The pixel shift is too limited.

[vladislav-kostin]

Where is the parameter that determines the amount of pixel shift? Idk the correct technical term, but I mean the amount of pixel shifting for objects that are near, the current range does not allow me to display near objects at all. I can at least try to make things appear smaller, closer to real scale and nearer, but the only way is the ipd adjustment and it is not it. It doesn't allow me to create video with up close objects of irl size because there is not enough shift. My eyes still hurt from an attemmpt to scale things down by setting ipd adjustement to +40 :)

I understand that this parameter is limited probably to avoid large scale artifacts, but I'm ready to pay that price for being able to create immersive irl scale content, besides my idea is to try to inpaint the missing content, I think I have an idea that might work well for that.

I'm gonna dig into source, try to find it, and experiment with it. Will post my results if I find it and if it works.

[vladislav-kostin]

So turns out that what I wanted is divergence and it is called 3D Strength in the UI. I have to go far beyond recommended range and well into unsupported. So I had to switch to forward_fill method. Which is I presume a more primitive one, but still the results are better than what I was getting in other tools (in comfy nodes).

With divergence 20-40 and ipd adjustment of 30-40 I'm starting to get the characters that are correct size and are close enough. In the image below I could lean forward reach out and grab that drink in VR. I think this would correspond to the position of the camera.

This is still pretty eye-straining likely due too many inaccuracies.

I think one of the reasons I have to stretch the value so much is because I have low fov shots, but such is my use case.

[nagadomi]

Since 3d strength is defined relative to the image width, so that vertically oriented image might need four times the strength compared to a standard 16:9 video.

Also, the perceived depth depends on the screen size and viewing distance in the video player or image viewer.

In my experience, even a 3d strengt value of 5 feels too high, but I know that some users prefer using very strong 3D strength settings.

[vladislav-kostin]

If the divergence is not measured in pixels or dpi and is based on physical size and distance, then these values should be consistent for all images with same fov. At least that's what I think should be the case. The image borders have to be padded to account for the missing fov though.

The values I got on this image I believe are "physically accurate" in a sense that real camera would cature this much shift at this distance, so to me it's not a preference it's just the correct value to display the subject at the irl scale. It is a preference in a sense that there are two different use cases. One is for 3D movies, and the other is for immersive real scale content. For movies this stuff is preference based, for irl scale content it is physycal units based.

So I wish all units were a mode with real world scale so I wouldn't have to guess that much. But for now I would still have to, because I don't know the shot fov. I might have a setup where fov of the image/video will be known, the depth is metric, so in theory I would never have to manually adjust any values (unless there is depth error) and my stereoscopy will be always physically accurate.

[nagadomi]

With divergence = 2, the maximum pixel shift is 2% of the width.

SBS image viewers display images fitted to a fixed-size screen, so the magnification varies depending on the aspect ratio. Also, the screen size may differ between software. VR90 output may eliminate these dependencies.

Fov can be estimated from the image in DepthPro or MoGe.
iw3 does not handle metric depth correctly.
Probably, the one closest to your idea is
https://github.com/calledit/metric_depth_video_toolbox

edit:

magnification varies depending on the aspect ratio

This applies when either the width or height is fixed. I'm not sure how each software handles it.
You can fix the output aspect ratio by setting the padding option to 16:9.