Comparison with Depth-Anything-V2 demo

[gituser123456789000]

I pushed for Tiling.. and I assume we got some sort of BoostingM0nocularDepth added in.. but it's causing issues. We're getting worse maps than default Depth-Anything v2

Comparing side by side, back and forth, the huggingface demo results vs iw3's depth map output at the same settings.. there are many issues and all bad

I used a 1000x1333 image. The huggingface demo gave me this native resolution as the result... but putting in 1000 resolution in iw3 gave a result of 1008x1330.. which is 8 pixels wider, but 3 pixels SHORTER (instead of slightly taller to compensate for the extra width that was introduced. That right there is a big problem. Now the depth map doesn't align with the original image.

Beyond that.. the maps look very similar.. you can see they're made from the same model.. but the shading, the colors, the graident is off in many areas. The nose is slightly brighter, which means it will stick out more.. the eyes have more definition, which looks better at first glance, but that can mean the eyes will be sunken in.. other areas, where the body should be relatively uniform or just slightly different in shade, it's instead too much difference in gradient, so basic body or object shape is distorted.

As far as BoostingMonocularDepth (assuming we got this or some similar adaptation of it).. I never got an impressive looking result from that when I downloaded it locally and used the included MiDaS v2 and Leres conversions. Their demo looks impressive, but I got nothing impressive from it. I chalked it up to MiDaS 2 being an old model, but I think BoostingMonocularDepth may just not be great or accurate. It may produce nice looking detail in maps, but not accuracy.

I pushed for Tiling, but in this case, whatever implementation this is.. it's not working as I hoped.

We need default, native resolution, straight from the Depth-Anything v2 models. It looks impressive, because it seems to already be the best.. but iw3 is currently altering the output to be something that's less accurate.

Native resolution, direct result should be the default option.. or some check box to process in native resolution, direct from the model... and whatever scaling / tiling / BoostingMonocularDepth type action is happening, should be an option at best, but really, we should consider removing it if it's only causing inaccuracies.

I'm just delivering factual feedback to be clear. We appreciate your work very much, nagadomi. Thank you for attempting to implement something here as an improvement, but in this case, it's not performing as we may have expected it to.

[nagadomi]

First, iw3 does not use anything like BoostingMonocularDepth.
There may be some differences in implementation, but the process is basically the same as in the demo.

Regarding depth resolution,
DepthAnything model requires that the input size is a multiple of 14 for height and width.
1000 and 1333 are not multiples of 14, so it is resized to 1008x1330 before input to DepthAnything model.
When the depth is actually used, it is resized back to its original resolution 1000x1333.
The demo simply outputs the depthmap after this resize back, so it matches the original resolution.
iw3 does not resize back to the original resolution until it is actually used. Because it is a waste of space. In export feature, it wastes several to dozens of times the storage size.

Regarding the different gradients of depth,
the demo outputs 8-bit grayscale (256 level) and iw3 outputs 16 grayscale (65536 level).
At the moment I think that's the difference, but I'm sure many changes have been made recently so I will look into it again in more detail.

EDIT:
If you are testing on the official demo, it is always running with Depth Resolution=518 (and resize back to the original resolution). So if you want to compare, you need to use Depth Resolution=518 in iw3.

[gituser123456789000]

I realized it today, but the difference between 8-bit grayscale, 16-bit grayscale, and 32-bit grayscale may not be visible on the PC monitor.

I'm still in bed, so I haven't looked fully into this, but here they apparently eliminated the banding

https://www.reddit.com/r/StableDiffusion/comments/1dmypej/making_3d_basreliefs_with_depth_anything_v2_16bit/

[nagadomi]

iw3 already outputs in 16bit. If you check with Blender in the same way you will see that there is no banding.

[nagadomi]

I checked with Blender. Note that this does not use Displacement, but simply uses Bump (I can see the difference enough).
Left is demo ouptut(8bit), Right is iw3 Export output (16bit). Click to view larger image.

Both depthmaps look the same on the PC monitor.
demo outout(8bit)

iw3 Export output (16bit)

[nagadomi]

I have changed the title to the suitable title. Please edit if you have an alternate title.

[gituser123456789000]

I have changed the title to the suitable title. Please edit if you have an alternate title.

all good.. that's a fine title

I'm just always looking for better quality if it's possible. Like right now I want to see how good that Giant version of the model is.

I know you said iw3 was already being processed at a higher bitrate, but I have to learn and understand why I'm seeing what I'm seeing.

This way in Blender you posted, shows it nice and clearly.

Our monitors are lying to us by only being 8bit in most cases.. so it's hard to initially believe our eyes are wrong when just looking at the depth maps, but there's extra, additional hidden information in the higher bit files that we just can't see in that way, unless we do certain projects, like shown in Blender, or until we maybe get higher bit depth monitors. I think 10bit monitors or VR headsets are probably already a thing... or will be soon. It's also interesting to know that content we make now might already have the ability to look better in the future, on more advanced displays.