Neuro Symbolic Sketch Synthesis
Generative diagrams that reason before they draw
Why this name
“Neurosymbolic” is the precise term for what makes Lumina different: a neural image generator (the neuro half) is steered by explicit subject knowledge, geometry rules, physics laws, and syllabus definitions (the symbolic half). The sketch isn't just plausible; it's provably correct.
The problem
Some ideas are almost impossible to grasp from words alone, such as a force diagram, a geometry proof, or how a transistor switches. Students understand these fastest when a tutor draws them out, step by step. Most learners never get that tutor, and pure image models draw pictures that look right but are mathematically wrong.
What we're building
Ask “show me why the angles of a triangle add to 180°”, or snap a photo of a problem, and Lumina composes a labelled, exam-style diagram whose every element is checked against the underlying maths, then replays it stroke by stroke so the learner watches the reasoning build.
Key capabilities
Text-to-sketch: plain-language prompts become clear, labelled diagrams.
Symbolically verified: angles, proportions, and labels are validated against subject rules, so α + β + γ really is 180°.
Progressive rendering: the sketch animates into existence so students follow the logic, not just the result.
Whiteboard aesthetic: clean exam-diagram style, not glossy AI art.
Editable: learners and teachers can nudge, relabel, and extend any diagram.
Research focus
Conditioning diffusion / transformer image models on symbolic constraints: p(sketch | concept, syllabus).
Grounding generated visuals in curriculum knowledge so diagrams are correct, not merely plausible.
Measuring the learning gain of watching a diagram build versus seeing it finished.
Where it fits at ILM AI
The visual layer for Ilmino, MathPilot, and the Physics Lab, extending “Snap & Solve” from answers into pictures.
Lumina: every concept, drawn to be understood.