Measured on RTX 4090, DGX Spark and a 16GB Mac: speed, memory, and what worked on paintings
#11
by jooioooo - opened
Thanks for releasing Pixal3D. I ran it on three machines and on objects cut out of public-domain paintings (Van Gogh, Cézanne, Vermeer, Seurat, Heda), and I'm sharing the numbers here in case they help others pick hardware and settings.
Van Gogh's chair from The Bedroom (1889), the 3D front, and the same 3D from the side. The painting's perspective carries into the shape.
Speed and memory
| Machine | Settings | Time per model | Memory |
|---|---|---|---|
| RTX 4090 24GB (62 GiB RAM) | low_vram, resolution 1024 |
115 s median (83–200 s, 12 runs) | GPU 19.4 GiB; system RAM filled our 59 GiB cap |
| DGX Spark (GB10, 128GB unified) | low_vram, resolution 1024 |
169–420 s | worker RSS 89.6 GiB |
| DGX Spark | low_vram off, resolution 1536 |
255–305 s (no warmup) | about 47 GiB for the whole machine |
| MacBook M2 16GB | community 8-bit port (pixal3d.cpp) | about 34 min, one image | not measured |
- 24 GB needs
low_vram. With it off, both 1536 and 1024 ran out of GPU memory on the 4090. With it on, system RAM filled the 59 GiB cap we set, so on a 64 GB PC each server start left at most 3 results before the cap stopped it. RAM did not keep growing across generations. - On unified memory (DGX Spark), turn
low_vramoff. It appeared to use more memory, not less, and the upstream default (low_vramoff, 1536) ran fine. Our setup used torch 2.7 (cu128), which needed a CUDA 12.9 toolkit for the extensions and Triton 3.5.1 for GB10 (sm_121). - 1536 vs 1024 looked about the same at render size. Both reduce the output to 1M faces and a 4096 texture.
What we saw on paintings (one run per input, seed 7)
- The front lines up with the painting, and brushstrokes carry into the texture.
- The painting's perspective becomes shape: Van Gogh's chair leans back when seen from the side.
- With the background left in, 5 of 7 judged inputs got the floor or background attached as a flat slab.
- Objects held in a hand came out with the hand attached. Unpainted backs were sometimes filled with a simple blob.
Full report with every number, renders from four angles and the setup (EN/KO/中文): https://github.com/MindrLabs/image-to-3d-pixal3d