@Cyberdelia I still have a lot of issues with hands and feet, but I noticed that keeping 'hands' out of the prompt—both positive and negative—ends up helping a bit, I think.

@Cyberdelia Lately, I’ve started using 'face_yolov9c, foot-yolov8l, hand_yolov9c'—it’s been helping a lot.

I feel like I'm having the same issue... adding 'overexposure' to the negative seems to help

@Cyberdelia Sadly i cant upload them as it uses a celebrity lora , however i can text the prompt if that'll help

@Cyberdelia using the same seed for both V10 and V11 sometimes it gets overexposed. V10 is much more stable. I am using SDXL loras and refining it during adetailer with cyberrealistic XL V57

@Cyberdelia I was having the same issue, and I think I was over-weighting something in the prompt... I think in the negative. I haven't narrowed down what word was the issue, but if it happens again, I'll upload it

@Cyberdelia I think I've figured it out. It seems to happen when a keyword it used that it doesn't quite recognize, that and the CFG being set too high.

I use Prompt_bit_sorcerer's v2 workflow with the cfg skimmer, and it seems that raising/lowering the cfg/skimmer can make the image over exposed. And depending on how I weight up a keyword. Annoyingly, I've deleted the examples, but I'll try to recreate it

Yeah, minimal change from catalyst 2. Much better, than v10.

@Cyberdelia Still remember for the release of v7 Christmas Edition. I was so joyful, couldnt stop commenting:))

@Cyberdelia I love the Catalyst versions also! Keep them coming! :-D

Something for something. Still the most balanced model. Niche(or not so niche) concepts, like furry, anthro never worked. Try Ponyrealism and enjoy the 3d view:)

Have you taken ganxta pills? Whats the problem?

v.11 is a whole nother level compared to v.10. I don't know what your problem is to lash out like that, but it is obvious your own fault.

@bobby888 
Lash out? Where, how?

If you read the notes that were provided, and comments people left about v10, you'd see there was a lot of complaints about v10, and people apparently kept pestering Cyberdelia to just release Catalyst 2.0 as the new main release version. What I said, was a joke relating to all that. Might want to get that stick out of your ass, and develop a sense of humour.

(Back by popular demand....ok, that's somewhat a fabrication lol, but I love comparisons; & I'd like to think others do, too. It's just no one rly does them because they're kind of a pain in the ass to do, & you kinda have to have all the models).

So far I'm personally really liking V11 as a contender for the best model version yet (Psst, maybe v10 is coming across as weakest of the bunch in the image comparison here, too). Do we have any V10 fanboys in the room? ;) Make some noise!

That said I'm going to post more image comparisons between all the different versions of CyberRealistic Pony soon, so stay tuned....it's just kind of late atm :/

@ShadowCell Yeah, thanx. Pure 1girl comparison doesnt show big change. Changes are much bigger with lora usage, sex pose, lighting etc.

@massisaatvik581 I'm planning on doing the most barebones comparison first for the sake of objectivity/fairness, then I can try some other stuff; it's just harder (& less fair, I feel) to start throwing in too many random LoRAs, etc, & still have it make sense without massive data/image comparison sets. No matter how good a checkpoint is; it's still going to throw out garbage generations every now & again....adding in too many variables increases the odds of that happening, thus seemingly massively distorting perceptions for versions, I feel.

I usually use a bunch of LoRAs myself, etc.....but I'm not really sure that's how most new-ish users (that probably find these types of comparisons most helpful) probably use them though.

@ShadowCell I mean a strong character lora(can be 2.5d, 3d) with same strenght(i usually use them at 0.5-0.6) will show more difference. My test pose( side view doggystyle) has problem with 90% of realistic checkpoints(cyber can make it very well, but there are failure percentage between versions).

@massisaatvik581 Another factor is that LoRAs will show massive bias to whatever checkpoint/version they were originally trained on, so this also largerly distorts fair comparison.

I have noticed myself, that at times I have intitially been really disappointed with some new checkpoints in the past....but not because they're actually bad; but because a tendency exists for people to just import all their best prompts, best LoRAs, etc, that have worked well for them elsewhere so far, & expect them to play nice with something new & different.

I have found with time, that some checkpoints/versions I didn't originally like much, have massively grown on me once I've shifted towards a different prompting/LoRA paradigm that is more well suited to that particular version.

The newest PonyRealism particularly comes to mind in this regard, given it was a larger depertarture from the last few versions. With v2.1 to 2.2, you could just use the same prompts, etc....but with v2.3 I found it required changing things up quite a bit to get good results. Once you did though, the quality ceiling & potential increased over older versions.

@ShadowCell Thats the point of the comparison, same prompt, lora, pose, settings.

Anyway you are right, models handle loras differently.

@massisaatvik581 The problem there, is you'd need a massive meta analysis of images for that to be meaningful, or fair.

The way I see it, there's like 10,000+ LoRAs out there.....an avid user may switch back & forth mostly between lets say 20 personal favourites for the most part.

My favourite ones, in all likelyhood, may be LoRAs people have never used, & will probably never use. & the resulting image generations I end up doing could be amazing with those; it's just that the other 10,000 (minus those 20) could be all be pretty terrible, & not play nice at all. So it's a hard thing to balance, without it devolving into 'here's 5 LoRAs it works really well with, that most people will probably never use' type thing. I mean, I have zero interest in furry content, but for someone else, that might be all they use checkpoints for. On the other hand, it may generate furry images amazingly, it's just it's not going to be meaningful to me, because I will never use it for that, so it kind of renders it entirely moot if it does that well. I just feel the odds of me showcasing something very particular like that, & others also being exactly interested in that for it to be meaningful, rather unlikely.

In your particular instance, nothing else may work with that particular position/LoRA....but a model checkpoint can't really produce things that it was simply never trained on. (I dunno how interested you are in this particular topic, because I have a tendency to go into exhaustive detail well beyond anything most people will be interested in btw....lol), but when you have a checkpoint that wasn't trained for a particular thing in its data (lets say; doggy pose), then a 'doggy LoRA' & checkpoint will end up just fighting against each other for the most part, trying to do the two different, indepedent things that they were trained on; whilst also trying to make them fit together somehow.

In other words, LoRAs will work way better when there's already some of that content in the data to merge the new content in the LoRA with. I've used good checkpoints before, with basic prompts for example (like "working out in a gym"), & it's just competely failed to return anything of that nature, no matter how weighted or detailed the prompting, because it's just not in there. It wouldn't be fair to say it sucks because of that either though, because there's like a million different things people could prompt instead (though it's probably mostly just porn if browsing civitai is anything to go by...lol).

(Anyway, sleep time) I ramble too much when tired lol.

@ShadowCell I never use loras for poses, thats pure prompt, which should work with every pony, illustrious model, but usually they sacrifice booru adherence for realism. Bad XL merges, too many XL cause this.

I only use character loras (i can save them with prompt and pic).

My favorite method is to use earlier versions (I like 6.4 or 7) for the base image, then cherry-pick the best ones and use v11 for img2img/upscaling, with around .25 to .35 denoise.

Newer versions are more realistic and have better details, but have lost a lot of knowledge and flexibility. Try prompting "1girl, shuten douji" in different versions and watch as it goes from perfectly accurate horns/makeup/body shape in the earlier versions to generic chonky demon girl cosplay in the newer ones. Same for most characters that are baked into base Pony. You need a Lora for everything now. Swapping versions for the second pass tends to give the best of both worlds.

maybe model is corrupt , try re dl it

@Cyberdelia A111

@Cyberdelia what censorship ?

bad workflow

If you're using ComfyUI, try a HighRes-Fix script which should give better control over the output. It'll help prevent duplication errors often seen when outputting images larger than the training input image sizes.

@pixel8er , ok thank i will search that, was not aware that existed, ok i see, but it s not interesting it's just an upscaler

did you do lora's before or it s the first one ? i'm also doing my first one right now, i just double the learnig epoche because the sample where very bad, also increase learning rate , and resolution now it looking a bit better but still not good enough, i keep it running, not sure i have the right settings

@carlosurportable419 not my first lora, i've trained more than thousands loras right now, but this one seemsnot to work at all, i've tested a lot of different params, even training for 20k steps, and nothing seems to work

@Nephilim, ok so you saying, the problem comes from the checkpoint, so maybe i should try an other one first

@Nephilim  looks like my is on the good way, i tested today, i restarted the training yesterdays, looks like change setting in the middle of training is a bad idea, it got corrupted face, so i started over yesturday, i did some test now, it's already not bad, but training is not even half way, i only have 1080ti , it s gonna take 3 or 4 days to complete unless i stop before, so i guess it should work on this checkpoint

actually it's disappointing, looks like it did not make any improve sins yesterday, same as you it s still bad quality, only when you got a close photo from close of the face it s looks not bad, but from distance it s very bad, like the character standing on the beach, it s not even close to what it s need to be, what other checkpoint you got good results on ?

@carlosurportable419 big love was the "cleanest" ones

@Nephilim i getting good results , i start over again, tried Prodigy but the sample was always the same, then get back to Adamw @ 1600x nothing special in the setting but i think you probably need more longer training , mine is not done yet, but it's start to get a cleaner image , also have around 30-40 pics with 10x repeats

@carlosurportable419 i'm already at 350(3 repeats for 1800 steps = ~2100 samples) images, still getting bad results haha, i just think it does not work the same way on these photo models as in anime models

@Nephilim ok, maybe i speak to fast, result are not that good, i did a few more test, lot of pics it looks like plastic skin, not that great, i guess i gonne need to find an other checkpoint for my next lora

@Nephilim  i stopped my training, looks like i was already overtraining, but you right this checkpoint is not good for the loras but i have good news, the lora works very good on ponyFinalcut_v11BF16 i trained in 1600x and give very good result in that resolution and the quality is good, the skin don't looks plastic, i tested on some other checkpoint but that give me some weird stuff lol

@carlosurportable419 gonna test on it too

got worse results with it, slightly worse

Great

@Cyberdelia Thank you ☺️. I’ve also been running some tests and the “nipple problem” isn’t as universal as I originally thought.

@SoopaTroopa 
I've seen nipples too, but it seems related to the clothing, and state of the character. That is if she is wearing something thin, and especially if she would be considered to be in an aroused state, nipples. Although if what she has on up top is thicker, and especially if she would not be likely aroused, she will likely not have visible nipples.

I should also note that I saw this in versions prior to 11 as well, so it's not new. I'm guessing you were just using a type of clothing you may not have before.

Agreed. 11 overpowers so many loras and eliminates their influence from the prompts. Though it seems v10 is offline now.

@Full_Of_Love v10 is semireal.

Adherence come, when a model is close to the source. Generally anime, than semireal, finally realistic models is the order.

About nipples. Just put the word into negative.

@carlyleihsan355 It's been there for months. Only works half the time. 😄

@Full_Of_Love 
One thing I've noted is that the generation tool used, can play a noticeable factor in how the model behaves. Like when people were complaining constantly about V10 looking cartoony, I only saw stuff like that when generating images with it on this site, while when doing so with the tool on my computer, I never had any issues, they were all completely realistic, without need for excessive prompting to push it to that direction like was needed here. Same with some other things, I could take the prompts that give me a consistent result from here, put it in the one in my computer, and get decidedly different results. It could even render the same things differently, like if I generated a shower here, it likes giving me one with a kind of marble dark brown tile, while on my computer, with the same prompts, it defaults to a sort of marbled light tan tile.

To add, and repeat what Cyberdelia has said before, different versions of the model, handle things differently, so you may need to adjust how you do things to get the same kind of results. This is even when using the same generation tool, adjusting the weight of the LoRA, prompts, and weights of your prompts may be required.

@Cyberdelia Honestly if you keep version 8.5 or 10 available then I don’t mind what happens in future versions of it. Either of those two serve me well.

@Cyberdelia I meant without having to download it and instead using it as a model. If keeping these versions up costs you money however then don’t worry, I can manage without them. You’ve already done a lot for me on this site :)

@Cyberdelia thank you! I would greatly appreciate something being done about the nipples in version 11 though, as they’re unfortunately quite common.

As you should be able to see from the images below, that's not a universal issue with this model. Different models, and rendering tools, can handle the same things differently, also the LoRA you're using, can have various undesired impacts on your images, especially when combining different ones.

Best I could suggest is to first make some images without any LoRA at all until you get the lighting, and so forth the way you like it, changing the prompts, as well as their weights until you get the desired results. This will give you a baseline to work from, now you can start adding some LoRA, but it would be best to add them one at a time, so you can see the various impacts of each one. From there you may need to change the LoRA used, as you may find a specific one has too many undesired affects, and/or just adjust the weights of the LoRA, the prompts used, and the weights of your prompts.