Universal Image Node

Self-initiated solo R&D: a modular, node-based image workflow system where natural language is the primary control layer, while the model resolves concrete node and channel parameters at runtime.

Problem

Typical node-based image pipelines trade power for configuration cost—many small knobs, many node types, and brittle graphs. The motivation here is to lean on model-level generalisation so users can describe intent in language and still reach broad, reusable image operations without rebuilding heavy manual configuration every time.

Approach

• —

  Node graph: Each node is a configurable image transformation or generation step.
• —

  LLM at runtime: Output channels and behaviour can be shaped by an LLM (via Gemini API) as the graph runs, not only by static wiring.

Data flow (core): inputs and node configuration → graph execution → runtime-defined output channels.

Demo showcase: Spatial Narrative exists only as a presentation case for demos and the portfolio video—it is not part of the core system.

Technical implementation

• —

  Stack: Next.js · React · Gemini API
• —

  Control model: Natural language carries high-level intent; the system resolves concrete node and channel parameters at runtime.
• —

  Design tension: Modular workflows are the goal; the hard part is multi-step stability—each step can inject stochastic variation, so the system needs guardrails against drift while staying flexible.

Role & status

End-to-end ownership: concept, system design, and implementation. Primary user today is the author; longer term, the direction is a tool for creative technologists and generative-image workflows.

Proof

Use the on-video controls on the demo reel at the top of this page for pause and scrub (same asset as the home / work preview).

What’s next

Extend this toward a modular content engine: an intelligent configuration layer behind the graph so nodes can adapt instead of staying on fixed wiring, and so interaction stays at the level of intent—what you want the system to do—rather than only low-level node setup. Multi-step drift and stability remain the main engineering pressure as graphs get longer.