From research to experience
A natural interactive system needs to understand what the user said and why they said it, decide how to respond and turn that response into coherent sound and images. An application also has to account for waiting, interruptions, feedback, devices and available resources.
The projects at OneZero Labs are organized around these stages. They have different roles and are at different stages of maturity, and together they form a chain of work that runs from exploring capabilities to validating experiences, and on to evaluation and improvement.
| Project | Primary role | Questions it addresses |
|---|---|---|
| OmniSoul | Social intelligence model | Context, personality, emotion and multi-turn interaction |
| OmniFlow | Real-time interactive character world model | Continuous input, character performance and long-term consistency |
| OmniX-1 | Video generation research | Visual expression, cross-modal control and iterative optimization |
| SociaBench | Social intelligence evaluation | Scoring methods, performance by dimension and evidence for reproduction |
| Amicura | AI companion and real-time interaction app | Conversation, voice, visuals and user experience |
OmniSoul: social intelligence
OmniSoul focuses on social intelligence. It studies how a system understands the people, relationships and intentions in a given context, how it keeps personality and expression consistent across multi-turn conversations, and how it responds appropriately to emotion.
We pay attention to subtle differences in natural language, as well as to behavioral stability over long interactions. Evaluating these capabilities requires real contexts, diverse scenarios and failure cases, not just a set of pre-scripted opening lines.
OmniFlow: real-time characters
OmniFlow covers two lines of capability: real-time video interaction and high-quality character generation. Real-time interaction focuses on continuous input, low-latency response and motion control during generation; character generation focuses on voice-driven performance, lip sync, identity preservation and long, continuous footage.
This direction involves latency, temporal consistency, instruction control and system scheduling at the same time. We want this research to serve real interaction, keeping the picture continuous when the user changes what they want the character to do, rather than having to start over every time.
OmniX-1: video generation
OmniX-1 focuses on video generation and iterative optimization. It studies how to express subject identity, motion and scene structure, and how control works across text, images, video and sound.
Recursive self-improvement is a research topic that must be validated with care. Any process involving automated evaluation, data regeneration or model self-optimization needs to watch for error amplification, evaluation bias and the need for external review. Specific results should rest on formal technical materials.
SociaBench: model evaluation
SociaBench organizes evaluation of AI social intelligence around dimensions such as role-play, personality consistency, emotional understanding and social reasoning. It gives researchers and application teams a way into understanding how models differ.
Interpreting an evaluation depends on model versions, sample coverage, prompts and judge settings. When reading a leaderboard, look at the methods and reproduction notes as well, and avoid treating an advantage on one leaderboard as an advantage on every task.
Amicura: real-world interaction
Amicura explores product experiences around AI companions and real-time interaction, bringing together text, voice and visual expression. Continuous conversation, user interruptions, different languages and proactive communication give research more concrete experience problems to work on.
Findings from application projects can help adjust models, data and system design, but a single experience is no substitute for systematic evaluation. Access conditions, available features and related terms of use for the product are set out in Amicura’s own pages and documents.
Availability
A research direction describes a problem we want to solve; a research preview shows a capability we are exploring; a beta version is used to gather real-world feedback; and a production service requires a clearly defined scope of features and terms of use. Each status has its own meaning, and they should not be used interchangeably.
The brand visuals and concept imagery on this website are there to help explain our direction. Whether you can call an API, download a model, use a particular feature or deploy commercially depends on the corresponding release page, license or an agreement between both parties.
Understanding, generation and evaluation, working together toward more natural intelligent experiences.
About us