Reads
Cloud depth-estimation model
Cloud Refinement · +1 Credit · True 16-Bit Depth
Depth that reads like anatomy, not brightness.
AI Depth is the cloud depth-estimation model that automatically refines a Relief Generator generation, instead of computing height from grayscale tone. The result is closer to a 3D scan of the subject — it understands that a nose comes forward and an eye socket recedes, regardless of how the generated image happens to be lit. It isn't a manual engine you point at an uploaded photo; it runs as an optional refinement pass on whatever the Relief Generator just sculpted from your prompt. The model returns genuine 16-bit depth data — 65,536 gray levels, not the 256-level estimate cloud depth models typically hand back.
Two modes, one cloud call
AI Depth uses the cloud model's output directly. AI Depth + Detail takes that same base depth and adds an additive local high-pass texture layer on top, controlled by the Surface Detail slider — this is what puts fine skin, fabric, and hair texture back into an otherwise smooth AI-estimated surface. Surface Detail at 0% returns exactly the raw AI Depth base with no texture added.
Both modes call the same cloud model, so both cost the same on top of a Relief Generator generation: +1 AI credit, whether or not you add Surface Detail afterward.
Depth precision
True 16-bit (65,536 levels)
Surface Detail
0–200%, default 60%
Cost
+1 credit / generation
Full Precision, End to End
Real 16-bit depth, not an 8-bit estimate stretched thin
Most cloud depth models — including earlier versions of this one — normalize their result and hand it back as a plain 256-level grayscale image, so any fine gradation the model actually computed is thrown away before it ever reaches your browser. AI Depth's cloud model has been extended to also return the depth data before that step, at full 16-bit precision, and Depth2Beam decodes and carries it through untouched: past the initial pan and zoom of your working canvas, straight through to every export that reads the height map — the standard Height Map, the true 16-bit Height Map, the STL, and every layered or stacked export built from it.
The Lit Relief Smoothing slider (default 12px) is still there for its own reason — the raking-light shading in Lit Relief preview can visually amplify ordinary surface noise regardless of source precision — but it's a preview-only lighting control, not a fix for banding. It never touches the exported height map, STL, or any other preview mode; Height Map, Slope Map, and Compare are unaffected either way. Set it to 0 to see the raw, unsmoothed data.
Ready to generate one?
The Relief Generator builds a sculpted relief from a text prompt alone — describe the subject, then switch on AI Depth refinement for anatomical accuracy.
See Relief Generator →