Start here
- Speed
- 11.7 ms
- Memory
- 0.18 GB
- Video
- Keeps up
What Oku3D sets out of the box. Change it once you have seen a reason to - not before.
Look here
Depth Anything V2
Depth Anything V2 Small FP16
Depth Anything V2 Small Q4F16
Depth Anything V2 Base FP16
Depth Anything V2 Base Q4F16
Depth Anything V2 Large Q4F16
Depth Anything V2 Metric Indoor Small FP32
Depth Anything V2 Metric Indoor Base FP32
Depth Anything V2 Metric Indoor Large FP32
Depth Anything V2 Metric Outdoor Small FP32
Depth Anything V2 Metric Outdoor Base FP32
Depth Anything V2 Metric Outdoor Large FP32
Depth Anything V3
Depth Anything V3 Mono Large FP16
Depth Anything V3 Mono Large Q4F16
Depth Anything 3 Small FP32
Depth Anything 3 Large FP32
Depth Anything 3 Large 1.1 FP32
Depth Anything 3 Giant FP32
Depth Anything 3 Giant 1.1 FP32
Depth Anything 3 Nested Giant-Large 1.1 FP32
Depth Anything 3 Metric Large FP32
Lotus
Lotus-D FP16
Lotus-D v1.1 FP16
Lotus-G v1.0 FP32
Lotus-D v2.0 Disparity FP32
Lotus-G v2.0 Disparity FP32
Lotus-G v2.1 Disparity FP32
Depth Anything V1
Depth Anything V1 Small Q4F16
Depth Anything V1 Base Q4F16
Depth Anything V1 Large FP32
MiDaS
MiDaS v2.1 Small FP32
MiDaS DPT-Swin2-Tiny FP32
MiDaS DPT-Hybrid Q4F16
MiDaS 3.1 DPT BEiT-Large 512 FP32
MiDaS 3.1 DPT BEiT-Large 384 FP32
MiDaS 3.1 DPT BEiT-Base 384 FP32
MiDaS 3.1 DPT Swin2-Large 384 FP32
MiDaS 3.1 DPT Swin2-Base 384 FP32
MiDaS 3.0 DPT-Large 384 FP32
Distill Any Depth
Distill Any Depth Small FP32
Distill Any Depth Base FP32
Distill Any Depth Large FP32
ZoeDepth
ZoeDepth NYU+KITTI FP32
ZoeDepth NYU FP32
ZoeDepth KITTI FP32
GLPN
GLPN NYU FP32
GLPN KITTI
Apple Depth Pro
Depth Pro FP16
DINOv2
DPT DINOv2-Small NYU
DPT DINOv2-Base NYU
DPT DINOv2-Large NYU
DPT DINOv2-Giant NYU FP32
DPT DINOv2-Small KITTI FP32
DPT DINOv2-Base KITTI
DPT DINOv2-Large KITTI
DPT DINOv2-Giant KITTI
UniDepth
UniDepth v2 ViT-L FP32
UniDepth v2 ViT-S FP32
Metric3D
Metric3D v2 ViT-Small
Metric3D v2 ViT-Large
Metric3D v2 ViT-giant2
Video Depth Anything
Video Depth Anything Small
Video Depth Anything Large
DyFN
DyFN
MoGe
MoGe ViT-L
MoGe-2 ViT-L
VGGT
VGGT 1B FP32
Marigold
Marigold Depth v1.0 FP32
Marigold Depth v1.1 FP32
Marigold Depth LCM v1.0 FP32
Marigold Depth E2E-FT FP32
E2E-FT
Stable Diffusion E2E-FT Depth FP32
Iris
Iris FP32
TIPS v2
TIPS v2 Base DPT FP32
TIPS v2 Large DPT FP32
TIPS v2 SO400M DPT FP32
TIPS v2 Giant DPT FP32
About this comparison
- Models
- 76
- Scenes
- 25
- Model and resolution settings
- 343
- Depth maps rendered
- 8,575
Every model at every inference resolution on every scene - 8,575 depth maps. As far as I know the widest such comparison anywhere, on scenes built to be difficult rather than pretty.
Oku3D turns flat video into 3D by estimating depth on every frame. Which model does that decides how much detail you get and what it costs your graphics card - and the honest way to choose is to look, not to read a score.
Oku3D ships 13, picked for what they do to a picture, what they cost, and what their license allows. All 13 are here, plus 63 the app does not ship: a shortlist is worth what it was chosen from.
Know a subject that breaks depth models? Write to jens@oku3d.com - same for a model that belongs here.
How this comparison works - the scenes, the measurements, and why there is no accuracy score.
Speed and memory
The 13 models Oku3D ships, measured inside the app at every resolution they offer, on one graphics card. Speed and memory are the two things only the app can produce - the wider library is compared by picture everywhere else here.
Click any row to make it the reference; every other row then compares against it. Starts on the model Oku3D uses out of the box.
| Model | Resolution | Speed* | vs. reference | Memory* | Video | Download |
|---|---|---|---|---|---|---|
| Depth Anything V2 SmallFP16Oku3D's default model - what you get without changing anything. Fast enough to follow video on most graphics cards. | 252×252 | 3.5 ms | - | 0.08 GB | Keeps up with video | 47 MB |
| ⋮ | 364×364 | 6.7 ms | - | 0.11 GB | Keeps up with video | |
| ⋮ | 518×518 | 11.7 ms | - | 0.18 GB | Keeps up with video | |
| Depth Anything V2 SmallQ4F16Same model, compressed weights - a third of the download for a slight loss of detail. | 252×252 | 3.9 ms | - | 0.05 GB | Keeps up with video | 18 MB |
| ⋮ | 364×364 | 7 ms | - | 0.08 GB | Keeps up with video | |
| ⋮ | 518×518 | 13 ms | - | 0.15 GB | Keeps up with video | |
| Depth Anything V2 BaseFP16Noticeably cleaner edges than Small, still comfortable for video. | 252×252 | 6 ms | - | 0.24 GB | Keeps up with video | 185 MB |
| ⋮ | 364×364 | 11.2 ms | - | 0.31 GB | Keeps up with video | |
| ⋮ | 518×518 | 20.2 ms | - | 0.44 GB | Keeps up with video | |
| Depth Anything V2 BaseQ4F16Base quality at a third of the download. | 252×252 | 5.7 ms | - | 0.13 GB | Keeps up with video | 69 MB |
| ⋮ | 364×364 | 11.2 ms | - | 0.20 GB | Keeps up with video | |
| ⋮ | 518×518 | 20.5 ms | - | 0.32 GB | Keeps up with video | |
| Depth Anything V2 LargeQ4F16The largest V2. Finer separation between objects at similar distance. | 252×252 | 14.9 ms | - | 0.34 GB | Keeps up with video | 223 MB |
| ⋮ | 364×364 | 27.9 ms | - | 0.47 GB | Keeps up with video | |
| ⋮ | 518×518 | 55.4 ms | - | 0.73 GB | Too slow for video | |
| Depth Anything V3 Mono LargeFP16Best overall detail - thin structures like hair and railings survive. | 252×252 | 12.9 ms | - | 0.74 GB | Keeps up with video | 655 MB |
| ⋮ | 364×364 | 18.1 ms | - | 0.91 GB | Keeps up with video | |
| ⋮ | 518×518 | 32.5 ms | - | 1.09 GB | Keeps up with video | |
| ⋮ | 728×728 | 63.4 ms | - | 1.64 GB | Too slow for video | |
| ⋮ | 1036×1036 | 149.8 ms | - | 1.88 GB | Too slow for video | |
| Depth Anything V3 Mono LargeQ4F16V3 detail at half the download and a noticeably smaller memory footprint. | 252×252 | 12.4 ms | - | 0.44 GB | Keeps up with video | 336 MB |
| ⋮ | 364×364 | 17.5 ms | - | 0.60 GB | Keeps up with video | |
| ⋮ | 518×518 | 30.9 ms | - | 0.79 GB | Keeps up with video | |
| ⋮ | 728×728 | 64.3 ms | - | 1.34 GB | Too slow for video | |
| ⋮ | 1036×1036 | 153.8 ms | - | 1.57 GB | Too slow for video | |
| Lotus-DFP16Exceptional on thin detail such as wires and branches. Built for photos - only its smallest setting keeps up with video. | 256×256 | 34.5 ms | - | 1.99 GB | Keeps up with video | 1815 MB |
| ⋮ | 384×384 | 42.4 ms | - | 2.48 GB | Too slow for video | |
| ⋮ | 512×512 | 67 ms | - | 2.50 GB | Too slow for video | |
| ⋮ | 768×768 | 149.7 ms | - | 4.44 GB | Too slow for video | |
| ⋮ | 1024×1024 | 290.4 ms | - | 4.53 GB | Too slow for video | |
| Depth Anything V1 SmallQ4F16The first generation, kept for comparison. V2 Small is better at the same cost. | 252×252 | 4.2 ms | - | 0.07 GB | Keeps up with video | 19 MB |
| ⋮ | 364×364 | 7.7 ms | - | 0.12 GB | Keeps up with video | |
| ⋮ | 518×518 | 12.8 ms | - | 0.23 GB | Keeps up with video | |
| Depth Anything V1 BaseQ4F16First generation, superseded by V2 Base. | 252×252 | 6.1 ms | - | 0.13 GB | Keeps up with video | 71 MB |
| ⋮ | 364×364 | 11 ms | - | 0.24 GB | Keeps up with video | |
| ⋮ | 518×518 | 19.7 ms | - | 0.41 GB | Keeps up with video | |
| MiDaS v2.1 SmallFP32The fastest option on the list, and the coarsest - depth reads as a smooth gradient. | 256×256 | 2.3 ms | - | 0.13 GB | Keeps up with video | 63 MB |
| MiDaS DPT-Swin2-TinyFP32Older architecture, fixed at one resolution. | 256×256 | 4.9 ms | - | 0.30 GB | Keeps up with video | 157 MB |
| MiDaS DPT-HybridQ4F16Older architecture with more detail than the other MiDaS builds. | 384×384 | 25.2 ms | - | 0.47 GB | Keeps up with video | 112 MB |
* Measured on an AMD Radeon RX 7900 XTX. Speed is the fastest of 40 to 100 timed runs per setting, so a best case, not a typical one; memory is a middle reading, not a peak. Other hardware shifts the numbers; how the settings compare carries over. Video marks what estimates depth inside one 24 fps frame here - slower settings still play, they just refresh depth less often.
Which one should I use?
If your graphics card is busy
- Speed
- 3.9 ms
- Memory
- 0.05 GB
- Video
- Keeps up
The least memory of any setting that still keeps up with a frame here, leaving the most for everything else.
For video, on a newer card
- Speed
- 55.4 ms
- Memory
- 0.73 GB
- Video
- Slower than the picture
Past the frame budget on this card, but users on newer ones report it keeping up. Worth a try before settling for less.
For stills
- Speed
- 290.4 ms
- Memory
- 4.53 GB
- Video
- Slower than the picture
The one I reach for on a still - a judgment, not a measurement, so put it beside another and see. The wait does not matter there; the memory might.
For stills, on a smaller card
- Speed
- 149.8 ms
- Memory
- 1.88 GB
- Video
- Slower than the picture
The finest grid on offer that fits on an ordinary card. Far too slow for video, which a still does not care about.
Measured on an AMD Radeon RX 7900 XTX. Other hardware moves the numbers; how the settings compare carries over. Not a quality ranking - see for yourself above.