The lora creates an second image that can be combined via image stich (with the orginal) to create an image that can be viewed in VR.
It works well with 3D, 2.5 and photos. 2D images can be an hit or miss, due to them missing in the dataset.
I recommend generating atleast 2 imagepairs, with clip strength of 2 and model of 1.5 (varies for type of image and wanted effect depth)
V1.5
Around 1600 images in dataset
Better perspective handling
Better reflections
Sometimes less depth then V1 and unruly
Experiment for V2
V1
Around 3600 images where used in the dataset, with mixed content.
The 4000 steps version is a bit more random, and can be found on huggingface (eggnog/Stereogram_Kontext_LoRa).
Description
Initial release
FAQ
Comments (9)
Thanks for making this LoRA. I've only played with it a little bit but my first impressions are very good. As with most image generation, you have to roll the dice a few times and pick the best one, for sure.
Can I ask a little bit about the dataset used? Were all the images used from "real world" stereographic camera photos, were they artificially generated (e.g. via iw3 or similar) or a mixture of both types?
In my opinion, existing methods for creating 3D images (like iw3) do a good job of calculating depth for the image but aren't capable of creating reflections and refractions in the image e.g. the light reflecting off a leather jacket or light being bent and scattered through a glass ornament.
I feel like I am getting some outputs from your LoRA that do a decent job of representing reflections which hints towards the source data containing real-world 3D photos, but let me know if I'm right.
Hi, thanks.
> were they artificially generated (e.g. via iw3 or similar) or a mixture of both types?
The dataset doesn't contain generated images (~70% real, ~30% render).
I find that iw3, Stereocrafter ... always have some drawbacks, that i don't want to introduce.
The idea to include more images with reflection is nice, if I'm going to release an v2 ill do that.
cihog Great - looks like you've taken the much harder but hopefully more rewarding approach. That must be a lot of work. Did you find some sort of batch process to split the stereo images into left/right pairs and auto place in the "dataset" and "control" folders for training?
Either way, I'm impressed by your approach and can't see that the 30% 3D render content would be a problem for training. It probably just keeps the LoRA more flexible for animated/CGI images.
Having had a bit more of a play, I've certainly seen evidence of realistic shine on hair and clothing and even the odd twinkling eye on characters. That's a huge step in the right direction for creating realistic stereo images. I'll definitely keep an eye out for any new versions you might release as well :-)
thank you for this concept!
it also works nice with AI generated content.
(and for the ones who remember the "3D" images from "way back when": just squint a bit and you will see it on your display. no VR equipement needed))
If a VR film can be produced, it will be a new milestone
this is really really cool, thank you for this. If you ever get the urge to make a Wan I2V lora of this kind, I won't be stopping you ;)))
You dont need a lora for that brother, you can use a depth map then stereoscope in ComfyUI. it might take a minute or so but at least you wont need to wait for or train a lora
fusaki333 does that mean that I also don't need this lora to make stereoscope images?
Radyschen It depends on what kind of effect you want. Stereoscope shifts the pixels with a depth map and doesn't create new details (to my knowledge).