Image extrapolation for stereo photography?

I believe there are tools to REDUCE the interaxial (reduce deviation) in a stereo pair, but not INCREASE it (very well), due to the guessing required for the missing obscuration that would emerge.

Paul Houle wrote:

I first got into stereo photography because my son had found some stereograms on the web and an early thing we tried was making a stereogram from some anime art by creating a depth field and sliding the pixels around accordingly.

This failed, of course, because in a stereogram slightly different fields are visible to the left and right eye and it showed up in that attempt as black fringes around the edges of objects because I didn't have any data to put there. Here is an image for one eye:

What it comes down to is that different things are obscured for the left and the right eye

The term for that is "occlusion". It's a well-known problem in converting 2D images and videos to 3D (2.5D is a more accurate, but less common term) stereo images.

so there is really more information in a stereogram than an ordinary photo. Maybe some of why emotional scenes in the cinema work well in stereography is that you get to see a little bit more of the expressive muscles on the sides of the face that are normally obscured.

I think the greater emotional effect is in the separation of characters from background.

For the anime art this can fixed by emulating the "multiplane camera" that is used to composite cels, if I have separate character and background cels it is straightforward to generate images for the left and right eye... But I need either a complete background or enough of the background to fill into those areas that are black.

The algorithms that generate that are referred to as "occlusion inpainting" or "disocclusion" algorithms. Google those things and you can find all sorts of papers and experimental software.

If you're only concerned about manually converting a relatively small number of 2D images to 3D, try the inpainting features of your photo editor, such as PhotoShop "content aware fill", Affinity "inpainting brush", or the G'MIC plug-in in Gimp.

I shoot stereograms today with a QooCam Ego and often get good results. I wish though sometimes the interaxial distance was not fixed or that I could otherwise manipulate the depth map of the image.

Disocclusion is less of a challenge for video, because you can build up the background as foreground objects move. There are a dozen different under-$100 2D to 3D video conversion programs such as Video Solo or 3D Combine.  My Sharp 3D TV and Sony 3D blu-ray both have real-time 2D to 3D conversion, but it's nothing to write home about.

Some similar scenarios would be: making a stereogram with a normal photograph plus a depth map (there are good algorithms for guessing a depth map from a single frame), making a stereogram from a light field camera, etc.

With all the image synthesis stuff being developed today such as diffusion models and NeRF it seems you should be able to make something up for the pixels that are obscured in the original frame.

If you want that sort of thing, there are Stable Diffusion plugins for PhotoShop, the Gimp, Krita, etc.

Usually I find that stereograms work better if the left and right images aren't too far from each other and in that case the fringe areas aren't very large and not so hard to interpolate. I've also read that they had tools in the movie industry by 2008 which could apply depth maps to an ordinary movie to "depth grade" it and make a stereo movie. I don't know how much labor that takes as Hollywood pays a lot of people to touch up frames for movies.

A lot of labor. But not as much as shooting in 3D, especially when you compositing stereo video with CGI.

So I'm wondering: what kind of tools are out there for extrapolating pixels to set stereo photos free from a fixed interaxial?

I think I already answered that.