by the way, can i ask you where to put those inpainting models? (auto 1111)

@epinikion thanks, could you tell me what specific topic did you try to fix with this new z version? and why should i use it? (already downloading)

right before the weekend too lmao

or Z+ Z++ Z1+...

Several languages have additional characters, and UTF-8 has plenty of symbols after Z :D

and looking forward to it~!

No way to know, it's impossible to load one model after another to check their differences, if any, unfortunately.

@Zenyth There is a way, Download them all, set them all to xyz plot script and launch it, you will see how photos change. Dont forget to turn on the draw a legend fc

@Zenyth I assume that the model creator knows what they did in the new training

@epinikion XD... What else? I mean, is there any improvement in Z over X, or they are just different?
Different dataset? Fine-tune of X/Y?

@epinikion Thanks!

@martinvanis2241 Yeah, I was kissing... er, what? autocorrect? I mean, I was kiDDing! Anyway, epg V1 is my favorite but I haven't tested yet XYZ versions yet, but I tend to do those kind of "my favorite version of a model" with "the best stuff added to the most recent one" merges myself, so it's great to see epinikion doing the same to create models that are the best of all worlds.

@epinikion not the embedding. Even the woman with the red shirt and white ballons looks corrupted:
1. https://orchestration.civitai.com/v1/consumer/images/846C28FF5AE07F4263933CFEFA4F10EEB533D8F5A706A274EB686A08ECDADE93

2. https://orchestration.civitai.com/v1/consumer/images/3CEA66700533BA85D92E3DA95FF30A1884DC0E8CB5605B6FFCA435E578C1A662

3. https://orchestration.civitai.com/v1/consumer/images/48AA820BB9E2EA5A699A069DB08FB741382C2BD16A71007218C5F564689DA1D9

4. https://orchestration.civitai.com/v1/consumer/images/F7CDCD77D08858FD29971055731E7CBCC00C521AB8EE43E8EC7AA36FCDA2ACB8

Much worse than yours... What am I missing?

@epinikion do you use Automatic1111 for them? If so, can you please share your configuration? I don't get the same quality even there

@epinikion I don't know, that's what I did, it doesn't look the same. The woman with the red shirt and white balloons looks slightly different with broken fingers, and the raccoon example is... nsfw?... https://ibb.co/wpDG3jD
What am I doing wrong?

@epinikion VAE is loaded (per terminal). Both CPU and GPU seeds look different. I will continue to investigate. Thanks

Btw the raccoon prompt contains a typo. Should be "raccoon" not "racoon" 😅 (didn't fixed it in my tests though)

I normaly enable in the Automatic 1111 Webui the "SD Vae" Select Field and change from automatic to vae-ft-mse-840000-ema-pruned, this gives on my prompt-sets on all my tested models (300+) the best results. Sometimes the 560000- vae gives a little more crispy tattoos or fashion-details .. but these are nearly impossible to detect. If a model is to "oversaturated" it sometimes help to switch to "none".

fantastic, this is what I was already using. thanks guys.

I use either mse-840000 or klfanime2, despite 'anime' name its actually good for photos

Same. I had lots of weird limbs in Version Z. The earlier versions seem to have more consistently good limbs.

Is it possible that it's because X is earlier?

@bomb4u never!

with hands its all in the prompt and dont expect adetailer to save you either , instead set its denoising very low like 0.15 low and just let it clean up the lines and add detail. you have to prompt for the hands to be how you want or , one thing that also can work is like find a fantasy western style art not digital or anime lora( you set its weight low like 0.2 but you prompt the hands heavily like (((insane detail fingers))) they tend to drawn hands with a lot of detail but in the same position a lot. This gives SD better options it seems. its still a crapshoot in some ways but if you set a batch and let it go you will get some with very nice properly dine hands with quality detail. Outside of that inpainting meticulously one finger at a time supposedly works but im too impatient for that .

Controlnet is best for consistent hands.

@lizardbeth21908 But what can I do that with all the tricks I keep getting hands like: https://ibb.co/mNCvq8C

@Nobodybutmyself with openpose? I also get broken hands even with higher weight for the ControlNet (if the object is far and the hands are relatively small)

@nrofis I never use openpose. I'd highly suggest finding a picture you want to use or an idea that you can build a pose off of to use other options. I only mainly use Lineart/Depth/IP adapter. Stacking these controlnets can be super strong to getting guaranteed hands no prompt required at all. Even alone they seem to fair well, but they pretty much guarantee all 5 fingers to be there at all times as long as map picture is good enough and you have the correct weights.

@Nobodybutmyself yeah, that's obvious. I use that. But when I try to generate "new image" (without reference one) I got stuck with the fingers

This is just a joke. Do not take it seriously! I'm just trying to be funny.

Maybe don't generate hands? ;)

Okay, really bad dad joke over. Personally, I try not to focus so much on hand generation. I ignore hands in general, mostly because I'm not extremely attentive to AI art but I am extremely attentive to AI art. What I mean by that is, I pay attention to other details.

I prefer proper proportions over good hands. I think AI training needs to get better to where hands are a staple rather than a feature of correction or hit and miss. (Not placing blame on generated models. The blame falls on original training data, i.e og versions such as SD 1.0, SD 1.5, SDXL, etc.) Happy to have the technology sure, but its those finer details that are normal to us that matters.

We enjoy excellent symmetry, we're humans. So we see a symmetrical person, instinctively, they just look better. At the same time, we're human. Symmetry is all over the place for people. We ignore the tiny imperfections on people. We'll probably focus in on it more in a still shot rather than a living breathing person, but for the most part, we're very forgiving for real people.

But we tend to notice off things in images, such as, crap generated hands. For me, its an extremely unique phenomena. One of the determining factors of attraction for me is the state of your hands. Basically, you either have ugly hands, average, or pretty. So as you can imagine, hands really stand out for me. I've trained my brain to ignore generative hands, because its not really a staple feature of SD.

When it is a stable feature of an SD version, then I'll start getting picky. Cause hands matter.

But for now, Eye symmetry, and body proportion is my main concern. (Sometimes shoulders come out to wide or I get some fat legs that don't fit the body. Just weird junk like that.) But that's my take on hands. You can completely ignore this message after all. It's just my personal preference to ignore hands unless they're distressingly bad or ill proportioned.

@superjosh250 I value your opinion but disagree with :) I think generating good hands is not less important than generating a good face and body. Just to make it clear, I'm not talking about ugly or pretty hands in terms of proportion or texture, I'm talking about extra, missing, or broken fingers, like these: https://ibb.co/fYnHcg7 https://ibb.co/rtDQzst

@nrofis No I feel you on that. Um, probably best to just mess with your prompt to negate bad hand generations as best you can. I don't really use controlnet as some suggested unless I want something very specific. In my experience it's fudges with my results. Could just be cause I'm using an AMD GPU or something else.

A couple of things you can do is one, get a bad hands negative prompt. That helps. Second, look at all sorts of models and look for negative prompts and specifically anything with limbs. You don't need a whole paragraph like some people do but a few keywords could go a long way. I'd write some of the negative prompts I use but you need to find what works best for you. Look at prompts here, Realistic Vision, RealPixar or RealCartoon... And others that pique your interest. Grab pieces of the negative prompt regarding anatomy and then go to town with testing. See what's consistent and keep it.

You can start by looking at my post and how I prompt. Check EpicPhotogasm v3.1 to see my post. Doing a Z run now, but trying to perfect my prompting to get what I want. But you should be able to borrow from my negative prompts.

@superjosh250 Tried so many negative prompts. It is not immune. That's why I'm asking here for another methodology. Because I have flaws in all the methods I tried. I hoped that someone has a robust solution to this problem

@nrofis Oi... well sorry that my suggestion didn't help at all. Hopefully there's a more experienced prompter that can help you.

I will say in my personal results, my images are quite consistent with the odd ball of a really glaringly bad hand or extra limb here or there. Out of 10 images rendered, I might have 2 with bad hands or extra limbs.

Last thing I could suggest, is try not to artificially bloat your negative prompt so much or render batches instead of one at a time and sift through the images for the best one. And maybe in the positive prompt try things like waving or arms crossed. That might help out a bit. Other than that, I've got nothing left in the tank. You might need to use the all-powerful and wise google/reddit for suggestions. OH! And keep in mind, that you want consistent results not perfection. SD is far from perfect right now. So maybe have the goal of 6 out of ten images have decent hands. Any less than that is unacceptable.

Okay, no more ranting. Hopefully later on today, I might post some decent Z rendered images. (Probably without hands. ZING! ;)

Just use positive and negative prompt that generate great hands. U cant just rely on a model to do everything. Even if 1.5 do everything quicker and better than XL, dont suppose u dont have to write a good prompt.

I just wanna mention that negative prompts like "bad hands,extra limbs,extra fingers" and similar kinda stuff do nothing in terms of hands problems or any sort of anatomy problems.

You just putting trash in your negative prompt that would in many cases be visible on your rendering quality.

Diffusor actually is not even aware of this,and everything is just a pure luck.

ControlNet can help but its still not perfect,same as embeedings and LORA-s.

Some models are better then other and this is what it is.

SD models have problems with anatomy and there is not much you can do,asides from inpainting.

SDXL models are much better,but they also have they own fair shair of bugs like speed and other stuff.

With that said always render at least 5 images in the batch and then you can select which one you wanna use or do corrections in inpainting booth.

I asked the same question. This is the answer: https://civitai.com/models/132632?commentId=283854&modal=commentThread&highlight=283999

@nrofis doesn't really help other than Z is the best version? If I am reading the reply right

from experience i'd say z is best and most flexible

@DevilLady now we have Last Unicorn... I can't keep up

@epinikion yeah, am not downloading them all just to compare to then delete the ones I don't want

@epinikion ho WOW I didn't know it is so simple. Thanks!

What are your settings? I'm interested that you are getting artifacts. I've noticed on certain samplers this does happen. The DPM++ 3M sampler seems to create artifacts. But I've found artifacts are immensely reduced with a higher step count.

You're better off going 512x768 and then using the 4x-UltraSharp upscaler. An ADetailer can also help keep the faces flawless.

@EryxAtys nope, sd upscalers are mediocre and take a lot of time and i found that it's way better to just upscale with topaz gigapixel for example

@brassen250 i usually go for ~90 steps for that res and it works really good. I mean that's pretty much sdxl size so some problems are to be expected but still this model is really impressive 

@DevilLady Yeah, I wish I could do that. Running an AMD card. Going past 768 crashes SD instanty. I could do 90 steps but it'll take triple the time. Gotta work smarter on AMD side. If I pass it through the hires fix, I can upscale it by 1.5 max. After that I run out of VRAM. Sucks to be on AMD side for AI art. Love my card but not so great for quick dirty render runs or extreme detailing.

@brassen250 well that sucks indeed. You could always consider selling your current gpu and buy for example rtx 3060 with whopping 12GB VRAM. It has pretty decent price right now

@DevilLady I had thought about that I'd rather just save for a second top tier GPU in the first place. I'm getting crazy good gaming performance with my AMD card. So worth the investment. Besides that, I have a secondary PC that could use it. I'd just convert that PC to my main render machine. Besides, I'm sure with some more time, SD will get even better for AMD GPU's. Maybe not on the same level as NVIDIA but better nonetheless.

@DevilLady just curious, how long does it take you to fully render a one realistic high quality image and what is your gpu. I have a rtx 4070 and it takes me about 5 minuets and I don't know if that is slow.

@DevilLady Try '8x_NMKD-Superscale_150000_G.pth' for upscale. Works great for me. Olivio Sarakas recommends it in this video. https://www.youtube.com/watch?v=TrcwBSlczfQ

@bearsdfg I use a 3080 10GB with custom liquid cooling (I doubt that actually helps this performance) but at 80-120 steps at 1152x1440 I usually run about 4-6 minutes. I think the 3080s run a touch faster than 4070s, but I've been out of the benchmark loop for a while. I don't run hirez fix or upscaling, that's just base resolution for SD1.5. Sounds crazy, but I've found keeping standardized aspect ratios (16:9, 8:10 for portrait, etc) and playing around those values seems to generate better. I start at 20 steps, see if a particular resolution makes a decent workable image, then crank up the steps and tweak prompts from there. After that, inpainting and adetailer if necessary, then upscale once.

@DevilLady I've run over 100 images through Gigapixel. It totally botches faces and hair. Respectfully, you have no idea what you're talking about. Get a better graphics card, and then we'll talk.

@bearsdfg my rtx 2070s takes around 6-8 minutes at ~1400x1024

@EryxAtys no need to be so high up your butt dude, Respectfully you have no idea how to use topaz because it looks slightly worse than ultimate sd control net tile upscale and it's 20 times faster. I'm perfectly happy with my good old rtx 2070s