Benchmark · Tanks and Temples

From 301 photos to a 3D scene: the Train benchmark

Bring the locomotive’s shape, markings and texture from photographs into a 3D scene. Train’s PSNR, SSIM and LPIPS improve on the cited 3DGS paper values on all three tested platforms, with LPIPS about 30% lower.

MakeWorlds team · Tested 2026-10-01

MakeWorlds render of the Train locomotive’s front and body — Original photoMakeWorlds render

Original photograph of the same Train front and body view — Original photoMakeWorlds render

MakeWorlds render of the Train scene from an oblique viewpoint — Original photoMakeWorlds render

Original photograph of the same oblique Train viewpoint — Original photoMakeWorlds render

MakeWorlds render of the Train locomotive’s side texture — Original photoMakeWorlds render

Original photograph of the same Train side viewpoint — Original photoMakeWorlds render

Look closer at the details

Detail of the body markings in the MakeWorlds render — Body markings and paint texture

The same body markings in the original photograph — Body markings and paint texture

Detail of the steps and railings in the MakeWorlds render — Steps and railings

The same steps and railings in the original photograph — Steps and railings

Both detail comparisons use the front-and-body viewpoint with identical zoom and framing. The original photo is on the left and the MakeWorlds render is on the right.

Compared with the 3DGS paper baseline

Train improves on the cited paper baseline in PSNR, SSIM and LPIPS on all three tested platforms. LPIPS measures perceptual image differences; lower is better. In this round it is 29.5%–29.9% lower than the cited paper value.

Train · Medium · 30,000 steps · means over all evaluated test views
Δ = MakeWorlds − paper. Higher PSNR / SSIM and lower LPIPS are better.
Result	PSNR ↑ / dB	SSIM ↑	LPIPS ↓
3DGS paper baseline	21.097	0.8020	0.2180
MakeWorlds · Mac	21.227Δ +0.130	0.8463Δ +0.0443	0.1537Δ −0.0643
MakeWorlds · Linux	21.282Δ +0.185	0.8479Δ +0.0459	0.1528Δ −0.0652
MakeWorlds · Windows	21.205Δ +0.108	0.8468Δ +0.0448	0.1530Δ −0.0650

The baseline is the 30,000-step row in appendix Tables 7–9 of the 3DGS paper. Metrics cover all evaluated test views; images are selected display views. Read the paper · Tanks and Temples

Test conditions and sources

Input: 301 photographs, all cameras successfully registered
Settings: Medium tier, source resolution, 30,000 steps
Hardware: Mac: Apple M4 Max; Linux and Windows: NVIDIA RTX 4090
Mac evaluation: 19 fixed test views, using cameras solved by the product.
Images: Selected Mac test views view-000009, view-000057 and view-000249, pairing renders and original photographs from the same evaluation.

Keep reading

Selected public scenes tested on Mac, Windows and Linux