This is for generating the see-through movie style 'holograms' as seen in Bladerunner, Etc.
Trigger is "Transparent Hologram Woman". If it's not clear this is a biased towards generating Holo-Waifus. If the response is good, I'll make a husbando one next. Guyfu?...(If that's not a word it should be)
So, this model is both over-trained and under-trained at the same time. I know that's confusing. But it's a It will respond well enough with small prompts but start describing more and more and it will stop reacting.
I'm including 3rd, 2nd, 4th, 7th epochs, as I'm testing them.
2- Does glowing effect well without artifacting, gives transparent-ish results sometimes, but very subtle.
3 - Starting to artifact but also will give a transparent result most of the time.
4 - The artifacting is readily apparent at times, but will always generate a transparent figure.
7 - CRISPY AF. For use with stubborn prompts/overtrained checkpoints(Dreambooth, etc...).
It's still a work in Progress, but it's good enough for me to be happy enough with releasing it. Dataset was 23 images I made in Photoshop. I'm looking to use this training set to get a more diverse and varied dataset to improve the training later. And hopefully people will post some images I can use to improve the dataset.
Technically I think they are called "Volumetric Display Projections", which is just a start of why making this training was so difficult. SD does not like making people not occlude the background. And it shows. It also doesn't like an object that's transparent but doesn't refract light. This does both of those things 'okay' despite this.
Description
This is gives the best all around results with most checkpoints, without the over-training being that obvious. Keep prompt length small.
FAQ
Comments (6)
This is neat. Rather than one that does women and one that does men, it would be cool if you were able to do one that could do men, women, animals, objects, etc. Make a hologram of anything in a scene. Assuming that's realistically possible.
I love the idea.
Yeah, that'll be something I'll be working on. But it's really hard to create a dataset that it will learn as the visual concept of 'transparency' without refracting the background. Really hard to get good stock images, as it's a hard thing to search for. I had to make every image in Photoshop which leads to my bias which is not good.
I think base XL does transparent objects pretty well, though.
This was only my first attempt and a single training run, to prove the concept was viable.
@Rysk Yeah, that makes sense. It's a very cool looking concept, so nice job.
Hey, love this concept, im trying to use it to make ghosts, but when i put it in the Lora folder nothing shows up in A1111, broken?
This is an XL LoRA. Are you using a SD 1.5 checkpoint?
Hi, loved this lora. I used it a couple months back to make a character on an SD model(not SDXL), as I prefer to make my characters like that then upscale them, and it worked perfectly fine, even though it seemed to be the same version, but now it seems to only work on XL models anymore? I was wondering if it would be possible to make it available for SD again? At least as another version?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.