The v4 is the BEST model I've ever used. It starts to look a bit like NAI for me.

GREAT work!

I tried to use some LoRA's but it generates pics very slow. Does anyone know how to fix it?

My setup: RTX 3080 12 Gb VRAM, 16 Gb RAM

This model is simply extraordinary. This is next level compared to pony XL based models.

While it cannot exactly reach the highest possible render quality of pony or other models, it's extremely fast and exceptionnally good at following prompts, imitating specific art styles and overall it's way less difficult to prompt than pony.

Truely incredible work!

Update: If you have it, the best results im getting with the sampler euler a cfg++, scheduler beta, 12 steps, 1.0-1.25cfg, 1024x1280.

Yeah, this is crazy. Im trying around with NoobAI v-pred 1.0 since it came out now, and getting somewhat close to making actual images now with it (sometimes), with a lot of fiddling around, and then someone told me about this.

Right away i can say, many of the points i thought exclusive strengths of noob, this model can do too, sometimes close enough... and sometimes even better.

Prompt understanding: Weaker then noob v-pred... but sometimes stronger then noob eps from my tests. Despite being only able to use tags, if you know that and work right with them, you can make MUCH more complicated scenes then you first think.

Multiple characters: Not as exact as in v-pred, but better then any pony, and even most illustrious, models. I dont always get exactly what i wanted, but often it gets close enough. If you can live with the model being a bit creative with the characters, you can get pretty good scenes. Anatomy errors are kind of on par with IL models, much worse then in noob v-pred, but much better then in any pony model. I got lucky more often then not, which is usable enough in my book.

Artist styles: Biggest plus of this model for my by far. Just as both noob versions, this can use EVERY single artist style ive tried, perfectly, ESPECIALLY FURRY, and can even mix them extremely well. And i really mean mix them, not use one sometimes, and another one sometimes. This point makes this far far beyond anything pony could hope to achieve.

Ease of use: Biggest reason for this to be a viable option to noob. With noob v-pred, you have to prompt almost EVERY single thing you want in the image, and describe the scene, lighting and details exactly, it has some creativity, but almost never enough to count on. This model has a big part of the creativity from some IL finetunes, where with just a few prompts, you could get a full scene with many details. Backgrounds can be a bit worse then IL, but so far much better then noob, even eps.

And now for the reason why the average person might prefer this to noob for quite some time to come: You set the settings it tells you, and- Todd Howard kicks down door "IT JUST WORKS!" In noob, you have to fiddle with extensions you didnt even knew existed, cfg rescale, latent modifier, thresholding, use ui's that even support v-pred, try around the cfg and steps for every single new idea you wanna try, etc.

Ive said something like this in another comment already: Noob v-pred is an unpolished diamond... but almost noone knows how to polish a diamond. This one is a already polished quartz, which might not be as beautiful, but close enough for most average people to not see much difference on the surface.

Heyo, so because i REALLY love and care about this model, i have to talk about some negatives and ideas for improvement now, especially after trying a very similar model in the meantime.

But first again, this model has almost all strengths of noob, but can actually make good pictures easily, this kind of noob-mixes are probably THE future of 2-2.5d models rn, getting ever closer to perfect.

I have now tested this model and another one a lot, which might be one of the, if not the, closest models to this one ive found so far.

https://civitai.com/models/946205/konpaevo-mix-noobai-eps

And i have to admit... it seems to do MOST things better then chroma, but not all. Chroma still does 2 major things the best imo: Pure 2d, while perfectly adapting an artist style, especially furry. (It can help a lot to put up to 1.5 weight on a name to really get a pure style, but this model can do it.) And the other thing is extremely close-up details, most noticeable with "portrait" and, well, "close-up". The details (at least for furry and scalie characters) i get with this are still barely unmatched even with konpa.

But now for what i think is the main problem for this model: The few steps needed. Sure, this increases gen time and works very well actually, however after testing a lot now i think i can say that this is the main reason anatomy can get screwed, and background diluted quite a lot of times. For complicated concepts, which this model can DO, it seems like a 50/50 if the image is usable just from the anatomy, which reminds me of using SD1.5 and having to hope for rng.

Testing the konpa model ive realized, the only reason it seems better in many cases is, first because its mainly trained on IL and just merged with noob, still giving it a lot of strengths, but also much stronger base style, which prevents it from fully using artist styles like this model. But IL also gives it better background, lighting and slight realism, which isnt always a good thing if you like me, want pure 2d sometimes. And the second reason is, because it uses (by default) 24 steps, which also isnt that much, but double as much as this one (by default). And because i have the image preview while generating set to every step, i can see how the image changes with each one. Which shows me that the main reason konpa does anatomy and most other things better (most times, it still screws up pretty often too), is just because of the more steps, it just has more time to think and correct itself. At around half of the steps, i can see that it also makes the same mistakes as chroma (and most other models), but then has more time to correct itself, becoming more consistent overall.

Long story short, if this would get a version that just works with more steps, i would still use this over even konpa for most things, even just because of the better style adherence. That said, i dont know if around 24 steps would be enough either, since thats also the reason konpa still screws up sometimes, its still not a lot of steps. Obviously, using either model with just more steps, isnt a good idea, because theyre not made for that, and it will ruin images.

Btw yes, i have tried both models with turbo loras like lightning, PCM and DMD2. While they can work pretty well actually, they so far always proved the point to me that less steps equal worse anatomy. AT LEAST for furry art, which does not have as clear-cut anatomy as humans, models that are just for making waifus which 99% of SD users are interested in, will probably have less problems with fewer steps.

So yeah, i think the biggest possible improvement for this model would be a version specifically build around more steps, imma just throw out the number 36, just because 24 seems still not quite enough for at least other models, but thats mostly just an uneducated guess. It doesnt even really have to be trained more ore differently imo... the pure knowledge and capability of this model is already so advanced, that idk how much it could, or needs to be, improved by that rn. (Probably v-pred would be the next biggest step for that.)

Lastly i wanna say, im just someone who tries very complex scenes using just a model, just to test and push it. And this model so far, was able to actually do most of them, with multiple characters. And i can tell, if it had more steps, it could have actually finished most of them right. Right now it just has to stop at a point too early to get there often.

Thanks again for this model, im not smart enough to work on things like this myself, all i can do right now is share experience from tests, to try and give infos to the people that actually can improve things. :'D

A nice improvement when coming from an older SD 1.5 model. More robust with multiple characters, anatomy and poses. Can do 1536x1024 out of the box. Also, very fast! But quality and style is just mediocre. There are much better looking models like the Nova Furry/Animal/Anime ones.

Here are my issues:

- Rather mediocre looking stock style, some characters like Simba or Toothless look off

- Using artist styles doesn't impact results as much as hoped

- Eyes mostly look bad, though results vary per character and distance

- Cities/buildings look pretty garbled/blurry and in general very monotone/grey

- Characters and objects in mid and far distance look garbled/blurry very quickly

- Often adds too many toes

I moved on to models like Nova Furry XL, that give higher quality results (but need more steps). Maybe with a future update, we can get cleaner results at the cost of (slightly) more steps? I think 12 is just too low. I also have to keep the CFG lower than 2.5 otherwise my results look messy. Tested on both A1111 and Forge.

Any way to reduce the grainy/noise?

Even though I'm not a fan of furry - this model is something else. with my potato the biggest advantage is the minimum number of steps and a noob as a base. I want to thank author from the bottom of my heart and really hope for future updates and improvements.

Beautiful just beautiful.

The model is great! However, when I use this model to train LoRA, the output results are very blurry. The material I’m using is 1024x1024.

This is easily a worthy competitor to NoobAI based versions of Indigo Furry Mix XL, I say this because while that model needs at least 25 steps to generate an excellent image, with this one is possible to generate an image with just 10 steps.

CompassMix and BananaStrike were kick ass model merges. Love em.

But I gotta know...

What are the original models merged into any of these versions? I'm trying to figure out the source model causing all the artifacts and text/logo issues.

All other people's merges with these issues seems to traced down the line to this model.

Consider, different models to merge and/or merge types. This model has a lot of potential structurally but is suffering during the first few IN blocks and the very last few OUT blocks of the unet.

Id love to help if needed.

I assume you've used a LCM/Turbo/Hyper/Lightning LoRA while the elaboration of this model.

Which was that lora?

This model looks promising, only one slight issue; I’m on mobile, and for some reason I can’t use it, is this an issue on my end, or a device specific thing, thanks if anyone responds.

Apparently the major part of the users of this site are allergic to use models in the CivitAI's generator that are able to generate an image in less than 20 steps.

I say this because because is unbelievable the fact WAI-CheapFast-ANI-NoobAI and this model aren't winning in the auctions for the ability of being used in the Image generator of this site.

sdxl lightning lora doesn't work with this (the images come out blurry or messed up), but it generates images in low steps (~12?) without it just fine.

Anything for the actual humans?

I just want to say that since January, this is still one my top go-to models for SD. So incredibly good =D

It's a great model, but it takes a lot of time to learn how to use it. I'm so used to the features of Illus / Pony that even the tags are used differently here.