Would appreciate comments, posts, reviews, and requests
if you'd like to pay for a commission or just send a dollar or two -> PAYPAL or Kofi
V2: Not sure how satisfied I am with it, hence the "alpha", anatomy seems worse and more hit or miss, I got it to look better but it wouldn't work with Loras as well and changed the style to an even more generic slop style so decided to not upload that version and work on this one some more.
my base prompts, feel free to just use your own and it should work.
Pos: masterpiece, best quality, newest
neg: lowres, worst quality, low quality, bad anatomy, bad proportions, ai-generated, old, oldest, simple background (add monochrome and greyscale to get more vibrancy)
dimensions I usually use: 1536x1536, 1280x1280, 1846x1280, 1280x1846
Cfg: 4-6
IllustriousV0.1 and Noob loras seem to work from my testing
V1:
Images were made using swarmui (only V1)
Illustrious-V1.0 merged with an unreleased Noob1.1 (eps) model of mine. (on the tensor page it states merges can be shared)
I've found that 1280x1280 is the best all-round resolution (if you want to go above 1024), 1536 works but depending on the prompt it might have a hard time focusing on the subject. 1024 and all the other resolutions work. Images after more testing and others using the model we've found that genning below 1536. unless doing 5:9/9:5 (I think 1216 x 1824 or vice versa) looks much worse, especially with loras.
LORA Training info ( do not train on this, train on base illustrious 1.0)
On the note of loras TRAIN THEM ON 1536 THEY TURN OUT MUCH BETTER THAN TRAINING ON 1024, ALSO RAISE EPOCHS OR USE A HIGHER LERNING RATE (we mainly use 1 repeat, so 20-25 epochs on illustriousV1.0 for characters (unless you raise learning rate instead, in that case lower epochs, 5-10 I believe are the goal, though I myself choose the higher epochs. The comparsons look more or less the same so it's preference) and styles seem to work fine at the normal epoch counts 15-50
Natural language does work and can be used in conjunction with tags.
I use euler A and Align Your Steps and my prompts are the same as other models, so your usual positive and negative should work fine.
works with noob1.1 and illustrious0.1loras
it has a tendency to generate greyscale or monochrome images if at least 1 color isn't prompted, even if said tags are in the negative
Positive: masterpiece, best quality, (optional: very aesthetict newest,)neg: lowres, worst quality, low quality, bad anatomy, bad proportions, fused fingers, missing fingers, watermark, username, simple background, censored, unclear eyes, blurry eyes, ai-generated, lowres, jpeg artifacts, white background, old, oldest, Description
FAQ
Comments (52)
We sailing the high seas with this one boys
10$ for a downloading a finetune of a model, madness.
while I understand the importance of paying those who put work into this, this is not really the way in the current marked...
first case of ai piracy lmao
@EBIX Technically we were pirating ever since 1.5 model, after NAI model got leaked
You will download the car!
@JohnnyB1 no problem with paying but they promised open source and the model isnt even on noob's level and noob is free...
Greatly Merged. Your "merging" skill is Impressive!!
Based Open Source enjoyer
Damn, you started a revolution xD
I knew this was gonna happen as soon as I saw them allowing merges, but I didn't expect same-day-delivery. Thank you, good sir!
Respect to you who sacrificed 10 dollars!
Truly a hearbreaking moment... knowing he sacrificed all those potential candies to give us a chance to produce Open-Source art with this Sniff Godspeed.
new zukiCuteILL waiting list here
Shut up and take my buzz, you magnificent bastard!
Is it possible to create "realistic" images from this model?
from the little bit i tested I don't think it looks good
@bloodsplash btw, seems like eyes have some distortion and deformities, do you think you could fix in a uptade?and do you recommend any upscaler?
@P_Universe I just use 4xultrasharp, as for the eyes I can try.
nova Animals has IL 1.0 and makes realish images.
@bloodsplash thanks!
Thank you for sharing and supporting original creators of Illustrious 1.0.
Always the best
good understanding
very fast
very good results
This is the new Standard for Anime pictures
Recommended prompt:
(anime coloring:1.2), saturated, Moody Lighting, intense (HDR) (intricate details) newest, masterpiece, best quality, absurdres, highres, (cel shading:1.2), 4k resolution
do you have any evidence that intense, (HDR), (intricate details), and 4k resolution do anything besides introduce token bloat? have you determined without a shadow of a doubt that they meaningfully impact generations, and you're not just falling for placebo and confirmation bias?
@chuunikaede Im not studying stable diffusion. If you dont like the results feel free to change it and move on bro
@Satsuma9977 it's not a matter of "liking" the results, you're recommending a prompt which has multiple bogus prompts in it.
Wow, Illustrious v1's natural language understanding seems quite a thing. By far the best in SDXL based anime models, maybe?
And this merge, seems to make it quite more stable. I love this one.
Posted some XYZ grid images of natural language prompts.
This is actually the first anime model that is as good as your V3 model or maybe even better.
Can't find your post,where is it to see your results?
I'd love to see if there's anything you can do with it, i like the style of this new model but there are a few things i've noticed yours is better at than it is
Funnily enough, IL 1.0 has no natural language training, at all (see page 8) https://arxiv.org/pdf/2409.19946
I have tested a bit and it does seem to work to a degree, but... just as much (or worse) then some IL 0.1 finetunes before, probably mainly just understanding the tags used in natural language.
@Bigboyblaziken might be, it'd be a pretty stupid thing to market if it's easily disproven though lol. I do think it's probably just the better tag understanding as beyond simple things like girl running at viewer it doesn't work great
@bloodsplash "A pretty stupid thing" describes the entire IL 1.0 release lmao. :'D But yeah, actually many more things they claim right on their civitai site, is misinformation.
https://civitai.com/models/1232765/illustrious-xl-10
"First-of-its-kind Stable Diffusion XL model to natively support 1536px resolution."
Since you can only use this model (at least on civitai, and this is the description, from civitai, so only that matters), with on-site generation... which still does NOT offer this resolution... you actually cannot use the model like its supposed to, where youre supposed to. Biggest just straight up lie already.
"ideal for creators who want to apply their own fine-tuning (e.g. LoRA training) to achieve particular art styles or specialized outputs."
You cannot download this model (from this site, which this description is from, again only that matters), you cannot make loras with it. The second biggest false claim, right in their own official description.
And there are more, but i dont wanna get too deep into that here, if so id write tomorrow still. Just wanted to give some immediate examples that show, yes, they do just straight up tell stupid lies, and will face no repercussions for it, because truly... nothing ever happens. (Id write a whole ahh review on their page... but they have limited comments to only people who have access to the model on civitai, and i for sure wont pay for it to not be able to use it correctly on here, so... thats some nice censorship.)
Uh, back to the topic tho for at least one thing. Ye, its probably just the better tag understanding, but IL 0.1 had that already, so that part isnt new, but the "natural language" they claim for 1.0, seems to be just another lie. OR they lied in their paper... which wouldnt really make it better. But it for sure doesnt feel like this truly has some real NL training either way.
@Bigboyblaziken Noob also has better "natural language" understanding, but it's also shit, they didn't market it how Illu1.0 did. I haven't tried using natural language with Noob but i've heard it's more or less comparable to illustrious1.0. The thing that really gets me is the drip feeding method of release, V2,V3 and V3.5 are ready to release and are supposedly game changing models and yet they're keeping them locked away and plan to release them every few months (they're not even sure about the schedule). By the time those come out, we might already have a new Noob which will be superior. They claim to want to help the ai community but holding back models that would progress us isn't helping,
@bloodsplash Ye noob is much harder to use, and not as visually impressive most times, but some concepts i can feel just working better, and it has much less bias towards certain things then IL models. Which ofc its also based on, but changed so much that... its not IL anymore. Noob imo handles natural language just as much as even IL 0.1... very likely since its based on it. What i hope for is that either IL gets it together and makes their next versions downloadable again (wont happen), or Noob gets it together and makes their models much more easily usable, for better visual quality out the box (still hopeful). Or that another player enters town and just steamrolls everyone with a perfect new base model, which i wouldnt mind tbh.
@Bigboyblaziken Illustrious's devs on their site and have said that all future versions will be open source and downloadable
@bloodsplash Hmm, i guess well see where that is going... for some reason i cant quite bring myself to take them at their word. C:'
GOAT
based piracy
YARR FUFCKKCK HYEAAAH WE ARE THE PIRATES HARR YARRR🏴☠️🏴☠️🦜🦜🦜🦜🦜☠☠☠☠
Get cucked onomaai
This model is absolutely incredible and works perfectly well at 1280x1792. It would be great if you could do the same merge with each new version of Illustrious (1.1, 2.0, etc.) if they release them.
this one was just a freak miracle, but I'll try when those release
@bloodsplash do you think you could do a v2 version with a merge of vpred 1?
@P_Universe v-pred architecture is very weird to merge with other ones, even noob eps. I'll try but don't get your hopes up
@bloodsplash I know right,it's kinda unstable,maybe if you try for example noob pred cyber fix can help more
this model might be "pirated" and that's fine, but maybe it's a good idea to support it because this is really kind of sick
It's no worse than making money from a model trained from images they didn't have the rights to, imo.
free as it should be, i rather support you then scummy devs high on own farts after initial illustrious success.
Any plans to update for ILxl2.0? Your model is still my favorite IL merge.
Yeah, I've been out of the ai loop for a month or two, didn't even know it released
@bloodsplash Looking forward to it!
Details
Files
Available On (4 platforms)
Same model published on other platforms. May have additional downloads or version variants.