← Lab bench
EXP-104Result

We rebuilt our 3D character with a newer generator: clothes came out sharp, the face did not

80px

the face's height in the input image: too few pixels for a 3D generator to rebuild a face

An AI-generated woman in a green T-shirt, grey trousers and black boots, standing in a full-body pose on white
1/4The input: one full-body image, 600 × 900 pixels. The face is only about 80 pixels tall.
  1. 1 · Input
  2. 2 · New 3D
  3. 3 · The weak point
  4. 4 · September

We re-ran our September photo-to-character test with an open, commercially usable 3D generator released since. The same AI-made full-body image went in, and about two minutes later a textured 3D figure came out. Clothes, boots, folds and body shape were far cleaner than in September. The face and the back of the hair were not, and the reason is worth knowing before you try this yourself: the generator can only rebuild what the picture shows clearly. The person shown is AI-generated, not a real individual.

What we measured

MeasureSeptemberNovemberNote
Input1 portrait; body invented by an image modelthe same full-body image, 600 × 900
Time to a textured 3D figuren/aabout 2 minutes68 s to generate, 55 s to export
Outputrigged mesh, 3 motionstextured mesh, 11.6 MB, not rigged
Clothes and bootssoft, blurred foldscrisp folds, laces and seams
Facesoft but readablewaxy, pale, low detail
Back of the hairclosedholes

What went wrong

  • The face was about 80 pixels tall in the input. Clothes are large, simple shapes the generator can rebuild from few pixels; a face is small and full of detail it cannot invent well, so it came out waxy.
  • The back of the head was never in the picture. The generator had to guess it, and on curly hair the guess left holes.
  • The free GPU allowance ran out after one run, so we could not try other seeds to see whether the face was bad luck or the input.

What happens next

  • The lesson, and our next test: give the generator a sharp full-body image with the face at least 300 pixels tall, and run several seeds at the top quality setting.
  • A back view of the same person as a second input, so the hair is seen rather than guessed.
  • Rig and animate only once the face holds up. Until then, the September character stays.
Cover of the book Gaussian Splatting in 2026
The book behind the lab

Get the book today: learn Gaussian splatting and apply it to real projects.

People keep asking how we build what you see here. The book walks students, educators and practitioners from a drone clip to a measured 3D in the browser, with hands-on labs and the code.

Launch price until 3 December.

Teaching a course (Level 5, Level 6 and up)? Email junaid.malik@simamdigital.com about a class discount.

Next experiment
One portrait to a walking 3D character →
Want this for your data?