorrect anatomy of prehistoric animals etc. With each successive model family the renders become more accurate and the anatomy of subjects and contexts gets better (No mean feat considering no raw footage exists and no video materials to train from for the most part, even artistic or cinematic renditions are also hard to come by)
The inference models, the image models and the video models must independently understand the anatomy, physics the accurate period, timeline and context.
Samsar now supports Seedance 2.0 Video model by Bytedance. This is perhaps the first-ever, reliable production-ready integration of Seedance2.0 within the context of an agent harness. It is enabled in the production app and can be configured in the community edition via Fal, GMI Cloud or Samsar-js credentials.
For each of the subsequent demoes, we used Seedance 2.0 as the video model, GPT 5.6 Sol for inference, and GPT Image 2 or Seedream as the image model settings.
Text-to-video using this setting incurs 40 credits / second in Vidgenie, the Agent.
Let's now showcase creations of the agent, with selected model settings, We will be creating 1 minute duration portrait videos from text prompts. (User can create upto 3 minutes duration video in one-shot)
For the first demo, we travel back to the Late Cretaceous North America, approximately 78 million years ago when the Deinosuchus hunted on the banks of muddy rivers taking down small herbivores and large dinosaurs alike -
Seedream + Seedance 2.0. GPT 5.6 Sol Inference.
For the next showcase, let's venture deep into the Pisco sea where the apex aquatic hunters of the late Miocene era battle over carrion of a whale.
This demo has no voice-over, and this was by (prompt) design. If you want this style, you can simply specify in the prompt. See video description for the exact prompt to achieve this render style.
Seedream + Seedance 2.0. GPT 5.6 Sol Inference.
We travel to Paleocene Colombia where we witness the majestic Titanoboa hunting for lungfish in the immense rainforest wetlands approximately 59 million years ago.
GPT Image 2 + Seedance 2.0. GPT 5.6 Sol Inference
Fast-forward to the late Ice Age 30,000 years ago, a group of Mammoths is braving the ice-blizzard during their annual winter migration.
GPT Image 2 + Seedance 2.0, GPT 5.6 Sol Inference
In the final demo we journey deep into the Late Jurassic European and meet the Ophthalmosaurus. Hundreds of meters below sunlight, this hunter of the deep sea roamed the oceans surrounding Europe approximately 155 million years ago.
Seedream5 + Seedance 2.0. GPT 5.6 Sol Inference.
You'll notice that none of these videos have any character dialog scenes and do not switch between present day and past timelines. This is also by design, you can influence the direction of the video and the genre as well with keywords in your prompt, for example if you want educational style documentary which switches between speaker scenes in current day and the past scenes from the eons gone by you can specify in the prompt itself. Check out some other renders for pre-historic fauna which are more ed-tech style i.e. switching between present and past timelines.
Curious to try out the platform, but not sure which models to pick for your use-case?
Contact us at hello@samsar.one
Or join the Discord for discussions and to find out which model settings best suit your use-case.