Hibrid Illustrious-SDXL- This version could work with some of my loras and some NOT.
V0.97 - New version, I think it's more realistic and now can create Asian women more easily.
V0.95 This version works better with my LoRAs and it's the Recommended.
Commissions Open.
Description
FAQ
Comments (31)
I am having trouble generating with the latest version, I want to compile a summary of your configuration:
- VAE: Illustrious.bin?
- Sampler: DPM++ 2M SDE Karras?
- Upscaler: 4x Siax 200k?
- Key tags: smile, film photography or smile film photography?
Hello I just re-uploaded it, it should work fine now.
Not sure if it's just me using an older forge version or something.
But the v0.99 version is outputting washed out images with white spots on it for me.
v0.95 worked just fine for me so thats odd.
Hello I just re-uploaded it, it should work fine now.
big +1 for training on a good dataset of random girls and not just (no offense intended) asian dolls or really uncanny plastic surgery victims lol.
Sad it does heavily train over what illustrious is good at; fictional characters. I can't even get the popular characters. I hope someone will outdo Pony Realism eventually..
most realistic illustrious real checkpoint
excuse me i don't know about Illustrious ? what can Illustrious do when added to a realistic sdxl ? benefits ?
Illustrious should be more "versatile" than SDXL in some aspects like nsfw things.
@Hikarias anyway, my opinion about sdxl when it comes to generate persons pictures is it's good at making a persons standing like status even other than front view is a challenge for it , it su-ks at everything else. people just add some porn poses at their merges , but sdxl is very limited , may illu or others can help , but base still limited . idk why people not seeing this as an issue . it only good at simple shape,poses , nothing else (natural like or realistic).
@amazingbeauty Of course everyone sees that as a problem, it's nothing new. It's just that there are many limitations with the current models, I mean sdxl since it was released in 2023, has improved a lot and has "moved forward" compared to what it was back then.
There is a point where you can't keep improving SDXL models and maybe we are close to the best that can be done with SDXL, but maybe not, we can't be 100% sure. Maybe some new trick or "improvement" will come out that will take SDXL another leap forward.
Illustrious is based on SDXL, but with a different training dataset and a few more tweaks.
That is why it has some improvements in certain aspects but also certain deficiencies in others such as "Realism" compared to SDXL-based models.
Flux is very good at prompt understanding, but unfortunately it is a distilled and censored model in many aspects. There have also been attempts to improve it and I hope it continues to do so.
Let's hope we get something better in the coming years.
Outstanding model! Been using the 0.95 lora compatible one. Please keep working on it. Again, great model.
Thank you!
@Hikarias Dumb question, what VAE do I use with it? Been getting weird finicky performance from auto1111 all day.
@olternaut It should be baked, and it's just the sdxl vae.
auto1111 hasn't been updated in 7 months. You could try using Forge.
@Hikarias Oh, way ahead of you there. I was hoping Forge would solve the burnt images issue. But it didn't. I think it's the loras I'm trying to use....Desi_Tabela, it's for Indian body types (pony), and also another lora a skin color slider. Hoping the author will come out with a illustrious lora specifically.
@olternaut Yeah this is a Ilustrious model , not a Pony model, So pony LoRAs would not work.
@Hikarias I suspect that by attempting to force the model to use pony loras, repeated attempts might actually corrupt the model file. Because I'm trying to use it now without loras and it's still burning images. I'm going to confirm by deleting the file and re-downloading it.
So I've learned the hard way to not use Pony loras with illustrious checkpoints. Both the checkpoint and the pony loras get corrupted. Furthermore, if I attempt to use the corrupted pony loras with pony checkpoints, then those pony checkpoints get corrupted too. I had to delete and re-download a lot of files. Lesson learned.
I dunno man, just to make sure everything is okay, can you check to see if your models are okay? Something may be corrupt with them. Just to be on the safe side I mean.
@olternaut Sorry to hear it, But I have re-downloaded it, Tested and it seems that there is no problem, I tried it on two different PCs.
And it seems that no one else apart from you has the problem you mention.
Could you tell me all the settings you use?
@Hikarias Well at the moment, I re-downloaded 0.95 and am running it with no loras, euler a, karras, 60 steps, CFG of 5, and since I'm running forge now I've left the VAE drop down box blank since there is no "automatic" setting. I assume the VAE is baked in? Anyways, still getting some burn marks on faces with weird artifacts. At least the entire screen isn't burning anymore. So I tried again selecting the sdxlVAE_sdxlVAE and still getting burn marks. Varying the CFG to higher or lower settings doesn't seem to do anything either. Also, I'm running python: 3.10.6 and torch: 2.1.2+cu121 .
Changing prompts seems to have helped. Running on a beach seems to cause burn marks on faces. When people stand still then it's fine. lol this AI model is finicky.
@olternaut I have tried your same settings and I still not getting any weird things. Show me the image to see what kind of artifact are you getting.
@Hikarias I solved it basically. Apologies, I just had to re-download the models again that's all. You've made a great model that can do different ethnicities without the need for loras.
Does anybody have any idea why some image generations start to burn even if you are trying to regenerate a successful image using all the same parameters and seeds?
got the same problem
@konygebony I can't remember who, but someone suggested that if you are changing checkpoints and loras around, the UI may not actually be unloading the old models. It's suggested when that happens you have to reset the entire WebUI. I haven't tested that yet though.
Best Realistic model I have seen! Thanks!
I really like the realism, they look like real pictures. Wind if you know a better model then by all means tell us.
In love with your model. Absolutely in love with it. It's finicky though. It's too sensitive to prompting and one wrong word can burn an image or create artifacts. But you really have something here and your hard work shows. You must of had a great dataset to work with. Looking forward to version 0.98. lol
great model, if you're having issues with images frying i'd try switching to "DPM++ SDE", I'm using a1111 and it defaults to "DPM++ 2M" but noticed the sample images use SDE and got a lot better (and consistent) results. great job!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.