Artifact

I was making artwork for the Vantaa Underground track «Вайбер» and used Gemini Nano Banana to explore abstract background material. The image was a useful start, but I wanted to vary it deliberately, reproduce a promising direction, and export an equirectangular environment map for a Blender scene. That gap between a surprising image and an editable process became the first Artifact experiment.

Effects accumulated, the experiment became reusable, and the scope moved from one cover to a procedural editor for covers and posters. Vantaa Underground incubated the product by providing a real need and its first production work; Artifact now stands on its own. It is aimed at independent artists and designers who want to explore quickly without giving up an editable, reproducible result. The works show that progression while crediting the external inputs behind each one.

Artifact opens into Layers because that is the shortest path from a blank document to a useful variation. An author can combine images, text, fills, seeded emoji fields, procedural noise, geometric arrays, 3D primitives, and more than thirty GPU effects. The RAND action can produce a starting point, but the document seed and per-layer offsets keep chance inspectable rather than disposable.

This is the core creative iteration loop: make a starting point, change structure or effects, compare variations, and keep the one worth developing. Four aspect ratios and ×1/×2/×3 PNG or JPEG output let the same system move between square covers, portrait posters, stories, and landscape artwork. Portable .artifact packages preserve the editable document with its imported assets. Opening another Artifact document is staged before replacement, and the current work is kept as a recovery copy until the author confirms the change.

I wanted a node-based creative workflow without depending on a product centered on AI pipelines. Artifact’s Nodes view stays focused on artwork: sources, merges, color correction, repetition, and a visible export path. It is an optional power-user surface, not the price of entering the editor.

Both modes edit one serializable CanvasDocument. Stack mode infers a linear graph; Nodes exposes the directed graph. Both resolve through renderDocument. Pixel-like parameters share a 540-pixel reference, while 3D preview and export share one scene and committed camera state. That keeps composition consistent across previews, saved work, and exports even when resolution and antialiasing differ.

The first prototype was promising until Nodes began triggering React’s “Maximum update depth exceeded” error. I nearly stopped there. Tracing repeated React Flow measurements led to a focused fix: discard no-op dimension updates and document clearer state ownership. This was a correctness bug, not a reason to add workers.

Later sluggish interaction had a different cause. Manual profiling recorded heavy graph edits at roughly 8–12 FPS. Moving deterministic procedural-noise generation into a Web Worker raised the same kind of interaction to roughly 57 FPS or higher. CPU image transforms followed; React Flow, PixiJS, live Three.js viewports, and compositing stayed on the main thread.

The performance benchmark separates frame timing, long tasks, thumbnail work, document renders, and GPU phases so the next optimization can begin with a measured bottleneck.

The first versions could live entirely in the browser. As image generation, cloud projects, and repeated use arrived, I kept that local-first authoring loop and added remote boundaries only where they earned their cost. Signing in adds private cloud projects; it is not required to begin making or to keep an editable document.

Provider-backed jobs now cross an API and worker boundary instead of placing secrets or long operations in the editor. A separate backoffice exposes account tiers, usage, cost, and recovery controls without exposing prompts, artwork, or project documents. The release path then verifies the frontend and services as one revision before production promotion.

The v0.41 release record documents the current account, backoffice, and exact-SHA production boundary. Billing, public sharing, collaboration, and project history remain planned rather than shipped claims.

I own the product end to end: direction, interface, rendering, local and cloud projects, testing, releases, and account-backed AI operations. The remaining question is not whether Artifact can ship as a product, but which of its more ambitious workflows are ready to lead the experience.