Fox sculpture
One mesh. Three layers.
Neutral shape / base color / base + normal
Inspect every view.
Choose a photo. Slide across its matched render.
Turn the model ↗Camera 01
MATERIAL RENDERSOURCE PHOTOThe display composite uses this photograph's own mask. It does not enter either metric calculation.
Where the pixels disagree
Red: candidate increased error. Blue: candidate decreased error. Black: outside the owner-mask interior.
Read the metric contract and its limits
The photo audit control uses the neutralized-color atlas and mesh normals under the fitted light. The candidate uses its revised color atlas and normal map. This control is separate from the untextured clay state in 3D compare. The local report compares full-resolution 531 × 800 saved camera renders inside the owner-and-mesh region. The Nebius audit resampled the photo/render pairs to 360 × 542 JPEG and eroded each source owner's mask by three pixels, without the exact mesh-hit intersection. Those contracts differ; a small change of sign is a diagnostic sensitivity, not a material acceptance result. The original source masks remain bound to their own photos.
Texture studies
Source zones, repeat methods, and material previews from two licensed photographs.
Peeling paint
Paul Stevenson · CC BY 2.0
Weathered wood
sagesolar · CC BY 2.0
Illustrative display lighting and assumed roughness. Paint uses the inferred normal prior; the wood height prior is shown separately and is not applied as measured displacement. Physical PBR remains unqualified.
Reports
Qualification, measured Nebius usage, and the source record.
12-view CUDA pixel audit
aijob-e00he632g06bw1kxtvSource photographs, their own masks, and both saved camera renders were packaged by view. The outputs were downloaded and SHA-256 checked.
Read the cloud evidence report →Nemotron 3 Super
One bounded diagnostic review of deterministic facts. It selected improve material validation and left asset approval with the saved gates.
779 input · 911 output
Cost estimates
Billed amounts have not been verified. These calculations use rates recorded on 22 October 2026; the H100 instance rate may differ from Serverless Job billing.