This lora create SBS 3D images that can be viewed by the crosseyed method
The phrases: "Stereoscopic SBS side-by-side 3D," At the start of the prompt helps a lot.
Check image prompts as examples.
V2 Update: V2 is leaps ahead of V1. I refined the dataset by getting rid of images with minimal effect and parallel view images that slipped through the first time. I also tested the produced loras a lot and came to the conclusion that 2400 steps lora is the most ballanced of all. For some reason, even after 3600 steps Krea 2 refuses to increase the depth of the effect. So these are the best settings I propose in order to have the best results:
Lora strength 1: minimal 3d effect, and minimal NSFW in SFW prompt.
Lora strength 1.3-1.7: Very good 3d effect and minimal NSFW on SFW prompt.
Lora strength 1.8-2: Best 3d effect and more NSFW (nudity) images on SFW prompt. Possible appearence of text in corners depending on the checkpoint.
The level of NSFW on SFW prompt depends on the model. Heavy NSFW models tend to undress subjects more. Censored models keep things clean.
In NSFW prompts you can use Lora strength 1.3-1.7 with excellent results.
Description
Better parallax effect, greater depth, better concistency
FAQ
Comments (11)
Thanks for posting version 2. The example images seem to really have a strong (but not too strong) 3D effect so it'll be fun to test this one out. Interesting that you did find a few stray parallel format images in your dataset. I wonder if that did account for the occasional mixed format in version 1 outputs(?)
Thank you. I think it is was a complex problem. The dataset enhancement helped a lot but Krea2 seems to be the major issue. I wonder if i could make it even better by training for more steps a more ballanced dataset (more sfw images and less nsfw). When I won't need my pc for a couple of days i will try to train for 6000-8000 steps to see if I can overcome the Krea2 pretraining obstacle.
Great - I've made sure to "follow" this page and will be very happy to test any future iterations of this lora. I did notice that version 2 leans very heavily towards nsfw. My use case is a bit awkward as I often do pics that are clothed female glamour shots but if I use this lora it will definitely push them into the nsfw category - if it can make clothes disappear, be pulled up or open, it will ;-D
Did you consider splitting your dataset into sfw and nsfw and training separately? Ideally, I'd rather use a sfw lora for most things but you may find a specialised nsfw 3D SBS lora would be the more popular (it is Civitai, after all!). Anyway, I'm certainly not demanding anything and am just grateful that someone is putting out good quality 3D SBS loras.
EDIT - I just noticed your detailed prompts in some of the "clothed glamour" images in your examples. I wonder if that's what I'm missing at the moment e.g. detailed descriptions of the parallax effect at different distances. I will copy some of your prompts and try them myself. If this allows me to lower the lora strength but keep the strong 3D effect, that could be a good outcome. Thanks.
@bennyboy_77 By refining the dataset I changed the ballance between sfw and nsfw images in favor of nsfw. I have already collected some more sfw images with good 3d depth and in v3 I will change the ballance for a more sfw outcome. p.s. You can set the lora to 1.3 without missing much of the effect and keeping the image relatively non nsfw.
@mpamais Good work. I'm sure you've got it all in hand but if you need anyone to help with training, let me know as I've got a reasonable local setup here (RTX 3090 card) so am happy to pitch in and do some training. Maybe I could try to train a dedicated SFW version based on your dataset (?)
I've had a little chat with ChatGPT and it points out that it would be fairly straightforward for me to use either Stereo Photo Maker or create a Pillow/Python script to chop and swap stereo photo images to make a new dataset of "parallel" SBS images from an original "cross-eyed" dataset. If you want, I could treat this as a little project to supply you this adjusted dataset or even fully train it and you can post it here on your page.
My other thought is that I reckon there may be more parallel data in Krea's native understanding of 3D SBS and you may find it adopts and integrates the trained data more smoothly than your "cross-eyed" dataset. I've no proof this is the case but most SBS I've encountered across the years has been parallel images.
Apologies if I'm overstepping the mark. No pressure, but the offer's there if you want me to do anything mentioned above. If not, no probs and I'll look forward to version 3, while having a play with your higher quality prompts I mentioned before.
I think 3D SBS has a lot of promise in Krea 2 but it has to be damn near perfect or it looks a bit "off" when viewed on a VR headset. I guess the human brain is incredibly visual and good at spotting any discrepancies. Damn our super powerful human brains!
@bennyboy_77 Thank you for the offer. I indeed try to work with a 3060 12GB card and it takes much time to train, but i have plans to upgrade soon. I have already created an app in python to do exactly what you said. it crops the images and exchanges the right and left side to create parallel view dataset from crosseye and vice versa. I already have the new dataset in both formats and I plan to train this after I finish the crosseye versions first. I'd be happy to share with you the app if you'd like.
@mpamais Amazing. Sounds like we're on the same wavelength with that one! If it's no trouble, it would be fun to try out your app. I get the impression that Stereo Photo Maker can do it fairly easily but possibly only one image at a time. Not really a problem, if not, as I don't have my own dataset but I'd probably just use it to make "parallel" versions of my own "cross-eyed" creations.
I upgraded from a 3060 to a 3090 last year and the difference is incredible - that extra VRAM makes a huge difference so it's still a strong option if you can't get hold of a 5090. I think some people swear by the 16GB cards too. Anyway, you've probably got that all budgeted and planned for.
Thanks again, and I'll not pester you! :-D
@bennyboy_77 I will probably go for the 5070TI since the 5090 cost is VERY huge. I will upload the app to cloud later and send you the link.
@mpamais I'm pretty sure that card is the one that is considered the sweet spot for value and availability so it looks like a good choice. Thanks (in advance) for access to the app.
It works man lol but the last thing i need is to be staring at my computer for hours crosseyed.
Thank you. I have been trying to get it right for a couple of weeks. If it took any longer i would have problem keeping my eyes straight LOL