Idol

Single image to 3D object

Idol turns one image into a 3D object made of Gaussians. A photograph, a render or an Iris-3B generation goes in; a splat you can turn around comes out, including the sides the image never showed. It can generate each object at several levels of detail, from 32,768 to 1,048,576 Gaussians. With 0.66B parameters, it runs in under 8 GB of GPU memory.

Try the demo Code Weights

Alejandro Cañada^†Chema GarabitoSperidlabs · † project lead

Gallery

Every object below came from a single image.

From a prompt

Iris-3B draws the image from a prompt. Idol builds the object from the image.

Prompt

“a small robot duck in sunflower-yellow plastic, with orange feet”“a vintage mint green Italian motor scooter”

Iris-3B image for: a small robot duck in sunflower-yellow plastic, with orange feet — Iris-3B · Image

Idol · 3D object

Adjustable levels of quality

From 32,768 to 1,048,576 Gaussians.

32,768 · 0%

1,048,576 · 0%

‹›

Object

Left

Right

1,048,57612.2 MB

Benchmark

Mean scores. Bold is the best in each column.1 × H100 per model

Metric	Idol	TripoSplat	TRELLIS	LGM
Input view
PSNR ↑	12.25	11.76	11.89	12.46*
SSIM ↑	0.726	0.712	0.716	0.722
LPIPS ↓	0.213	0.225	0.235	0.237
IoU ↑	0.862	0.860	0.845	0.834
New views
DINOv2 ↑	0.781	0.743	0.754	0.575
CLIP ↑	0.885	0.868	0.868	0.784
Sharpness distance ↓	0.81	1.03	0.97	1.22
Request time ↓	6.8 s	9.9 s	7.1 s	1.1 s

Model	Input view				New views			Request time ↓
Model	PSNR ↑	SSIM ↑	LPIPS ↓	IoU ↑	DINOv2 ↑	CLIP ↑	Sharpness distance ↓	Request time ↓
Idol50 steps, 262,144 Gaussians	12.25	0.726	0.213	0.862	0.781	0.885	0.81	6.8 s
TripoSplat50 steps, 262,144 Gaussians	11.76	0.712	0.225	0.860	0.743	0.868	1.03	9.9 s
TRELLISPaper setting, 50 + 50 steps	11.89	0.716	0.235	0.845	0.754	0.868	0.97	7.1 s
LGMDefault, 65,536 Gaussians	12.46*	0.722	0.237	0.834	0.575	0.784	1.22	1.1 s

* LGM has the highest PSNR despite visibly blurry results, because PSNR averages squared pixel error.

Across GPUs

One image to 262,144 Gaussians in 3.4 s.

Request time at 50 steps.

B300 269 GB3.4 s
B200 179 GB3.6 s
H200 140 GB6.6 s
H100 80 GB6.7 s
RTX 5090 32 GB14.1 s
A100 80 GB16.3 s
L40S 45 GB16.6 s
RTX 4090 24 GB19.1 s
RTX 3080 10 GB46.2 s
L4 22 GB54.6 s

Median over 8 images, with the same code and settings on every GPU, one GPU per machine.

Cite this work

@misc{canada2026idol,
  title  = {Idol: Single Image to 3D Gaussian Splats},
  author = {Ca{\~n}ada, Alejandro and Garabito, Chema},
  year   = {2026},
  url    = {https://speridlabs.com/research/idol}
}