About OmniVeo
OmniVeo is an independent community lab documenting the evolution of logic-native video generation systems. We focus on practical prompt design, instruction-based editing workflows, and frame-accurate evaluation methods.
What we publish
We publish technical notes, reusable prompt templates, and case-study videos to make model behavior observable and reproducible. Our publications include detailed frame-by-frame analyses, prompt engineering guides, and comparative studies across different model versions. Each publication is designed to be immediately actionable for researchers and practitioners working with logic-native video generation systems.
How we evaluate
Our reviews prioritize temporal coherence, spatial anchoring, causal consistency, and instruction adherence at frame-level granularity. We employ a multi-layered evaluation approach combining automated metrics (optical flow analysis, text recognition accuracy, motion consistency scores) with expert human review. This dual approach ensures we capture both quantitative performance indicators and qualitative reasoning failures that automated systems might miss.
Logic-Native Video Generation
Core Principle
Logic-native generation means the model does not merely interpolate pixels — it maintains an internal symbolic representation of the scene. Mathematical expressions, spatial relationships, and causal chains are treated as first-class citizens rather than emergent visual patterns. This enables true instruction following: when you say "divide both sides by 2", the model understands the algebraic operation and renders the corresponding visual transformation.
Evaluation Axes
- Temporal Coherence — Symbol identity and stroke morphology remain stable across frames. No spontaneous glyph mutation.
- Spatial Anchoring — Objects and text maintain absolute world coordinates even when temporarily occluded by hands or camera motion.
- Causal Fidelity — Physical and logical consequences of actions are correctly rendered (e.g., erasing a term actually removes it from the equation).
- Instruction Adherence — Natural language directives are executed with semantic precision rather than surface-level visual similarity.
Research Methodology
Every case study is decomposed into micro-segments (typically 2–4 seconds). Within each segment we isolate a single capability: symbol continuity under occlusion, multi-step algebraic manipulation, or agentic layout planning. We publish both the raw video and the synchronized agent trace so researchers can correlate model behavior with internal reasoning steps.
Why Frame-Level Analysis Matters
Most video benchmarks evaluate at the clip level. OmniVeo insists on frame-accurate inspection because logic errors often appear only in a single frame (e.g., a coefficient that momentarily becomes unreadable). By surfacing these micro-failures, we accelerate the development of truly reliable reasoning video models.
Key Research Areas
Prompt Engineering for Reasoning
We systematically explore how different prompt structures affect the model's ability to maintain logical consistency. This includes chain-of-thought prompting, explicit step decomposition, and meta-instructions that guide the model to plan before rendering.
Instruction-Based Editing
Beyond generation, we study how models respond to iterative editing commands. Changing a variable mid-derivation, inserting new constraints, or requesting alternative solution paths reveals the model's true understanding of the underlying logic.
Agentic Spatial Planning
Advanced models demonstrate the ability to pre-allocate screen real estate for future derivation steps. We analyze how the model decides where to place new equations, how it manages visual hierarchy, and how it maintains consistency when the derivation grows beyond a single frame.
Cross-Modal Symbol Grounding
We investigate how models map between textual mathematical notation and visual handwritten forms. This includes studying stroke order, character topology, and the relationship between symbolic meaning and visual appearance.
Technical Architecture
OmniVeo Lab operates as a transparent research platform. All published case studies include the original generation parameters, the exact prompt used, and frame-by-frame annotations. This level of transparency allows the community to reproduce results and build upon our findings.
Our evaluation pipeline combines automated metrics (optical flow consistency, text recognition accuracy) with human expert review. We believe that true reasoning capability cannot be reduced to a single scalar score — it requires qualitative analysis of failure modes and edge cases.
The lab maintains a growing library of prompt templates that have been validated across multiple model versions. These templates serve as both practical tools for researchers and as diagnostic instruments for understanding model capabilities.
We publish raw generation parameters alongside each case study, enabling researchers to reproduce our results and extend our methodology to new models and scenarios.
Join the Research
OmniVeo is an open research initiative. We welcome contributions from researchers, engineers, and practitioners interested in advancing the state of logic-native video generation. Whether you have a novel prompting technique, a new evaluation metric, or an interesting failure case to share, we encourage you to participate in building a more rigorous understanding of these emerging systems.
All published materials are released under open licenses to maximize their utility to the broader research community. We believe that transparent, reproducible research is essential for the responsible development of increasingly capable video generation models.