{
  "version": 2,
  "input": "starting-render-v1.png",
  "inputSHA256": "0aa4000f5a545f09854f013f7052c3702414350b8b52b1fff5b62109289847ae",
  "prompt": "Edit the supplied 3D renderer frame into a natural, unretouched landscape photograph of this same single sandstone mesa. Preserve the 16:9 camera framing, horizon height, broad mesa silhouette and position, nearly flat top, and empty desert setting. Preserve sunlight from upper left/front, the cool right-facing cliff shadow, and the cast shadow extending right. Keep the sky clear and cloudless. Use authentic irregular fractured sandstone, eroded sedimentary beds, restrained ochre and tan minerals, broken ledges and local fallen rock. The key requirement is physically believable ground spatial frequency: broad, smooth, low-contrast variations in dry alluvial soil dominate the foreground and middle distance. Allow large quiet patches of fine compacted sediment interrupted by sparse, irregularly clustered small stones and a few shallow erosion channels. Do not blanket the ground with uniformly dense pebbles or uniform high-frequency grit. Stone density and size must vary naturally over space. Fine sand grains should mostly be below the pixel footprint, not individually resolved throughout the picture. Ground details must diminish in size, contrast, and sharpness with distance, blending into soft aggregate tones toward the horizon, as a real lens and sensor integrate subpixel detail. Use natural optical low-pass behavior without blurring the whole photograph. No repeating tiles, checker patterns, periodic ridges, regular stippling, cloned rock clusters, crunchy sharpening, halos, procedural noise overlay or equally sharp foreground and distant gravel. Talus should be localized around the mesa base, not spread evenly across the plain. Use realistic sunlight, readable blue-sky fill in shadows, subtle aerial haze, neutral photographic color and modest contrast. Keep the mesa in focus with believable landscape depth of field. No extra mesas, mountains, plants, people, roads, buildings, clouds, text, UI, watermarks, sunburst or lens flare. Output one convincing landscape photograph, not an illustration, comparison layout, CGI render or texture sample.",
  "comparisonScope": "One revised image per model from the identical renderer capture and wording, using provider-specific output settings. Actual dimensions are recorded for each result. The original OpenAI reference used an earlier prompt and is a historical baseline, not a controlled model comparison."
}