@Autismix_anon, sometimes I see people don't add "score_6_up, score_5_up, score_4_up" to prompt and even add it to the neg, why is that? Just a random discovery or these tags are somehow furry related?

https://civitai.com/models/298707/25dprony

PonyXL (especially V6) itself supports WAYY more than ponies already, if you havent tried it

Adetailer seems to work in forge couldnt get it to work in 1111 if anyone has the same issue

@senile If the problem you're having is weird artifacting when using adetailer, this will probably help: https://civitai.com/articles/4044

Also could be bad VAE

@Autismix_anon So there is no baked in VAE?

20H on 3060? man i think i will be rent a gpu for training lora 😅 mine is 3060 too

@Mikayori001 I'm no expert so maybe I'm doing it wrong. I tried again with 2 times less steps and epoch... with only 4 epoch 5 steps per img it was done in 6 hours. Not as good as the previous training but still very nice.
I actually want to use online GPUs but that shit is super hard for me and not straightforward at all, it's as if they think everyone is a developer. I'm fine paying for a service but at least make it easy to use for anyone ffs... x)

Did you used gradient checkpointing? mine is also 3060, 12GB... training for Pony diffusion XL taking like 1 hour

Someone compiled a list of innate artists/styles/characters here: https://rentry.co/ponyxl_loras_n_stuff
Mostly LoRAs but that too. I also personally wouldn't follow some of their advice. And prepare to see objectionable content.

@Barjoe01 yes I saw many tutorial recommending gradient checkpoint, even thought I don't understand what it is. Are you telling me I would get the same result without it, and like 10 times faster ? x)
I'm going to wait for your answer before doing another one, this is important.
By the way I was probably tired and I was wrong I didn't use AutismMix like I intended to at first. I probably thought using PonyXL instead would give more compatibility and then I changed to that then forgot xD
I'm still going to do another traiing with AutismMix because it's superior I think.
I do have another important question though, as I'm training for a style that already exists in the checkpoint, should I use the trigger in the captions or not ?
While I already get good result in AutismMix it looks more like older style from the artist not the recent one, and so far the training with PonyXL did make a difference that worth it. It did not work at all to use that Lora trained with Pony, applied to Autism. Hope that's not too confusing sorry lol

@maurice_ Thanks a lot that's exactly what I was looking for !

@Barjoe01 I just tried to remove gradient checkpointing and it was 50s/it, I stopped it right away hahaha that would have take me 72 hours to finish xDDDD

@tetsuoo Sorry for the late response, actually with gradient checkpointing enabled, it should be faster… I guess there’s some problem in the config file…

Try this config : https://civitai.com/models/281404/lora-training-guide-anime-sdxl

@Barjoe01 only 2 hours, that's pretty fast to me not late lol
The computer is in the middle of a training, 8 hours left but I will test it later for sure, thanks a lot ! ^^

@Barjoe01 what optimizer did you use ? prodigy ?

@Mikayori001 I've tried prodigy and AdamW (I believe, quite forgot), both of them work maybe 1 - 2 hours mostly, using the same config as I pointed above

@Barjoe01 How many images you had in your dataset ? For me it's 250-270 images, with 10 or 5 repeats... This could explain that x) It was recommended to use just 1 repeat in the method you pointed

@tetsuoo 100 images, 1 to 2 repeat, set it higher than that make it overtrained on me

@Barjoe01 this config is wrong. I finally understand why, after lot of testing x) Learning rate is at 1: it's not going to learn anything, it needs to be minimum 0.0005. Being at 1, the result is nicer images but nothing learned at all.
Another thing is Network Dim at 64 with Network Alpha at 1 : this gives a big file 4 times bigger than my previous trainings, I'm not sure if that makes any difference.... I use Dim 16 Alpha 8, could be 32/16 why not, I think more then that is not necessary.
At the very least using Prodigy for the optimizer is the only thing I needed to change, it makes training 50% faster, totally worth it ;) Now I'm not sure about when should I use cosine or constant, that's gonna take a lot more painful testing to figure that out... Not sure I want to do that, I've been working on this for 5 days now x)

@tetsuoo ah, that's great! glad you managed to sort it out, I hope it can be more improved

@Barjoe01 Update : Turns out ".txt" extension for captions wasn't specified, so basically I lost 2 days making failures because of that, it was training on nothing hahaha no wonder all it could do is draw better omg never again

@Autismix_anon I second that. I also have a 3060/12 and there's no way I could use SDXL models with A1111 WebUI, it would be very slow and upscaling would give OOM Out Of Memory error so I had no better choice than ComfyUI or Fooocus. Forge UI changed the game, huge difference

@tetsuoo I have a very lean install and yet SDXL is still painful. I feel you. Similar story for upscaling, it doesn't give me the courtesy of an error message and instead just hangs indefinitely. I didn't know about Forge, which could be a real game changer! Trying it now

@Autismix_anon Thank you so much for the suggestion. Works amazingly, runs below 9s/it. That's using --cuda-stream and --pin-shared-memory but I'll try to get away with not using those. Mind you I'm used to around 50s/it best case scenario with several optimizations enabled. You sir are a life saver. Generating images this good in about 3 minutes will probably never stop being insane to me.

https://rentry.org/ponyxl_loras_n_stuff#innate-characters

not a comprehensive list, i can gen Haru Okumura from persona despite not being listed, but not Dixie Kong from donkey kong. You might have to see what you can gen manually

@HimikoHonda Yeah that list is missing a lot probably. I've been able to generate Esdeath from Akame ga Kill and Nejire from MHA even tho they're not even listed on there

this probably isnt the right way but for me i've just linked a checkpoint loader and lora loader in ComfyUI into a Save Checkpoint node and it works, you can save other things like ModelSamplingDiscrete or FreeUV2 into any model with that

Supermerger works, i have a tutorial on YT but hold fire cause i need to try and make one WITHOUT music lol

If PonyXLv7 ends up being able to mix artists as well as NovelAI then yes AutismMix will probably also be as good

I feel like this and Pony have already blown NAI out of the water. Picture them on SD3. Whew. Porn shall never be the same.

@crawlingrat I do agree, when it comes to doing NSFW concepts or even understanding prompts Pony is really good, the only thing that's bad about it is a lack of innate artists and not being able to mix artists together (it can but it's not as good as NAI). Let's just hope SD3 won't be as censored as SD2 was cause if it is then there's no point in making Pony on SD3. That's mainly the reason why SD2 failed.

@GogetaSSGSS3 The fact that I can train my OC on it is the main reason why I feel it does better then NAI. I to hope SD3 isn't destroyed like 2 was. This new model seems to understand things so much better. Maybe I could finally use two different character loRA without them fusing together- or needing extensions like regional prompt!

@crawlingrat Yeah it definitely has its strengths and weaknesses, there's still a lot to improve which the creator of the model said himself, that's why he's working on PonyV7 currently.

Also I think SD3 should be released pretty soon. I was in the PonyXL discord and they were talking about it there, apparently it should be released in a week or two, not sure if that's true or not.

I saw some images of it and it's capable of generating pretty good anime images, even tho it was only a small amount that I saw but still, I have high hopes.

@crawlingrat Yea but it's not JUST about porn though, think of those of us who ARENT NSFW focused, sure I do things that CAN BE"NSFW" leaning (Because comics woot woot) but largely I freak out trying to do NSFW lol.

@duskfallcrew It's a NSFW/SFW model, it can do really good SFW stuff as well from what I've seen, you just gotta put "rating_safe" in your prompt box instead of "rating_explicit" or "rating_questionable" (these 2 are for NSFW/Ecchi images), so you should be able to get pretty good results with that tag.

Considering PonyV7's dataset so far is confirmed to be over 25 million images it's gonna be able to do SFW stuff even better if all goes well.

@GogetaSSGSS3 I'm ware pv7's coming, i'm in the creator program xD i'm in the same discord as pony's creator. I am yes also aware but always forget (so thank you for the reminder) how pony works, and i'm new to it in general, but i've started offering MORE loras as much as i can get ported over to pony lol

@GogetaSSGSS3  Geez I'm just getting so excited thinking about PonyXLV7. Does the creator need any sort of support? This entire thing can't be cheap to create. I'd be happy to donate to a patron or something.

@crawlingrat I think there's a link to where you can donate to a site in the latest announcement post he made in the ponyxl discord. There's also some info there if you want to share you GPU and provide some of its power for a certain amount of time. Just message the creator (AstraliteHeart) and he'll let you know about the details of that

@GogetaSSGSS3 Me messagine Astralite: YOU WANT MY GPU?
Her: What kind
Me: <_< piss poor imac GPU

community: CACKLES DYING LAUGHING

(note: This convo didnt happen, it's just a meme on the fact i have 8gb system ram AND 8gb vram XD)

Because you need to have Pony's OG diffusers version, OR create a diffusers version of this. Colab may not like downloading Civit files, and if oyu're using the Kohya SS one, it needs to be a diffusers version for some reason

@duskfallcrew Thank you! THANK YOU!

@crawlingrat Not a problem <3 Always happy to help, I'm currently training stuff on PonyXL as well, largely you CAN train on merges but it is hard, so the likely thing is test and see? PonyXL is so far blatantly amazing.

It's a prompt "trick". It ends the current chunk, good at making clear separation of your description to the model when your prompts get long

@JimmyCarter I didn't know that, thank you Mr. President.

Is this best to put at the beginning or the end of the prompt?

@JimmyCarter I've heard of this trick, and it's said to slightly improve the quality of the generation.

But all I know is that BREAK is always placed at the end of the prompt, not somewhere before, so I don't understand why he put it here.

@Fox009 end

@create023 that's because score_9 in itself contains 20+ quality tags. I've read about this somewhere, maybe it was on the PonyXL discord, I can't really remember, but regardless that's why people use BREAK right after using score9 or after the whole string of score tags, because the score prompts take up a lot more tokens.

The model uses the score tags instead of tags such as "masterpiece, hd, best quality..." so it makes sense why the score prompts contain so many tokens.

@create023 My understanding is that you want to keep the chunks to a length of 75, so apply BREAK when your prompt is approaching that token length. Beginning, end, wherever. I haven't fully tested it myself to really get it down, but some people say it helps and it seems to help my prompts.

You don't need Pony. They are two seperate models. Not sure about the exact workflow, but afaik any normal XL workflow should work, just switch the models. Comfy got too complicated for me when XL came around. Using Fooocus currently and it works like a charm with that.

@mrblue42 Ah I see. The info section on here said "This asset is designed to work best with the Pony Diffusion XL model, it will work with other SDXL models but may not look as intended" so i was a bit confused. Thought both model needed to be merged or use this as a refiner of some sort. But thanks for the heads up! I'll have to check out Fooocus myself.

@Autismix_anon can you tell more about uniform samplers?

@Autismix_anon Works pretty well indeed, thanks for that hint.

SDXL Default VAE (aka SDXL VAE) works fine for most folks.

Just do 10 repeats or maybe 15

S.D1.5 loras don't work on this since it's an SDXL model. Other SDXL loras might also not work that well with this model since it was trained on Pony Diffusion, which is very different from other SDXL models.

So basically use only PonyXL/AutismMix loras if you want good results.

That's kinda a skill issue on your end, most LoRAs are not compatible with PDXL. Check to ensure that the LoRA says "Pony" for compatibility.

@GogetaSSGSS3 @Autismix_anon  Yes, sorry for not being specific enough, i ment SDXL models in general, ones that were trained on probably models that are not PDXL.

@besughino They won't work.

PonyXL is its own base model, just like how SDXL is a base model for many other models like Animagine or Kohaku and so on. AutismMix is based on PonyXL, so you can only use PonyXL loras with it.

What you're trying to do is basically like trying to use a SD1.5 lora with an SDXL model, it just won't work.

have you tried putting earrings in the negative prompt?

@GenericGiraffe yes. I tried putting earrings, earring even with power :1.5, accessories. i tried putting no_earrings in the positives, but nothing. it keeps generating earrings every now and then

you can put jewelry in negative prompts too

Ponies were basically just used to create this model. Much like a bottle of glue

This a model based on PonyDiffusionV6. It's a model that was made for pony art but it turns out it was also trained to do a whole lot of Anime/Hentai stuff. It recognizes a lot of Anime/Hentai characters and artists. It also does SFW/NSFW concepts way better than any other SDXL/SD1.5 model.

It's kind of crazy how furries basically made the best Anime model for us.

@GogetaSSGSS3 We gotta owe them what's owed. Furries (and Bronies in particular) allowed us to step further into the AI tech.
So yeah, I guess Friendship is magic for science xD?

@The_rule_of_Thetra That's very true, unlike other anime fans who have way too big of an ego to work with eachother in order to make a good model they actually decided to work together and make a really good or possibly the best local model for Anime.

PonyV7 is in the works and it already has a dataset of over 25 million images, so it's a pretty big project. And this time they're gonna fix a lot of the issues that PonyV6 has, lets just hope everything goes to plan.

@GogetaSSGSS3 although i am invested im making AI art i lack much of the technical skills. i do wonder if 25 million images. how impressive is that for a model anyways? like whats the average training dataset.

@spheregod I'm not really sure since I've never really tried doing something myself, nor do I have the knowledge or resources for it but considering PonyV6 was trained on 2.6 million images it is a HUGE difference.

I remember seeing the creator mention that he spent more than $20,000 in order to be able to create a model like PonyV6, which is not a surprise considering the amount of GPU's/resources he needed in order to train a model on so many images (and the electricity bills were probably really expensive as well lmao). It's gonna take a couple of months before Ponyv7 is gonna be released, not to mention they still haven't decided whether they're gonna train the model on SDXL or the new SD3

@GogetaSSGSS3 It's not only a matter of images, it's also a matter of resizing (for the best results the images must follow some specific resolutions: if they don't fit you either resize them (losing quality) or crop them by trimming the edges with unnecessary parts (and that requires manual labor). There is also tagging, which is not that hard as a task to add, but it's daunting when you want to recheck the tags (there's ALWAYS a good chance some tags are wrong\missing).

@The_rule_of_Thetra not sure how that works exactly but I do remember seeing some of the admins of the PonyXL discord talk about that and apparently they went through thousands and thousands of images and they had to tag and check them manually, until they somehow automated that, I respect that honestly, that takes some serious determination.

Also, they did mention that out of the 25 million images they were gonna pick 10 million good quality ones for V7, which is still almost 4 times as big as PonyV6. Either way i'm excited to see what they're gonna do with this next version

Make sure your stack is up to date. What is your hardware and software setup?

Probably controlnet model is not for SDXL

@Volnovik it is the controlnet that comes with forge, But the models I downloaded were the same ones I used with SD, I couldn't find if there were specific models for the XL I used the ones from this link https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/main

@alucardhellsingn3119  Thats 11 for 1.5. SDXL are here:

https://huggingface.co/lllyasviel/sd_control_collection/tree/main

Try a different vae

@The_Generator I figured out the issue and it wasn't the vae.

@MadlyLaughing what's the issue then?

@SomethingWithAi I was using v1.6. I set up a parallel install for v1.7 then updated to 1.8 and suddenly everything is working. I suspect there's a support problem for Pony on v1.6 or a conflict with an extension I have installed on my 1.6 version. I'm still playing with it to figure out the issue, but it works fine now.

This is my list of non-anime femael characters
Also on this repository I added anime female characters that PonyXL recognise
Both lists are not full

characters_list/characters.txt at main · ertyased2/characters_list (github.com)
characters_list/anime_characters.txt at main · ertyased2/characters_list (github.com)

@ertyased thanks a ton!

@AshWasTaken if you know some characters that are not in the list I would be greatful if you tell me

@ertyased For sure if I find any.