Vantage phase 1: the hero planet, and the god rays that weren't

What shipped

The planet. Three concentric spheres: a PBR surface (MeshStandardNodeMaterial with colorNode, emissiveNode, normalNode and roughnessNode), an unlit scrolling cloud shell, and an additive BackSide fresnel atmosphere, all over a milky-way starfield and lit by a single sun DirectionalLight whose direction is the one source of truth every shader reads.

The post chain is scene pass, then bloom, then radialBlur god rays sourced from a black-planet/bright-sun occlusion scene, composited in linear HDR before ACES. All 40-odd knobs live in src/data/planet.ts. Textures fetch by script from Solar System Scope under CC-BY, credited in CREDITS.md, and never touch git.

This was the phase most likely to fail, so it went first.

Decisions

The research round earned its keep twice.

The small catch: bloom lives in three/addons, not three/tsl, which is the kind of thing that costs an hour of confused importing.

The big one: r185’s built-in godrays() node is volumetric shadow-map raymarching. That is not the effect the design wants. What I want is the classic screen-space disc-occluded shafts, and the built-in is simultaneously the wrong look and the most expensive option on the menu. Replacing it with a radialBlur over an occlusion pass gets the right look for a fraction of the cost. It’s kept in the notes as a documented upgrade ceiling in case the project ever wants true volumetrics, which as it turns out it eventually did.

The radialBlur composite was the one piece research couldn’t pre-verify, so the wiring that actually worked is written down in src/render/post.ts rather than left to be rediscovered.

Tone mapping stayed ACES after a same-frame AgX comparison. AgX is what the research recommended for emissive highlights, and on this scene it read flatter and slightly milky, while ACES gave deeper ocean blues and an inkier night side. Decided on pixels, not on doctrine. Both curves stay one tunable away.

What broke

Solar System Scope ships its normal and specular maps only as Adobe-Deflate TIFF, which no browser will load. Rather than take a dependency for one file format, the fetch script grew a roughly 90-line TIFF decoder (zlib inflate plus horizontal predictor) and a minimal PNG encoder, verified pixel-correct against known geography.

The FPS verification detoured through a hall of mirrors. Headless Chrome pinned rAF at exactly 57 fps regardless of resolution, which looks exactly like a failed performance gate. Dropping the render resolution changed nothing, which was the tell. It was a frame-pacing cap the whole time, and --disable-frame-rate-limit exposed the real headroom underneath.

That one’s now a standing lesson: a suspiciously uniform frame-time distribution is a pacer, not a workload. If the number doesn’t move when you make the work easier, you’re not measuring the work.

Numbers

Median 119 fps steady-state at 2560x1440 with every layer on, on an NVIDIA Ampere with WebGPUBackend confirmed rather than the WebGL2 fallback. The reviewer cross-checked around 890 fps on uncapped submission and confirmed resolution scaling behaves, which is what turns the 119 from a number into evidence.

The first five seconds after load run at 10 to 40 fps while pipelines compile and 8k textures upload. One-time cost, worth knowing about.

Textures are 22 MB on disk: day, night, clouds and stars at 8k, normal and specular at 2k, since SSS offers no 4k tier. Build is 884 KB raw, 242 KB gzipped. Screenshot checklist 7 of 7.

Next

Phase 1b is the texture authoring pipeline: procedural bake to equirect maps, a PlanetDefinition bundling texture set, region-ID map, terrain map and metadata, plus channel-packing and KTX2 for the hero pass. Then the fictional planet replaces Earth, and “looks as good as Earth did” becomes the bar to clear.