Kinda curious, what's exactly the differences between normal SD with SDXL aside of well, XL producing bigger results?
"What does the Sui Sei?"IRyS
-
🔓 Momosuzu Nene has returned from hiatus and explained her house was broken into and vandalized. 🔓
🎤 vHoeShow has announced debut times for their new JP wave. 🎤
🥜 God's Strongest Orc Vtuber, Randon, has put out an urgent call for assistance. 🥜
🕎 Idol Corp's Yena Youngblood has managed to delete her own Youtube channel (it has been recovered). 🕎 -
I'm in the market for an alternative server host. Please see here for details.
AI Waifus and You 101
- Thread starter Watamate
- Start date
Man, I was ready to boast about the size of my own LORA folder, but it's only a third of yours. That's a lot of futas.So I stumbled on a OC character I liked... But for some fucking reason the person seemed to ruin it by making it a futa. So I tried (I think successfully) to recreate it. Obviously, I'm not going to post the original picture. But I do wonder, is there any vtuber out there with a similar vibe? Not really NSFW, but still.
View attachment 67676View attachment 67677View attachment 67678
It was also my first time using an SDXL-based checkpoint, and I do quite like it so far. Besides the fact that none of my old Loras are compatible and I have to start from scratch... But that's fine right? ...
View attachment 67675
Can't say any vtubers come to mind there. Not exactly to my tastes.
On that note, I have once again spent way too long making another LORA that ended up being somewhat usable, but far from perfect.
This one has multiple outfits that actually work this time.
Onolumi - 3 Outfits - Pony | Stable Diffusion LoRA | Civitai
NEW PonyXL Version: Better looking and much more consistent when prompting for different outfit variations. Does struggle somewhat with the face wh...
Again, Civitai auto-rated most of the images R for Ridiculously Haram because they dared show bare stomachs, so here's all the pictures:
Pretty much yeah. Checkpoints using Stable Diffusion 1.4 through 2.1 as basis are trained on images that are natively 512 x 512 (2.0, 2.1 768 x 768), whereas ones using Stable Diffusion XL are trained on images 1024 x 1024. This has a couple of advantages and disadvantages.
The main disadvantage is obviously higher hardware requirements, although this can be circumvented if you don't mind long render/generation times. And also Loras are to my knowledge not cross-compatible, but there are already a lot of Loras specifically made for SDXL and Pony (which is generally a better SDXL based Checkpoint for Anime related images).
Advantages are, it's better* at details that Stable Diffusion 1.4-2.1 struggles a lot with, like fingers and the like. You don't have to rely as much on upscalers because you can just render at higher resolutions instead**. Whereas 1.4-2.1 would just start doing very weird things if you natively generated larger images, like multiple bodies/heads/etc. Higher dynamic range as well. You're also less reliant on negative prompts, but you're still free to use them. And less reliant on prompts like "masterwork, high quality image, 8k resolution", etc.
*It's definitely not perfect and still struggles, just less.
** You still want to use an upscaler if you want really large images
I've been mainly using these two.
XLsusjelomix [Pony/AutismXL based] - v1.0 | Stable Diffusion Checkpoint | Civitai
XLsusjeloMix A model mix of the AustimXL confetti + Cream Cod's style by r3c If you want to support me buy me a coffee or ask for a comission I'm o...
AutismMix SDXL - AutismMix_pony | Stable Diffusion Checkpoint | Civitai
Mix of pony with some stuff. It's an attempt at making pony more predictable and less dependent on schizo negatives without removing its comprehens...
Man, I was ready to boast about the size of my own LORA folder, but it's only a third of yours. That's a lot of futas.
Can't say any vtubers come to mind there. Not exactly to my tastes.
On that note, I have once again spent way too long making another LORA that ended up being somewhat usable, but far from perfect.
This one has multiple outfits that actually work this time.
Onolumi - 3 Outfits - Pony | Stable Diffusion LoRA | Civitai
NEW PonyXL Version: Better looking and much more consistent when prompting for different outfit variations. Does struggle somewhat with the face wh...civitai.com
Again, Civitai auto-rated most of the images R for Ridiculously Haram because they dared show bare stomachs, so here's all the pictures:
No Futa, these are my largest folders. (I have Civitai Helper which greatly inflates the actual amount of files)
That's a damn accurate Onolumi Lora though.
Last edited:
Oh God, I just realized I haven't looked at the size of my folder in almost a year. I've just been slowly collecting checkpoints, so the fucker is bloated as fuck. When each checkpoint ends up just being anywhere between 2 GB to like 8 GB, it gets out of hand really fucking fast. There's only like 3 checkpoints I actually use regularly. Ended up deleting like 140 GB of shit I've never touched.
Yea.. I ended up buying a 500GB SSD just to hold Stable Diffusion files, and I've still ended up with ~150 GB of overflow on a spare drive now also dedicated to SD. I don't think I'm prepared for SD XL or any newer things yet...Oh God, I just realized I haven't looked at the size of my folder in almost a year. I've just been slowly collecting checkpoints, so the fucker is bloated as fuck. When each checkpoint ends up just being anywhere between 2 GB to like 8 GB, it gets out of hand really fucking fast. There's only like 3 checkpoints I actually use regularly. Ended up deleting like 140 GB of shit I've never touched.
If people want to try out XL/Pony. There's a new Checkpoint that caters to Hololive/Vtubers made by a /vt/ user. Haven't tried it yet, since it only came out about 30 minutes ago, but looks good from the images.
civitai.com
Ponymagine 9.1 - Gamma | Stable Diffusion Checkpoint | Civitai
Works with both Pony and Animagine LoRA And it retains most of the Animagine character and style knowledge Versions Gamma Enabled support for Pony ...
There have been a few times now that I've seen a LORA I wanted to try, only to find out it is SD XL, or whatever Pony is. I should take another look at it at some point.. I did try SD XL quote some time ago and found the first few results were no different (or maybe worse) to 1.5 so I never bothered looking further.If people want to try out XL/Pony. There's a new Checkpoint that caters to Hololive/Vtubers made by a /vt/ user. Haven't tried it yet, since it only came out about 30 minutes ago, but looks good from the images.
Ponymagine 9.1 - Gamma | Stable Diffusion Checkpoint | Civitai
Works with both Pony and Animagine LoRA And it retains most of the Animagine character and style knowledge Versions Gamma Enabled support for Pony ...
Ok, when I get home I guess. Who am I to refuse a face like that?
In the mean time, would you mind posting a couple of comparison images of the same prompt with an SD1.5 / XL checkpoint, maybe also with / without LORA?
Pony is pony diffusion, which is a SDXL finetune. Basically the issue was that despite XL being better than 1.5, hardly anyone ever made the jump because there was a fuck ton of content already made for 1.5 and the jump from 1.5 to XL wasn't that noticeable. However a bunch of ponyfag coomers were upset with how inefficient the current tech is and trained their shit off of millions of images (unsurprisingly a lot of it is porn).
The end result is that Pony Diffusion is significantly better at understanding naturally written language prompts, doesn't really need negative prompts, and doesn't rely on quality modifiers like "masterpiece" or "high quality" to produce good results. It's basically the most intuitive and consistent tech and to top it all off, it's the best thing for SFW and NSFW art. It's even trained off of a variety of characters and series, so it will better understand things like "Fubuki from Hololive" without really needing a specific lore (it helps but you don't need it). It also knows a lot of artist styles by name too. So now all the creators are either accommodating Pony by making pony versions of their work or are making the jump altogether. The only real issue is that everything works better with score prompts, so all your prompts have to start with "score_9, score_8_up, score_7_up".
People have been merging Pony diffusion with their own special type of checkpoint and one of the most prominent ones is Autism Mix SDXL.
I was under the impression it was a whole separate thing to Stable Diffusion, but it's just a regular old checkpoint?
Also, thanks for letting me know the lore behind it. I very naively thought that Pony was just the name they come up with for some reason and it had nothing to do with those ponies. How foolish of me.
Might take some getting used to, writing natural prompts, since I have only ever used booru tags in mine. Also the quality modifiers have never been an issue since I have always just had the same string of words applied to the start of ever prompt by default. Now that I think about it, it's been a long time since I ever thought about changing that up to see if I got better results...
You don't need to use natural prompts but I've noticed it helps in narrowing down the results. You can check out the images that people make in civitai to see what type of prompts they're using to make their stuff. But yeah, it's an incredibly effective checkpoint that people are using or merging with their checkpoints, which has really allowed SDXL to show how its better than SD1.5.
Going from PonyXL to 1.5 requires some rewriting of the prompt, along with longer gen times for PonyXL but I'll try to give a few examples. In full transparency, the 1.5 Checkpoint I'm using is a personal mix, and the PonyXL Checkpoint is AutismMix SDXL. No upscaling was done.
PonyXL, Botan Lora. Seed 2051917411 | PonyXL, No Lora. Seed 2051917411 |
1.5, Botan Lora. Seed 2051917411 | 1.5, No Lora. Seed 2051917411 |
PonyXL, Mio Lora. Seed 700742927 | PonyXL, No Lora. Seed 700742927 |
1.5, Mio Lora. Seed 700742927 | 1.5, No Lora. Seed 700742927 |
Tried to keep the seed the same for comparison's sake. The three-legged Mio can be easily avoided.
Well I can defiinitely see the improvement there. I have just spent the night downloading a few models and messing around with a more realistic one. Will probably be another day but I'll post my own comparison pics as well (can you believe I had also planned to do a Mio beach scene...)
One thing I noticed straight away though is my card is struggling to produce 1024 x 1024 images to the point where it's not feasible to use. Dropping it down to 1024 x 768 is more bareable, but I don't know if going below 1024 like that will impact the quality too much?
One thing I noticed straight away though is my card is struggling to produce 1024 x 1024 images to the point where it's not feasible to use. Dropping it down to 1024 x 768 is more bareable, but I don't know if going below 1024 like that will impact the quality too much?
I did try some smaller images and they do work, but you lose noticeable detail. Might be worth testing out how upscaling works with the smaller images though.
Unrelated picture.
Looks like most of the SDXLs and PonyDiff's LORAs and stuff is accessible on Civitai's on-site art generator. Even if they have to use credits per generation now than 'completely free' like they touted beforehand, it's still a nice thing for someone who just want to see how the LORAs work before they download it to their local machine or can't be arsed to set one up.
Shishiro Botan with Disco Elysium's portrait artstyle LORA. I tried it on Civitai's own AI art generator on their site since most of the required stuff (PonyDiffusion, the character lora and the artstyle lora) is usable in said on-site generator. Took some tries but it was pretty decent.
Might give other artstyles a try, probably on PixAi since unlike Civitai, it has hires.fix/face.fix which some of the 'artstyle' lora is required or needed for best results.
Shishiro Botan with Disco Elysium's portrait artstyle LORA. I tried it on Civitai's own AI art generator on their site since most of the required stuff (PonyDiffusion, the character lora and the artstyle lora) is usable in said on-site generator. Took some tries but it was pretty decent.
Might give other artstyles a try, probably on PixAi since unlike Civitai, it has hires.fix/face.fix which some of the 'artstyle' lora is required or needed for best results.
Oh hey! I downloaded those models t- I mean that's neat where did you find that haha?
Once again, I said I'd post some pictures of my findings a week ago, but like an autist I got carried away downloading every checkpoint and LORA in existence and comparing them to eachother, and trying to get the bad ones to look good, instead of just sticking with the good ones. (I even bought a new 1TB SSD to hold the models..)
What I've been finding is some of the models look terrible, ad when I try to recreate the images on Civitai, see that they also used a load of extra style LORAs on top of the model, which is just cheating if you ask me.
Maybe I'll post some nice pictures Tonight™
Yeah... I definitely spend more time downloading LORAs that I'll never use over actually genning pictures. I've resorted to moving my LORA folder to an external HDD, which I feel has a less noticeable effect than having your checkpoints on one.
Checkpoint wise I've honestly been mainly using Animagine and AutismMix, depending on the LORA compatibility. Which is still my main annoyance, since Loras trained on XL are a mixed bag with Pony ones like AutismMix, and Pony trained ones are a mixed bag with XL ones like Animagine. But more recently I've definitely been leaning more towards Animagine, since I've had a better experience with that so far for purely Vtuber-related things. One thing I only realized far too late is that Animagine still very much still uses the old style of prompting, which early on gave me a very warped perspective (IE. thinking it was shit).
Alright, so what I have found so far is Pony is very good at creating very nice images from almost no prompt, while stil getting what you want. You no longer need to have a mish-mash of words in the positive and negative (they're calling them "schizo prompts", which I find amusing). It's especially good at making the most un-seiso kind of pictures without needing additional LORAs for different positions and stuff. And of course, one of the main advantages is being able to start with a large image right off the bat without requring upscaling. Like SD1.5, you're going to get a much clearer image of a subject up close, than one further in the background.
Here are some comparision pictures to take a look at. Links to lots of very large images ahead. Small thumbnails to save the forum.
All comparision images were made using the same prompt and seed, only altering the first line of quiality tags when switching between Pony, XL and SD1.5
(Some pictures made use of the Regional Prompter extension. Speaking of which, I spent a whole night trying to get the FuwaMoco one to work naturally but always ended up with the male faces merging onto one of the girls, or having only one of them. In the end I cheated by prompting just the girls and table, and inpainting the people in the window. Pretty pleased with how it turned out. Also I cheated on the Ayame one by adding the tattoo in Photoshop. )
All that being said, don't let the benefits of PonyXL distract you from the fact that I made a LORA of a fucking monkey
Civitai
Here are some comparision pictures to take a look at. Links to lots of very large images ahead. Small thumbnails to save the forum.
All comparision images were made using the same prompt and seed, only altering the first line of quiality tags when switching between Pony, XL and SD1.5
Starting with PonyXL at 512x512, the results are awful, but that's to be expected as it is not supposed to be used at this low a resolution
Adding Hires Fix and ADetailer improves it a good amount
I never really used ADetailer before, but it looks like everyone was using it on their model examples and recommending it, so I think that's the way to go now.
Adding Hires Fix and ADetailer improves it a good amount
I never really used ADetailer before, but it looks like everyone was using it on their model examples and recommending it, so I think that's the way to go now.
At 1024x1024 We suddenly see a big improvement. I've also compared two different samplers here. With SD1.5 I had always used DPM++2M Karras, but it seems like these days everyone suggests Euler a on their Pony models, so that's what I've been using. They say Euler gives a bit more of a smooth anime look. Can you see it? I can't really say one is better than the other.
With Hires Fix & ADetailer
With Hires Fix & ADetailer
Going back to good old SD.15, you can still get very nice images. You might even say nicer than Pony. I do think the backgrounds especially are much better detailed.
SD1.5 actually came much closer to the picture I had in my head for this scene. It took many tries with Pony to get something close, but SD1.5 gave me something I liked straight away.
It really does require that Hires Fix and ADetailer
SD1.5 actually came much closer to the picture I had in my head for this scene. It took many tries with Pony to get something close, but SD1.5 gave me something I liked straight away.
It really does require that Hires Fix and ADetailer
Strange things start happening when using SD1.5 at a higher base resolution. Suddenly I'm getting two people when I asked for one, and sometimes extra limbs. This is why you need to start small and then upscale. Still, those are some nice scenes, right?
With Hires Fix & ADetailer
With Hires Fix & ADetailer
I never really used SDXL. As I mentioned before, I tried a couple of models when it was still new-ish, didn't like it for whatever reason, and went back to SD.15.
Here's one comparison
With Hires Fix & ADetailer
Here's one comparison
With Hires Fix & ADetailer
Pony is supposed to be good at producing known characters without the need for a LORA. I'd say it does a pretty good job - some models better than others, but I'd still always go for a LORA if one is available.
With Hires Fix & ADetailer
With Hires Fix & ADetailer
With Hires Fix & ADetailer
With Hires Fix & ADetailer
SD1.5 does not specialise in knowing characters without LORAs, but it does an OK job in this instance
With Hires Fix & ADetailer
With Hires Fix & ADetailer
(Some pictures made use of the Regional Prompter extension. Speaking of which, I spent a whole night trying to get the FuwaMoco one to work naturally but always ended up with the male faces merging onto one of the girls, or having only one of them. In the end I cheated by prompting just the girls and table, and inpainting the people in the window. Pretty pleased with how it turned out. Also I cheated on the Ayame one by adding the tattoo in Photoshop. )
All that being said, don't let the benefits of PonyXL distract you from the fact that I made a LORA of a fucking monkey
Civitai
Last edited:
Similar threads
