|
|
Post by Queens of Absolution on Aug 29, 2023 6:07:44 GMT -6
Like everyone, I use Stable Diffusion or some form of it to create my AI images. We're all using them now - whether we're doing so through a website, a Stable Diffusion companion, or the core itself. It has helped us enrich our stories and make bouts more interesting. For me, I had a problem: Stable Diffusion's dataset didn't properly render Elyse Levesque who for all intents and purposes, is the leader of the Queen's fighters. So I hit the proverbial books and learned the ins and outs of Stable Diffusion which is a really powerful AI tool that I now think we are all blessed with ability of having. What I came up with? The ability to 'train' new datasets for Stable Diffusion that allow the creation of AI imagery that it wouldn't make before - in this case Elyse Levesque. Or at least as close to her as currently possible for me with training the AI datasets. It's not perfect but I thought I'd share comparison results with you all. My next project is Tiera Skovbye. But that will be a little while - learning all of this took a while but in the end it only took 10 minutes (quite literally) to create the Elyse Levesque dataset that the AI now uses when I want to make an Elyse Levesque image for the FCBA. I'm curious, how do you all get the results you desire. Some of you have pretty basic AI images and others very advanced? I'd like to know how the training compares.
Elyse Levesque (Reality):  Elyse Levesque (AI):   |
|
cbr
New Member
Posts: 31
|
Post by cbr on Aug 30, 2023 3:45:15 GMT -6
|
|
|
|
Post by Dave Knowles on Aug 30, 2023 4:58:26 GMT -6
![]() Alesha Dixon Attachments:
|
|
|
|
Post by The Dollhouse on Aug 30, 2023 5:51:21 GMT -6
I don't use Stable Diffusion because I'm too lazy to learn it in depth, and because I'm getting good results with MidJourney and Mage.Space. Examples: imgur.com/a/FQgcwoLThe issue is that some celebs are rendered pretty accurately and others not at all and it seems random. I'm getting terrible results for almost all the girls in my stable while other random celebs turn out amazing. MidJourney now also has a new plugin that lets you swap faces and it's pretty f*cking good. So even if the original AI pic gets the face wrong, you can use any pic of a celebrity's face to replace with the ai generated one, the results are pretty good, for some celebs more than others, and sometimes you need to try a bunch of face pis until you find one that 'fits' well into the AI generated head. (also, this plugin is not free.) |
|
|
|
Post by Queens of Absolution on Aug 30, 2023 7:08:02 GMT -6
cbr - I don't think AI will have problems reproducing something close to reality. It's best labeled on some Stable Diffusion documentation - it's like looking through a lens of a camera to take a photo from a mirror. At this juncture, AI imagery isn't 'fully' there and it will take a lot more training and datasets plus accessibility to get it there. Dave Knowles - definitely an interesting AI interpretation. I'm guessing you had some sort of animated VAE still on (or were using a model intended for 3D work) without knowing. Still, great work! I don't use Stable Diffusion because I'm too lazy to learn it in depth, and because I'm getting good results with MidJourney and Mage.Space. Examples: imgur.com/a/FQgcwoLThe issue is that some celebs are rendered pretty accurately and others not at all and it seems random. I'm getting terrible results for almost all the girls in my stable while other random celebs turn out amazing. MidJourney now also has a new plugin that lets you swap faces and it's pretty f*cking good. So even if the original AI pic gets the face wrong, you can use any pic of a celebrity's face to replace with the ai generated one, the results are pretty good, for some celebs more than others, and sometimes you need to try a bunch of face pis until you find one that 'fits' well into the AI generated head. (also, this plugin is not free.) Haha, totally understand and get it. I use MidJourney too (it's what created Late Night Before the Bell with Natalie Dormer's 'hero image') and MidJourney is definitely a separate AI which comes up with fantastic results. Mage.Space on the other hand, is just a well developed web UI for Stable Diffusion to make it more accessible. See the screenshot below for the model listing showing SDXL, Stable Diffusion v1.5, etc.  I completely agree with you about the rendering. It's an issue. If I recall correctly, given it's not on your stable page which is at least outdated from November last year as the Queens have Paige now, you have Bella Thorne. She should, in my experience, come out pretty accurately on virtually any site and in MidJourney. I can however, not get her to render correctly in Mage.Space which is actually my least favorite Stable Diffusion art creation site and I try to get everyone off of it as quickly as possible. In MidJourney she comes out decently although I did this quick so there's no refinement:  Haven't tried this plugin you speak of; it does sound pretty good. I haven't been keeping up with MidJourney too much but I know using "-v 3" or "-v 4" at the end of your prompt without the quotes might yield better results as it uses the Version 3 and Version 4 engines respectively. I don't know, I'm not too fond of the Version 5 engine that's default now if you don't specify. Now, the album you provided, imgur.com/a/FQgcwoL - contains some brilliant work. Which did you use to come up with those images as those are impressive. On a sidenote, Natasha Calis is one of the few from my Stable that couldn't render right - I trained a model of her for Stable Diffusion, yesterday which only takes about 10 minutes on an A100 GPU: 
|
|
|
|
Post by The Dollhouse on Aug 30, 2023 9:29:39 GMT -6
In MidJourney she comes out decently although I did this quick so there's no refinement: I don't think this looks anything like Bella Thorne lol, maybe 20%. I've tried making Bella in every AI I could get my hands on and it never came out accurate enough. :\ Now, the album you provided, imgur.com/a/FQgcwoL - contains some brilliant work. Which did you use to come up with those images as those are impressive. Thanks. All images in the album were made on Mage.Space and I got tons more. Why are you trying to get peopel off of it? It's the website that got me the best results for celebs by far. MidJourney is not even close in that regard. And yes I know it uses Diffusion, but the "models" they have there to pick from help a lot. I use the "Realistic Vision 3-5" model for celebs. If I try making anything with the default SD model, it produces complete garbage. Not sure why but it does. P.S - Wait what...? you have a $6000 GPU...?  I'm still rocking a GTX 1070.  P.S 2 - MidJourney face swap FTW. Now it's Bella Thorne!  |
|
|
|
Post by Queens of Absolution on Aug 31, 2023 2:09:59 GMT -6
I don't think this looks anything like Bella Thorne lol, maybe 20%. I've tried making Bella in every AI I could get my hands on and it never came out accurate enough. :\ For an AI like Stable Diffusion or MidJourney to create a representation without a specific dataset in relation to whom you want to create, it's like I said: the AI is "drawing it" looking through a camera lens into a mirror that reflects another mirror. It won't be perfect for anyone doing that. But she can be made to be passable - at least if you trace Bella's looks from her starting days to today without the dataset. Thanks. All images in the album were made on Mage.Space and I got tons more. Why are you trying to get peopel off of it? It's the website that got me the best results for celebs by far. MidJourney is not even close in that regard. And yes I know it uses Diffusion, but the "models" they have there to pick from help a lot. I use the "Realistic Vision 3-5" model for celebs. If I try making anything with the default SD model, it produces complete garbage. Not sure why but it does. P.S - Wait what...? you have a $6000 GPU...? I'm still rocking a GTX 1070. P.S 2 - MidJourney face swap FTW. Now it's Bella Thorne! I've just had bad experiences with Mage.Space compared to virtually any other AI site. But it was my first too before I moved to MidJourney and now onto more advanced AI usage. MidJourney is good for an 'artistic' impression. Stable Diffusion sites are good for a more exacting approach. I agree, I see the models there and I tried one of the Reality ones as I couldn't find your exact one and it made Halston pretty decently although not without caveats. To the first "P.S." - I have a Titan Xp and I bought an A100 from Amazon. The A100 aren't designed for video output but AI workload. So, subtract that recent purchase, I'm in the same boat as you when it comes to GPUs. I just don't see or feel the need to upgrade the Titan Xp - it does everything graphically admirally. For the second "P.S." - Wow, that's a useful feature; I will have to look into it. |
|
Anyway, here's my take on the subject:
I'm still rocking a GTX 1070. 