🍥｜anime

[nova remnant]

Insane tech for the moment

[near elbow]

indeed

have some native 2048x896 fun

[nova remnant]

But sounds so outdated already

[near elbow]

yes, AI is moving at a crazy pace

[nova remnant]

(meanwhile me using a 2018 gpu)

[near elbow]

you need a 40990f fast computer graphics card, says the AI here

[nova remnant]

40 fast commputer

[near elbow]

or the 90cm long 4900?

[nova remnant]

40 :0

Big grabby

[near elbow]

gotta jump, cu tomorrow

[nova remnant]

@native halo you need this PC to generate the biggest longest spikier fake elf girl hair

[nova remnant]

I'm ready for class

[nova remnant]

Dragon teacher took all of her.. Personality

[acoustic locust]

Does anybody have an idea of when SD3 is coming out?

[nova remnant]

No

[native halo]

before SD4

[native halo]

https://tenor.com/view/overheat-computer-explosion-boom-spark-gif-17323916

[jagged summit]

[junior quest]

wut is dat modal?

[keen hollow]

oof I dunno about the yellow teeth...

[keen hollow]

is this that same pony one or whatever?

also, btw, the model specification in the prompt list worked! 🙂

ran a set overnight, and running another set now 🙂

[lyric zealot]

yes, it's the autismmixSDXL_autismmixConfetti, a pony mixed model

[lyric zealot]

that's great

[keen hollow]

I love that her shirt just says "emo"

[keen hollow]

i should run this through a prompt set and test it out

[lyric zealot]

please do, just know that you need to follow the prompting style of pony model

basically the score tag and maybe source_anime

[keen hollow]

oh... oof.. I dunno what that is. 😄

oops, didn't mean to combine that post and reply haha

[lyric zealot]

you can read the pony page for explanation

but basically, there's a set of tags that were trained on, and you should almost always use it for pony mixed model

and the tags are score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up,

link here if you wanna read more https://civitai.com/models/257749?modelVersionId=290640

^ might have nsfw pic, so take note

[keen hollow]

so images are limited to tag definitions?

seems like it might limit creativity in prompts

[lyric zealot]

those tags are trained as quality tags, so if you wanna put masterpiece, best quality, then use those tags instead

but well, you can do whatever if you want

it's just high quality pic were trained with those tags

[native halo]

even with all those tags still cant add that shiny look from the older models,seems like no one has trained that style

[nova remnant]

lick the shiny

[lyric zealot]

im helldiving, so will replay late

reply*

[keen hollow]

So do you add all those score tags? Or just the one that's like the cutoff point for what you'd want? Not sure i get this based on a quick read of this description.

[lyric zealot]

explain please

help explain to Omnia

[nova remnant]

[lyric zealot]

[nova remnant]

imagine you go on the site and can search by the highest rated art

thats what those tags kinda are

their score

when the model was training

and it was supposed to be able to do each one individually but during training the model just learned that score_9 to 4_up is one giant tag that makes good images

just a big tag that works like (high quality, best, quality)

[nova remnant]

https://tenor.com/view/zargoth2-gif-2289670819019502421

I love strawberries

👍

[native halo]

what about cherries?

https://tenor.com/view/enma-ai-anime-eat-berry-spit-gif-16294670

[vital raptor]

Actually, they’re not one big tag

They are individual and you don’t need them all, or any for that matter. You tend to get better results if you use some though

[nova remnant]

ugly anime

[native halo]

https://tenor.com/view/super-punch-punching-gif-15341848

[nova remnant]

`It's Training time
We now have annotated data and can finally train the actual Pony Diffusion. Let's keep showing the model images and our captions containing the score tags so it also learns which of the score tags correspond to which images, giving us better versions of "masterpiece".

But wait, turned out I messed up a bit! What I described above is how PD V5.X used to do things, in V6 I wanted to also be able to say - "hey, give me anything 80% good and up". But score_8 tag would only give us images in range 80* to 90%. Perhaps using both score_8 and score_9 would work but I wanted to verify that, so I changed the labels form simple score_9 to something more verbose like score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up and score_8 toscore_8, score_7_up, score_6_up, score_5_up, score_4_up. In reality I exposed myself to a variation of The Clever Hans effect where the model learned that the whole long string correlates to the "good looking" images, instead of separate parts of it. Unfortunately by the time I realized it, we were way past the mid point of training, so we just rolled with it (I did try to use shorter tags after the discovery but due to the way we train it didn't have as strong of an effect).

So, to summarize - we used a model trained on human preferences to label all data with special tags and then trained an text to image model on this labels allowing us to ask model for "good" images via use of these tags

Do I need to care
Maybe, in some cases.`

what the author said

[supple raptor]

As far as I know, training is still in progress. I suspect there will be a closed beta test of some sort before too long, and public launch some time after that. I would think maybe a 1-2 months until public launch.

[native halo]

this prompt is a mess,lemme clean it for u 🧹

[nova remnant]

yes

https://tenor.com/view/kok-cat-gif-25438805

[junior quest]

https://tenor.com/view/toge-gif-21923909

need a girl with that collar

too small collar

[vital raptor]

right, but the way auto1111 and other tokenizers work, it splits a phrase into recognizable words and converts them to tokens. While the entire string may in fact be trained as a single 'token' to be recognized, all the individual parts are also known. I tested it myself and each individual tag has it's own affect, along with the entire string

my point was that you don't need the whole thing, reguardless of how it was trained or what went wrong

all of the images i've been posting only use
score_9, score_6_up, score_5_up, score_4_up, source_anime,
because in my testing, score_8 specifically, and score_7 to a lesser extent, had more influence on an image than any of the others, and it wasn't a style I liked

others might like the way it looks, and that's fine, but you simply don't have to use them all, or any of them

if you're using a particular artist style or character name it usually matters even less as it's going to use the examples of that artists actual work it was trained on for reference, which in theory is already of the highest quality

but, as always, it depends on your taste/preference/prompt/settings/model/etc

and in reality, the only reason I argue the point is the same as the whole copy/paste word dump prompts people use on 1.5.. if you have a good model, it's already "masterpiece" quality, you don't have to prompt for that

[lyric zealot]

holy

[vital raptor]

with my custom 1.5 merge i've never used those keywords, i had a couple negative embeddings, no extra negative prompts and no lengthy prompting at all. IMO, if the model needs all that garbage, it's a bad model at this point.

[native halo]

(perfect hands,beautiful hands,5 fingers,not deformed,perfect slim and magestic hand:2.5)

[vital raptor]

score_9, score_6_up, score_5_up, score_4_up, source_anime,
1girl, petite, pointy ears, devil horns, black hair, purple eyes, choker, mischievous smile, elaborate gothic victorian dress, dark cathedral, dynamic pose, smug, looking down on viewer,```

```Negative prompt: sweat, maid,```

it does hands pretty well without prompting for them

[nova remnant]

bad negatives

You want sweaty maids 👍

[vital raptor]

lol

well, apparently maid outfits count as 'victorian dress' in pony/autism

and with the particular artist style combo i'm using, it adds sweat on her face in every image

[native halo]

thinking 🤔

[lyric zealot]

oops meant to paste the pic

[vital raptor]

this is without the negatives, sweat drop on her cheek and a maid-ish looking outfit

the sweat isn't as bad in this particular prompt, but i've just built it into the preset I have for that artist combo 🤷‍♂️

[lyric zealot]

rare guy generation

[cerulean bane]

someone had to do it.

[lyric zealot]

I will take on the job, for like 10 minutes

[nova remnant]

not grabby

[lyric zealot]

okay im done with guy generation

[lyric zealot]

grabby?

[native halo]

grab these gifts instead

[lyric zealot]

@nova remnant where you at?

Horror + deformed face

[nova remnant]

farming in LE, don't wanna see ai anymore

You killed my eyes

[lyric zealot]

[junior quest]

model?

[native halo]

flatpiececore

[terse ore]

What SDXL models generate eyelashes that are the same color as the hair?

Cuz, usually the eyelashes are black regardless of the hair color

[native halo]

try with just a prompt like ( white eyelashes:1.6), and (black eyelashes:1.5) in negative

[native halo]

my honest reaction

[supple raptor]

Why is Rem being arrested?

[native halo]

who is rem

[loud falcon]

howdy, is the lora folder for comfyui different than for "stable"?

[bold ether]

https://comfyanonymous.github.io/ComfyUI_examples/lora/
put them in the models/loras directory

[loud falcon]

tyvm

[loud falcon]

does comfyui prompt textbox accepts embeddings just like stable?

[steady grail]

no wait, you need to add stuff like embedding:name of embedding, or something like that

but there are custom nodes to autocomplete that

[loud falcon]

could you show an example?

[steady grail]

found it, it's confirmed, the text is embedding:name_of_embedding

[teal lake]

午后，明亮的阳光，房厅中，一个男孩(看书)，(摇椅)

[loud falcon]

ty

[junior quest]

hmm I cant create some characters in autism confetti

sadge

it doesnt have honkai starrail characters

A..nd I merged animagine v3 and autismix confetti, ehh but images are so bad for sm reason

[supple raptor]

此消息是人工智能翻译的。这里没有机器人。我们大多数人在自己的计算机上运行稳定扩散。

[supple raptor]

Pony based models (like Autismmix) are not very compatible with other models, hence the separate category on CivitAI.

[native halo]

just need a char lora

oh wait u are on XL

https://tenor.com/view/tomo-tomo-takino-azumanga-daioh-osaker-laughing-gif-8912491713735942546

[acoustic locust]

👍

[gloomy scroll]

Oh no Yue is falling

[signal panther]

a glaz drakona superhero armor with nuclear symbols on it, cute, no helmet, girl, cute, ideal body, metal skirt, dtailed realistic Dungeon Siege III background, cute, beautiful, gorgeous, pink hair, seductive pose

The prompt)

It's so hellish

[vital raptor]

she'll catch yue!

[gloomy scroll]

[vital raptor]

man, when regional prompter works, it's so good

but when it doesnt...

[gloomy scroll]

What happened to Yue?

[vital raptor]

🤷‍♂️

i'm sure it's completely innocent

[gloomy scroll]

I agree

[vital raptor]

the regional prompter just blending them is adorable 😄

[gloomy scroll]

Just reminder that Yue likes to hang out with girls so take good care of her

[vital raptor]

she's in good hands 😄

[nova remnant]

Imagine not having a 4090 or something like that to train xl Loras

[nova remnant]

Which one

[vital raptor]

hard to imagine tbh 😏

[nova remnant]

Smh my head

[gloomy scroll]

[sage mountain]

damn stochastic rounding a game changer for bfloat16 training

[vital raptor]

man, change one word in a prompt and now i can't post anything here

[nova remnant]

You can send to me

[vital raptor]

😄

[nova remnant]

for science

[loud falcon]

how much strength model/clip you guys use for loras with comfy?

[steady grail]

depend from how many lora i load

i try to maintain a total ration of 90 100% and than play with any ratio so how much i can push them without burning the image or confusing the lora

except for lora like lcm, that need a certain weight to do their job

[vital raptor]

depends on the lora and model really, some work best at like 0.3, some are best at 1.2, depends on how well it was trained

[junior quest]

anyone

genshin characters are fine but starrail nope

[nova remnant]

Just train a lora then

👍

Or don't gen those characters

They are stupid

[junior quest]

loras not good when I train them

[junior quest]

souka

[nova remnant]

Sasuke

[junior quest]

tried to generate that purple girl (kafka), but autismmix dun know her

[nova remnant]

@pure oyster

[nova remnant]

I can help her stare into the sky

[void galleon]

i typed "Deathstare"

[lyric zealot]

[gloomy scroll]

❇️

[junior quest]

https://tenor.com/view/dr-strange-doctor-benedict-cumberbatch-gif-23561975

[junior quest]

https://tenor.com/view/sparkle-hanabi-hsr-star-rail-penacony-gif-14796476064034012533

[acoustic locust]

have any of you tried the new ideogram model?

it's better than dalle-3

[hidden lotus]

Does anyone know how to get animal ears on the side of someone's head? Like is there a lora or special prompt for that?

[nova remnant]

really

I played with ideogram a bunch before they put a queue into it and a limit

and it was okay, some good text

[acoustic locust]

I definitely think text is it's strong point

[nova remnant]

new model doesn't know 2b

no more 2b from behind, life is bad

Its definitely really good

miles better than before

can't say if its better than dalle3

dalle3 is so bad to use that I think it is lmao

[gloomy scroll]

You seen Mooey's Dalle images?

[hidden lotus]

that looks cozy

[gloomy scroll]

Asserting dominance

[nova remnant]

that is through the api right

I'm pleb user

[nova remnant]

too smol, no dominance asserted

[gloomy scroll]

I believe that's just made with gpt4

[hidden lotus]

making miss library from misao without a lora is going to be a painnnnn...

target and result

[supple raptor]

https://tenor.com/view/touhou-yell-anime-gif-5047799

[terse ore]

What's the difference between a regular model and an inpainting model, for inpainting?

[nova remnant]

Yes...

something about information in them that makes them better for inpainting I have no idea I don't inpaint I just click generate again

I'm too to notice a difference in them

[bitter latch]

I think batman might wanna turn around

[native halo]

no skills?

[nova remnant]

If she looks at me like that ...

https://tenor.com/view/kok-cat-gif-25438805

[junior quest]

her forehead is big

[nova remnant]

Just put a paper bag over her and it's good

[native halo]

thats the point of the meme,junior

[junior quest]

[nova remnant]

Needs more tightness

[native halo]

like this?

[nova remnant]

[junior quest]

shes a child

so nope

[native halo]

https://tenor.com/view/youre-gif-20173955

[junior quest]

Im?

[native halo]

my 1060 is older than you

https://tenor.com/view/joeyy-movie-1man-wopn-jxxyy-gif-27313120

[junior quest]

aniki

[nova remnant]

Me when the tea I make tastes nice

[junior quest]

now thats a win

[junior quest]

gib me prompt

[south sable]

||`Morrigan Aensland,

1girl, bangs, black background, black jacket, black shirt, black shorts, blue eyes, blunt bangs, cutout, clothing cutout, crop top, eyeshadow, fangs, gradient hair, grey hair, hair rings, jacket, light purple hair, looking at viewer, makeup, multicolored hair, navel, open mouth, red eyeshadow, shirt, shorts, sleeveless, sleeveless turtleneck, sleeveless turtleneck crop top, solo, stomach, turtleneck, turtleneck crop top, two-tone hair, vampire`||

[vital raptor]

i need to make an emote or something for prompts like that

dunno why it triggers me so much

[south sable]

Idk why you would care either

[vital raptor]

1girl, bangs, black background, black sleeveless turtleneck, croptop, black shorts, blue eyes, blunt bangs, (clothing cutout:0.9), midriff, fangs, gradient hair, grey hair, hair rings, light purple hair, red eyeshadow, two-tone hair, vampire```

probably something to do with my IT career and doing so much tech support and customer service over the years, it bothers me when people get misinformed, and then end up spreading the same information and methods to other new people

Not saying it's i8ntentional or anything, youtubers are mostly to blame since none of them bothered to learn anything before making a bunch of videos on it

[south sable]

thats a lot of assumptions. I'm just sharing the prompt. I dont recall spreading any misinformation.

im using old models, old prompts, and old workflows. idk. the guy wanted to prompt

[vital raptor]

I'm not saying you are, just that it has happened

[south sable]

always awesome to get pointed at when having fun...

[vital raptor]

even with older models, repeating the same words multiple times doesn't do anything

[south sable]

okay, thanks.

[vital raptor]

Like I said, I'm not saying you're at fault

I'm saying prompts that are structured that way bother me because someone, at some point, has spread misinformation on how to prompt

and it's not helping other people to create what they're trying to create, it just makes things harder

[south sable]

next time i'll let them know I'm just going to cause them more problems if I share the prompt.

[vital raptor]

ok, i guess just be offended then 🤷‍♂️

I never said not to share the prompt, i said the prompt bothered me

Either you learned something, or I'm just some random dick on the internet that made fun of you

doesn't matter to me really

[south sable]

yeah

i just hate coming on this thread because there's always someone who has to do that while I'm just trying to share. but its cool. I wont spread misinfo and ruin people's gens.

[vital raptor]

maybe you shouldn't be so quick to get offended over it, it has nothing to do with you or what you're sharing, sorry for the attempt at being educational

[south sable]

yeah

[native halo]

it would be better if you stop typing

[south sable]

https://tenor.com/view/bocchi-bocchi-the-rock-bocchi-the-rock-gif-sad-worried-gif-14499929766418002003

[native halo]

https://tenor.com/view/anime-azumanga-daioh-shocked-chiyo-gif-18944897

[signal panther]

Prompt, please.

[native halo]

1girl,pink hair,very long hair,green eyes,(latex:1.5),nurse outfit,thighhighs,garter straps,(arms behind back:1.4)

[signal panther]

Tahnx

[bitter latch]

bing

[supple raptor]

New version of kohaku-xl is nice. I think it's getting released very soon.

[zinc flax]

I think I've fully switched to pony xl now lol
It's super nice with style loras

[silent kraken]

Lora? Or model. It feels like it's a style specific model

[silent kraken]

Ohh it's gradient hair😂😂. Here I am yapping about models and loras

[supple raptor]

dark scenes and backgrounds are my biggest complaints about Pony XL

[zinc flax]

I always prompt white backgrounds so I can't comment on that lol

[supple raptor]

Kohaku-XL-Delta is released now btw https://huggingface.co/KBlueLeaf/Kohaku-XL-Delta

[zinc flax]

So many xl models coming out lately

[supple raptor]

Hopefully Kohaku's training methods will catch on because they were able to train this model on 2x3090's. Pretty much all of the other big XL models have needed A100/H100's to train in a reasonable amount of time and most of them still don't handle less common tags as well as kxl-delta.

[gloomy scroll]

Wouldn't sharing this be better since it tracks downloads and reviews?
https://civitai.com/models/332076/kohaku-xl-delta

[supple raptor]

I suppose so. Kohaku used both links in the announcement posts they made in other servers.

[gloomy scroll]

Yeah I understand why some prefer hf but imo for casual users CivitAI is better for visibility

[zinc flax]

Oh it's a merge

[gloomy scroll]

Nice

[zinc flax]

Well, trained base model + merge

[supple raptor]

It's a merge in the sense that the bulk of the training was done on a LoKR and that was merged into the model instead of training it directly into the model which is less efficient. That's the trick that allowed them to pull it off on relatively weak hardware.

[native halo]

from the sample imgs it looks wobbly and the background looks less detailed prob due to the low training

[wet elm]

Hey guys, ive been using SD 1.5 for a long time but i want to try out anime SDXL, but i dont know how to use it, so i have some questions:

• what is the refiner? Is it needed for anime
• do 1.5 loras work on sdxl?
• do 1.5 negative embeddings work on sdxl?
• is the tagging style the same as 1.5? Like booru tags
• Is hires fix necessary in all images?
• Is the VAE the same?

Tysm for the help.

[zinc flax]

• I don't think you need any refiner for the newest xl models
• 1.5 loras don't work on XL
• I don't think so
• Newest trained anime model use booru tags fortunately
• I use 832x1216 and it already looks great imo, no need for hires
• Most models already have the VAE baked in so no need to download

[supple raptor]

It's got a ton of artists trained. Want a different style? Use a different artist tag. I haven't tested them all yet, but there are probably more than 400 that work well.

[native halo]

yea its here,still looks wobbly

[supple raptor]

That's hardly any of the trained artists. I've used ones with as few as 500 instances with good results. Haven't tried lower than that yet. https://huggingface.co/KBlueLeaf/Kohaku-XL-Delta/blob/main/docs/artists-kxl-delta.json

[native halo]

well if u can gen an img that looks clean and has good lineart post it here to test some prompts with it

[bitter latch]

I have the day off i'm going back to bed

[wet elm]

Ty! Is there a way for me to separate 1.5 loras from XL? like different folders idk
Also, is using 1024x1536 bad for XL?

[zinc flax]

I don't think you can separate the lora folders in webui. And yeah, that reso is completely fine

[steady grail]

yep, inside lora folder, you can place the lora in subfolder

[gloomy scroll]

[keen hollow]

: wave :

my gpu has been running for like 48 hours 😛

[signal panther]

Prompt, please.

[bitter latch]

anime, cartoon style, Nurse with pink hair, japanese nurse uniform, hands on hips, smiling, detailed character sheet

[nova remnant]

That's what she said

[void galleon]

so i got a bunch of ghosts in my new picture 😦

#1019361238234443776

Spooky

[nova remnant]

Watchu doing

[nova remnant]

Would eat

[grizzled moat]

[nova remnant]

more food

Jelly bunny

[grizzled moat]

[lyric zealot]

Just to add on, if you have a SDXL based model selected, then in auto1111 webui Lora tab, you will only see those Lora that are for sdxl based. And also, I saw you asking about refiner, basically refiner is a concept of changing model mid generation. You can even try doing that with sd1.5 model. Example, use one model to generate good human anatomy, then use second model as refiner to change style to that model. The refiner is based on ratio, so if it’s 0.5, means at half the steps you set, it will use the refiner model you set.

[nova remnant]

https://tenor.com/view/trigger-cat-gif-18390951

Refiner is very cool

Can make good looking... beach landscapes

👍

[lyric zealot]

[nova remnant]

[lyric zealot]

[gloomy scroll]

Only if loras have proper metadata

[wet elm]

what is the best webui to use with SDXL anime? I started using xl today and im using auto1111

[wet elm]

cool! is there any config to change from 1.5 to xl? like clip skip

[lyric zealot]

Yup true true

[lyric zealot]

As far as I know, clipskip doesn’t matter for sdxl

To change to using sdxl model from 1.5, it’s probably as easy as changing the model and vae

[wet elm]

tysm, i thought i needed to change a lot, is there a way to speed up the gens a bit? im getting like 3 mins per img (i have a 3060 12GB)

[lyric zealot]

On 1.5? Or sdxl?

You probably need to use xformers

[wet elm]

xl

[wet elm]

im using it

[lyric zealot]

If it’s xl

Might be you need medvram for xl

[wet elm]

hmmm yes that could be it

maybe if i lower the resolution too

[lyric zealot]

What res are you genning at?

And hires scale?

[wet elm]

im genning at 832x1216 no hires, i just copied some random image in civitai to test.

[gloomy scroll]

Forge or Comfyui/StableSwarm

[wet elm]

Forge? never heard of that, is it too different from auto1111?

[gloomy scroll]

It's just more optimized a1111

[lyric zealot]

Basically a1111 but faster

[wet elm]

nice, does it have extensions like adetailer?

[lyric zealot]

You can add it like a1111

[gloomy scroll]

It uses a1111 extensions although not all work but adetailer works

[lyric zealot]

Should not be 3 minutes for that then

[wet elm]

nice! can i transfer all my extensions and stuff to it? (sorry for making too much questions lol)

[lyric zealot]

Maybe close to 1.30

[wet elm]

yeah now i got 1 min 20 secs, i was running some apps in background

[gloomy scroll]

You could probably just copy paste your extension folder but haven't tested myself

[lyric zealot]

I don’t think that will work, not sure tho

There’s quite a few built in extension like controlnet and such

[wet elm]

im going to try it out, tysm!

also, do you guys have any recommendation for models, loras, or negatives i could start with? idk if my old 1.5 negatives will work neither do i know how to prompt in sdxl lol

[gloomy scroll]

I think webui just detects each folder inside extension folder and loads it if it has initialization script

[gloomy scroll]

Also neko Yue requires pats

[lyric zealot]

hopefully so, but if you wanna try that, need to remove stuff like controlnet and such

[gloomy scroll]

maybe not since it disabled extension if it has integrated version

At least multi diffusion can't be enabled

[lyric zealot]

currently, ponyDiffusion is a pretty popular SDXL based model, despite the name, it's pretty good at anime

like 1.5 model, every SDXL based model might need different kinds of negative prompt

the model im using is autismmixSDXL_autismmixConfetti, which doesnt really require much negative. The one im using is (worst quality, low quality:1.2), 3d, censored, abdominal muscle, artistic glow

[wet elm]

i saw one called AutismMix and it looks great, im going to try it.

[lyric zealot]

ohh, then might be worth trying it

[wet elm]

oh nice, tysm!

[lyric zealot]

can confirm, it's my normal model now, even when i only have 8gb vram

[wet elm]

how long do your gens take? also what webui are you using

[lyric zealot]

but since they are based on PonyDiffusion, you need to follow a specific tags as it's trained on that

[lyric zealot]

forge. for 768 x 1024, hires fix 1.5, 24 steps normal, 10 steps hires, about 50 seconds

[wet elm]

damn thats fast

[wet elm]

what tags?

[lyric zealot]

They made a mistake in training, and instead of just score_9, it's trained on score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up instead. So use the latter

and there's rating_safe, rating_questionable and rating_explicit

also source_anime

[wet elm]

oh so like quality tags

last question (i hope so) Is the prompting too different? should i still use like
1girl, blonde hair, long hair, blue eyes ?

[lyric zealot]

yup, for most SDXL model, it's still really danbooru tag

and no worries about asking question

[steady grail]

guys, i'm introducing myself in the world of training lora, someone know a good guide?

[wet elm]

i saw this in the forge github page, is it too difficult to do this? I wanted to keep everything (sorry for the ping)

[lyric zealot]

I didn't use this method, so not sure. If you install forge normally, you can opt to use models/lora/embedding and others from automatic1111 webui

by changing the webui-user.bat, which is what most of us did

[wet elm]

thats what i want, could you give a brief explanation on how to?

[lyric zealot]

once you got forge downloaded, and ran it once. You can open up forge webui-user.bat and edit it, mine look something like this

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=

@REM Uncomment following code to reference an existing A1111 checkout.
set A1111_HOME="D:\Downloads\A1111 Web UI NEW\stable-diffusion-webui"
@REM
@REM set VENV_DIR=%A1111_HOME%/venv
set COMMANDLINE_ARGS=%COMMANDLINE_ARGS% ^
  --ckpt-dir C:/Users/User/OneDrive/Desktop/Stable_Diffusion_Models ^
  --hypernetwork-dir %A1111_HOME%/models/hypernetworks ^
  --embeddings-dir %A1111_HOME%/embeddings ^
  --lora-dir %A1111_HOME%/models/Lora^
  --vae-dir %A1111_HOME%/models/VAE^
  --realesrgan-models-path %A1111_HOME%/models/RealESRGAN^
  --esrgan-models-path %A1111_HOME%/models/ESRGAN^
  --no-gradio-queue^
  --cuda-stream^
  --pin-shared-memory^
  --api

call webui.bat

obviously use your own path to your own ckpt-dir from auto1111 and such

you might not need no-gradio-queue and below, but that's my own stuff

[wet elm]

so basically you do --ckpt-dir "path to the model folder"
--hypernetwork-dir "path to the hypernetwork folder"

and so on?

[lyric zealot]

yup

if you open up your forge webui-user.bat there should already have a format for you

but there's some stuff lacking, like if im not wrong, vae and below is added by myself

[wet elm]

ok, tysm again!

[lyric zealot]

[nova remnant]

@native halo

[wet elm]

im having a problem, when i run the "run" file, it gives me the problem of unrecognized argument

[lyric zealot]

show me your webui-user.bat

[wet elm]

`@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=

@REM Uncomment following code to reference an existing A1111 checkout.
set A1111_HOME= F:\Stable Diffusion\stable-diffusion-webui
@REM
@REM set VENV_DIR=%A1111_HOME%/venv
set COMMANDLINE_ARGS=%COMMANDLINE_ARGS% ^
--ckpt-dir %A1111_HOME%/models/Stable-diffusion ^
--hypernetwork-dir %A1111_HOME%/models/hypernetworks ^
--embeddings-dir %A1111_HOME%/embeddings ^
--lora-dir %A1111_HOME%/models/Lora^
--vae-dir %A1111_HOME%/models/VAE^
--realesrgan-models-path %A1111_HOME%/models/RealESRGAN^
--esrgan-models-path %A1111_HOME%/models/ESRGAN^
--no-half-vae ^
--no-gradio-queue ^
--api
call webui.bat
`

[lyric zealot]

try without --no-half-vae

oh

and i think

since your A1111 home directory there's a space, you need to enclose it in " "

set A1111_HOME= "F:\Stable Diffusion\stable-diffusion-webui"

[wet elm]

i tried and got this error:

[lyric zealot]

did you try putting it in " "

[wet elm]

yes

[lyric zealot]

another thing to check, see if your have those folder in A1111

[wet elm]

yeah i do

[lyric zealot]

[wet elm]

wait, maybe its because i have \ in the directory but its / in the webui-user.bat

idk

[lyric zealot]

send your current latest one again

[wet elm]

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=

@REM Uncomment following code to reference an existing A1111 checkout.
set A1111_HOME= "F:\Stable Diffusion\stable-diffusion-webui"
@REM
@REM set VENV_DIR=%A1111_HOME%/venv
set COMMANDLINE_ARGS=%COMMANDLINE_ARGS% ^
--ckpt-dir %A1111_HOME%/models/Stable-diffusion ^
--hypernetwork-dir %A1111_HOME%/models/hypernetworks ^
--embeddings-dir %A1111_HOME%/embeddings ^
--lora-dir %A1111_HOME%/models/Lora^
--vae-dir %A1111_HOME%/models/VAE^
--realesrgan-models-path %A1111_HOME%/models/RealESRGAN^
--esrgan-models-path %A1111_HOME%/models/ESRGAN^
--no-gradio-queue^
--api
call webui.bat

[lyric zealot]

I don't see what's wrong

[wet elm]

me neither

[lyric zealot]

what's the error, can you copy and paste here?

[cerulean bane]

oh no, did you play around with the code and now the code keeps playing you?

[native halo]

https://tenor.com/view/anime-gif-16572778167231286334

[nova remnant]

If I code a cute goth girl

Will she play with me?

That sounds like a good plan

[cerulean bane]

https://tenor.com/view/big-titty-goth-gf-bongo-cat-choices-end-world-hunger-end-world-poverty-gif-16193827

[wet elm]

ok, i will be sending in private just so it does not clutter the chat

[keen hollow]

just running a boatload of prompts against different models, tryin' things out

[nova remnant]

[lyric zealot]

So SD3?

[native halo]

after we get proper controlnet tile and shuffle in XL 👍

[lyric zealot]

wtf is controlnet tile and shuffle

[native halo]

https://tenor.com/view/amen-gif-23345354

[lyric zealot]

https://tenor.com/view/noted-anya-gif-9118962738426881203

[nova remnant]

We need less math and graphs

And more cow girls

[native halo]

i like the cowgirl

[lyric zealot]

[bitter latch]

Wolverine Miku Mouse

prompt: anime-cartoon Style, black outlined, vivid colors, soft shadows , far wide view, Anthromorphic Blu hair Girl-Mouse with Claws of Wolverine with body covered in silver cybertech chainmail, Miku Hatsune Anime themed background , extra detailed, ultra sharp, ultra science-fiction

[vital raptor]

if you haven't figured it out, you need a space between the path and the ^ character

[nova remnant]

rare comfy post

[gloomy scroll]

Comfy spreading greatness of sd3

[flat obsidian]

#🍥｜anime girl with dog face

[wet elm]

I figured it out, but ty anyways!

[gloomy scroll]

Pancakes for everyone

[junior quest]

waffles

[wooden coral]

Pfannkuchen

[junior quest]

gimme that pose

name

[steady grail]

training lora is hard 😦

i can't imagine how much harder must be to train models

[nova remnant]

I just told SD I wanted goth dragon girl

I'm sure there is a pose like that

[junior quest]

eh.. come on I need that pose... sadge

[bitter latch]

hmm it's fuzzy

there we go, A passender on the Star Rail train

used a TurboXL model

[steady grail]

tubo is a little washed out?

weird, when i used it on pony i didn't get too much of a difference

[bitter latch]

could be the custom model

[nova remnant]

booba

[nova remnant]

[nova remnant]

what model is that

looks similar to something I know

[void galleon]

A mix mash of 3 different models in different ratios

im testing a few of em

AnimagineXl3 As the Second one with a 0.7 ratio towards it

the first and third ones can be anything

with the basic Animagine Lora

[nova remnant]

@void galleon model putting the light where it matters the most

👍

[void galleon]

what helped the most was Finetuning prompts

and learning how

[void galleon]

(Anime style:3) is mostly how i get my style

[nova remnant]

holy

thats a lot of weight

[void galleon]

Honestly, needs it, theses ai's aint no chatgpt, they don't know their own weight

[nova remnant]

@lyric zealot you can fix her

[bitter latch]

ok turboxl sucks

back to normal models

[nova remnant]

@void galleon this model just breaks if you put 3 weight on something lmao

[supple raptor]

Not sure if this was related to the message above it or not, but if so, I increased contrast and lowered brightness on the sdxl vae specifically for Pony and AutismMix to help with them being a bit washed out. https://huggingface.co/nubby/blessed-sdxl-vae-fp16-fix/blob/main/sdxl_vae-fp16fix-blessed.safetensors

[steady grail]

no idea

i haven't deepened xl too much

[bitter latch]

suppose to be voltron, hmm looks like the base

[nova remnant]

Every mech is the same

I learned it from @native halo 👍

[bitter latch]

sighs, i'm getting bored with ai art

and games

just finished a 16 book series

so .. bored

[nova remnant]

@primal wharf you here

rare mike

[primal wharf]

Ricardo Milos

Wasup

[nova remnant]

[wispy canopy]

what miku do you identify with?

3-3 is very funky tho, love it

[lyric zealot]

I will fix her

[junior quest]

[pure oyster]

looks like a modded skyrim character

[bitter latch]

well that got random

ok that one style is jus random

in a good way

skeleton

[junior quest]

kinda like bleach but nah

[bitter latch]

there we go

and in his wake, his little brother

[nova remnant]

She's smol

[wet elm]

do i still need to put lycoris in a different folder than lora?

[junior quest]

nope

at least for me

[wet elm]

tyy

what sampler do you guys use for ponydiffusion?

[nova remnant]

these are all with 4 steps and cfg 1

with sd3 might not have to bother about steps and cfg

[wispy canopy]

depends.

dpm2m karras if i want the style to pop more, ddpm for more bodily consitency, euler SGMuniform if using lightning lora

[wet elm]

so dpm 2m karras for general use?

[wispy canopy]

yeah

[wet elm]

ok tysm, so after all it does not differenciate a lot from 1.5

[wispy canopy]

but sde, 2sa karras, 2m, 3m sde, all have a small different outcome

[nova remnant]

also most of the lightning models out there suggest using dpm++ sde

[wispy canopy]

and depending on cfg/other things, it might not work as well

[wispy canopy]

if you have forge -> try the SGM uniform samplers

[nova remnant]

yes i heard about it but i dont use forge

[wispy canopy]

or the turbo samplers

they're actually the same sampler, but different schedulers

[nova remnant]

i updated the a1111 some mins ago, but i dont see sgm uniform added yet

[wispy canopy]

because it's not in default a1111

[nova remnant]

its got some other stuff added

they included LCM w/o animatediff

[wispy canopy]

finally xD

[nova remnant]

yeah finally lol

[native halo]

u can just install a script to enable the use of different schedulers like uniform on any sampler

[nova remnant]

interesting, is there anything in particular that's special with uniform samplers?

so far im getting good results with dpm++ sde or dpm++ sde karras

and those are with lightning models

[native halo]

i dont know i dont use lightning🗿

[nova remnant]

those are along the line of lcm turbo

low steps and cfg

i had little quality loss with lcm which is not so much with lightning

[wet elm]

im still struggling to see whats the best resolution to use with pony

[wet elm]

ive tried 768x1024 1.5 hires but idk something feels wrong

ive tried 1024x1536 no hires and things get stretched lol

[nova remnant]

try 896x1152

[wet elm]

with hires?

[nova remnant]

you can but i dont use hires with that when using xl

[wet elm]

oh it actually looks nicer

yeah this is better, ty

[nova remnant]

you could also try the generic 768x1024

[vital raptor]

being sdxl, 1024x1024 is the 'best' resolution, different resolutions are going to be deviations from what the model is native to, that said 768x1024 works fairly well though

[native halo]

you can try hi-res fix at 1.5 it fixes some of the wobbly lines but not all of it

[wet elm]

i tried, but some details felt like out of place and some artifacts came with it

im still figuring things out tho, i only started using sdxl yesterday, so if any of you have tips for settings and parameters

[nova remnant]

just so you know, you can even use 512x768 then hires fix by 2x

XL would parse it fine

[vital raptor]

depends on the model, for pony based models cfg 7, 30 steps, 1024x1024 using euler a works fine for me

also depends on the style and look you want

[nova remnant]

for example this is the output i get with 512x768 and hires

[native halo]

if u try a bunch of different resolutions,different samplers and different steps and a minimal prompt and still comes out wobbly or deformed then its the model

[vital raptor]

unless it's pony, then it's your prompt 😄

[wet elm]

it was probably the resolution, looks fine to me now

[wet elm]

ty! for a do you use hires?

[wet elm]

yeah, i reutilized an old prompt from 1.5 for testing and it came out... kinda weird

i still need to learn the prompting of pony

[vital raptor]

when i use hires fix it's nothing special, 4x-Ultrasharp upscaler with 7 steps and 0.25 denoise

[nova remnant]

btw the reason i chose to use 896x1152 it kinda happens to give more content space

[wet elm]

how?

[nova remnant]

thats in comparison to using 5121x768

[vital raptor]

technically i could use fewer hire steps since i'm not using it to 'clean up' the image as much as just a clean upscale, but on my gpu the time difference is trivial

[wet elm]

so basically no upscale just to clean up?

[vital raptor]

other way around, i 2x upscale, low denoise so it doesn't change much

[keen hollow]

evenin'

[wet elm]

oh so you lower the resolution a lot

[vital raptor]

no..

[wet elm]

its almos 1 AM for me lol

[vital raptor]

i gen at 1024x1024 and then upscale 2x

[keen hollow]

super evenin'?

[wet elm]

oh ok, that makes sense

[wet elm]

yeahhh

[vital raptor]

i have a 4090, so i gen at the full resolution, otherwise there's no real point (for me) to using sdxl

[wet elm]

damn how fast is that?

[vital raptor]

~5.3s for a 1024x1024 sdxl gen

around 3s using TensorRT

[wet elm]

what.

thats so fast lol

[vital raptor]

15.9s with 7 steps of 2x hires

[wet elm]

my 3060 gens like 55 seconds for each

[nova remnant]

55 seconds for sdxl image?

i wonder what setup you are using

[native halo]

prob too many steps or using a1111 on forge is faster like 30secs

[wet elm]

yeah, probably because i normally have some stuff in bacgkround

[wet elm]

im using forge with 25 steps

[nova remnant]

well lightning models are super fast which you might consider

[wet elm]

what are those?

[nova remnant]

im using rtx 3060 and it takes me about 2ish seconds to gen 896x1152

[native halo]

trade quality for speed model

[nova remnant]

quality arent really bad

[native halo]

no its just less

[wet elm]

damn thats cool

any way that yall know to speed up the images a bit?

[native halo]

yea money

[nova remnant]

the recent lightning models from community are pretty amazing as far as quality goes

[wet elm]

lol

[wet elm]

but any like extension or anything

[native halo]

just tensor RT but u need to compile some files for each model and also only works with a single lora

[lyric zealot]

In early, unoptimized inference tests on consumer hardware our largest SD3 model with 8B parameters fits into the 24GB VRAM of a RTX 4090 and takes 34 seconds to generate an image of resolution 1024x1024 when using 50 sampling steps. Additionally, there will be multiple variations of Stable Diffusion 3 during the initial release, ranging from 800m to 8B parameter models to further eliminate hardware barriers.

[vital raptor]

only way to speed things up is to sacrifice something, speed, money, quality, or resolution

[vital raptor]

it works with multiple loras now, and you don't have to build a huge engine for each

but some loras won't convert properly for some reason, it's pretty buggy

[nova remnant]

im curious about the other variations that would supposedly eliminate the hardware barriers 🙂

[wet elm]

yeah i dont really need too much speed up, as i usually use SD in background when im doing something else

[keen hollow]

not sure if @native halo will understand this one...

[native halo]

yea praying nvidia adds that to their drivers instead 🗿

[lyric zealot]

hopefully details soon

[native halo]

im just a cook

[vital raptor]

it'll be similar to SD cascade

[wet elm]

anyways, im going to sleep now, ty so much for your help guys

[vital raptor]

just different low memory versions that have less quality/coherence

[keen hollow]

nice

[lyric zealot]

is tensorRT on forge or still bugged on it?

[vital raptor]

haven't checked in a few days, i doubt they've fixed it

[lyric zealot]

yup it's definitely not fixed

[vital raptor]

i opened the issue ont he git hub for it, so i'd imagine i'll be notified if it ever gets closed, or even replied to

a few replies, but not from the devs so i don't expect much

[lyric zealot]

[vital raptor]

from what i can tell, it's usage on a1111 is pretty small since the installer doesn't even work properly

most people give up trying to get it working

[brisk marsh]

Hey @vital raptor , are you available to help me with AI Grillz lora model training?

[native halo]

i like grills too

[brisk marsh]

yeah so I am gonna train the lora model to convert selfie to grillz image

[vital raptor]

🤨

[native halo]

fat grills

[brisk marsh]

Are you interested?

[vital raptor]

in grills? I like bbq, but not really..

[brisk marsh]

If you can, I am willing to pay for that

[native halo]

https://tenor.com/view/lets-party-lets-get-this-party-started-grilling-grill-dance-gif-12600028

[vital raptor]

grills 😄

wait, "grills" are supposed to have bling

[native halo]

(paul wall:2.5)

[vital raptor]

looks like the broken sdxl math was not fixed in 1.8 like it was supposed to be

so that's cool...

[vital raptor]

sorry, didnt realize i didn't actually answer you 😄
I might be interested if you have a dataset already, feel free to DM me what you're wanting trained with some examples. If you've got money to put towards it though you might look into training on civitai or something, it's not generally too expensive to train on better hardware than mine

I'm heading to bed now though, so I'll have to get back to you sometime between sleep and work

[supple raptor]

you want resolutions that equal roughly the same pixel count as 1024x1024 because that's the native resolution which all of the buckets were based on. here's a cheat sheet.

[junior quest]

candy house

[wispy canopy]

even on pony, eface will find a way to do lyrics prompting 😄

pony sucks at lyrics prompting

[wet elm]

yoo tysm, that helps alot

[glacial bane]

what style is this?

[wooden coral]

Anyone else switched or switching to forge?

[gloomy scroll]

I have been using forge for a while since it gave me ~40% speed increase with XL on my 1070 8GB

[gloomy scroll]

(by mdf an:0.8), kashu \(hizake\),

1girl, small breasts, long hair, straight hair, white hair, blunt bangs, azure eyes, pointy ears, long eyelashes, choker, blush, witch hat, black [dress|robe|dress], mage, (smug:1.1), smirk, fang out, exaggerated expression, outdoors, nature, standing, magic circle, holding staff,

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_anime, sketch, graphite \(medium\), rating_safe <lora:sdxl_edge_enhancer:1>
Negative prompt: long hair, sweat, 3d, slit pupils, source_cartoon, source_furry, source_pony, painting, (worst quality, low quality:1.2), lowres, jpeg artifacts, scan artifacts, (abstract:0.9), chromatic aberration, unfinished, blurry, monochrome, censored

[wooden coral]

My speed improvements on a 4090 are not huge but still there. And it loads up way faster

[gloomy scroll]

Well start up is faster since it loads model after you click generate first time

[wooden coral]

Maybe its also cause my A1111 install is over a year old and I never started fresh

[gloomy scroll]

If it works don't fix it

[wooden coral]

I installed torch 2.2.0 on A1111 a while ago and that was already a pain in the butt

So yeah I dont live up to that

[bitter latch]

seems FB is currently having major issues

[native halo]

cool it should spread to twitter/youtube/tiktok

[bitter latch]

shrugs

according to down detector seems to have

seems to be mainly meta services

ahh ok seems there was a leak of a database that exposed alot of data

facepalms

well someone's having fun and it aint u

us

[native halo]

https://tenor.com/view/standing-goat-goat-my-honest-reaction-my-honest-reaction-meme-stupid-goat-gif-25856650

[nova remnant]

Almost looks as if you don't like them

[bitter latch]

gives me idea's 😄 evil grin

facebook haha

facebook

"my facebook has no face"

Miku Book

facebook right now

hehe baby miku face book

[wet elm]

what model/style is this?

[cerulean bane]

why no meta data

[nova remnant]

from what I know orax inpaints images and stuff to make them cooler

so no metadata

[steady pumice]

yes

[steady pumice]

it's autism mix and the granblue fantasy lora

[nova remnant]

@vital raptor cleanest prompt ever
You might like it

[supple raptor]

Mmm... Maybe don't send prompts that are intended to generate cheese pizza.

[nova remnant]

💤

[vital raptor]

?

[supple raptor]

it was a poorly written prompt that got deleted because of the type of content that it was intended to generate.

[vital raptor]

ah, i see

[keen hollow]

hope everyone is having a nice day

[vital raptor]

I'm home sick with a fever and it's been snowing for ~16 hours, but otherwise 👍

[supple raptor]

pretty good so far. haven't been up for very long yet.

[nova remnant]

@native halo what do you think about bunny girls

[vital raptor]

bunnygirls are ok

[terse jacinth]

i came here to bump @wispy canopy ~

[wispy canopy]

😮

[terse jacinth]

https://tenor.com/view/azusa-k-on-smug-pathetic-gif-19369332

[wispy canopy]

what did i, or did i not do ? 😮

[terse jacinth]

how are things friend ?

[wispy canopy]

I'm alive 😮

well, semi

have been on a sickness streak as of lately

[terse jacinth]

ah shiiiet son

https://tenor.com/view/kaede-azusagawa-kaede-gif-head-headpat-gif-13284057

[wispy canopy]

doing some testing on pony models

[wispy canopy]

pats received

[terse jacinth]

how are things otherwise, anything new and fancy ?

haven't spun SD in months

[wispy canopy]

we got sd forge, and the best nsfw sdxl in the form of pony 6

and all it's spinoffs

[terse jacinth]

no news on SD3 i suppose ?

got an email

i need real-time generation 😛

[wispy canopy]

nah, they just announced the papers, i think

i mean, for a hentai model, this IS pretty impressive

[wispy canopy]

realtime high quality yes

playing a lot of helldivers 2 lately tho, amazing game

[terse jacinth]

i mean real-time as-in 24FPS 😛

[wispy canopy]

yeah, and then have that hooked up to a game, which will feed regional prompting to that 😄

yes plz

[wispy canopy]

love the style by mdf an

@gloomy scroll if you have prompt fusion installed, you could try by [ mdf an:kashu \(hizake\):0,0.4]

you'd have to juggle the 0.4 a bit, but in that case it interpolates each step instead of sudden switch over, might even work better

[vital raptor]

doesnt a1111 do that natively?

[wispy canopy]

no, it's a modification of the behaviour

the OG a1111 spec just switches over abrubtly at the specified point

wiht prompt fusion (the difference you need to specify 2 numbers), it will interpolate from A -> B

so step 1 = 100% mdf an

step 2 = 75% mdf an/25% kashu

step 3 = 50/50

uhh... percentage

[vital raptor]

ah, when I read interpolation I thought of step 1 is A, step 2 is B, step 3 is A..etc

[wispy canopy]

that's the [this|that]

specification

[vital raptor]

yeah

[wispy canopy]

also, there's a modified version for that one too

[nova remnant]

I'm boredom

[wispy canopy]

well, i just like playing with the artist names and getting some cool stuff out that way

[vital raptor]

yeah, i was thinking of making a text file with a ton of artists and using wildcards to randomly mix them to see what interesting things i get

[wispy canopy]

gotcha fam

all currently discovered WORKING artists for pony i have

some might be false positives tho

[wispy canopy]

especially the japanese names

another very strong option with that wildcard file is [__knownponyartists__|__knownponyartists__] for blending artist styles

[vital raptor]

looks like one of the artists on the left has the same sweat problem one of the ones i use has 😄

[wispy canopy]

i think it's kashu

[nova remnant]

throw some random japanese characters in your prompt to see if the japanese names are working or not

big brain

[wispy canopy]

mdf an causes a lovely sketch style, but has kinda diluted contrast

that list took me quite a while (gpu hours and comparing images) to make

and still have 17k images to check :/

MADNESS

well -> time to go to bed

zzzzz

[nova remnant]

gremlin stealing the cookie jar

[gloomy scroll]

That was just something I stole from Nubby

[nova remnant]

stole cookies to give to smol elf

[vital raptor]

thats what cookies are for

[nova remnant]

@gloomy scroll

[nova remnant]

[grizzled moat]

u steal from the dragon now ?

[gloomy scroll]

Dragon doesn't use Nai

[grizzled moat]

[native halo]

https://tenor.com/view/oreimo-lick-lips-licking-lick-kirino-gif-17887497

[grizzled moat]

so thats Nai

[gloomy scroll]

This is

[nova remnant]

its raining here

probably gonna have no energy and die of hot temperatures

[grizzled moat]

Cat exists > rains > dies from hot temperature

[nova remnant]

if energy goes out I won't have my fan

and ded

dragon bunny ||butt|| 👍

[native halo]

https://tenor.com/view/oh-my-goodness-gracious-gif-25726284

[nova remnant]

[nova remnant]

a dog in tuxedo walking a cat.

[nova remnant]

https://tenor.com/view/sus-cat-cat-stare-stare-bruh-cat-really-n-gif-25597234

Something wrong is not right here

catdog

hotdog monster

[nova remnant]

do [cat:dog]

[nova remnant]

generates cat only

it does each at a different step

a monster that's half cat and half dog

[terse jacinth]

in my case, more for real-time audio-reactive visual generation

SDXL will bring us there, someone been working on that project