← All Automations COMPUTER VISION + GENERATIVE ART

Pose Lift

I paint superhero covers as a hobby. This is the AI project that grew out of that — strike a pose on webcam, it gets rebuilt as a real 3D mesh, and that exact pose drives a generated superhero portrait in my own painting style. Calibrated against genuine computer-vision engineering difficulty, not a thin wrapper around a generative-image API.

WHAT IT DOES

Your pose, real geometry.

• A single webcam frame gets reconstructed into a true 3D human mesh — 6890 vertices, real joints and geometry, not a 2D skeleton or landmark overlay
• That mesh renders into a depth map, which conditions the image generation directly — the output pose is pixel-locked to your real geometry instead of a text description guessing at it
• Pick a superhero and a stylized comic-cover portrait generates in your exact pose, rendered in the same painting style as my real physical artwork
• Runs as a live web app — capture, style pick, and generation all happen in the browser, no local setup required

HOW IT WORKS

Real math, not a guess.

1. 3D reconstruction via ROMP/SMPL

   A webcam frame runs through ROMP to produce a genuine SMPL mesh — real vertex and joint data, not pseudo-3D landmarks. Getting the gated SMPL model registered and converted into ROMP's expected format took real work — this was the hardest technical piece of the build.

2. Depth map, then depth-conditioned generation

   The mesh renders into a depth map, which conditions a Replicate flux-depth-pro generation. Early attempts at unconditioned text-to-image guessed the pose wrong every time — switching to real depth conditioning got a correct pose match on the first try.

3. Serverless GPU backend

   ROMP and pyrender run headless (xvfb + EGL) on Modal, exposed as a rate-limited API. A Next.js frontend on Vercel handles capture, style selection, and the generation flow end to end.

4. Style locked to real paintings

   The generation prompt is tuned to match my actual physical comic-cover paintings — Amazing Spider-Man #258, Giant-Size X-Men #1, Silver Surfer #4 — so the output reads as "my style," not a generic AI look.

WATCH IT RUN

From paintings to pixels.

TRY IT

Live, open source, ready to run.

Strike a pose, pick a hero, see yourself rendered in the style of my real paintings.