UNITH's Video Synthesis Pipelines

unith operates two in house video synthesis pipelines thai and deva 1 both power the same digital human product and are fully compatible with all conversation operating modes, but they differ significantly in how they generate video this page explains what each pipeline is, when to use one over the other, and how to get access thai thai is unith's primary video synthesis pipeline and the one that powers the majority of head visuals available on the platform today it has been in development and production use for several years and is considered stable, broadly supported, and well suited to a wide range of deployment scenarios thai works by compositing speech driven lip sync over a pre recorded idle video loop the digital human plays a looping idle animation when not speaking, and transitions to a response loop when generating a reply this approach is highly reliable and computationally efficient, making it a great fit for most use cases including virtual assistants, customer facing agents, and embedded web experiences when to choose thai you need broad availability of head visuals and voices your use case does not require highly expressive or naturalistic facial movement you want a well supported, production hardened pipeline deva 1 deva 1 is the result of unith's ongoing r\&d investment in next generation video synthesis rather than compositing over a pre recorded loop, deva 1 synthesizes every frame from scratch in real time using a 3d gaussian splatting renderer this means the digital human's appearance, pose, and expression are fully computed per frame, driven by the audio signal the result is a more lifelike, less rigid presentation in the future, deva 1 is designed to address some of the expressive limitations that are inherent to loop based synthesis micro expressions, subtle head movement, natural gaze dynamics, and prosody driven facial changes are all possible with deva 1 in ways that thai does not support by design deva 1 is under active development head visual availability will grow over time, and quality improvements will be shipped continuously when to consider deva 1 you are working on a longer horizon project and want to build on unith's forward looking pipeline visual naturalness and expressivity are a priority for your use case you are building an experience where the rigidity of loop based video would be noticeable deva 1 is currently available to a limited set of clients as part of an early access program it is not enabled by default see access and availability below comparison thai deva 1 generation approach loop based idle + lip sync compositing per frame synthesis via gaussian splatting expressivity idle and response loops micro expressions, natural head motion, prosody driven movement head visual availability broad — majority of unith visuals limited — expanding over time maturity production stable active development / early access conversation modes oc, doc qa, plugin, ttt oc, doc qa, plugin expected latency low comparable to thai availability all clients selected clients (early access program) conversation mode compatibility both pipelines work with all standard unith conversation operating modes open dialogue (oc), document / knowledge base (doc qa), and plugin (plugin) thai additionally supports text to video (ttt) generation the pipeline choice does not affect how you configure conversation behavior, prompts, or integrations — it is a property of the head visual, not the operating mode access and availability thai is available to all unith clients and requires no special enrollment deva 1 is currently enabled only for selected clients participating in unith's early access program if you are interested in enrolling or if deva 1 better fits your use case, reach out to us directly interested in deva 1 access? get in touch with the unith team and we will assess whether your use case is a good fit for the current program