STARTING POINT → VISUAL TARGET

One mesa. The same light.

Left: the actual browser renderer. Right: an OpenAI image edit of that frame into a photographic target. The edit preserves the broad composition and sunlight direction; small silhouette and framing differences remain. It establishes visual direction, rather than a pixel-matched ground truth.

Actual WebGPU starting render of a single mesa — **Actual renderer / starting point**1920 × 1080 · explicit GPU readback · procedural mesh and scanned materials. Capture evidence

OpenAI-edited photographic reference of the starting mesa — **OpenAI image edit / photographic target**Generated from the starting render using the built-in image generation tool. This is a reference image, not renderer output. Edit prompt and provenance

01 / Cliff structure

Replace smooth, regular terraces with irregular resistant beds, large fractures, broken edges and local overhangs. Judge the silhouette and the shaded face first.

02 / Ground and talus

Break the smooth apron into angular fallen stone and graded debris. Add believable ground scale without repeating texture or evenly distributed noise.

03 / Sun and sky fill

Keep one clear lighting setup. Match warm direct light, cool readable shadows, contact shading, restrained haze and photographic exposure after the forms read correctly.

Previous scene before narrowing the scope

The earlier cloud scene is retained for context. The current working scope is the single mesa and sunlight.

Open previous scene

The mesa renderer is a first study, with photographic quality still open. A saved image does not establish sustained 60 FPS. Browser and bounded performance checks.