🍥｜anime

[nova remnant]

the model i uploaded had filename_version.fp16.safetensor and it stripped the fp16 completely as a result the images i was generating and uploading didnt' show up in the model section

[steady grail]

lol

i can't find convetion

[nova remnant]

convention.. typo

[steady grail]

i only find a lora

[nova remnant]

interestingly when clicking the model download button, it shows up as modelname_version.safetensor

the version i had on pc was v1 and at civitai it was v10

it lets you specify versions when uploading like v1.0 and once uploaded they change the filename to v10

[steady grail]

charming

[steady grail]

this model iss so cute

the results are very charming XD

[green temple]

indeed

[steady grail]

what model is this?
i'm always there for cute stuff

[nova remnant]

chroma neo ft something something

[cerulean gazelle]

＃

[steady grail]

can't find it 😦

[nova remnant]

how do you specify age in prompts?

im getting decent results with -> 20 year old and such, but some suggest using age:xx

[steady grail]

rarelly models reponds to age

[nova remnant]

age:xx doesnt work for me

[steady grail]

some rare cases does

[nova remnant]

but i get nice results with xx year old

its more reflective on realism models than anime

[steady grail]

dilf\milf old man\woman, young man\woman, young girl\boy usually work well enough

[nova remnant]

neat

also mature girl, mature woman

[steady grail]

yep

[nova remnant]

matue woman would be more aged

[steady grail]

the rest is done by the situation and the bias of the model

[nova remnant]

difference in mature girl and mature woman :p

[steady grail]

yep

girl\boy tend to be more young

i suppose that comes from booru

[keen hollow]

this photo from a ladder? lol

[nova remnant]

from above

[supple raptor]

that just comes from normal language. 1girl on booru means literally every girl at every age. in normal language girl is normally used for younger females and woman is used for adult females.

btw the booru tag for older looking females is mature female. 1girl would still be used in those cases.

[nova remnant]

lots of cool variations with cute, mature / woman, girl

[steady grail]

1girl represent just the number and sex of the focus

[keen hollow]

messin' around with @nova remnant 's new one...

[nova remnant]

looks neat!

[green temple]

https://huggingface.co/zzzAI19/ChromaFT

[green temple]

mine is the NEOFT v1 pruned version

[steady grail]

i should look more often on higgingface models

sometimes, rare perls can be found

[keen hollow]

@nova remnant - some interesting ones but nothing mindblowing yet

[keen hollow]

how'd I do?

[supple raptor]

reimu from wish.com

style is nice though

[keen hollow]

i dunno who that is, but i tried 😄

[supple raptor]

here's the prompt without the style garbage 1girl, solo, hakurei reimu, touhou, brown eyes, black hair, hair tubes, hair bow, red bow, red shirt, sleeveless shirt, detached sleeves, red skirt, yellow ascot, arm up, armpit focus, medium breasts, (sarashi:1.1), smug, from side, black background, simple background

[keen hollow]

@nova remnant- this new v2 heavily skews toward a more yikes outcome... 😐

[supple raptor]

basically the main character of touhou and the 2nd most popular character on danbooru (only behind miku)

[nova remnant]

unstable?

[keen hollow]

just a lot of nsfw on very simple prompts

[nova remnant]

oh

didnt plan on censoring.. . the v1 was a bit of struggle to get nsfw at times

[keen hollow]

just gotta guide away

[nova remnant]

looks beautiful. if explicit stuff comes out too often hope you can adjust the prompts to negate them .. in general its kinda sensitive to prompts

[green temple]

@nova remnant that`s your negative tho ! so nice

[dapper ore]

What kind of models are these?

[near elbow]

samurai jedi?

right!

[green temple]

wow i really liked it !

"birds friendly " i guess..

[near elbow]

can't have too many of them

too bad they ate some of her fingers 🙂

[green temple]

lol

noooo 🤣

[near elbow]

she's masking it by having a bird in each hand now

[green temple]

so cute !!😍

[near elbow]

tell that to this guy 🙂

the trick for cute is "chibi" btw, but you probably knew that already

[green temple]

🙃

[near elbow]

"chibi cthulhu"

chibi terminator t800

ok, bedtime 🙂

cu

[green temple]

yes i know , but i`m trying to make it like that

[steady grail]

the first version of that model is the better one

i don't know why he is going for that plastic style

😦

[nova remnant]

girl

woman

adult girl

[pure oyster]

I got the courage to do an update with a git pull, and now I bricked everything lmao.

[nova remnant]

was there supposed to be something new?

i updated a day ago

[pure oyster]

probably not I just went from '1.6.1' to '1.8.0-RC'

now my xformers is preventing me from generating, and I can't launch without it

[nova remnant]

RC.. those might not be stable

mine doesnt say RC btw

[pure oyster]

you also have cuda and I don't

[nova remnant]

cuda is for nvidia

[pure oyster]

also with an amd card half these components aren't even compatible for me

[nova remnant]

oh thats not good

[pure oyster]

yea, I got fucked for trying to save $800

I see why Nvidia has no competition now

[steady grail]

yep

it's just that amd is focusing on a different sector

[nova remnant]

you could get a decent nvidia gpu for around $400

$800 would be too much for a gpu

[pure oyster]

gotta go hard or go home

got my 7900 xtx when it was about $850

[nova remnant]

Did you git pull a Nvidia version sd

[supple raptor]

Well.... xformers is Nvidia only so I'd wager, yes. Not sure if Zluda has xformers working or not.

[merry tendon]

emad resigned?

[supple raptor]

yep. will be interesting to see how things go from here under new leadership. I hear that a significant portion of the staff has walked away as well.

[merry tendon]

damn i did not really like emad but he's better than most other people in the AI scene

[supple raptor]

As someone with basically no knowledge of the inner workings of the company, he didn't really come off as much of a CEO. I'm not really sure if that was a good or bad thing.

[merry tendon]

yea i wasnt really active in the SD community for some time so yea

did not get to see what happend

[supple raptor]

There wasn't really anything public that I saw that pointed to him preparing to walk away. A week or so ago, there was a Forbes article that painted SD as basically being on the verge of going under, but I always take news from Forbes with a grain of salt so I didn't put much stock in it at the time.

[merry tendon]

hmm oke, anything interesting happend with SD in like the last 2 months?

[supple raptor]

SD3 is on the verge of maybe releasing, assuming the new CEO doesn't decide he doesn't want to release it. It's basically an LLM that generates images.

[merry tendon]

Like instead of clip some llm?

[supple raptor]

Many people have switched from A1111 to Forge which is a fork that uses comfy's generation code is much more efficient, but sacrifices some extension compatibility.

[haughty lynx]

SD3 uses T5 iirc?

[merry tendon]

ive tried forge and it's really awfull

[supple raptor]

It uses both of the CLIPs the SDXL uses as well as T5.

[merry tendon]

tf is CLIPS and T5

[supple raptor]

XL uses OpenAI's CLIP (same as SD1.5) as well as OpenCLIP (better version of the one used in SD2.x and without garbage training). SD3 uses both of those as well as T5.

[merry tendon]

t5 from google?

[supple raptor]

yep

It's a beefy model and will be released in multiple parameter counts. The biggest is 8b which should have approximately the same performance and vram requirements as a 13b LLM from what I understand.

[merry tendon]

the question is tho...

can you quant the llm

like you do with llama.cpp

or exl2

[supple raptor]

people much smarter than me have said yes, though I don't know how much image outputs are effected

[haughty lynx]

wait isn't the largest 8b or they just made a bigger version

[supple raptor]

yes, i'm a little drunk and my brain was going faster than my fingers.

[merry tendon]

should work yea, both exl2 and llama.cpp have multimodal support

[supple raptor]

Yeah, Kohaku, mcmonkey, comfy, and neggles were talking about the potential for quantizing it just after it was released.

[merry tendon]

why do people like forge, just because it's fast?

[supple raptor]

Unless your hardware is shit, it isn't really faster because it insists on doing some cursed model switching when not needed. It's mainly just more efficient on vram, especially with XL models.

[merry tendon]

ive tried it and i think it's really weird

i don't like it personally

[supple raptor]

VAE passes in both directions are slow AF which is especially painful if you do hires fix + adetailer because that means 5 vae passes.

t2i > decode > upscale > encode > i2i > decode > crop > upscale > encode > i2i > decode > downscale > stitch
enjoy ~10s latency on each decode and encode operation before it even starts.

[vital raptor]

do you normally use comfy or something other than a1111? it's mostly the same front end wise with some extensions built-in

[merry tendon]

sd next

well i have used it

before i quitted

xd

[vital raptor]

I don't use it personally since it doesn't support TensorRT and I have a high end card so it doesn't have any speed benefit for me

[merry tendon]

yea i don't really care about speed

never did

[vital raptor]

yeah, tbh on a 4090 speed is pretty much the last thing i worry about with my gens

[supple raptor]

SD Next is kind of ass too. Vlad likes to commit shit to main before it's ready so it's a frequent battle with just getting it to work at all. Unless you have AMD on Windows or need Diffusers support, I wouldn't recommend it. I'd point to Forge if you want to run XL because Auto's implementation is garbage, or A1111 if you only care about SD1.5 because Forge is filled with it's own set of curses that are often worse.

[merry tendon]

ive tried vlads thing yesterday with sdlx again and it barely works, quality is really shit

[supple raptor]

Comfy has become my main UI when I'm at PC since getting used to using Efficiency Nodes which helps to condense things like regular t2i and hires fix down into just 3 nodes.

[merry tendon]

idk what im doing worng

[supple raptor]

I launch Forge if I need to use my local UI from mobile since I've been using XL

StableSwarmUI is another solid option. Think of it like SillyTavern for ComfyUI. For most tasks, it will create your nodegraph for you, but still gives the option to load a custom workflow if you want/need to.

[merry tendon]

nah i just wanna get sd next to work and do my stuff

but yea

i already said

"ive tried vlads thing yesterday with sdlx again and it barely works, quality is really shit"

[vital raptor]

i just use a1111 because it works for the XL gens i'm doing and i don't really do anything fancy, been meaning to try swarm though, sounds interesting

[supple raptor]

Vlad takes the worst of A1111 and often makes it even worse somehow, but wraps it in a slightly more pleasing UI.

[merry tendon]

well can anyone help me now?

im kind of stuck here

[supple raptor]

What's the error you're getting?

[merry tendon]

no error

just shitty gens

even with recommended settings from the model creator

[supple raptor]

Can you share one with PNG info?

[merry tendon]

yea

[merry tendon]

need to spin it up real quick

[merry tendon]

this is one that completely failed, this shit happens everytime i set clip skip to 2. clip skip 1 is just plain bad quality image

[supple raptor]

and no errors in console? I've seen NaN errors on XL produce grey images like that before.

[merry tendon]

no errors

[vital raptor]

i don't think jpg files have the pnginfo in them

[merry tendon]

they do

[vital raptor]

i guess i'm dumb then because it's not showing in the metadata or in the pnginfo tool in a1111

[junior quest]

wut bout -2? I saw somewhere

[vital raptor]

clip skip doesn't matter with XL anyway

[merry tendon]

does not work in sdnext

[merry tendon]

well its giving bad images and idk why

[vital raptor]

well based on the filename of that image (the only info i can get from it your prompt is score_9_source_anime_white_background_blush_covering which looks about right

i'm assuming you have more prompt

[merry tendon]

yea

[supple raptor]

Let me load up webui and see if I can get PNG info. My prompt inspector bot wasn't able to find anything for that image.

[merry tendon]

maybe discord removing info again

[vital raptor]

a1111 doesn't get anything either, it doesn't have the png metadata because it's a jpg

[supple raptor]

Technically it does. Just not on Forge/A1111 (not sure about SD.Next). You just shouldn't use it.

[supple raptor]

it's definitely not. the 3 separate instances of the prompt inspector bot I use is in 3 of the servers I've posted in over the last 30 minutes and all of them are working as expected on my own images which don't use stealth png info.

[merry tendon]

ok wait

[supple raptor]

yeah, no png info there. did you drag and drop from the folder into discord? copy paste from browser will strip it because basically all browsers strip metadata on copy.

[merry tendon]

i always drag and drop form folder

[supple raptor]

Ok. Let's just try some of the simplest possibilities first. Which VAE are you using?

[merry tendon]

the one provided with the model page

[vital raptor]

standard sdxl vae then

[merry tendon]

yep

[vital raptor]

what's your full prompt?

[supple raptor]

Try this one. Let's rule out them using the shit official VAE that overflows at fp16. https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors

[vital raptor]

pony is very specific and usually needs most of the score tags to generate anything useful, not just score_9

[merry tendon]

score tags?

[vital raptor]

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up,

[supple raptor]

I feel confident that the prompt isn't the issue in this case. You might get bad outputs with a bad prompt, but they shouldn't be fully grey.

[vital raptor]

they shouldn't but if the prompt is score_9_source_anime_white_background_blush_covering like the filename suggests, then it's exactly what was prompted

so, just curious

[supple raptor]

check the Important information section in the description https://civitai.com/models/257749/pony-diffusion-v6-xl
basically score tags are used in place of the typical best quality, masterpiece, etc tags that are normally used on most models, but are even more important here because PurpleSmart made a fuckywucky when training and noticed too late to fix it.

[vital raptor]

put the rest of the score tags i pasted at the begining of that prompt and it'll probably be better results

that image looks pretty typical for that prompt

[supple raptor]

I recommend puting them at the end. Front loading them has been consistently worse for me.

[vital raptor]

personally, i'd get rid of all the negatives as well since it's not needed

mine seems to do fine front loaded, but it could be because my artist tags are first 🤷‍♂️

this is also autismmix, but i didnt have any noticable issues on pony either

i also don't use all of the score tags, but that's a personal preference

[supple raptor]

@merry tendon let's try this. this image is one I generated in Forge with Pony. Try putting this into PNG info and sending to txt2img, then generrate to see how it comes out.

Make sure to open in browser first then you can either drag and drop into PNG info or download/save the image and manually select it in PNG info.

[merry tendon]

k wait, will take some time because my shit is stored on an hdd

[supple raptor]

feel free to toggle off hires fix and adetailer. we just want to confirm that you can get an image that's similar with all of the parameters I used.

[nova remnant]

hdd holy

[merry tendon]

yea lol

[vital raptor]

I don't think i've used a hdd in over 10 years

maybe even 15

[merry tendon]

https://tenor.com/view/i-identify-as-poor-chris-rock-chris-rock-selective-outrage-im-not-wealthy-im-broke-gif-7068350681723523282

[vital raptor]

😄

fair

when i was broke, i also worked in IT jobs so I got leftover hardware a lot 😄

[supple raptor]

i use hdd too. 4tb ssd is expensive

[nova remnant]

playing path of exile on a hdd makes you slowly turn into a psychopath

[vital raptor]

yeah... i have 8TB of nvme in this PC... it was expensive

[supple raptor]

That's why my games are on my SSDs

[merry tendon]

??????????????

[nova remnant]

lmao

[vital raptor]

speaking of which, i need to do some cleanup

running out of space

[merry tendon]

please..

[vital raptor]

lol, i'm being sarcastic

it's like when i complain about my SDXL gens being slow and taking over 20 seconds!

[nova remnant]

I have a 500gb hdd for random crap, 250sd for windows and 1tb nv for games

[supple raptor]

Ok, that looks like similar to what I've gotten at CLIP skip 1 in Comfy/Swarm where that's respected. Try it again with CLIP skip 2, which is what A1111 uses for SDXL.

[merry tendon]

yea thats where we come into this problem

[supple raptor]

And I assume you updated your Sd.Next?

[merry tendon]

i installed it yesterday

and updated 20 minutes ago

[vital raptor]

for what it's worth, this is a1111 using clip skip 1 and 2, they're identical

with pony

[supple raptor]

yesh. A1111/Forge does not respect CLIP Skip settings for SDXL models. It uses 2 regardless.

[plain vapor]

I was about to say that I couldn't see the difference at all 🤔

[supple raptor]

Does SD.Next have an Emphasis Mode setting? A1111 has some cursed code which causes emphasis to become completely fucked sometimes which can result in images like that. A fix was added to A1111 1.8 which fixes that. No Norm here is the fix. I don't know if Vlad implemented the fix on his end or not.

If he didn't, then this extension will probably work to fix it. https://github.com/hako-mikan/sd-webui-prevent-artifact

[merry tendon]

i have this

[vital raptor]

wtf, emphasis mode? no wonder i couldn't find that setting

[merry tendon]

changing anything does not work

[vital raptor]

i deleted that extension when i upgraded to 1.8 because i heard they fixed that problem but then could never find the setting

then re-added it because it's still broken with default settings

and it's not emphasis specific, it happens with no emphasis too

[supple raptor]

Try fixed attention

[merry tendon]

nope still broken

[supple raptor]

I'd set it back to default and see if the extension works then. Vlad's opinionated changes are predictably unpredicable.

[merry tendon]

still broken

im just gonna give up

[supple raptor]

Do you normally do much other than t2i?

If not, you should try Comfy. I can give you an easy to use workflow with hires fix + face detailer (a better version of it than you can get on any of the A1111 variants)

[nova remnant]

why don't you use a1111 or forge

if you just installed it

[supple raptor]

a1111 is shit for sdxl and forge is just kind of shit overall aside from vram efficiency.

[nova remnant]

idk what you do in forge it looks good to me

[vital raptor]

🤷‍♂️ i've never seen any actual comparisons to prove why comfy is so much better at image generation, but people keep saying it is

aside from supporting some new features or exotic workflow, i've never seen a real difference so I can't help but think it's just people liking one over the other and then calling everything else shit

[nova remnant]

https://tenor.com/view/keyboard-type-fast-jim-carrey-speed-typing-gif-12747003

[supple raptor]

In a very general sense, A1111's emphasis normalization makes it better than Comfy. SD wants emphasis across the prompt to equal close to 1.0, which is what A111 does. Comfy doesn't normalize so you need to reduce emphasis on other tokens when increasing tokens more than ~1.5. TLDR: this means that A1111 is often easier to get good images from.

[supple raptor]

I'm drunk. My typing is slow

[vital raptor]

for most people easier = better

not everyone wants to be a pro, they just wanna generate some anime booba

[nova remnant]

small booba

[supple raptor]

yes, I agree. Comfy's lack of normalization makes sense for his vision of his UI, but I don't like it from a user standpoint.

[vital raptor]

yeah for me, comfy was an interesting visual tool to help understand what i had already read about how images are generated, but largely generated the same images i was already getting, so i just use the easier interface because i don't care to tweak things so much that i need to build up a custom workflow

[supple raptor]

here's my main reason for liking comfy currently, aside from the ease of using NAI API. face detailer using bbox (like adetailer) and then segm to only inpaint the actual face instead of the entire bbox detection like a schizo (which is what adetailer does.)

[vital raptor]

but that's not 'comfy', that's a custom node (extension) and could just as easily be something that exists for a1111

if the reason comfy is 'better' is because it has better node/extension availability, then fair enough

[supple raptor]

Adetailer is also an extension, which is the equivalent of a custom node.

[vital raptor]

but that's only if you need or use those tools

[supple raptor]

Both are a terrible experience without extensions/extra nodes.

[vital raptor]

yes, and there's like 5 other extensions that do similar things

my point is that comfy is not 'better' because of a node/extension being a thing

it might be why you use it, but it doesn't make the UI itself better

[supple raptor]

There isn't an A1111 extension that does bbox + segm detection for face detailer.

[vital raptor]

i don't use face detailer 99% of the time, so i wouldn't know

[supple raptor]

face detailer is beneficial 99% of the time

[native halo]

animatediff is also better on comfy,u can combine it with several controlnets,do vid2vid ,the a1111 sucks and doesnt let u do all of that

[supple raptor]

yeah animate diff in a1111 is and probably always will be absolute garbage that should not ever be used.

[nova remnant]

I only have ever used adetailer on smol's pc

[supple raptor]

It's absolutely possible to gen good images without adetailer, but almost every image will be better with adetailer because of how the human brain works. We are biologically programmed to pay the most attention to faces, which adetailer imprroves.

[vital raptor]

i was using it a bit with some other XL anime models because it improved things significantly, with pony/autismmix I haven't felt the need for it and my realistic model doesn't benefit from it at all in most cases

it's long been my opinion that if you 'need' an extension or specific interface, lora, whatever, to get a good image from a model, you probably have a shit model

or at least a model that's not good for what you want to get from it

my old 1.5 workflow required a handful of negative embeddings and a lora to get the results i wanted, I didn't like that so i just trained a better model

[native halo]

for making videos,adetailer is a most otherwise u will get a different face every frame

[supple raptor]

Fuck it. I'm drunk enough tonight to say that most of the images posted in this channel are not very good. There are certainly exceptions. But almost half aren't even remotely anime.

TBH that's why I'm not here very often any more.

[vital raptor]

I have a model that produces the same face every time, it's very over-fit on a particular character intentionally

the non-anime images that get posted in here are pretty annoying

whether or not an image is 'good' kinda depends though, it's highly subjective

[supple raptor]

it's somewhat subjective

[native halo]

tbh most of images u get with AI are not very good,u get a really good img but its rng or if u use controlnet

[vital raptor]

it's all about what you want out of it

[supple raptor]

I will say that the barrier for entry is significantly higher in the Touhou AI server where I spend most of my time.

[lyric zealot]

now we just need to post more feet pic for nubby to stay here

[vital raptor]

cute? ✅
anime ✅
not deformed or weird proportions? ✅
good? 🤷‍♂️

[supple raptor]

https://aibooru.online/posts/65359

[supple raptor]

It doesn't fit the criteria of what I was actually complaining about, but if I'm being completely honest it feels very "soulless" to me as much as I hate that description.

[vital raptor]

I imagine if this server/channel was dedicated to specific characters and style instead of an open place for beginners the overall quality would go up a bit

[vital raptor]

or hell, if it was actually only anime it would improve a bit

[vital raptor]

maybe, but it fits all of MY criteria, which is MY subjective opinion

just like everyone else posting their maybe-not-good images

i don't like some of what gets posted here, even if it IS anime, but that doesn't make it bad

just not my style

[native halo]

if sd would make contests like they used to,im sure we would get high quality imgs but since they abandoned contests,not worth the effort

[vital raptor]

yeah

[supple raptor]

I'd suggest joining the Touhou Project AI server and taking a look at the images from ctd (moon fan) she pretty consistently posts high quality images.

[vital raptor]

personally I just get bored and post whatever gens I get that are cute or interesting, they're not always great and I can't usually be bothered to spent time improving a particular image when i could be generating more

[supple raptor]

Anyways, time for me to go to bed. It's approaching 6 am.

[green temple]

nice hands lol

[nova remnant]

pat pat pat

[nova remnant]

anime

ugly

[native halo]

https://tenor.com/view/osaka-azumanga-daioh-low-tier-god-gif-14014891512407492266

[green temple]

indeed

[near elbow]

how to avoid the finger issue 101 🙂

[green temple]

🙃

[keen hollow]

mornin'

[near elbow]

so it seems

[keen hollow]

indeed indeed

doing a little auto/forge benchmarking this morning to see where things are falling these days

[native halo]

towards earth,last time i checked

[keen hollow]

benchmarks complete - Forge won in all categories. (enabling xformers gave it further boost at high-res 1.5, oddly)

the optional optimizations didn't really do a ton

[nova remnant]

[keen hollow]

have any of you gotten 'photomaker' working in forge?

[grizzled tartan]

这是什么地方

Where is this

[native halo]

chattanooga

[grizzled tartan]

Are you in America now?

[grizzled tartan]

HELLO

[native halo]

no,pangea

[grizzled tartan]

I dont understand what you are saying

[native halo]

https://tenor.com/view/ranka-lee-macross-frontier-kira-idol-pose-gif-27490676

[junior quest]

https://tenor.com/view/cat-smiling-smile-cat-orange-smile-cat-gif-23133359

[grizzled tartan]

[nova remnant]

@supple raptor do pony prompt tags affect sdxl?

or rather have any effect on sdxl models that are not pony

tags like these score_9, score_8_up, score_7_up, score_6_up,

[keen hollow]

@nova remnant - these are all your v2 model

[nova remnant]

looks nice 🙂

[nova remnant]

any suggestion which you'd like to see improve?

[keen hollow]

not really... I'm justhappy to generate like 250 images in 30 minutes on an xl model

[nova remnant]

lol ok ty, but suppose you want something more in terms of quality in the images?

btw omnia im messing with pony models too

[pure oyster]

pfff get on my level, I do 20 images in 30 minutes on 1.5

[keen hollow]

ouch

[keen hollow]

not really? I dunno .. seems like it works as well as anything else 😛

[nova remnant]

cute image .. is that from sd1.5?

[nova remnant]

okay... let me post a couple of example that im trying to achieve consistently

[pure oyster]

yes

[plain vapor]

Pangea is the oldest known continent, before it split apart (over billions of years!) and turned into the 7 continents we know today.

[nova remnant]

@keen hollow

those are v2 but im not getting consistent results with every prompt

[keen hollow]

I'm certainly no expert in prompt consistency 😄

dangit, what's the command-line parameter to specify upscale model location? 😐

[nova remnant]

its kinda more about how the model would react consistenty

experimenting with it for now....

[keen hollow]

answered my own question... --esrgan-models-path

[zinc flax]

booba

[nova remnant]

middle fingers you

[keen hollow]

i'm gonna run some of these in widescreen and see if I can get a decent ultrawide wallpaper, torqx...

takes about 30s per image though, so... here we go

err... probably 50s

[nova remnant]

https://media.discordapp.net/attachments/857955668829011988/961381752805425222/kok.gif

[near elbow]

with animagine XL, if you don't add a lot to the prompt, the result is quite chaotic (this is just "neko fun" 🙂 )

guess the tokens 🙂

[nova remnant]

can you send me your prompt

whole thing, negative too if using

[keen hollow]

i like the color contrast for this one

[nova remnant]

the house is bleeding yellow

[near elbow]

of course: "anime artwork neko fun chibi sakura mecha gundam . anime style, key visual, vibrant, studio anime, highly detailed"
Negative prompt: "photo, deformed, black and white, realism, disfigured, low contrast, loli,underage,teen,child"

[nova remnant]

cool glasses

[near elbow]

ah that was for the mecha - the neko chaos was "anime artwork neko fun . anime style, key visual, vibrant, studio anime, highly detailed" (same negs)

[nova remnant]

really wide ... !!

[keen hollow]

you know how there's a Japanese word for everything, like Zettai ryōiki? I need a word that describes skirts that drape like that when sitting... there's gotta be one.

that one is kinda short but same idea

[nova remnant]

japan isn't real

[cerulean cosmos]

https://tenor.com/view/gandalf-gif-21901728

[dire bridge]

Me:

[keen hollow]

eh i looked, couldn't find a word/tag for it

[nova remnant]

[near elbow]

wide and EVIL 🙂

wide and NEKO

it's the nekopalypse

ok last one

[keen hollow]

that's a lotta pink explosion

[nova remnant]

[nova remnant]

@lyric zealot hands are coming out bit weird on pony model, could i be missing something in the prompt?

the examples i picked up on, im using this for negative....
score_5, score_4, (worst quality, low quality:1.2), muscular, censored, monochrome

[lyric zealot]

what's your full prompt?

[lyric zealot]

is there pnginfo for this?

[nova remnant]

that negative and this in positive..

score_9, score_8_up, score_7_up, score_6_up,

beautiful girl, orange t-shirt, white shorts, blue eyes, detailed face, pose, dynamic angle, cinematic, green background,

source_anime, Koyama Shigeru

sure .. one sec

has metadata

wondering if i should be prompting bad hand

unless scores cover it

this came out fairly good

2 days into pony model.. baby steps

[lyric zealot]

maybe try specifying a hand pose

like peace sign

or something

[nova remnant]

👀

[nova remnant]

whats a good tag for generic hand gesture?

something similar to dynamic pose?

or dynamic angle

dynamc expressions

[lyric zealot]

i don't think there's such thing

no idea

[cerulean cosmos]

its probably openposes or something

[lyric zealot]

but you can try something like hand on chin, hand on lap

or stuff

[nova remnant]

ok

[nova remnant]

you suggest using open pose?

[cerulean cosmos]

unless u have a lora, yes

[nova remnant]

i usually use depth if i have to fixate a pose

[cerulean cosmos]

i havent tried depth yet, im sure its good as well

[nova remnant]

depth anything is pretty powerful

and line art

its funny how bad the images came out using pony model at first

until i specified artist style

and im not using any lora

[cerulean cosmos]

im starting to get better results

[nova remnant]

pony model?

[cerulean cosmos]

nope, this is 1.5

[nova remnant]

1.5 doesnt have the same level of prompt coherence as sdxl

[cerulean cosmos]

im looking forward to trying it when new computer arrives

[cerulean cosmos]

ive seen some increadble sdxl pony images

[nova remnant]

yeah thats how i got pulled into it lol

but you wouldn't believe how many times i deleted and downloaded pony at first

i rendered some aesthetic lewd stuff too.... was curious and they came out pretty good

girl focused.. nothing too harsh

[cerulean cosmos]

nice

[nova remnant]

@lyric zealot any idea if score tags have any impact on regular sdxl models?

[lyric zealot]

don't think so, unless it's a merged of pony

since those tags was trained on pony

[nova remnant]

let me share the image with you

can you take a look at the metadata, im bit confused, and the model i used had nothing to do with pony merge

[lyric zealot]

what about it

[nova remnant]

i used regular sdxl model but used pony tags

[lyric zealot]

try do a plot xyz

and see if there's any major differences

[nova remnant]

there is

but im not sure how that could have any effect

[lyric zealot]

well

sometimes random tag that doesnt make any sense can have a positive effect on the model

in this case

[lyric zealot]

maybe the word "score" is having an effect

[nova remnant]

it is obviously but can't figure out how or to be able to have control over it

[lyric zealot]

i think you can try to see if the effect is positive on your non-pony model

and if so, is it consistent

[nova remnant]

not consistent across different prompt style, but has brilliant effects on certain prompts

[lyric zealot]

well

if it works

then good

[nova remnant]

lol didnt post it on civitai.. doesnt seem logical to be using pony tags on regular sdxl models

im amused with autismmix

and this guy here Koyama Shigeru

[cerulean cosmos]

oh ye

[nova remnant]

cool effect

she also has good symmetry except her smile is bit big which is intended i suppose

[cerulean cosmos]

haha yeah im messing around with krita

[nova remnant]

isnt that paid service?

[cerulean cosmos]

nope, completly free, unlike adobe it seems

[nova remnant]

krita or krea.. hmmm confused

[cerulean cosmos]

i think its swedish, so krita

[nova remnant]

ahh ok

[cerulean cosmos]

https://github.com/Acly/krita-ai-diffusion

its increadble for a free editing

[nova remnant]

oh damn

a full blown app like graphic software

[lyric zealot]

careful

[nova remnant]

Dont worry

leave her to me

👍

[nova remnant]

[nova remnant]

too exposed?

[lyric zealot]

i think so... but i don't know

[nova remnant]

ok

[lyric zealot]

i don't think it's as strict as before

or just mods aren't around

[nova remnant]

idk either, but i dont think clothing like beach outfits should be censored

[nova remnant]

banned just for saying that

[lyric zealot]

well, that one might be borderline okay, just don't go further

[lyric zealot]

[cerulean cosmos]

dont go further u say ?

[nova remnant]

smol sus you dont need to tell me about borderline, no offense but i know what's blatantly showing and what's artistic

[lyric zealot]

well, nude stuff can be artistic but it's not allowed here?

[nova remnant]

damn she buff

[lyric zealot]

so i don't really get your point

[nova remnant]

surely i wouldn't be posting revealing body parts

[nova remnant]

that was not nude

maybe bit suggestive

but even then its focused on arstistic expression

[lyric zealot]

yup

that's what im saying

artistic stuff sure, but it can also not be allowed here

right?

[nova remnant]

idk why you are talking about stuff that im not posting

[lyric zealot]

like example nude stuff that's artistic?

im making a point

[nova remnant]

you dont have to

[lyric zealot]

well sure

[nova remnant]

i can acknowledge the policy against blatant nudity in here

the world is already bit too messed up with woke culture

dalle censors on short skirt

please avoid this madness

[lyric zealot]

avoid what madness?

avoid me telling you to be careful not to go too far?

[nova remnant]

getting touchy over something that is not obscene

[lyric zealot]

well

if you think so then go ahead

[nova remnant]

you are picking up an argument for no reason

[lyric zealot]

i am?

[nova remnant]

is there a reason for it?

[lyric zealot]

wait i am the one picking up an argument?

[nova remnant]

you attacked me for nothing that is offensive

[lyric zealot]

let me read thru again and see where i started the arguement

when did i attacked you?

[nova remnant]

can you quit twisting it?

[lyric zealot]

brother what are you on about?

[cerulean cosmos]

https://tenor.com/view/come-at-me-bro-fight-square-up-squirrel-angry-gif-16242592

[nova remnant]

sheesh

all this fuss over an image that is not even nude and lecturing how nudity is not allowed

that was an attack

[lyric zealot]

oh sorry for attacking your poor soul

im sorry

my bad

[nova remnant]

and you prefer to mock it

[lyric zealot]

okay im genuine sorry, my bad

didn't know you would feel that way

but okay

[cerulean cosmos]

HES OVER THERE OFFICER, THATS HIM RIGHT THERE

[nova remnant]

but okay... i didnt expect to have that ugly fuss over something pointless or something i'm aware and didnt do

[cerulean cosmos]

if u guys dont brake it up, i will get into it, and then there will be 3 people fighting , thats a mexican standoff

and we dont want that, do we?

[nova remnant]

yeah lets quit this woke thing

[lyric zealot]

ugly fuss over what? i literally said "careful" and that's it

[nova remnant]

careful about what exactly?

[lyric zealot]

then i explained that you shouldn't go any further than that

[nova remnant]

https://tenor.com/view/guy-arguing-guy-talking-to-wall-talking-brick-wall-gif-18667615

careful about what is that a threat ?

[lyric zealot]

here

you can go reread what was said

[cerulean cosmos]

u gotta be careful about future crimes u might or might not commit, capiche!

[nova remnant]

and that wasnt nudity

[lyric zealot]

and?

[nova remnant]

wtf is and?

you are going on about this like its a game

[lyric zealot]

i think we are just not talking on the same frequency

so whatever

[nova remnant]

we are not

[lyric zealot]

you do you

and have fun

[nova remnant]

you pushed your woke sentiment on me with a threat

[cerulean cosmos]

https://tenor.com/view/cats-fighting-fighting-cats-anniehasleftoop-gif-23317332

[nova remnant]

i dont think i can appreciate that

[lyric zealot]

people really love using woke and stuff huh

[nova remnant]

thats another word play

[lyric zealot]

[nova remnant]

and you haven't explained to me what rubbed you wrong to tell me careful

[lyric zealot]

i don't think i want to, since you will say im attacking you or something

or like forcing some woke sentiment on you

[nova remnant]

thats a problem and if you dont see it then we have nothing to exchange

[lyric zealot]

yup

we have nothing to exchange

[nova remnant]

good lets end this convo

[cerulean cosmos]

guys, for real, you are both fine

[nova remnant]

i didnt start anything for that outright threat... i didnt like this one bit

[cerulean cosmos]

in a sense, he was trying to help u, by not posting anything soucier then what u where doing

[nova remnant]

oh i know

and i wasn't posting anything of that sort in here

[cerulean cosmos]

it was a little girl in underwear, but enough about that

[nova remnant]

anime

[cerulean cosmos]

yes its fine

[lyric zealot]

[nova remnant]

its tragic how society is nose diving

[cerulean cosmos]

https://tenor.com/view/ryan-gosling-ew-wtf-no-disgusted-gif-16635152

im going to get food, u guys better behave while im away

[nova remnant]

lol have fun, im gonna go get some tea.. not arguing about it

[nova remnant]

hey

be careful

https://tenor.com/view/cat-running-away-escape-getaway-bye-gif-16631286

lol

i dont think i ever posted anything obscenely revealing here

not a single image w/o anyone telling me about it either

and when someone gags your throat for expressing something that is not violating anything then it feels aggressive

[lyric zealot]

iPeach ?

[nova remnant]

grabby

[cerulean cosmos]

sd3 eh?, lovely

[nova remnant]

i have second thoughts about giving up sdxl

we'll have to use sd3 along side sdxl 🙂

the pony model from sdxl is pretty dynamic

comparing pony with regular sdxl feels like some key element missing in sdxl

[keen hollow]

I haven't ventured into any "pony" things yet.. seemed complicated.

[nova remnant]

tagging is pretty simple if you make a template to just click and select into positive and negative box in a1111

im bored

as for the key part of pony that i found hard to cope was the aesthetics, but i got an artist list from nubby which i curated with female model focused art via chatgpt

now i just love pony model

[keen hollow]

aren't there a bunch of random keywords you have to use all the time?

more from your v2 model there torq

[nova remnant]

I made an easy template for the pony generic specfication

and i click them in their appropriate box and i start prompting

[nova remnant]

sweet!

[keen hollow]

yeah ill hae to lookup a quick start or something...

ok gonna start this batch and walk away for a bit

[nova remnant]

@nova remnant kinda encouraged me to use pony model when it first came out... i've had a bit of struggle until i finally got it working how i wanted

[keen hollow]

evenin'

[nova remnant]

hey

omnia im wondering if its possible to make pony model lightning

[keen hollow]

no idea 😛

[nova remnant]

im gonna look into it, worth a shot

generally pony models have something like 24 steps and 7 cfg

[keen hollow]

maybe ill try one tonight

is there like.. a main one everyone is using?

[nova remnant]

yes

its the pony v6

[keen hollow]

is that xl?

[nova remnant]

https://civitai.com/models/257749/pony-diffusion-v6-xl

[nova remnant]

yes

im using https://civitai.com/models/288584/autismmix-sdxl autism mix

[keen hollow]

geez these pages are like landmines

looks like there is a lightning version already

[nova remnant]

lol i know

there is?

can you link it?

[keen hollow]

on the one you're using

[nova remnant]

oh right

yeah its something i forgot to download and try

im going for it

[nova remnant]

download speed sucks big time atm

[cerulean cosmos]

thats what im talking about

[slate narwhal]

[nova remnant]

@keen hollow if you are using a1111 you wont have the sguniform sampler types

i was trying the autismix with lightining lora just now, doesnt look as good as w/o lora

i wonder how is recommending to use Euler A SGMUniform on a1111

[keen hollow]

i have no idea what i'm doing with this thing yet lol

[nova remnant]

already downloaded?

[keen hollow]

yeah

[nova remnant]

let me share the template, this goes in your positive prompt

[keen hollow]

yeah the score thing

[keen hollow]

getting very... messy? outputs...

[nova remnant]

positive:

score_9, score_8_up, score_7_up, score_6_up,
-your prompt here-,
source_anime, artist-name-from-danbooru

negative:

score_5, score_4, (worst quality, low quality:1.2), muscular, censored, monochrome

yes that's fine, you need to use artist name for good images

[keen hollow]

oh interesting, I had 4 and 5 in the positive

[nova remnant]

try this artist in the positive section Koyama Shigeru

my curated few

[keen hollow]

mmm maybe need to test a few samplers

ill fart with it for a few mins

[nova remnant]

try dpm++ sde karras

euler a should work too but i prefer the first

and in a1111 there is no sgmuniform as the author suggests

[keen hollow]

lol i keep making it worse rather than better.. you should see some of these 😄

[nova remnant]

its tricky at first

once you get the hang of it you'll love it

[keen hollow]

much more reliant on tags, it seems

[nova remnant]

2 key elemetns to pony models... the prompt template and specifyng artist name

i would stress on artist name with huge importance .. its a build or break it thing

[keen hollow]

I couldn't name a single anime, let alone an artist LOL

did i do it right?

doesn't seem like anything special.. hmm

[nova remnant]

can i see your positive and negative prompt for it?

[keen hollow]

oops... already blew it away

[nova remnant]

here is an image with metadata you can load up

[keen hollow]

i made some tweaks lemme see what I come up with

what res is that?

[nova remnant]

768 x 1024

[keen hollow]

oh my gosh it came out with a foot...

[nova remnant]

you could try including this into positive just to be safe lol ...

rating_safe

add it after `score_9, score_8_up, score_7_up, score_6_up, rating_safe,

[keen hollow]

i'm trying to grab random example prompts from the site.. had to turn on blurring to even navigate the thing

[nova remnant]

there are basically 3 ratings that are used ...
_safe, _questionable, _explicit

[keen hollow]

wow this is tough

[nova remnant]

in the image metadata that i shared i had a mis type... it should be

source_anime,

[keen hollow]

ok maybe i need to simplify this first before getting crazy

[nova remnant]

do you have a template for the generic prompt?

that files goes into webui/ root folder

[pure oyster]

@nova remnant I got SD running again, but with zluda

[nova remnant]

but rename the old csv file

[nova remnant]

nice, but i hardly have any technilcal know hows on amd system

[pure oyster]

It's ok AMD doesn't either.

[nova remnant]

ha

btw omnia you dont wanna use default templates when using pony

i just lef the default and pony to swap when switching models

[keen hollow]

ugh

[nova remnant]

that kinda looks like low cfg

[nova remnant]

What res and vae are you using

[keen hollow]

832x1216, no vae.. I assumed it was baked

[nova remnant]

no, use sdxl_vae

[keen hollow]

ok

[nova remnant]

Do you have prompt padding enabled?

[keen hollow]

nope

very watercolory

[nova remnant]

[keen hollow]

literally just "score_9, score_8_up, score_7_up, score_6_up, source_anime, 1 girl, sitting, Kantoku"

[nova remnant]

it works bit differerently i think

[nova remnant]

Is it pony or autism

[keen hollow]

pony

[nova remnant]

first line : score_9, score_8_up, score_7_up, score_6_up, rating_safe,
2nd : prompt goes here.
3rd source_anime, Asanagi,

[keen hollow]

meta included

[keen hollow]

sure ill try it

got this

[nova remnant]

source filmmaker, sketch, lineart, chromatic aberration, unfinished, blurry, monochrome, censored, abdominal muscles
@keen hollow try this in negative and cfg7

[keen hollow]

sure, changing neg

[nova remnant]

@keen hollow load this to see how the sequence are laid out

[keen hollow]

ok

[nova remnant]

Also maybe you should try autism if you're looking for the more generic 1.5 look

Like torqx image

[keen hollow]

that new neg did not go well LOL

trying that image

[nova remnant]

It's what I use on pony

Something else is wrong

[keen hollow]

maybe it's a forge thing

[nova remnant]

Pony works on forge just fine

ok quick check, what sampler, steps, and cfg do you have set?

[keen hollow]

I ran your metadata exactly (though used pony, not autism)

seems like it got pretty close

[nova remnant]

ohhh

pony and autism are bit different how they listen to prompts

pony is more versatile and style can switch a lot

[keen hollow]

(and turned off hi-res just for the test)

[nova remnant]

They aren't very different with complex prompts

Things get difference with very little prompt

#🍥｜anime message

[keen hollow]

yeah that one is still downloading

[nova remnant]

the settings the author recommends for lightning variant are not available on a1111

im using regular autism mix

[keen hollow]

it's almost done downloading

[pure oyster]

Some of y'all really need to use adetailer.

potato ass faces

[nova remnant]

lol my adetailer is enabled by default

unless im doing very specific portrait with faceID then i turn it off

[nova remnant]

I'm just using a bot

[pure oyster]

Using AI to post AI, such degeneracy.

[keen hollow]

ok, autism is d/l'd

let's see

yeah I dunno, this is derpy...

[nova remnant]

nice, try the image i posted as a sample

[keen hollow]

I might just stick to torq's model LOL

yeah this is based on your sample

[nova remnant]

hmm i wonder

[keen hollow]

I just changed a couple words in the prompt

[nova remnant]

did you use send to txt2img from pnginfo?

[keen hollow]

ill run yours exactly

yeah I did

[nova remnant]

shouldnt be diferent

[keen hollow]

there's no planet where this is better than torq's model... 😛

[nova remnant]

lol

[keen hollow]

here it is before hi-res

[nova remnant]

Better in what way

[pure oyster]

Bread and lettuce for dinner? She should be crying from a dish so shameful.

[nova remnant]

very close

do you have adetailer on too?

[keen hollow]

not on

running hi-res now

[nova remnant]

trun on the adetailer by its default settings

[nova remnant]

I put potato in the prompt, it just gives a blend of food lmao

oh dont use hires yet

i rarely use hires fix on those image i generate

im wondering about exact sample output

the seed, settings and all

[keen hollow]

the example you gave me had hi-res on it... I'm running it exactly as it is 🙂

almost done

[nova remnant]

but @keen hollowpony/austism will vary from sdxl a lot from using artist styles

[nova remnant]

yeah i wonder why that happens i had hires fix off

a1111 turns it on by itself which ive noticed on my end too

[keen hollow]

oh ok weird...

yeah hi-res came out garbage because it didn't put the settings in right LOL

[nova remnant]

yep

some kinda glitche in a1111

[keen hollow]

yeah

maybe adetailer confuses it

lemme re-run with adetailer

(adetailer uses pickle models that set off all kinds of alarms, btw)

[nova remnant]

you are probably right about adetailer messing that

[keen hollow]

crazyeyes lol

[pure oyster]

is there a "download" for sdxl, or do you just download the models?

[nova remnant]

i left adetailer to its default except for Detection where i set detection model confidence threshold to 0.6, its usually 0.3 or something

[pure oyster]

this is my average sdxl image

[steel arrow]

A little bit of 'horror'

[nova remnant]

Hmmm.. Something is wrong

Maybe you need adetailer to fix it

[keen hollow]

yeah it carried all that over... weird

[nova remnant]

it took me like several weeks to get the use concept right for autism since i was thinking along the line of regular models, but to simplify it, its alll in the artist style

#1 the generic prompt templete
#2 artist name

[keen hollow]

yeah this is the same thing with a very simplified description

ill try adding an artist

[nova remnant]

the fun part you'll get to is when you will see how nicely pony/autism follows booru tags

[keen hollow]

i dont love using tags

I prefer natural language prompting

[nova remnant]

sure that works too

Some cases natural language completely doesn't work

Time to sleep I think

[nova remnant]

that actually looks good

[nova remnant]

nice!

[keen hollow]

eh...

[keen hollow]

feels uninspired.. and didn't really follow my prompt

[nova remnant]

oh ummm lol i thought the image looked nice,

[keen hollow]

i mean it's a fine image

just doesn't feel like i guided it

[nova remnant]

i think thats cause these models tend to listen to booru tags a lot

[keen hollow]

after using lightning this feels like it takes forever LOL

[nova remnant]

i mean i could write something like ... a woman sitting on a log under an apple tree, or i could type woman, log, apple tree, outdoors

[nova remnant]

a woman, sitting, on wooden log, under an apple tree, detailed face, cheery, sundress, green background,

[keen hollow]

yeah I dunno.. I'm not sold...

47 seconds with autismpony

[nova remnant]

cant blame you, it takes a bit of getting used to lol

[keen hollow]

5 seconds with @nova remnant 's model

[nova remnant]

but you will love more you get familar with it

[nova remnant]

thanks, regular sdxl are ok too, but pony/autism are more precisly artist focused

[keen hollow]

austism...

[nova remnant]

these pony/autism models come alive when you find the right artist style

[keen hollow]

torq's

those were with Izumi Tsubasu

[nova remnant]

yeah but i mean the one you will love

let me share a file here