How the VFX engine works
July 10, 2026
A peek at the pipeline that turns a flat PNG into a perspective-accurate 3D pass — depth estimation, parallax, and the trickiest part, occlusion.
The illusion is the easy part; the engineering is making it hold up under real parallax. This post walks through the pipeline:
1. Depth estimation — a small network predicts a per-pixel depth map from the input PNG. Accuracy matters: a 1px error here is the difference between a solid looking glued-on. 2. Mesh build — the depth map becomes a quad mesh tiled in screen space, with per-vertex depth baked in. No actual 3D modelling, no raycaster at runtime. 3. Parallax pass — at render time we shift the mesh in 3D based on the camera tilt. The trick is bicubic resampling so smooth gradients stay smooth at the edges. 4. Output — we composite the parallaxed mesh over the original flat art at the chosen DPI, then upscale to standard POD framing.
The hardest part is occlusion: real 3D objects hide parts of themselves when viewed at an angle. We do a cheap pre-pass that masks likely-occluded regions and trust the user to pick angles where the trick reads.
Up next