I'm having similar issues. Many prompts I've locked in from v34 look a lot worse for wear in various areas, not merely eyes. Of course I don't expect them to be identical, but the results are often more noisy, less crisp, and blurrier.

For me I couldn't get a single presentable image out of all of the experiments I ran, mainly because of the eyes being messed up. Many of them looked weird and bloodshot or like they'be been gouged out of the sockets. Though to be honest even aside from the eyes the results feel underwhelming, unlike like my previous experiences where it felt almost like magic. I feel like something got added to the mix at some point which is messing things up, maybe around the time where every female character I generated suddenly started looking like a chibi anime character, Some point after 3.2 I'd say.

@tombot What artist tag do you use? Maybe it's a problem of the artist tag. If you do not enter an artist tag, you will get a picture with a random art style. The art style with bloodshot eyes is the style of artist 'dagasi', and since it accounts for a significant portion of the Kemono artist tags, this art style appears often.

Eh. It also seemed to me that large details (Scene, style, position of bodies) began to be generated better, but small things (eyes, vaginas, fingers) were worse.

@SuinSuin I'm generally of the opinion that a stable checkpoint should function on it's own without overly specific prompting, those weird bear uploads I make are just me testing what the default style of a checkpoint is. Also I think it's important that the style works at it's base because everything we make will be based upon that to some degree.

@tombot This model has many differences from the NAI model, so I don't think it has its own unique style. Like that the dataset is based on E621 and uses Clip skip 1. So I think it is necessary to fix the style with an artist tag. Of course, this is my subjective opinion and may be wrong.

are u changing the advanced settings on the engine drop down TensorRT to set those resolutions, and they don't cause artifacting?

@TheCyanFox  you have to generate an engine for each specific resolution so artifacting doesn't occur, it actually converts the model into ONNX but I haven't noticed anything different quality wise.

Man this might bring me back to A1111 from ComfyUI... Edit: Wow. It's a bit finnicky to get going but once you do it's off to the races. Holy cow.

Can someone tell me what launch options I have to disable to use TensorRT? I'm usually running with --xformers and --medvram, but that seems to cause conflicts now: i cant even generate the "engine", just getting errors.

medvram has issues with it at the moment @rantas

I wouldn't count on it. There is little interest in finetuning SDXL from furry model makers. Biggest problem is the cost. XL requires a lot of vram to finetune which means it's expensive and/or slow. Beyond that, XL itself is simply a pain to work with. There are 2 text encoders to worry about and the base SDXL model has offset noise in it. I've seen more interest in aligning a better text encoder to 1.5 than I've seen wanting to finetune XL.

Thanks for the info mate! I'm not so into the technical side of SD so I have always believed that newer(XL) = better without worring of these other aspect...
Well this means I will continue with 1.5 for yiff works!

@Mobbun Fine-tuning 1.5 is ideal today, but hardware to actually fine tune XL is a lot cheaper than you'd think if you buy the hardware yourself today. It's only going to get cheaper as time goes on, so there's a high probability of this happening in the future with the right collaborations.

The majority of the problems are based on ventilation and space for me, but that'll change soon enough.

@AbstractPhila I did say expensive and/or SLOW. There are finetunes off XL, but those use a smaller dataset and are thus much faster to train. Some don't even do a proper finetune and just merge a lora into the base checkpoint. The top furry model makers are using somewhere around 2.5 to 3.0 million images for their model. Even if they did attempt to finetune with just 24gb, it would take a long time for the model to be properly trained. There WAS interest in the beginning from the creator of Fluffyrock. They even designed a script for sharding the model(because XL takes so much vram), but XL proved too much of a pain to train so that's shelved for now. For other furry model makers, 24gb of vram acts as a limit for finetuning. Unless you have money for an a100, either from cloud services or the hardware itself, then you aren't going to make a big finetune off of XL. I've talked to the model makers, they've even ran the numbers, it's just too expensive/slow.

I do, but this has happened before. Feel free to stick with whatever version works better and try the next one if and when it comes out.

@tombot right that's fine then, i just wanted to make sure i wasn't the only one

if you have a 6GB video card then just use automatic1111 for example

A1111 on a google collab.

Thanks! I'll have to look into how to setup "automatic1111" for myself, since I'm on linux XD

Personally, I just use them as examples of styles and characters for prompts.

Wildcards can be used with the DynamicPrompt extension to generate random results.

Yeah it's just a cheatsheet to let you know what styles were trained to the model so you know what prompts have weight to them. Removes the guess work.

thank you so much

Probably due to your prompting, what prompts are you using to generate fur/scales and are you using any prompt that misguides the model to make skin. Reply with your positive and negative prompts.

@AlderAnalysisLabs ==(by taran fiddler), (by dagasi), (by iskra), (by Asaneman),

(masterpiece, 8k, 4k, hi res, high resolution, high details, absurd res), (best quality, high quality:1.4),

detailed background, hospital,

looking at viewer, facing viewer,

perfect anatomy,

(standing),

solo, (female, fox, (white fur:1.5)), very tall, tall, detailed face, detailed fur, (black lips, thick lips), (large tail, (multi tail:1.2), 9 tails), big ears, mane,

(perfect eyes, long eyelashes, thick eyelashes, pink eyes, detailed eyes),

((long fluffy hair:1.2), pink hair),

(gold dangle earrings, necklace, jewelry),(clothes, clothing, (white bodysuit:1.5), (long lab coat:1.5), shoes, thigh-high socks, black bra)==

==low res, lowres, blurry, bad anatomy, letterbox, deformity, mutilated, malformed, (worst quality, low quality, normal quality:1.4),

amputee, watermark, signature, unusual anatomy, username, sketch, monochrome, multiple views,

easynegative,

boring_e621_v4==

@LukeVale Could be the fact that you never actually say "female fox", "female anthro" or "female fox" therefore it may sometimes think that you want a female character anthro or human with ex. a white fox tattoo. I am unsure but the unclarity could be part of it. Try "(solo, fox girl, (white fur:1.5), tall, big fox ears, etc.)" as the character part of the prompt.

@AlderAnalysisLabs even with that it still generates meaty tanned like human hands. I can only assume it might be related to the artists, but the artists doesn't draw any humans.

@LukeVale I unfortunately don´t have an easy solution for you then, but i would recommend checking out this guide on prompting as it can really help in general and may help with the problem https://aituts.com/stable-diffusion-prompts/

@AlderAnalysisLabs thank you for the input either way! It's something that's not too big of a deal as is, since it's not every time, and it's not a deal breaker.

Do you frequently use those particular artists in your prompts? Some artists or combinations of artists will spit out very human details, while others will go in the other direction and generate things borderline feral.

The somewhat confusing thing, is despite how an individual artist draws, how the model actually perceives them might be totally different based on some unknowable factors hidden in the model itself. This is especially so when you mix artists together, it can connect some dots in the model that ends up looking human. I would try shifting the artists around, add or remove one, even just changing the order you invoke them can have a result.

"multicolor fur" is a powerful tag, and it'll do most of the legwork on it's own. If you go "(multicolor fur, grey fur:1.2)" it should pretty much by default give it a grey fur coat and a lighter (often white) belly.

@sansmia holy- THANK YOU! That'll be such a great tool!

tag "countershading" means specifically lighter color belly in animals. so adding it to "multicolor fur" will help.

Why is 22 the best?

I've noticed a trend, especially as I've spent a lot of time in 3.0+, it's getting harder to get 2d cartoony art out of the models as more realistic models and images are put in as a base.

These are the features of the operation of neural networks.

Imagine that she is standing in front of two roads. She will go on the left with a probability of 95%, and on the right with a probability of 5%.

Although the chance to go right is small, it is there.

Largely because you need to strongly define a style for it to know which of the two options you would prefer.

The v35 model seems to be temperamental with artist tags (with me at least), I've liked 34 much more (others will say 22 is best, idk). You'll just type the tag for them like "By Braeburned". There's a lot of artists, but some one's I'd expected to be included weren't. A list of them and demo images is here https://drive.google.com/drive/folders/14KBOBP1TVNY1DIvbzw2ljAM-pjEpsz3l

idk bro im drunked asf but you can try search in e621 wiki for a tag that match what you want (like convenient censorship), if you dont find it you should try a lora (i dont know if there is one)

@robloxsexman Hey, this is good and absolutely correct prompt, that does not work in Stable Diffusion at all. Thank you for help, anyway!

@solvenn7796 np man

I'm with @robloxsexman on this, 'convenient censorship' is the tag I've always seen for that, though if it's not well enough trained may not get the result that you want.

@PsySpy I solved it with "bubblegum and sticky tape". Cut and pasted object from another picture. Stupid, but it works.