Alright! Let's also focus on lower memory requirements! No more 'out of memory' even with 8GB and some 12GB cards! In 1.5, I crafted hundreds of wonderful waifus, dreamable ones, using a 4GB card + ULTIMATE SD with tiles in IMG2IMG. It works flawlessly so far.

I get it, yes as a base model, it is truly miles ahead of 1.x/2.x, but right now, trained/mixed 1.x models are doing amazing things while XL base has this strong uncanny valley thing going on, and serious problems with anatomy. Not to mention the system requirements, we are years out from common folks having the setup to run this (especially if you want do other things on your machine while it runs). 1.x has a strong lead at the moment, and for many folks that may never change.

Remove/disable the prompt-blending script.

are those textencoders implicit to the model/architectture or can we use them at will? I am not sure i understand tbh?

Yeah, I have the same feeling. The Base and Refiner from StabilityAI is all well and good but unless the Community is going to get behind it in a very distinct way v1.5 will probably still be the better one in the end. But we can only wait and see and hopefully with some upgrades to --xformers and whatnot the hardware requirements can be pushed back a bit!

But like SD v2.0 and v2.1 have proven in the past newer doesn't always have to be better!

How high of a hardware bar is it to you? I've got friends making batches of 24 1k pics with a 12gb video card.

The original model would get NaN in fp16 and render black images. This simply fixes that problem so you can generate with fp16 instead of fp32

The original release had a new 1.0 VAE baked in which degraded the image quality. Stability have now re-released with the 0.9 VAE baked into the SDXL 1.0 model.

@theally Where do I find the new release? or is this it?

For those who still experience NaNs error with sdXL_v10VAEFix like me, please use sdxl-vae-fp16-fix

Yes, I also have such experience.

Plus no LoRAs and no good checkpoints. And 4x slower on my 3080. And there's no ControlNet, so I'm back to 1.5, personally.

use comfyui, i always used to use auto111 but just tried out comfyui with these new models and it works pretty well for my 6gb vram

This error happens because you're trying to use old LoRa.

You can't use LoRa that weren't trained on SDXL.

@zym0x what this guy said

@zym0x I get this error just trying to load the model, not even generating images yet. is there a way to get it to work with auto1111?

I think I had a similar error. Make sure you change the VAE in Settings to Automatic. 1.5 VAEs don't work.

I strongly suggest to make an another install of A1111 at another folder. Some extensions are causing issues too. Then test one after another.

I think it's a refiner, I'll test it and let you know.

@danielmnb12620 THX

Too little VRAM. You have four options: Use without refiner, wait until model is further optimized, use Google Colab (they might shut it down halfway though) or a GPU cloud like Runpod/Lambda/Vast.ai

i have the same issue i dont understand

"...will retry using slow method." - most likely it means that additional memory on the disk will be used, because there is not enough VRAM, so the swap file forms! Therefore the solution is:

1) to organize a swap file on disk. in win 10 - settings -> system -> about -> advanced system settings -> performance - settings -> advanced - change => restart

2) in automatic1111 settings -> optimizations ->Cross attention optimization = "sdp-no-mem - scaled dot product without memory efficient attention" or similar like xformers.

It worked for me. I have 3090ti (24GB VRAM) + 64GB RAM! But I also know that it works for a 12GB graphics card as well. Also remember to update the nvidia driver (optional).

@gromov you can run an even on 8gb For me what produced the error is that I left some uncompatible LORAs by mistake

update your webui it will work, here is the link on how to update webui https://aituts.com/update-stable-diffusion-webui/

in my experience it's not slower. it must have to do with RAM limitations! maybe you can get it to work faster with some tweaks or try it under comfyui, which needs less vram.

I'm on a 3070 and gen times are around 20 seconds, 512x512 on 1.5 was 5 seconds, so I figure 4x the pixels for 4x the time is okay

Try setting medvram, that brought mine down from 90 seconds to 30

Also make sure you have the fixed VAE, that also added a long wait

I agree. It's just about 4x slower than my 1.5 models on my 3080 FTW and I think the images are worse in general than any of the 1.5 base models I use, such as Cyberrealistic and Dreamshaper. It does make better images with simple prompts, but I don't need help with that, so...

It's most likely the vRam, I have a 4070 with 12gb and it's slower, but very workable. Probably triple the time for a 25 step image

Of course I also have the same problem with SDXL that doesn't work!!!...better model 1.5....I'm on a 3050 card.

SDXL model is crazy!!!

I'm on a 3060 12gb and it works fine in ComfyUI but dire in A1111

@hailyeuhlman684 I can't get it to load on 3070 with 8GB VRAM. Can you tell me what params are you using for A1111 ?

Me too. I tried Comfy UI and it's 10 times faster than Automatic 1111 with the same settings and software configs.

Tried A1111 and it is slow, consumes twelve gigs of VRAM and additional 8 gigs of ram and results are meh... Big dissapointment

you have --medvram in you web-user.bat? I made a copy with thar argument just for SDXL1.0 models.
I have a Laptop RTX3070 and it works fast and very well.

not only the generation takes longer and needs more Vram but also the training of models and the creation of Loras. with my somewhat older Tesla graphics card I need much longer because I have to use a lower batch size and that despite 24gb.

and I'm not just talking about a little longer here.

I'm having problems with Automatic1111 it says it will take 10-15 minutes. It works in Comfy UI then i can render in 25-60 sec

slow and returns terrible hands and blurried deformed faces in group photographs... no improvements

@thisisthewaytosa3076 Is it 3x better?

your first problem is 8gbs on a old GPU.

Hi Bilbo! We have a guide which explains all the basics for SDXL inference! https://civitai.notion.site/SDXL-1-0-Overview-be7abd647f8d4446a615d0afc6f04296?pvs=4 It should answer all your questions!

@theally A checkpoint model that needs a Guide and videos to explain "The basics"???

@theally You are spitting to the air! Walking to the dark side. Measure twice, cut once!

@theally Thanks! TLDR for anyone else: SDXL 1.0 VAE got botched, and the previously released SDXL. 0.9 VAE is the new defacto SDXL 1.0 VAE. VAE "Fixed" really means 0.9 is the official solution.

Lots of other good info in there though, especially if you are new to ComfyUI :)

@BilboTaggins Please explain why SDXL generate low quality blurry faces of 4 or more people in 1024x1024!!!!!

@StableDifference I'm not sure, I am able to generate pretty good images in all sorts of dimensions.

Hi there! We have a guide which might help - https://civitai.notion.site/SDXL-1-0-Overview-be7abd647f8d4446a615d0afc6f04296?pvs=4 - I also recommend downloading the V1.0 VAE Fix + V1.0 Refiner VAE Fix (you need them both - SDXL is a two-step/model process!)

@theally II m not sure i understand the VAE fix > what "fix" means? i don t see that in the vae files on Stability Huggin face.

Have you tried updating Automatic UI. SDXL requires the latest version of it to work. I had this issue too but no issues after updating it.

Use ComfyUI instead, it works even with 4gb vram.

make sure you have the right VAE selected

@orwelian84 - How do you do that?

@amitb - Is there a way to update automatically?

@thexiled check your settings page. there's something in there that allows you to toggle VAE. With enough diligence you will find it!

@PCBuster1970 Have you considered user error?

Are you using LORAs? You can't use LORAs based on old models in SDXL. Had a similar error and that was the issue.

@aigentina_art still same i cant do it with 4gb vram

Yeah you can't use SD Lora with SDXL. Nothing to do with model. I ran into it an dfigured it out. So now when i see i just removethe lora, or if you wanna keep it remame to non existing name with XXX appended or something.

@thexiled I use Github GUI so I just use Sync Command and then launch the tool again. You might have to run the Git Pull request from the same folder. Once done, just start the tool with regular Xformer switch etc.

Does this work with NMKD's GUI? Thats what I use as I find A1111 really finicky and a metric ton slower than NMKD's interface.

Could do with more info on this page TBH.

@Gluthoric Don't use settings to change VAE. This takes ages.

Add : sd_vae to Quicksettings list

This way you can change the VAE on the fly.

@azcharia comfyUI can't load?

Try removing any LORA or TI from your prompt

This Checkpoint is USELESS

@jmkiii It wont even load, I get the exact same error. Then it auto switches back to the previous checkpoint and the CMD window is completely full of errors. I followed some instructions, downloaded refiner, model and example LoRA. They are all in the correct directories.

@gr3yh4wk1 Did you ever solve this?

Comfy doesn't slow anything down, in fact it runs faster than a1111. The ui takes some getting used to, but if you actually sit down with it, you'll realized the workflow is embedded into every single image.

@thisisthewaytosa3076 Model stabilityai/stable-diffusion-xl-base-1.0 time out

@thisisthewaytosa3076 changing setting sd_vae to diffusion_pytorch_model.safetensors: RuntimeError

Traceback (most recent call last):

@StableDifference I get size mismatch for model.diffusion_model.output_blocks.8.0.skip_connection.bias: copying a param with shape torch.Size([320]) from checkpoint, the shape in current model is torch.Size([640]) that downloading the model, refiner and Lora.

Check the VAE, SDXL needs SDXL VAE

I have the same problem and VAE is in place

Make sure you use good negative prompts. Also, play with the number of steps. I find >35 makes it look very cartoony. 20-35 is usually good.

I'm getting great results with the refiner. The prompts react a bit differently than SD1.5 so try experimenting a bit. Reduce steps on refiner if it messes up the anatomy. Less is better. Then fix anomalies with inpainting. Here are my results for a similar prompt, just base + refiner, no inpainting.

https://imgur.com/a/y6UrpaN

All settings for sampler and steps are same as in this ComfyUI workflow. Use it as a base for great results on all projects.
https://comfyanonymous.github.io/ComfyUI_examples/sdxl/

The release notes say SDXL has trouble with placement of specific details, like putting a red box on top of a blue table. So additional workflows are needed. I think the most efficient workflow to reach your goal would be to first:

1. generate base dragon.

2. use controlnet to keep shape of wings in place.

3. use inpainting to generate celtic patterns on the wings using a mask.

Hmm. It seems I'm using too many steps for the refiner. I normally use 30 steps for generation and I was using the same for the refiner. I guess that's not how the refiner is supposed to be used.

@jesse244 i personally use 5 as it seems to be the sweet spot my, what i do is 23 steps dpm karas and 5 steps on refiner. this really mae the output really good for me

Isn't there a 15GB version out there? this is parsed

Thats not the standalone VAE, its the whole model with the VAE embedded into it.

@jordonyoung337 There is no 15GB Model

@JesseDinkelberg 0.9 one was close to that

stupid it is the full fp16 model along with the half precision vae.

it was mentioned fp32 which is wrong

SDXL is the BASE MODEL for future fine tuned Models like rev animated etc. therefore you compare it to other BASE MODELS like sd 1.5 or 2.0 what is your problem ?

Just wait till we can compare rev animated to something like a rev sdxl version. This shit will be wild.

just test it and it works with EasyDifussion

I think at one point you had to be on beta for XL but don't think that is the case anymore. Just make sure you are not in low vram.

anyway, stop working with the last update

@Nadoiz if you have discord, hit them up on their server https://discord.gg/cMnf5Rez
The Devs are very active with their community and will help sort out any issues.

same thing happened to me ,the commander shows out of the memory

https://github.com/omniinfer/sd-webui-cloud-inference you can use this to offload load to cloud.

try adding--medvram --xformers or just --xformers I have a Geforce 1070 with 8G of VRam and 32G of Ram on my system. it takes about 2 minutes an image with 1.5V models, and 12 to 15 minutes with the SDXL models. Adding --medvram --xformers reduced it to 2 minutes for the SDXL models.

@ktmarine1999677 i have same card as him when i use medvram it slows down to like 3 minutes but with out medvram it generates images in about 30 seconds but i use automatic1111

glad that i got 3090 before it get expensive again

@The_one_and_only7723 You can use SDXL with 12GB, without even using --medvram? How?

@Vendaciousness i have 32 gigs of system ram not sure here what i used COMMANDLINE_ARGS= --xformers --api --autolaunch --no-half-vae

@Vendaciousnessi even made a couple loras i posted here with it

@Vendaciousness i had 16 gig of system ram but it took for ever to load the settings it would freeze my compute but would not crash start back up in a half hour so i got more system ram fixed the issue how much system ram do you have

They are the same size. At its most basic level, one GB is defined as 1000³ (1,000,000,000) bytes and one GiB as 1024³ (1,073,741,824) bytes. That means one GB equals 0.93 GiB. So civitai is using GiB (1024³), but huggingface is using GB (1000³)

nomenclature. 7gigabytes vs the mentally retarded other nomeclature of 6 and one half gibibites. The one 99% of people use? about 7gigabytes. ^_^

@crusherjoe20221217 This makes sense, thank you

@bugsbe Am I the only one who remembers when standard prefix + bytes was unambiguously based on powers of 1024? byte, * 1024 = kilobyte, * 1024 is a megabyte, * 1024 is a gigabyte etc. Sometime around the last 20 years hard-drive manufacturers changed it to follow SI standards instead of Comp-sci standard practices so they could inflate their advertised storage capacities and sometime around the dawn of the 2010s other people (like ISO) got on board with the distinction, and now they've started gaslighting us like it was always this way. Pepperidge Farm has forgotten, but 90s Kids will always remember. (slightly tongue in cheek, but it's a valid point that for many of us this distinction was a marketing trick, not a standard, for most of our lives.)

@aciarrai957 @aciarrai957 no that's me. I think only hard drive companies use the giberish names lol

@aciarrai957 Yes, the 1000x1000 was a trick used by hard drive and other memory/storage manufacturers to inflate the size of their wares, causing endless confusion to the unwary.

After multiple tries and updating A1111 GUI, restarts, redownloads and looking up guides I finally got it to load into A1111, only to generate yet another error when its running image generation. "RuntimeError: Input type (float) and bias type (struct c10::Half) should be the same". Honestly, seems like you need a degree in computing just to actually get it running!

Tried adding --no-half to webui_user.bat as that fixed it for another person. This locked my computer up for over an hour then bluescreened my PC after a screed of out of memory errors and other errors popped up in my CMD window (using RTX3060 with 12GB) and 16GB memory.

So, I think I need to completely reinstall A1111 from scratch after trying to get this running for over 6 hours.

I'm getting the same kind of errors and I'm running a similar GPU as yours. This model seems to require heftier hardware than what was previously required.

@LavaSplit Seems to be the case. A 40x series card is way over my budget so I'll have to make do with SD1.5...!

@gr3yh4wk1 It may be within my budget, but it's not a crucial thing for my set up and I really don't want to have to upgrade my power supply as well to run that kind of card. And my current card still works just fine.

I got it to work with a 3080 with 12gb vram and 32gb ram....I had problems at first as well, and went through a myriad of fix attempts. I have it working now, but I also did some things to troubleshoot other problems I was having. The thing that fixed the other problems was completely removing (using a gpu driver removal utility) and then reverting my NVidia drivers to a much much earlier version that had been suggested on a forum somewhere. Once again, I dont know if that's what resolved my SDXL issues, but it's working great for me now.

WAY better than 1.5, huh? Totally worth excluding everyone with less than 16GB o VRAM!

@Vendaciousness You can get SDXL working on 12gb now. I've also seen mentions of 8gb.

@dillon101 you can use it on 4GBVRAM and less according to the folks over at Intel. Just gotta have the right optimizations applied.

Also, I agree on images being overproduced. But that is a consequence of training on professional rather than amateur content. As most professional real life content is overproduced tat these days.

可能色彩效果不一样 加了vae色彩好一些

maybe you are using a lora trained with diferent sd version

Then why the name is VAE fix? It is very confusing

@helloansuman you should only care about what type of file it is written as.. if you see it's checkpoint trained or checkpoint merged then they are always stable diffusion base models. If there's separate VAE required then it'll be referred with a link below the base model download button. Also, VAE files aren't supposed to be gigabytes in size so you should be able to figure out.

Instead of portrait mode, try to generate the image at 1024x1024 and see if the problem goes away. If that does not work, please post the prompt so that we can take a look at it.

You seem delightful

Isn't every generic SD model absolute garbage? i feel like stuff only gets good when the autist community starts refining these models to perfection.

@placid That may have been true with SD1.5 and SD2.1, but SDXL base is supposed to be a good general purpose model. You can see many quality images generated with it, even without any LoRA.

BTW, did you mean (artistic or autistic) community 😁?

Like these guys work for free to give you free models, if you are not happy, do it yourself a 8 millions images training and share it!

I probably means that there are 24 nodes in the ComfyUI workflow. You can click on the copy icon at the end of the field and then paste that workflow into your CompyUI to replicate the image.

usually its due to the resolution being way high, or batch size or if there's a memory leak from other apps. Restart of SD fixes it sometimes, or at worst case scenerio, restart the pc. running on i9 13900k, 4090 here. fixes it

TLDR; Just use the VAE Fix version.

Long version: there were some problems with the VAE shipped with the 1.0 release (for example watermarks) so they release another version with the VAE from the 0.9 version.

@NowhereManGo So the VAE Fix has the fixed VAE embedded?

@saintbrodie Yes.

Same issue - did you ever get yours to work?

I had to install ComfyUI to use SDXL. Takes about 10min to do a 1024x1024 image on Automatic1111, and about 90 seconds to do the same thing with ComfyUI.

@rerewfwefw3 maybe your face isnt good

515 x 768 takes me 1.5 seconds, so it's probably your GPU.

I have 16gb ram and an rtx 3060ti and on sd 1.5 it took me on like 20 seconds on 40 steps and now it takes me like 4 min to render a picture on 1024x1024

Post your prompt so that other can see how it can be improved on.

Negative prompts are really important to remove these. For example: "extra limbs, bad hands, bad fingers." And you might need to group them in parentheses... I'm still new at this too...

It is a problem across almost every AI image generator.

Hands are actually really hard for artists! Consider: the angle of your hand changes the number of fingers visible, sideways It might look like 1 finger and a thumb, and they overlap in almost every view! Even faces domt do that.

Arms and legs have similar issues, that's why they're frequently mutant-looking.

fyi the part about adding things to the negative prompt to resolve this is untrue. for the most part, having the words "hands" or "limbs" in your negative prompt will, at best, hide the hands/limbs from view (the subject might have them behind their back instead of in front for a portrait), but the more common effect is that the production of hands/limbs in your photograph will now have less available training data on hands/limbs since you have hard-excluded much of it, and the production of hands or limbs becomes even more difficult for the sampler.

as a general rule, don't use the negative prompt. That's it. That's the whole general rule. Negative prompt bad.

If you really must, it's a last resort sledgehammer trying to push a thumbtack into the wall. and the wall is your image.

@shapeshifter83 

jc as to what you suggest is the best route for getting the closest anatomically correct hands + limbs? 【Besides using image control/OpenPose/reference image option, or including ➕ the [((best hands, good hands, perfect hands))] in prompt】bc these gnarly, mutated, so not anatomically correct “things” ai likes to consider as “hands”

໒꒰ྀིっ˕ -｡꒱ྀི১

@shapeshifter83 I have the exact opposite experience, adding prompts to negative will often get better result FOR ME. Don't push your personal preferences onto others and vote their comments down just because you don't agree and think your way is the only way.

@Madsen1981 it's not personal preference, it's an understanding of what a Stable Diffusion checkpoint model actually is and what it actually does. Downvoted. Your anecdotal experience doesn't trump an actual understanding of the underlying functionality. In a controlled test over a large enough set, your anecdotal experience would prove just that, anecdotal. Heck, my "personal preference" would be for it to be as easy as you claim, just throw things you don't want in the negative prompt and BAM problems solved; but unfortunately, that's just not how it works.

@pixeldustlux the answer is, first and foremost, the checkpoint model. Some have better training than others when it comes to small anatomical details like hands. the base SD XL is pretty good considering everything that has come before it, but is easily surpassed by some of the newer consumer-trained XL models and is thoroughly outclassed by more recent generative AIs, in particular DALL-E 3. My suggestion is to locate good hands by spamming batches, but a couple SD XL models like Copax Timeless XL and Nightvision XL have anecdotally served me well on hands. Also, as a general rule, the higher your CFG, the harder it is for the sampler to get hands correct. If you run CFG just high enough to avoid the washed-out look, you will probably have your best results regarding hands (somewhere in the 2-6 CFG range, depending on model). This is because the higher the CFG, the less options the sampler has. Positive reinforcement, ex.: "(hands:1.2)" can be helpful if your image is already guaranteed to have hands or needs to have hands, but if hands are optional in your image, you might have better luck letting the sampler have the option to keep them hidden. The reality is that we just aren't 100% there on hands yet. If you want hands and are okay with strict censorship of NSFW, go use DALL-E 3. It's an entire generation ahead of SD XL.

@shapeshifter83 This is correct. Few if any images in the training data had labels like "mutant hands", "extra fingers", or "bad anatomy", or any of that other nonsense people so often put in the negative prompt. If it wasn't in the training data it is not in the model and if it's not in the model there is no point prompting for it. When people put "extra fingers" or "bad hands" in the negative all you're effectively doing is pushing away from hand or fingers in general... which gives a false impression that it worked.

How many images in the real-world do you think have metadata like that? Almost none...

I also agree with you that some checkpoints have MUCH better hand/finger generation than others, at least in 1.5, but my best results are with ControlNet

@shapeshifter83 lol

Don't be too picky with the hands. If you look at various REAL pictures in google or instagram or any social media, the hands don't look good either. Most of the time, they look disproportionate.

you still want to use the refiner and you still want to use the VAE. nothing has changed on those fronts

You DONT need the VAE anymore. I've been using it without VAE for a month and it works and looks perfect. The refiner is always optional

to my opinion it's really something SDXL is missing. Compared to SD1.5 it's so hard to get correct inpaint result with SDXL.

Seriously? Get better at prompting...

Why don't you train your own model?

Try adding cartoon, comic and superhero to your negative prompts.

Not the actual issue, the issue seems to be when this model was made it over associated "Wolverine" with the famous character instead of the real animal, guess they really had a lot of Wolverine comics laying around, maybe? By any chance have you tried Stable Diffusion 1.5 with character names from "My Little Pony" (the newest pony tv show), for example "Rarity", "Fluttershy", or really of the main few ponies? (usually you end up with color schemes matching each pony's name as a prompt trigger even with the base SD 1.5 straight from Stability.)

Nobody knows a wolverine even is. If you were at an endangered animals convention addressing 100 people and you mentioned wolverines, 101 people, including yourself, would imagine Hugh Jackman before thinking, "Wait, no, they meant the animal."

thanks everyone for the feedbacks on my comment. The point I am trying to make, prompting or not, that a wolverine animal is not the same as a The Wolverine character. I have to heavy load negatives with anything related to super heroes or Marvel when i am trying to make an anthro wolverine to keep from getting random metal claws sticking out of the back of the hands, or fur coloration like the super hero costume colors. If going for a non-anthro actual animal, I can more easily get a wolverine but as soon as I try and get it standing upright or putting it in clothes, I get the Marvel character. Please everyone if my prompting sucks, give me some examples for generating an anthro wolverine that looks like a proper wolverine. Try to be helpful to a noob like me, and not just tell me I suck at prompts,

Also, take a look at some of the images I shared here. There are some pretty damn good wolverines generated using models trained for it.

It's AMD's job to do that. They were saying rocm was coming for consumer-grade cards back in 2019 for Navi 1 GPUs, which it didn't. Forget about any support for anything below 6800 in the future. AMD is nowhere near anywhere to compete with cuda. Nvidia is your only option for anything AI-related

Bruh, I'm using an AMD 6800XT with SDXL. I'm even experimenting with SDXL lora creation. Takes forever, but it gets the job done. @Dwanvea is right though. If you're sitting on a 5700xt, might as well forget about it.

PS: I've trained 1.5 Lora's in a fairly decent amount of time. But I also do all of this on Linux. If you're looking for Windows Support might need someone else's input.

@brassen250 At least for now... I went through it with my 6800 when 5.7 and soon after having my hopes shattered with 6.0.x. The thing is Nvidia just have so many years of being H(person)IC that cuda is like Kleenex. nothing else is taken seriously, BUT, I think they pulled an Intel and are too comfy. Those m2 chips are REALLY efficient at the things they can do, I am not sure about here but I know in the chat/instruct area they are crazy fast and low power. Intels GPU's are really good over there for a lot lof LLama2 based stuff. I am not very experienced in this graphical stuff. ZLUDA just dropped. I don't know how much of the performance issues on AMD's side are hardware vs software but I have seen software x100 a task with a patch. I even thought I had finally gotten ROCm going on my rx 6800 (non xt, I got it early before covid/silicon shortage) but I had just somehow got vulkan acceleration going. I thought it went opencl ->ROCm and Vulkan was for games and 3d stuff. I went from 3 -30 tokes/s on the llama2 based bot I was testing with. I want to say it was a pure or hybrid minstral 13B_q(4-6?)K_L_M

i got it running now even on windows with AMD. Try HIPSDK (instead of official ROCM), and search for ZLUDA on youtube. there are tutorials how to install zluda and its even quite fast. I have a 6900XT with 16GB VRAM and running the model works now always.

@themagro Thanks for the lead! windows 10 11? It has been awhile, I will check out the tutorials for a week or so first. Do you have adrenaline installed? I was actually pretty impressed with that software last time I used windows. The training alone is worth it, I also had issues with roop, insightface?, a bunch of things that seemed to stem from pytorch. as far as the stuff that works well on linux (just regular sdxl/pony 1024x1024 or SD1.5 with different loras) how does zluda feel compared to pure rocm?

Have a try at negative embeddings! That should give you something to poke around with <3

nipples, deformed, bad anatomy, disfigured, mutated hands, mutated fingers, cropped, censured, crop, extra fingers, cartoon, animate, Eye deformation, hand deformation, finger deformation, Worst quality, low quality, irreal, painting, Low definition, deformed face, Deformed eyes

For general use it's better to use it in ComfyUI not A1111, this because ComfyUI is way better with your PC's resources. Use this Workflow for easy and best results: https://comfyanonymous.github.io/ComfyUI_examples/sdxl/sdxl_simple_example.png It was created for general rengdering by ComfyUI's creator!

VRAM @ 100% memory load?

@Jaffa595 Well that depends on the VRAM you got at your disposal, 1.5 needs at least 4GB to run, SDXL should be able to run on 8Gb although it's slow it should work. If you experience VRAM issues try low VRAM settings or smaller renders.

@Graybles Sorry, my post was aimed at OP. From what info was provided, it definitely sounds like they are maxing their VRAM and thus making everything almost grind to a halt.

Same for me, it takes forever to load on chrome (but loads faster on Edge). Smallest image generations takes a long time for very disappointing results. I will Install comfyUi for my next step.

I own an RTX3080Ti 12gb Vram, 64gb RAM, 16core 1950X Threadripper

Every human on planet Earth should not be using Automatic1111 and should be using SD Webforge, it has integrated features like control net and everything else that blow Automatic1111 out of the water. It is also roughly 20% better at memory management among other things. Anyone experiencing problems, try Webforge.

@Graybles Chatgpt says a1111 is better

Try using the StableDiffusionXLPipeline instead. This should solve your issue.

https://huggingface.co/docs/diffusers/v0.20.0/en/api/pipelines/stable_diffusion/stable_diffusion_xl#diffusers.StableDiffusionXLPipeline

I had this too at the start but not anymore. I don't know the exact reason but I think it's because you have some VAE, LoRA, emending, or extension in your prompt but not in correct file in forge. I made sure to take 100% of those off my promps and it worked, then I added them back in only after i was 100% sure they were in the propoer file in forge and not my A1111

You are not the only one, its way better with difference...

Why would you use sdxl base?

Are you new?

Git Gud 😂

以bilibili上秋葉整合包为例子，把fp16改成fp8，显存存优化选择“仅sdxl中等优化”，这样你就不会爆显存了，就够用了

Hi, configuration for the SD XL model is far to low. Recommended minimum HW: 32 GB memory, 16 GB video memory. As example I'm running this model on an old AMD machine with 64 GB memory and an GPU with 16 GB memory (Radeon RX6900XT) and an Ryzen 9 5900x CPU...

Me, too.

Me, too.

Checkpoints have been baked with VAE since sd1.5 and possibly 1.4 - It was a viable way to make sure everyone was on the intended VAE for the model. This is less useful today as we all mostly use sdxl_vae, even on Pony models.

The same thing, just blurring and distortion.

openpose and controlnet for character

I tried a remix and got something remarkably similar to the primary ad image...

I had a similar result at first, looked like jpeg artifacts, but install an SDXL Vae and results are much better. I still have research to do but stay strong. You'll get there, fellow warrior.

Maybe learn to use it :D

under --> models/checkpoints

Likely something funky with your VAE or the decoding. The preview uses a seperate process to decode the latent that comes out of the sampling process into a visible image. A Checkpoint is basically three parts, the Model that has the weights and math for the sampling process, the Clip that processes the prompt into something "understandable" by the model to use in sampling, and the VAE that decodes the output into a visible image.

The image generation doesn't work with an "image", it's all just strings of numbers being processed based on input values like the prompt, seed, latent image or empty latent image, model weights and so on. Once the sampling is finished the latent that comes out is decoded using the VAE and turned into the image.

So if the preview seems pretty normal, but the final decoded image is just black, the Issue usually is at the VAE decoding step.

I just know some basics, and there isn't enough information in your comment to tell you more, but I would try out downloading the vae seperately and run a generation as a test in your ui using that one instead of the one that is included (if a VAE is included or the one included is used automatically depending on what you use.)

If you use any SDXL model that has it's roots from this original model, you most likely do not need a refiner. I haven't used a refiner since I broke away from this model. If you want realistic there's over 100 realistic ones that don't need refiners here on the site. Good luck!

Hola

I used EmptySD3LatentImage

HAHAHAHHAHAH!!!

As said the lady, surely not sight to expect.

i have the same question, please let me know too

What are your parameters? How are the dataset organized and tagged? It's impossible to answer that question with just this. I always train using base models (XL, Pony, Illsutrious) in kohya and they work with a large number of models that are based on it.

Just don't expect a character model trained on SDXL to work with Pony or Illustrious, and vice-versa. There's a reason why these models are put under separate categories.