I have been using 1.5 based models almost exclusively so far. Turns out I need about as much VRAM for a 1024x1024 gen with this XL model as for a 1.5 hires fix from 640x640 to 1600x1600. That seems like a bit of a deal breaker. And I'm not sure 8GB of VRAM are even considered low-end.

i agree. pony xl is better

What's worse is that the vram usage is much higher than for 1.5 based yiffymix models. I can't generate as hi res as i used to. Think I won't be using this pony/XL-based branch at all.

same for me i keep trying the different settings but nothing seems to work

male genitalia as well, a former strong point of all recent yiffy models

Hi, did you manage to solve the problem?

@Vetras Good afternoon, no way switched back to the 44 model. Maybe my laptop is not suitable for sdxl model or something else. And so I have a laptop with gtx 1060 6 gigabytes and intel core i 7 8750H processor and 8 ram. I would really like to solve this problem and try the SDXL model.
(if there's anything wrong or unclear I've written. I just used a translator.)

@chilon249 Good afternoon, I will try your steps to correct the errors. Yes I am using Automatic1111 maybe the problem is because of that.

@chilon249 so I guess I'm gonna have to put in webui-forge. At my speed, it'll take a day to download, but okay.

@chilon249 ,No, the most interesting thing. Everything worked well on Automatic, but Forge...

Funny. The secondary download solved the problem

it's not vpred so no

@chilon249 Thank You

Can;t wait to see what you cook up with SDXL :)

Hello, im also AMD user and i can use many SDXL checkpoints out there, if you have the VRAM (I have 16gb in my 6800XT) you can use zluda with stable diffussion to continue making pictures with AI, there are plenty tutorials that explains how to make it works. hope it helps.

@jesussalasgil995  I watched some videos about zluda, but my GPU is not compatible with it. But I'm good for now. Maybe I'll play around with sd 1.5 sometimes. Even so, thanks for the help

What GPU do you have? I've managed to run many SDXL models when I had an RX 6600XT with 8GB of VRAM (on Linux), but it's not pretty, you need many optimization options enabled and inference time is really long on high-res.

Heavy user here- I have 24gb vram (nvidia 4090) and I am rendering these at 1024x1024, highres.fix 2x resolution.
E6 tags work fine! some small differences between v.43/44 but otherwise the same.

3060Ti 8GB, 32GB RAM, 20GB Virtual Memory on SSD

1024x1024 work fine, take ~20-30 sec, HR up to 2x without scripts take ~8-10min, 1.5x - ~2-4 min

hi, 4060Ti with 16Gb VRAM works really well, (a bit slow compared with 4070Ti or higher, but works :D), 1024x1024, 1152x896 or 1216 x 832, i've got a 3040x2080 without script, (1216 x 832 upscaled x2.5) , e621 tags are okay most of the times but it doesn't always know some characters, like Susie of Deltarune. (sorry for my bad english)

RX 5700 8GB vram here, runs fine at 1024x1024 on --medvram, at the start it took ~4 minutes with Zluda but after tweaking some I got it down to around 2 minutes 40 secs per image

As an Addendum to my thoughts on V50; it's best feature is the faces, with or without Adetailer, it creates spectacular faces, it's weakest point so far is male characters- I can't tell if it's artist dependent but it just doesn't like male parts a lot of the time

Config file need only for v4 models. For v5 config file not needed.

but i am quite confused what are the v4 or v5 configs i see no one explaining it in detail so i can understand, i feel like i am reading somthing without context

@champagne1 If you download a v4 model (that's v40-v44 models), you also need to download a small config file and place it if same folder with model). You can find config file for v4 models in download section for each model. If you download v5 model, config file not needed. You just place model file in folder and use it as usual.

config files are for checkpoints using v-pred only.

v50 doesn't use v-pred

ok so i dont need the config file, and i have not used one for v50sdxl then why dose my comfyUI Yiffymixv50SDXL setup not work, i followed the guidelines showen, and i cant make it work if my life dependent on it at lest so the photos make sens or follow prompts, will try again later to troubleshot it or if you got any advice i be really happy,

@chilon249 finely made it work with the workflow i know understand it better, thanks for all the help

I am also having problems... I am on A1111 (don't know if that one needs the yaml or not, but I also can't find it if it does x.x) and even though I follow all of the specified settings, I end up getting a fried image. I've copied some of the examples users have posted exactly, and got the exact same image -- just fried. I'm using the correct VAE as well... Any idea? A1111 webui is up to date, and I just downloaded v50SDXL.

@LionFuzz so i have found out what was wrong from chilon249 and hes workflow for ComfyUI, but i allso manged to complet som on SD-A1111, and as i understand it you are not supose to use the VAE, just use checkpoint and then the same for refiner i think that is how it works for the A1111 as well, but i am not sure still struggling with getting the ComfyUI to do what i want in prompts as well as A1111 YIffymixv50

I gotta say it still retains quality in a way I never would have expected its pretty damn good,

CompassMix XL lightning model was an interesting choice and it performs great.

I dont know SHIT about model merging, but yiffy + DucHaiten could have crazy good potential!

Every new model you try (or version thereof) takes some getting used to, as well as a bit of relearning at first.

why the caps ?

No such luck, sometime its two females, or two males. even with the use of BREAK tag

Only with controlnet or inpainting, unfortunately

@C1yde I also use Regional prompt, which I have 50/50 luck of using it.

@koopa990 I've tried that a bit as well but it's just way too finicky in even the best cases and completely breaks when using LORA

@C1yde maybe in the close, close future, the machine can be capable of genning multiple people without a sweat!!

@koopa990 Haha I feel your pain, I can get it to somewhat work in SD1.5 IndigoFurryMix when prompting chars like Krystal and FoxMcCloud, but even then I have to use undies instead of genitalia, otherwise it mixes them up in the most silly ways, lol. I hope AI results becomes better over the next 1-2 years without complex add-ons or hacks!

same as me, i copy all the same settings and prompts. it also happens on pony checkpoints as well

if i change the character, or a few keywords, the output are is completely different

Same here, I've tried to get the newest version working for me off and on for weeks now, but I just can't seem to make it create a halfway decent image. Been sticking with v44 as of late.

are you using a SDXL VAE? because if you are using a different VAE it might be frying your images

@chilon249 Thanks for your comment. I did follow the instructions on the page properly. VAE set to none (tried the SDXL version vae too) and samplers set correctly. I am running a1111 with the arguments --medvram-sdxl --xformers --no-half - However, this has not impacted any previous versions of the model, they all still work great. Haven't quite fixed it yet!

I have this problem too. Approximate 3 of a 10 images is humanlike creatures and not furries. Well, i think that is a problem with SDXL model from which v50 was created.

If you're using the 4.x series, you need to use cfg rescale. For A1111, download this extension https://github.com/Seshelle/CFG_Rescale_webui, for ComfyUI, just use the built-in RescaleCFG node.

@CosmicElement Thank you so much, this fixed the problem for me =]

you can find it here https://huggingface.co/hakurei/waifu-diffusion-v1-4/blob/main/vae/kl-f8-anime2.ckpt. An alternative VAE for the 1.5 models is Furception https://huggingface.co/RedRocket/furception_vae/blob/main/furception_vae_1-0.safetensors. For sdxl models use https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors. Most models should come with their own VAE so you can just set the VAE setting to none or automatic