Same scene.
One more pass.

A lighter scene. Consistent walking speed.

The latest preview now adjusts its resolution automatically, reuses water reflections between updates and refreshes smaller shadow maps less often. Use the quality menu inside the scene to choose Auto, Smooth or High detail. The FPS counter reports that browser’s current frame rate.

The character previously slowed down whenever rendering dropped below 20 FPS. Movement now covers the same distance per second at simulated 10, 20, 30 and 60 FPS. Long stalls above 250 ms are capped to prevent a sudden jump; hidden tabs stop movement.

Same 1280 × 900 viewport and Chromium software renderer, tested sequentially. Auto reduces internal resolution under load, so this is a practical performance/quality tradeoff. Software rendering remains slow. These results do not establish FPS on your device. A server hardware-rendering attempt lost its graphics context and was excluded.

The courtyard’s assets and lighting are retained. This repair improves rendering cost and responsiveness; it does not close the remaining artistic gap with the original demo.

Frame-time measurements · Movement timing checks · Published controls checks · Mobile touch checks · Repair ledger · Source and timing record

What does one more pass buy?

A darker courtyard, warm firelight, wet reflections and a traveler with moving legs, arms and cape. The earlier scene remains playable alongside it.

Before · D, pass 3

Generated assets, static traveler

The completed asset-assisted experiment.

After · Additional polish

Courtyard & articulated traveler

Press Play, then use arrows or WASD / ZQSD. Click / tap to walk, X to stop, drag to orbit, scroll / pinch to zoom.

This is a separate polish exercise, not another controlled lane. We changed the visual target to the darker courtyard in the author’s demo and used one fresh Astra low-effort builder with Astra direction and verification. The generated arch is reused; the static traveler is replaced with an articulated model authored in code. No new asset-generation call was made. Historical A/B/C/D results are unchanged.

Time for this extra step

On the first local navigation, the first WebGL draw was submitted at 4.68s and the first draw with the loaded arch at 9.40s. The complete screenshot was saved at 25.02s. Draw submission is not a measured browser paint or GPU completion; the captured scene is the evidence of visible rendering. This browser used software rendering; hardware FPS was not measured.

Tokens & API equivalent

Input includes cached tokens; output includes reasoning. Standard Astra API rates are an equivalent, not a subscription charge. Existing assets and hardware costs are excluded. Cutoff: 2026-09-11T09:11:10.255436+00:00; later responses are excluded.

Subscription meter: 44% → 45%. This rounded seven-day reading covers the whole account, including concurrent activity. Exact consumption by this pass or model is unavailable.

Token & quota ledger · Timing evidence · Movement & pose checks · Preview controls · Mobile checks · Walking capture · Scene source

Press Play on a preview to control the scene, or open it full screen. Move with arrow keys, WASD / ZQSD, or click / tap. Drag to orbit, scroll / pinch to zoom, and X to stop. Fresh-load control checks ↗

Before · Lane C

Authored geometry

Full Luna direction, three fresh Astra passes, authored in-engine geometry.

Explore C ↗

Follow-up · Lane D

Textured 3D assets

The same fixed target and seed, with a generated archway and traveler supplied before the three passes.

Explore D ↗

Both captures use the same probe viewport and initial/reset view. Each run controls its camera composition, so this is a workflow comparison, not a pixel-aligned quality test. Runs took place sequentially.

Three passes, under Luna’s direction

Luna xhigh directs one fresh Astra low worker per pass. Each worker receives the reference and current scene; Luna verifies the browser between passes. No separate judge agents.

Pass 1

Build the scene

13m 40s from kit ready to capture

Explore pass 1 ↗

Pass 2

Refine the scene

28m 23s from kit ready to capture

Explore pass 2 ↗

Pass 3

Final visual pass

35m 11s from kit ready to capture

Explore pass 3 ↗

Original experiment: time, tokens & cost

Asset generation: 7m 38s. Kit inspected and ready: 13m 21s from task start. Luna’s complete three-pass handoff: 54m 15s from task start.

First paint or a partial first element was not instrumented. The first scene is evidenced by 27m 01s from task start; the final pass by 48m 32s. These are captured-render upper bounds.

Times are cumulative to the first completed capture in each pass, an upper bound on first visible rendering. The earliest capture can precede control fixes; the screenshots above show the validated pass. Tokens and model costs are incremental per pass. K = thousand tokens; input includes cached input, and output includes reasoning.

C total: $3.181 · D total: $3.173 API equivalent. These include each director’s recorded setup and finishing. C had no equivalent generated-asset preparation stage. The recorded lane costs are effectively unchanged.

Exact time, request, token and quota ledger · Preserved first-pass timing evidence · Original C pass figures

How close is this to the original demo?

The author’s initial prompt asks for an isometric fantasy scene in Three.js: realistic shading, reflective wet floors, a character, click-to-move, lazy camera follow, drag to orbit, scroll to zoom, and an environment with subtle motion. The requested frame rate exceeds 60 fps; no gameplay is required. The post separately reports $0.90 for assets; our $0 external 3D API fee uses existing GPU capacity whose cost is unallocated.

We now follow the asset-assisted workflow as well as the direct Luna-to-fresh-Astra structure. This uses Forge’s image-to-3D system and two inherited source images; it does not reproduce the author’s exact assets, 3D model, rendering setup or undocumented iteration history. Root handled asset preparation and publication separately.

The author’s follow-up correction puts savings around 40–70% across his threads and acknowledges differing techniques. This single follow-up cannot establish those savings or the earlier 80–90% subscription claim. There is no new Astra-only run with this same asset kit.

Validation and limits: All three desktop probes pass loading, movement, lazy camera follow, orbit, zoom and advancing animation. The final public checks are linked below. The assets are static, character movement is not skeletal walking, and the floor, vegetation and distant architecture remain visibly procedural. Lighting, reflections and composition still fall short of the reference. Software rendering does not validate the author’s hardware frame-rate target.

Independent public browser verification · Final pass interaction report · Final source archive