Do you use a1111?

@ayidencayman872 yep

@TheCAL Switch to Forge. Its same gui as a1111, but faster, and no need for commands, like half-vae and similar, Its automated.

Only downside, its not compatible with old extensions. If you use old extensions, i can recommend ReForge.

@Cyberdelia Reforge is on Gradio 3 and works with all a1111 extensions, no flux support. But very up to date. Have a ton sampler, which Forge doesnt have. Speed is similar, its on Forge engine.

@Cyberdelia  sweet I can't wait to try it out!

Which sampler and scheduler do you use, how many steps? Do you use custom Vae?

Для автора

https://civitai.com/models/1287309/natural-beauties-real

Suit yourself, but dont cry, if the results are not dreambabes.

@Cyberdelia Roger that, thank you! We'll be waiting :)

@Cyberdelia I'd kill for a CRP Inpainting model. I absolutely love all the models you've made! Been watching your updates for ages now. I hate that I have to use someone else's model for inpainting images made with CRP. It'd be soooo nice to just use CRP to do my "touch ups" for images made with CRP.

to create a lora you need 25-30 images

¿You try not use LoRas and put Tinkerbell directly in prompt?

For me gives better results.

Lol. Is my question really dumb or something? I'm just looking to refine the already good hands just a bit more. Any help would be appreciated, whether its with Meshgraphormer or something else. Thanks!

I think you can try something from SDXL Promax suite, but given this is PONY, you might run into issues. Controlnets with PONY are bit hit or miss.

@Mescalamba Thanks for the response!

6.5 looks more "vibrant" than the others... @Cyberdelia / @ShadowCell any idea why? Is it because of the images used to make this checkpoint or is it because of some different settings?
Just trying to understand as I'm a noob and didn't learn how to make checkpoints/loras yet :P

Interesting. What is first from top and second from bottom? Only those have unfucked eyes. Altho I suspect its question of how well it aligned with noise/solver/scheduler or VAE.

@Mescalamba If you click on each picture, it shows the resources used for each (the checkpoint version used). Anyway, it's v6.5 at the very top, then it works its way down to v8.5 at the bottom.

I should also point out that this isn't entirely a fair test... but a general indication. Sometimes a checkpoint will be great (generally speaking), but using the exact same settings, it will just randomly generate some junk images every now & again (kind of a roll of the dice). You'd really need a large meta-analysis of images being compared & averaged out for it to be fair, but that's not really practical for the most part.

@Cyberdelia I hate to say it, but I also kind of personally prefer the style of v6.5 the most also lol.

@ShadowCell Ah yea.. I was looking at "i" icon, forgot I could just click pic.

I think 8.5 needs different sampler or highres fix.

That said, 6.5 is sorta "real" realistic.

This specific sample of 8.5 apart those eyes is very "clean". Bit too much maybe.

@Mescalamba Admittedly, I would never really typically generate images like this just completely barebones, so it's not entirely reflective of the real world results I'd usually get...but the moment I start introducing different settings or LoRAs; I feel it becomes less of a test of the checkpoint, & more about cherry picking different random settings to get good results (so a comparison would be even less meaningful).

I try to do these kind of comparisons so people at least have a vague idea of what might be different with each update though. The eyes thing def sticks out to me, but after hi-res fix/upscaling, etc, then it pretty much ceases to be an issue across each version. v8.5 is prob overall the best so far I think (in terms of consistency, etc), but not everyone will be cranking the settings like me either, so I try to use very generic settings for tests to represent most users.

@Cyberdelia Lol, dont think there is something like "too much models". At least not in your case. This is one of most popular Pony models for quite a while.

@Cyberdelia It saddens me to hear you're hiding away a v6.5 type variant ;( which was my favourite style. I can understand your reasoning/aprehensions though.

Curiously, what are the objections; that it 'spams' notifications, overcomplicates things for other people, or...?

Seems like more choice would be better.

@Cyberdelia Thanks. I found out that any name doesn't work, it's good to try out existing characters. Some are too much trained and you don't have much control but some are trained, so you have full control but keep the face relatively consistent.

Well, apart training LORA (which is hard without consistent character :D), you could try getting one render of face and use it for via IP adapter to attempt somewhat close other angles.

I think you can even push multiple images in batch into IP adapter to increase consistency, got bit outdated workflow somewhere for that.

https://github.com/nerdyrodent/AVeryComfyNerd/tree/main/workflows/SDXL

Not sure how much it works these days, but worth a shot?

@Mescalamba I think a better aproach would be an adetailer with the lora and higher denoise. You can even run seperately the steps to have better consistancy.

@bobby888 Hm, you mean face detailer? Yea should work.

@Mescalamba exactly.

Update: I’ve tried using version 7 and that one works, however versions 8 and 8.5 still get stuck on “pending”

@Cyberdelia the issue seems to resolved for now!

Reduce the promp until the picture becomes more stable, according to personal observation, some lore does not respond adequately to duplicate prompta, as well as overestimated weights, i.e. different from the recommended weight values, and the picture can be very smeared on the same seed, artifacts appear, but then I use variability on low values.

_____________________________________________

Сокращайте промп до того момента, пока картинка не станет стабильнее, по личному наблюдению на некоторые лоры неадекватно реагируют на дублирующие промпты, а также завышенные веса, т.е отличные от рекомендуемых значений веса, и у меня на один и то же семя картинка может очень сильно размазываться появляются артефакты, но тогда я использую вариативность на низких значениях.

@CiviCreator0100110001 I tried again on CyberRealism, with the Euler Carras sampler, the result became much better, I need to improve it some more) Thx for information

try an empty prompt

_______________________

попробуйте пустой промпт

Nice

same problem ,maybe corrupted download

disable VAE , was the fix

Many SDXL models need a clip-skip of -2 (or 2, depending on the UI you use), otherwise it will be a static non-image. I thought a lot of the XL models I was trying were garbage until I accidentally discovered this. Now they all work great and as expected.

Any upcaler should fix eyes. Either with inpainting or highres fix.

Check out Pony Realism Slider. A strong weight like 3.5 does wonders

such as???

What is your setting(sampler, scheduler, cfg), do you use any other booster? Never gave me this. Actually its bad for furry, anthro.

@Cyberdelia I get they want to earn a bit more, but everytime it's these overly drastic changes where they completely screw up everything in ways that couldve easily been avoided. I looked at the stats. There were enough underused Checkpoints they couldve easily put up for Auction, leave 100-150 safe. There even was a spot held by a Checkpoint that was deleted! But no. They had to dump all 200 for no reason. And now wer are getting 200 meme Checkpoints in return. Theres nothing for users like me to render with next week, so no income for Civit either. I dont get these people.

At least I finally got a new gpu that can handle it.

@Cyberdelia comfy, render time is about 30 seconds for me.

@Cyberdelia tried all that. 1girl, 1woman and multiple 2girl 3girl in negative :/ doesn't work.

@Cyberdelia something like 1280x720

@Cyberdelia i think the world "SOLO" in prompt and "multiple" in negative helps it. Thanks.

Regional Prompter is a good way to prevent this.

am getting that too you on stable diffusion UI?

If you're on an nvidia card, check to see if the driver is offloading to system RAM - you can disable that in the nvidia control panel

Increase (face focus) weight and use face fix. That is what I do.

@ginlee__ should I put "(face focus)" in prompt?

@TONY_OFM Yes. You can increase by doing (face focus:1.3) and it will work like slider. Only parentheses is same as (face focus:1.1)

@ginlee__ thanks a lot buddy, another thing I tested and it worked a bit was to increase the empty latent image. It was 1080x1350 so increase it to something like 1344x1616

What do you mean landscape? Change resolution.

And why women or "bears" only? Otherwise it's great. BTW I don't get any troubles on full body. Just add "heels" or "feet", or etc you want to see along with "full bosy" - it'll make this happen quite easily

I already read this somewhere. Try custom Vae-s(there are here a lot).

What does the console write out?

If not else, change to Forge.

@Cyberdelia true, i figured it out tho, use this: https://huggingface.co/spaces/Sanster/Lama-Cleaner-lama

brush over the nipples and then they are gone, the nipples are nice, but i post the images on instagram so need pictures no nipples.. it is still a really good model though, much better than ponyrealism, thanks

@Cyberdelia doesn't work, tried it many times, even put it like this: (((((nipples))))), still model includes nipples poking through the shirt, i will say though that this mainly happens when i put "big breasts" or "huge breasts" in the positive prompt. if i leave out breasts then i think it'll sometimes not show nipples poking through shirt, the girl will have really small boobs though so that's not really good

Imagine being able to create your own wank material instead of subscribing to some thot's OF.

Save the safetensor file in a path like this stable-diffusion-webui/models/Stable-diffusion, reload the UI and then it should appear at the upper left corner.

@phantomkosmos thank you