You can use it by lowering the strength. Of course, some LORAs will not work very well.

alot of mine work, but of course some just dont at all or barely work

It depends. Just have to try it. A lot of sliders work on pony/illust/sdxl even if they don't advertise it. Most Pony char/artist loras are unnecessary on Illust. So it's mainly concepts and non-artist styles that need checking.

You could try using a highly weighted tag that only e621 has, like translucent clothing

I doubt it has much, if any. There's no ugly cartoon style bleed.

If you mean directly facing the camera, then the tag for that is "straight-on", not "from front".

@Zlia That didn't help with the strength either. But good try.

The tag for 'from front' is 'straight-on' as per booru tags go. But it's not usually needed anyway, because that's the default angle for most images. Unless, you're prompting something which is usually visible from behind, it should work. Also, instead of using "from behind, above and below", you should be using them separately, like "from behind, from above, from below".

If you can give an example, we might be able to help out more.

@nymical I disagree. Also, many others have included the "from front" tag, that is, more generally, not to take the back, if at all it mentions "ass" "butt". Then the prompt thinks "ass focus" is unfortunately negative at this point despite it.

R'le34, Santaku and elsewhere, for example, include "from front"/"from front position". So, I don't see any reason why this couldn't be taught. There are a lot of them. Even so many that there is enough viewing in them too.

@FujiHuji You're right that it can be taught, and it might even be the case if the model trainer have included 'from front' during training. I was just talking about the booru tags, as they are the most prominent ones.

Also, that's exactly what I meant when I said, "prompting something which is usually visible from behind", that 'from front/straight-on' is usually not needed, but it still have its uses in those situations.
I'm not asking anyone to not use them. I do use them when needed as well.

@nymical Well yeah. It could be something else, although I've looked at many examples and determined it accordingly, but really if I mention this to the prompt what FujiHoji said the AI ​​focuses on it the most. Sometimes it looks like it's from font, but more cursed. For example, it leaves the bottom part "from behind" and the top part "from font".

So, it's like its back could go 360 degrees. Very scary. XD

More bad lucky. In Pony it would have been an even worse mutation. That's where "from front" worked. :)

You should try v-pred mix version of model. It is much more responsive to prompt.

Also limited amount of "flux-style" can fit into prompt and translate into better result.
But not always. CLIP is pretty dumb.

What are the differences? I downloaded version 9.

May you use the tag-uncensord?

me2

I initially generate at a resolution of 1280x720(16:9). I get excellent results.

It knows cream the rabbit without showing a literal rabbit and cream. Instant favorite.

par for the course, also RIP r18.com

This is going to be really hard on less skilled artists that haven't made it up there yet. Companies will just use AI instead of hiring a junior artist.

Ah, so i could use it not on ComfyUI

@WAI0731 i have no workflow without an vae in confyui, i test it in a1111 seems to work

Pretty much any character with 100+ entries on danbooru should work decently.

https://danbooru.donmai.us/tags

-2

I find it doesn't really make a difference

Better support only means a higher chance of them working but is still a chance.

I have a pool of like 200 Pony loras waiting for an Illustrious counterpart, and after testing them 1 by 1 only like 30-40 of them work on this model, that number did not change between versions 7.0 and 8.0 of this model, with v9.0, I repeated the process and only like an extra 5 loras started working, so yeah is still only a chance.

You don't need any addons to try them on this model, just remember the formatting for loras on this model "<lora:filename>, (the comma at the end of the lora itself with no spaces)" and also you have to play around with the strength beyond what you normally use in a Pony model, like if on Pony the lora is used with a 0.8-1 strength, don't be afraid to crank that between 0.5-1.5 to get it to work on this model, my Pony loras that work almost all of them don't do anything unless I put the strength between 1-1.5 most of the time.

@Starvader Got it. Thanks for the comprehensive explanation.

You can directly copy someone else's generation data and then remove the unnecessary parts, leaving only the basic prompts. That should work fine.

Re-read the model description, so you can get more familiar with the specifics of this model compared to others, your CFG is slightly higher than the recommended (8 when the recommended is 5-7), personally I sometimes get high saturated gens when using 7, so I would play around with a CFG between 5-6.

Also, this isn't necessarily the case but try lowering the strength of that Shenhe lora to 0.8 from 1, like with my previous example sometimes this model will react weird to some loras at full strength, but they tend to behave completely normal when used at 0.8 or lower (or straight up drop it cuz this model actually knows Shenhe without the need for loras).

@Starvader thank you so much, it worked, I just was so used to using CFG 7-8 on pony

Been having the same issue, especially with certain trained Loras, the output images are either extremely underexposed and deep fried or extremely overexposed, happens on CFG 6 and 7 too

@yda4441279 I had cases like that and most of the time lowering the strength of the Lora to something below what is recommended helps, like using it at 0.5-0.6 when it recommends to not go below 0.8, but other times not even this helps, in my personal experience this mostly happens with "Illustrious" loras trained mostly or entirely on NoobAI instead of base Illustrious or mixes of it.

Another thing that can cause a deep fry out of nowhere that I've noticed is going too ham on the weight of prompts, like if I put weight on too many words even if its something low like 1.1 or 1.2 they will sometimes get fried, or if I only put weight on one or two words but go beyond 1.5 on them it happens too.

@Starvader Strange, yeah I'll try lowering the lora strength thanks
I haven't had the issue when using many super high weight prompts yet

I'm actually seeing it on loras I've trained myself, on the Illustrious base model, and I found that adding (oversaturated, overexposed, bright), (undersaturated, underexposed, dark) to negatives fixes it too

Look at websites like danbooru or gelbooru and find characters you like. If those characters have 200+ images then you can likely use them by simply prompting their name exactly as it appears in the tag list. The more images a character has of them, the higher the chances txt2img works without a LoRA.

This model has WAY more knowledge about anime characters then Pony v6.

Use escape from parenthesis commands, you can call up almost any character on Danbooru/Gelbooru with over 100 images (for the most part). put the "character name \(the series\)" So it would look like this:

"1girl, megumin \(konosuba\),"
or
"1girl, HK416 \(girls frontline\)"

I hope this helps, and happy prompting!

@BattleRabbitAIart I appreciate it thank you so much!

@ImJohnJohn Thanks!

@heavy0 I'm happy to help! I know learning that was an absolute game changer for my fan art

It has its own VAE baked into the model so you should use none, but you are free to experiment with others if you want to.

I would recommend you start with v9.0 if you have never used this model before, be sure to read the recommended settings and base positive and negative prompts too in the description.

I've been using this model since v7.0, and while a lot of people like that version over the newer ones due to personal preferences in how it handles body proportions compared to the newer ones, at least in my experience, v7.0 can be stubborn has hell when applying some Style loras, so if you plan to use styles outside of the ones native to the model you have to keep that in mind.

v8.0 is almost the same, it behaves better than v7.0 when it comes to Style loras and loras in general, but people report this is where the "proportions" issues started, I personally have not experienced these issues myself, at least not to the level where it is a problem (1 out of like 20 gens will have like a slight mismatch on head to body size or one hand to the other but is too rare to see it as a real problem for me), but a decent amount of people pop here and there claiming that so again, keep that in mind.

v9.0 at least for now is the best for me, I have not noticed any major problems with it, and while some people claim the proportions issues are still there, I've been getting those far less on this model compared to v8.0, plus is the first version in a while that got actual new training material added to the model instead of just trying to reinforce the current material and attempt to fix issues here and there (which it also claims to do).

So go for v9.0 if you want a taste of the model, and if whatever your gens are, you feel like body proportions are an issue, then try v7.0.

UPDATE as v10 dropped hours after this brief overview, after a few hours of testing, v10 feels like a straight upgrade over v9.0, everything I previously mentioned still applies, just change v9.0 with v10.

As a general rule, you should use the newest version. If you want to get a bit more insight scroll through the pictures of each version and look for the ones without any loras, that you like. With WAI I can say, that the newest version never disappoints.

@Starvader Excuse me, how could I use v9.0 on the site?

With the update the generator only allows me to use v10 to create. And the previous versions only appear me for download. (Forgive my ignorance, I'm still learning how to use the site and the generator correctly. Thanks)

@WAI0731 Oh, I see. Thank you very much for your replies!