🍥｜anime

[vital raptor]

and it just adds the hires steps to the existing image, so you aren't generating from scratch

[lapis orchid]

right now im really needing to inspect my images. ive gotten some i liked then upon closer inspection i noticed flaws in them

[vital raptor]

flaws are hard to avoid, you can always go back and inpaint them later if they're bad enough

[lapis orchid]

inpaints something im trying to get a hang of

like i can manage to have it fix hands once in awhile but cant seem to get the hang of major fixes

like in the ones i been generating i been having issues with the hair out of place

[vital raptor]

it's definitely a good tool to get familiar with, it can save a lot

[lapis orchid]

but the generations have gotten better now that ive lowered the steps

went from 50-30 on the samplar.

[vital raptor]

anything beyond a hand/foot/finger replacement or other minor things and I just generate a new image, it's usually not worth spending the time

[lapis orchid]

it might be because im new to this that i generate alot of images i like

like when its done ill show you the current one. its nice but her hairs out of place

i dont think hi res will fix it

[vital raptor]

yeah, at first it's like "omg, it's all amazing!" and then a month later you're deleting thousands of garbage images you made 😄

[lapis orchid]

like i just generated this one. didnt notice many flaws in it

i mean the continuity works

[vital raptor]

it's infected with white dot syndrome 😄

[lapis orchid]

white dot syndrome? do tell

you mean the particles?

[vital raptor]

like, on the wall above her feet

yeah

[lapis orchid]

i have light particles in the prompt

i could remove that

[vital raptor]

first you see them above her feet, then notice them on the wall to the left, and on the sheet under her

a lot of XL models will just generate those anyway, it's something to do with using negative prompts

a lot of my inpainting time is just removing those dots

[lapis orchid]

ok the next one im generating doesnt have that.

but it has a random lantern

it would be one to consider keeping otherwise. should i hires fix and up the denoising strength or send it to img inpaint?

[vital raptor]

up to you really, in that image they mostly don't seem out of place

when you don't ask for light particles but they're all over the image it sucks

that one is particularly bad for some reason, the 3 before it didn't have any at all.. a quick pass through inpainting and they clean up nicely

[lapis orchid]

ill show you what i mean here in a second

how would you go about removing them?

[vital raptor]

i just use inpainting and put a dot on each one

[zinc flax]

btw you should remove that score_9 stuff, it won't make any difference to the model you're using

It's illustrious base

[lapis orchid]

i dont need any of those then?

[zinc flax]

You don't

[vital raptor]

only for pony based models

[lapis orchid]

oh right i was originally using the pony one and started generating in illustrious to see how itd look there

[zinc flax]

score_9, score_8_up, score_7_up, source_anime remove these basically

[lapis orchid]

i think one of them was a prompt for a lora. i should see which one it was

actually it wouldnt matter cause i think that lora is for ponyxl anyways

so this is what i ment. it just painted a lantern floating in the sky. otherwise i think this ones pretty good.

[zinc flax]

Can it not be removed using negatives?

[lapis orchid]

so like hi res fix and add it to negatives?

[zinc flax]

Yeah just put lantern to the negatives

[lapis orchid]

ok ill try it

[vital raptor]

doesn't specifically have to be during hires fix either, you can just generate the same seed with small prompt changes and see what happens

[lapis orchid]

mmmm yes you are right

i forgot about the seed.

i could try that aswell

i may do so as i forgot to remove the scoreprompts anyways

[lyric zealot]

Higher steps usually mean higher details, but to a certain extent

It depends on your sampler, if you are using Euler A, or some others, it will not converge and always try to add “additional” noise to the image before denoising it, so as to “add” details

But most samplers will converge, meaning going past the converging point, additional steps will not change/improve the image at all

[lapis orchid]

i wonder if thats what it was then. i was still using euler a. i was watching some videos and i did noticed that dpm++ 2m with karras kept the original image pretty close

also i been keeping the schedule type automatic for euler a

idk if i should be using a specific schedule. i havent dived into what schedule does much

[lyric zealot]

Yup, so if you are using Euler a, then even at 1000 steps, you will still see slight changes to your image, as that’s how it works for that sampler

I can’t really remember about schedule type stuff

Is that the Karras stuff?

It’s been forever since I’ve generate stuff

https://media.discordapp.net/attachments/1182548209508302899/1182548212377194526/neuroYAY.gif

[lapis orchid]

i think for this batch i been using euler a automatic the whole time

so i did a hires fix at 1080p and the lanterns not as big of a deal anymore. i did 10 steps at hi res euler a and denoising at 0.5

[lyric zealot]

Then just set your steps to some decent values that you like, since it will never converges

[lyric zealot]

https://cdn.discordapp.com/attachments/1296776525563297802/1296785903653752953/ezgif-6-033fa713f6.gif

[supple raptor]

Align Your Steps, sometimes abreviated to AYS is usually my preferred scheduler for Euler and Euler_a. This scheduler is a bit different than others though. It is optimized for a specific number of steps. If no number is specified in the scheduler name, or if the name includes Nvidia, then that number is 28, which means you want to use 28 steps for both the low and high res passes for best results. Imo it's results are often the best.

[lapis orchid]

i see that in there now. the AYS. i dont see anything about nvidia though

also i removed the scores from earlier and added the seed and already got a pretty generation.

[supple raptor]

Which UI are you using?

[lapis orchid]

web-ui is thats how you say it

auto1111

[supple raptor]

I haven't run that one since AYS was added. If it only has 1 entry related to that though, I would assume it's the original 28 step version that Nvidia released

[modest fog]

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

[lapis orchid]

you are suggesting euler a with ays at 28 steps? atleast with using euler a

[supple raptor]

yep, that's what i use

not to say you can't get good results with other samplers and/or schedulers. just that i usually prefer that combination over the others for my own gens

[lapis orchid]

ill try it out. right now im just doing alot of experimenting and learning. im sure this will take time to get real good at.

[modest fog]

what is ays

[modest fog]

ugy

[lapis orchid]

align your steps

i do appreciate the info ive gotten. itll be more fun to learn with good teachers around.

[modest fog]

share some booba talk some stuffs

more fun

[lapis orchid]

hopefully i can share some good booba soon lol

[supple raptor]

Align Your Steps. it's a scheduler that nvidia released a paper for a little while back. basically instead of coming up with a formula for the scheduler that should produce good results in a targeted range of step numbers, they attempted to optimize the formula to give the best results for a specific number of steps.

[modest fog]

[modest fog]

so you always go 28 steps ?

[lapis orchid]

this was my image after the changes.

using the seed

[modest fog]

thats indeed an image

[lapis orchid]

it is an image indeed 🙂

[supple raptor]

with AYS, yes. it doesn't work well with different numbers of steps. some UI's, like reforge have some additional versions of AYS that are made by the community for different numbers of steps

[lapis orchid]

i think it should be ok at 30 steps

[supple raptor]

it will likely do better at 28 steps if it is the original version, but you can certainly experiment

[lapis orchid]

ill try it at 28 aswell see what happens

[vital raptor]

running x/y/z grids is your friend, makes it easy to setup a group of settings and then generate the same image using the different settings for comparison

like, you can run a grid using the same image in every sampler (x), and steps 20, 25, 28, 30 (y) and then narrow it down to the ones you like

or whatever grouping of settings you want to test

[supple raptor]

yep, i'm just setting up to test different sampler and scheduler combinations on noob eps 1.1 right now since i haven't tested them on it yet. my preferred settings from vpred 0.65 just worked™ so i didn't bother testing over the weekend.

it is a good idea to test across multiple prompts and seeds as well to make sure that your results aren't biased to a single prompt/seed pair

[modest fog]

If it looks good then it's good

👍

[lapis orchid]

getting better results after the tips. this one came out pretty good imo.

[junior quest]

model?

[bitter latch]

awwww

is that eyes inside eyes? what?

takes a nap

[bitter latch]

about to say " I have the power"?

last one, off to work

[modest fog]

@green temple tired bloosmol and books

[primal wharf]

RIP hands

[rough oxide]

using anlas? otherwise you could inpaint hands pretty well usually. i always use nai to inpaint my noobai images too even

inpainted both of these hands for instance, you can usually eventually get something that works

[also damn that prompt is concerning]

[fervent tulip]

Boobies

[fervent tulip]

f off

fuck

you

bitXc

lmao

[rough oxide]

wat?

[fervent tulip]

suck my d#ck

mf

@dull jackal

SUCK MY 🇩 🇮 🇨 🇰

@wheat falcon

Thanks

@wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon

@wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon

@wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon

@wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon

@wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon

@wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon

@wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon

@wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon @wheat falcon

[bitter latch]

oh look someone to block and prob get banned

[rough oxide]

is this a ban speedrun or something?

[bitter latch]

haha

[fervent tulip]

yup

someone dared me

[rough oxide]

maybe get better friends

[fervent tulip]

fine

[bitter latch]

kinda ironic since one of us here is named cs1o

[fervent tulip]

my bad

[bitter latch]

and prob has the power to ban ya silly self

[fervent tulip]

its 11am

@wicked lake

hello

[bitter latch]

uhh oh now ya done it

[wheat falcon]

Imagine if I would had discord notifications enabled 😮

[fervent tulip]

i dont give a sh#t

[wheat falcon]

Yea bye noob

[fervent tulip]

F

u

[bitter latch]

he's prob a smurf anyhow

[rough oxide]

i was thinking the same

[bitter latch]

ya know you been around to long when.....smurf

[wheat falcon]

Probably, since a few days its crazy how much spammers we have

[bitter latch]

hm now I wanna make smurf art

[rough oxide]

but uhhh cs1o since ur here

that's one damn concerning prompt

[wheat falcon]

Oh true we will look into that, thanks

[rough oxide]

<3

[bitter latch]

oh boy

hmmm...

time for a dns ban

hey @wooden coral

gee I love seeing blocked msg's. hahaha , sigh, lame, yawns

the wheels spinning but the hamster died

[rough oxide]

yeh pretty much

[bitter latch]

back to art

hmm wonder if they know I can't see their msg's... yawns

well thats not miku she-ra

tho looks like she took muku's hair as a trophie. uh oh

[rough oxide]

i wish medabots didn't die out, the games were fun

[bitter latch]

there more anime

[wooden coral]

Heyo

[bitter latch]

puts up the chat

this @wooden coral

haha

smurf miku cyborg?

spooky

ponders bed

she mad.. 😄

going to bed. here ya go

[junior quest]

model?

[mental marten]

#🍥｜anime messageyaml (quasarcake:1.1), (ebifurya:0.7), (wlop:0.9), 1girl, solo, konpaku youmu, white hair, short hair, black hairband, konpaku youmu \(ghost\), green vest, puffy short sleeves, green skirt, sitting under tree, sleeping upright, outdoors, nature, scenery, best quality, amazing quality, very aesthetic, absurdres Negative prompt: (multiple views:1.1), (bad:1.05), error, fewer, extra, missing, worst quality, jpeg artifacts, bad quality, watermark, unfinished, displeasing, signature, extra digits, artistic error, username, scan, (blurry, lowres, worst quality, low quality:1.1), ugly, (bad anatomy:1.05), artist name, (watermark:1.2), (patreon username:1.2) Steps: 28, Sampler: Euler, Schedule type: Align Your Steps, CFG scale: 4, Seed: 2593261862, Size: 832x1216, Model hash: 6681e8e4b1, Model: noobaiXLNAIXL_epsilonPred11Version, Denoising strength: 0.5, Hires upscale: 1.5, Hires upscaler: 4x_NMKD-UltraYandere_300k, Emphasis: No norm, Version: f1.1.0-v1.10.1RC-latest-839-gde03257f

[green temple]

Nice , I like the one on the right

[royal saffron]

cat run

[fallow shell]

how

[green temple]

[gloomy scroll]

detailed background, best quality, masterpiece, cinematic lighting, highly detailed,

BREAK
(outdoors, ground vehicle, tree, road, car, sky, motor vehicle, cloud, building, day, street, sign, power lines, blue sky, lamppost, road sign, city, utility pole, crosswalk, sunlight, plant, traffic light, fence:1.3),

BREAK

BREAK
1girl, solo, (small breasts, open hair, long hair, straight hair, white hair, hime cut, blunt bangs, azure eyes:1.3), (pointy ears), dark, choker, fangs, blush,(full body), standing, facing viewer,
(modern clothes, shorts, short white hoodie, sneakers),
happy, :D,

<lora:LoraMerge_dim128:1>
<lora:add_detail:-0.5>
Negative prompt: (naked, nipples, cleavage, skirt, serafuku, navel, dress, skirt, coat, Coat tail, trench coat, cape:1.3), verybadimagenegative_v1.3, (low quality, worst quality, bad anatomy, bad hands, normal quality:1.5), artist name,

(claws, holding, holding bag, black hair, animal ears, ahoge, twin tails, clothes writing, (hair ornament), hood, hat, ribbon, hair bow, bow, headwear:1.3), (mini skirt:1.2), (leotard:1.3), holding skirt, lifting skirt, clothing pull, animal ears, depth of field, houses, wizard clothes, witch clothes, dress, serafuku, navel,
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1029913699, Size: 1280x720, Model hash: b32ed1c950, Model: dpep-b64-2, Denoising strength: 0.5, Clip skip: 2, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate/erode: 32, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 0, ADetailer use separate steps: True, ADetailer steps: 10, ADetailer model 2nd: hand_yolov8n.pt, ADetailer confidence 2nd: 0.3, ADetailer dilate/erode 2nd: 32, ADetailer mask blur 2nd: 4, ADetailer denoising strength 2nd: 0.4, ADetailer inpaint only masked 2nd: True, ADetailer inpaint padding 2nd: 0, ADetailer use separate steps 2nd: True, ADetailer steps 2nd: 10, ADetailer version: 23.7.9, ControlNet 0: "preprocessor: none, model: control_v11p_sd15_openpose [cab727d4], weight: 1, starting/ending: (0, 1), resize mode: Crop and Resize, pixel perfect: True, control mode: Balanced, preprocessor params: (512, 64, 64)", RP Active: True, RP Divide mode: Mask, RP Matrix submode: Horizontal, RP Mask submode: Mask, RP Prompt submode: Prompt, RP Calc Mode: Attention, RP Ratios: "1,1", RP Base Ratios: 0.2, RP Use Base: True, RP Use Common: False, RP Use Ncommon: False, RP Change AND: False, RP LoRA Neg Te Ratios: 0, RP LoRA Neg U Ratios: 0, RP threshold: 0.4, Hires upscale: 2, Hires steps: 8, Hires upscaler: 4x_foolhardy_Remacri, Lora hashes: "LoraMerge_dim128: 45063301a4db, add_detail: 7c6bad76eb54", TI hashes: "verybadimagenegative_v1.3: d70463f87042", Discard penultimate sigma: True, Version: v1.5.1

Image itself has metadata so you can also just save it and drop it to a1111 or something else that knows how to read it

And I probably used this for cnet

[fallow shell]

how can i create an image with same background but different poses..

[gloomy scroll]

Not rly my area of expertise

[junior quest]

fooocus inpaint is better than a1111?

[steady grail]

i don't like a1111 too much

cam most of the stuff, but a little so so

hello

[modest fog]

You like the big one

[green temple]

just the shawl

[supple raptor]

fooocus dev moved on to making forge and will only update it to fix bugs. it's dead. if you want good inpainting either use invoke, which has by far the best built in inpainting, or a krita plugin that uses the comfy or a1111 api.

[modest fog]

steals her drink

Am hungry

[modest fog]

eats her

[dull jackal]

Good evening, everyone! How are you?

[rough oxide]

gud. just genning like usual.

[modest fog]

ugly frog

[rough oxide]

it's simply gekota, the mascot from railgun/index ;_;

[rough oxide]

them's fightin' words.

[modest fog]

green is always ugly nothing we can do about it

[junior quest]

krita plygin? this one? I got no idea how to work on this

[bitter latch]

Krita is a external art program https://krita.org/en/ ... so.. no.

[junior quest]

only work with comfy?

[bitter latch]

it's a art program

and I have no clue about the plugin

[terse ore]

What pony model, if any, is particularly good with (anime) hands? I want to use it to inpaint hands and make hand poses based on prompts, rather than using longer workarounds like controlnet. If possible.

[rough oxide]

don't think any model has really solved the hands problem.

[modest fog]

the more you train a model for hands

the more cursed it gets

[gloomy scroll]

Feeling very Padoru

[near elbow]

coming for you 🙂

[gloomy scroll]

Get in line

[modest fog]

@green temple bloo magic

[bitter latch]

long neck

[terse ore]

What model is good at making many straight small lines? Like a brick wall 🧱, window sills, wooden slats, tiled floors, etc

AI usually messes those up

[rough oxide]

probably flux/sd3.5l, they're way better at background details but you won't be able to get any specific anime chars aside from omega popular ones.

[junior quest]

nvm my bad prolly

[bitter latch]

elf!

there more animmayish

falls asleep

(snoozes)

[upper hemlock]

Got brainrot

[keen hollow]

ew.. guts

[modest fog]

Robot cables ?

x3

[modest fog]

Looks like some tifa clothes @keen hollow :b

[green temple]

guys how to delete stuff with inpainting 🌚 🥲

send an example pls

like these lines

[terse ore]

I won't have PC access until I'm home. But you can make a mask of where the lines are and describe what you would want in place of it. Ie if the line is over the sky then inpaint it and prompt "dark blue sky"

[supple raptor]

inpainting to remove things is usually pretty annoying since you need super high denoising, but then the masked area often doesn't match the rest of the image. i usually just draw over the areas as well as i can then inpaint at fairly low denoising to make it better match the style if needed.

[green temple]

oh i thought it was easy

np im trying

[green temple]

high denoising added a little shading but it`s not that visible ( it works) 👍

i tried

[supple raptor]

Nice, I often end up with the colors not matching up at the edges when i try that, especially in dark areas like the blue in the background. raising mask blur can help with that somewhat, but it's often still noticeable if you look for it

[green temple]

yes it is , but it`s ok

[supple raptor]

a trick you can do if it bothers you is to just run the whole inpainted image through img2img at like 0.2-0.3 denoising. it'll change the image a bit, but will usually hide the areas that you inpainted

[modest fog]

draw over it yourself with similar colors and try inpainting or img2img back its usually easy just takes a little tiny bit of time

[green temple]

nice i`ll try

[green temple]

smol drawing is hard

[supple raptor]

just post this frog

[supple raptor]

https://tenor.com/view/suwako-post-this-frog-touhou-fumo-touhou-fumo-gif-26420843

[modest fog]

Is not

[modest fog]

Pink

[green temple]

the left w

[modest fog]

w as in something or as in win

[green temple]

as in win

what`s the "something" ?

[modest fog]

I dunno

[green temple]

ok

[modest fog]

Maybe something looked like a W there

[green temple]

💀

[modest fog]

I'm just confused cat

[wooden sail]

Nice hat that was

[modest fog]

@green temple happy smol to forget problems

And a bnnuy

[terse ore]

Recommendation for a pony model that has very simple aesthetic? Like simple flat colors, basic design

[supple raptor]

is there a specific reason you want pony? noobai is just better in every way unless you're specifically going for that more western art style that pony has by default

[terse ore]

Can noobai use anyline/misto controlnet? I may as well use that then. But I just need something to inpaint backgrounds in a simplified anime style

[supple raptor]

noob has it's own controlnets that are actually trained for it, i don't remember which types of controlnets they have out right now since i rarely use them and they are still working on them, but i know they have several controlnets as well as an early version of an ip-adapter that they're working on for it which already works quite well

https://huggingface.co/Laxhar check the collections for the relevant sections. for now, i'd skip the vpred models as they have some weirdness, but hopefully that will get better once they finish training it.

the ip-adapter is here https://civitai.com/models/1000401/noob-ipa-mark1

controlnets are also here if you'd prefer civitai https://civitai.com/models/929685?modelVersionId=1097585

[terse ore]

Ok thanks 👍 mainly I was having issues with older stuff being unable to make many straight lines like a brick wall, tiled floors, etc AI usually don't do too well with those

new models too, but they're probably better

[supple raptor]

flux would probably be your best option for that but requires higher specs to reasonably run it, has far less flexible styles without using a lora, and basically 0 character knowledge (if you care about that)

oh, it's worth mentioning, that noobai models are not aesthetically finetuned, meaning you should not expect the style to just look good by default. it is expected that you will use artist tags to create the style that you want. the model is trained on all of danbooru's 2024 dataset up to ~October and responds to tags with as few as ~80 uses on danbooru pretty reliably.

[terse ore]

Oh then maybe some kind of flux model. I'll have to try it too. There are different kinds. I'd have to look up simple artwork styles and find artists I like to try using those styles I guess

For noobXL

[supple raptor]

you can check the gallery for the model on civitai for some ideas of different styles. many people include their prompts https://civitai.com/models/833294?modelVersionId=1116447

also basically everything i've posted since early october has been from one of the NoobAI versions or Illustrious and should contain the metadata so if you've seen something from me with a style you like, you can check the metadata to see what i did

[terse ore]

Thanks for your vast contributions ☺️

[supple raptor]

Coke

[desert island]

fun... but man I'd have to do a lot of inpainting if I even wanted to clean all this up. ☹️

[green temple]

2koma , 4koma negative will work

[modest fog]

Sooma helps

[green temple]

smol help

[modest fog]

@green temple

[green temple]

[modest fog]

but

Am tired

[trim current]

https://tenor.com/view/bonk-gif-26414884

[bitter latch]

holy heck .. civitai has went full retard on ads.. wow

and the scripts they use now is crazy

[green temple]

split background ?

[desert island]

it's a lora, and workflow that uses your character in ipadapters.

[green temple]

oh

[desert island]

Okay last one but this was fun 😄 thanks @green temple

[green temple]

welcome

[green temple]

what happend to your dreamshaper+ animaginexl mix ?

[desert island]

this is still it.

[green temple]

oh ok

i thought you changed it

[desert island]

nope I love it I just change the weights a bit sometimes between what's more important from realistic to anime... but this particular weighting seems to be able to kind of switch between reality and anime... ish a little better. depends on what sort of images I use in the ipadapters and prompting.

[desert island]

the eye flare effect is cool plus I love the pattern on the blouse

[green temple]

thx

[desert island]

well my Jinx attempt was not great ... 🤔

😁

[steady grail]

depth cn?

[desert island]

yup and pose data

[steady grail]

gj

[acoustic swallow]

Hi everyone~! =D

[steady grail]

hi

[desert island]

and the counter

[terse ore]

Is flux1-dev-Q8_0.gguf the correct one to get for comfyui?

[desert island]

depends on your workflow.

[terse ore]

It's 12gb and my GPU is also 12gb vram. So I am not sure if this is the one

[desert island]

shrug don't know. I have a q_4 version of flux on my system but I don't want to load it up since Flux for whatever reason causes my system to crash from time to time.

[supple raptor]

reminder of why you should probably still be using xformers at the end of 2024

[green temple]

doesn`t matter with forge , right ?

[supple raptor]

it does. differences between the two should be comparable across any UI

also, when installing xformers you have to be sure to include the index url for the pytorch branch you're using, if you just do pip install xformers you will end up with torch and xformers that are for cpu only. happy to help if you have questions on how to do it right.

if anyone here is using ROCm 6.1 on linux, i'd be interested to see comparisons with their new experimental xformers wheels for ROCm since i don't have an amd gpu to test

[green temple]

WARNING: xformers: A matching Triton is not available, some optimizations will not be enabled.
Error caught was: No module named 'triton'

[trim current]

Stability Matrix is a godsend if you enjoy multiple uis

[supple raptor]

normal. triton is linux only. you can just ignore that

[supple raptor]

seems like a pretty neat idea, but i already hate the portable installs of all of the standalone UIs. a portable install of a UI that manages all of those portable installs sounds like a nightmare in practice to me

[trim current]

should try it for yourself and see

[supple raptor]

does it allow you to easily manually install/update dependencies or is that all done automatically for you?

[trim current]

all done automatically

[green temple]

is that ok 😅

[supple raptor]

yeah, i wouldn't like that. i prefer to be able to manually update dependencies. torch and xformers get updates ~monthly and new versions tend to be slow to be officially adopted by all of the uis

[supple raptor]

i don't see anything problematic here. temp is safe and you're not exceeding your dedicated vram

[green temple]

👍

[trim current]

then dont update

when something new comes out

??

lol

also its a frontend, nothing stopping you from going in the venv of where u want and manually do stuff there

[desert island]

I think that's what they were implying they would do 😁

[green temple]

👍

[supple raptor]

that's precisely why i don't like the portable versions, they don't use venv which makes manually updating those packages more cumbersome and means when something goes wrong, the simple solution is to delete your entire install and start over from scratch instead of simply deleting the venv and rebuilding just that

[trim current]

then i guess i was wrong its not portable

venv at the bottom there

[green temple]

i like the lighting

[desert island]

golden hour with bokeh

[green temple]

👍

[supple raptor]

oh cool, in that case this seems much more interesting. still quite a bit of work to switch everything i have over to it, but if i need to re-install stuff in the future i'll definitely give it a try

[trim current]

everything is symlinked too like models and vaes etc

[vital raptor]

I just use only 1 UI, keeps things nice and simple

[trim current]

comfy for more complicated stuff that yields in better quality, forge for flux [comfy too], reforge for day to day

[vital raptor]

i just use swarmui, it has comfy built-in if i need to get advanced

[steady grail]

symlinked? like mklink?

[trim current]

swarm is nice need to learn it more

[vital raptor]

to be fair though, i also don't have any hardware limitations to worry about

and the few problems i've had with it, i just go into their discord and talk to the dev about adding the feature, or just add it myself

[trim current]

i use a 3060 so i cant train sdxl and up but i can use flux and illustratious etc no issue reallt fast

flux not so fast tho, its like 15sec or so a step

[vital raptor]

yeah flux is slow, even on a 4090 it takes ~15s for a base generation compared to 2s for sdxl

[trim current]

right?

neat tool

[green temple]

how much xl takes for 1 image gen with highres ?

[trim current]

with highres? a minute but i only hires stuff i like

[green temple]

nice

[vital raptor]

2x hires sdxl takes me around 10s.. it's horrible 😏

[modest fog]

Cookie

[green temple]

🌚

[trim current]

4090 life

[modest fog]

Don't listen to captain, he's bad person

[vital raptor]

[green temple]

with no highres takes me 3 minutes

[trim current]

thats awful

[desert island]

what resolution is hi res?

[green temple]

hires off

832x1216

[desert island]

ohhh hmmm I always gen in 1920x1080

[modest fog]

@green temple bloo and cookies

[desert island]

well... to be fair I start in 720p but then upscale to 1080

[green temple]

[vital raptor]

that's what hires is

it's the upscaling

[desert island]

I was just trying to figure out if hi-res meant 4k+ or not

[vital raptor]

oh, no, it's what people use to refer to upscaling because a1111 called it 'hires fix'

[desert island]

got it... I only spent a little time with automatic before switching to comfy so I wasn't aware of the lingo thanks for the edification 😄

[supple raptor]

theSaddest

[desert island]

kill it with fire!!!!

[supple raptor]

hires fix is just the common term used for the process of generating at a lower resolution (usually the model's native resolution), upscaling the image, then running img2img on the upscaled image

[trim current]

need to get comfyui live inpainting set up and ifgured out

[desert island]

is there such a thing? I just paint masks on images and re roll on that image... is there a better way (like some node that does it in a better way)?

[trim current]

https://github.com/diStyApps/ComfyUI-disty-Flow

[supple raptor]

i wonder if there's a live inpainting thing for invoke? their inpainting ui is a lot better than the options in comfy

[trim current]

im gonna need to download and mess with invoke

[supple raptor]

i don't really like using it for normal generations, but inpainting and outpainting in it are really nice

their regional prompting is also way better than doing it in comfy or swarm even with swarm's new region ui thing

[trim current]

i see it on matrix so deff be easy for me to mess without without needing to do much

[supple raptor]

idk if matrix's symlinks avoid needing to do it, but when i installed it, i had to let it scan the models and loras i wanted to be available to it, which was a very slow process (like 3 hours) but i have ~1tb of models and they're on a slow HDD so YMMV

[vital raptor]

i should update swarm apparently, didn't know there was updates for the regional prompting

my home internet has been out for 5 days so i missed that

[supple raptor]

yeah, there's a + button next to the prompt field now that can be used for a few things, one of which opens a UI to visualize the regions

[vital raptor]

neat

that reminds me, after some struggles i was having with regional prompting recently it made me want to reinstall a1111 and try to do the same thing to see if it's a prompt/model thing or if there's something in the code that just makes it bad

like, 90% of the time just ignoring most of the prompts

[supple raptor]

i've had the same struggle for the most part with swarm as well as previous attempts in a1111. with invoke It Just Works™ and uses drawn masks instead of vague bounding boxes

[trim current]

interesting how this FluxGym trainer has 12gb support but i cant get xl training to work with 12gb

[summer night]

can i update forge without breaking my extensions now?

i've been using the same build of forge before extensions broke this whole time because auto1111 is still too slow

[vital raptor]

man, swarm's inpainting interface is not intuitive at all, but after playing around with it enough I think I finally got it figured out

once i realized i could just erase bits of the original image completely, then mask over the empty area it made things so much easier

original image for comparison

[vital raptor]

Here's another one, wanted to remove most of the pink stuff behind her so I started with selecting the 'image' layer and erasing everything i wanted gone, then filled the empty area with the paintbucket tool and masked over the same area and generated a new image. Then masked a few smaller things to clean up other details

then used the 'refine image' button for a 2x upscale using 0.2 control percentage (hires fix)

don't mean to spam, but I figured this might be useful for anyone trying to learn inpainting in swarmui 😄

[green temple]

Smol inpainting

[vital raptor]

[humble current]

my shortest prompt yet

"A pouty girl with too many cookies to eat"

[gloomy scroll]

I don't have onesie cookie muncher so have this instead

[desert island]

cookies and milk for everyone!

[supple raptor]

A situation that you don't have the perfect Yue for?!

[humble current]

bring your best dress

[vestal ferry]

https://tenor.com/view/doro-dorothy-nikke-gif-4582005177308983234

[green temple]

no cookies for smol haters

[modest fog]

[green temple]

tall smol

[modest fog]

It's a long shirt

(really, I cut the image even)

She stole my shirt

[green temple]

no it`s @vital raptor shirt

[modest fog]

@green temple what kinda of challenge am I putting smol in

[green temple]

why is smol in water

[modest fog]

I just wanted water

Is blue

@green temple 🔵

More bloo

[modest fog]

Gift for me?

[green temple]

yes

[trim current]

anyone decent w controlnet

obligatory @supple raptor ping

[supple raptor]

I don't use it very much, but I might be able to help

[trim current]

ill move convo overf to dms

[dull jackal]

love the painting style for these

[dull jackal]

this is super cute, too

[green temple]

thx

[vestal ferry]

https://tenor.com/view/march-7th-m7-doro-doro-doro-hsr-gif-2288130168749029939

[desert island]

very pretty flowers

[green temple]

thx

[modest fog]

Grabby

@green temple

[trim current]

there, vae was being weird

[desert island]

you really gotta zoom all the way in to appreciate it. Kudos 👍

[trim current]

[green temple]

[bitter latch]

ooh

[bitter latch]

meh that last one was crap

[bitter latch]

hmm temple of miku?

[modest fog]

Red

Scary

[tawny wasp]

my Creation 🙂

[green temple]

better than half of my gens but booba 👎

[tawny wasp]

hi, i can't understand what you're saying. do you like it or not?

[green temple]

😂

[green temple]

i like the style only

[tawny wasp]

ahh,ok 🙂

[green temple]

😅

[keen hollow]

My ultrawide Aerith and Tifa from yesterday are gone... guess someone didn't like those. 🤔

Thought they were just dynamic.. maybe I missed something? 😐

sorry if so! 😬

[bitter latch]

interesting style.. what is it .. chalk?

chalk crayon smudge?

[slate narwhal]

I don't know exactly; it looks like there is chalk + minimal. The main thing comes from a specific MJ sref trained on Flux, then my own Lora mixed in. I think the chalk effect is from my Lora. Here is a comparison: First is the original sref Lora; second, the sref Lora + my Lora

Sref + My Lora, same prompt as the original, with the cat in:

[vital raptor]

i like those last ones

[green temple]

me too

[slate narwhal]

The problem is my Lora creates lots of variations by itself and has several styles mixed.

so things can get really crazy;

[bitter latch]

ahh

hmmms.. hope there's a door dasher av for my taco bell I wants it!!!!

yes, there's one. hot damn taco bell tonight.. farts expected with hopefull no more

😄

some for now, some for breakfast!

https://tenor.com/kZ5kRPyg3Su.gif

[slate narwhal]

Yes, these two are variations of my Lora. It enhanced the quality of both. The original Sref Lora is too raw to achieve this effect

[bitter latch]

hmmm . now I gotta try with 1.5 😄

loads up forge. or is it reforge now?

mixed chalked smudged crayon and ink 😄

[bitter latch]

hmm umm.. ok...

[bitter latch]

gives up and goes back to reading

[unkempt cape]

Is it just me who fucking sucks at noob ai prompting

[zinc flax]

Cookie

[lyric zealot]

that's sick

[keen hollow]

glad you like that one

got a batch going, i'll see if there's any winners

[lyric zealot]

can i have meta data for that?

[wooden sail]

A creepy monster and a female elf

[keen hollow]

yeah of course

here's some from the set (different style, same description)

here's some with same style you liked... (wlop is the key here)

here's this one, metadata should be intact

[modest fog]

Ugly elf

https://tenor.com/view/rrane-battal-rrane-battal-kızgın-battal-kızgın-gif-25355985

[modest fog]

No

[lyric zealot]

thank you

[lyric zealot]

https://tenor.com/view/cat-side-eye-gif-5016118126108679217

[keen hollow]

But now, you gotta post some generations 😎

[lyric zealot]

it's been like forever since i gen some stuff. Been playing a lot of poe2

[keen hollow]

Today's the day!

[trim current]

you mean ruthless 2

[lyric zealot]

oh and also, my whole swarmui is kinda fucked

[trim current]

whats up with it

[lyric zealot]

I don't get the reference

no idea

[trim current]

oh poe 2 is hard af

ruthless af

[lyric zealot]

oh

i mean, yeah, but i kinda like it hard

[trim current]

i bet you do

[lyric zealot]

[trim current]

https://tenor.com/view/heladisita-the-rock-my-honest-reaction-memes-gif-45273446769020606

[lyric zealot]

i didn't really play much of poe, so wasn't expecting anything from poe2

[trim current]

oh poe 1 is easy as like

night and day difference

lol

i like the new difficulty

[lyric zealot]

yup yup

i like that you have to dodge properly, and kinda pay attention to how the boss works. Obviously in endgame, you can probably steamroll it tho

[slate narwhal]

Nice one 🔥

[modest fog]

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

[modest fog]

scary

[steady grail]

align your steps scheduler is cool, but on 1.5 it gens so many mutations 😦

[modest fog]

@green temple you are banned to outer space by space smol

[summer night]

hmm... noobai can do more characters and styles than pony, but it struggles more with hands

[supple raptor]

https://tenor.com/view/frieren-beyond-journey's-end-sousou-no-frieren-frieren-emuach-flying-kiss-gif-1166051893005797096

[supple raptor]

are you using the correct number of steps and with it?

[modest fog]

Kill it

[supple raptor]

https://tenor.com/view/anime-frieren-フリーレン-fern-フェルン-gif-7393526705896726552

[steady grail]

ehy, hello, correct number? the min number of steps is 10

i don't know if it have some other parameter to stabilize stuff, but must of the problems cames in the raandom stuff in the bg and the hands

the hands, they became horrors

without it, perfect hands, with it, mangled hands

[vital raptor]

i think AYS is specifically designed for 28 steps exactly

[steady grail]

oh, really?

didn't know that

[vital raptor]

pretty sure, searching for @supple raptor talking about it in this channel might have more info

[steady grail]

it was added by default on comfy or it's an extra? i can't find his custom node package

anyway for now i'm using at 0.10 denoise with 100 steps as second step

somewhat imroves images, without genning too many mutations

it's a core add

let's look online

[steady grail]

https://www.youtube.com/watch?v=7usmv4L8WQ8 pretty interesting

[bitter latch]

thats better

[bitter latch]

grinz

[supple raptor]

Yep. Some people have made versions that work with other numbers of steps, but it's a safe bet if the implementation you're using doesn't indicate a different number then you should be using it with exactly 28 steps.

[bitter latch]

makes her more anime

[steady grail]

i'm using the core node of comfy

[steady grail]

suggested sampler?

[supple raptor]

Euler or Euler_A

[steady grail]

oh...

[bitter latch]

face is funky

[supple raptor]

It sort of works with some others, but it was made for Euler_A

[steady grail]

i should try with a first pass with euler 28 steps and a 2° however i like

wait a second

[steady grail]

euler a cfg++ is connected in some way? or can improve stuff someway?

[bitter latch]

Im liking Dpm plus plus 2M SDE HEUN

stupid keyboard

[steady grail]

still not digging this sampler, no idea what's the difference

[supple raptor]

It does seem to do a pretty good job with the CFGPP variants of both of those samplers in my experience

[steady grail]

oh ty

i'll try, ty for the suggestions

[supple raptor]

👍

[steady grail]

anyway the image up there was done with a second pass with ays and eulera++ 100 steps .10 noise

[steady grail]

mutations are relatively low, still improving the prompt response

in my experience, it's hard to gen correct wings

[bitter latch]

is in full random mode :D

oops to random

Sailor Jedi Miku? wtf ??!?!?!

[valid flame]

Do any of you know how to make a girl and boy in a picture without having the fuse or be awkward

I need a prompt

[bitter latch]

I think Duo

[valid flame]

Can you tell me a prompt please

[bitter latch]

Duo , 1girl, 1boy .... I think

[valid flame]

Ok I guess I'll try this

[valid flame]

What does break do

[steady grail]

it's a thing of auto, or like concat on comfy

[steady grail]

let's see how to explain

[valid flame]

Yes please

[bitter latch]

split personality, miku harley ?

[steady grail]

the prompt is split in steps of 75 tokens

after this 75 tokens, the token ocunt turns to 0

[valid flame]

Tokens are texts between commas right

[valid flame]

Oh so it refocuses on something else

[steady grail]

after reading all the prompt, the model average all the tokens and you get a long sentence that makes sense for the model

[valid flame]

So it's essentially 2 prompts when you break it

[steady grail]

every word have his token count

one work weights as 3, another as 3, and another one as 12

[valid flame]

Why does she look like a man

[steady grail]

what break does is filling the missing tokens after the word break with ,,,,,,,, until it reach 75 tokens

usually you do it to weaken the connection between successive tokens

with the tokens before

[valid flame]

Why

What's successive tokens

[bitter latch]

shrugs

[steady grail]

all the toens before break and the one after

the tokens influence the tokens after them in a cascade effect

[bitter latch]

hmms and tries a differnt way

[valid flame]

Oh I know that

[bitter latch]

cackles

[steady grail]

if you add red as first token, that color will leak in most of the successive tokens

[valid flame]

So break starts fresh right

[steady grail]

unless the other tokens overpowers the color red with a different inflence

[valid flame]

With brackets

[steady grail]

something like that

[bitter latch]

puts in random breaks

oh gawd

[valid flame]

I usually just do 8 steps that wonder what'll happen if I break

So I can break two characters that might fix my issues

[bitter latch]

umm ok then

[valid flame]

Omg duo worked

Fuckin finally

[bitter latch]

ooooh I like it 10 steps

[steady grail]

cute, but mangled hands

[valid flame]

Hey it's 8 steps

Now I'll add a Lora and make him gojo flying with me

[bitter latch]

20 steps

[steady grail]

i spent a lot of time on a workflow and a model that fixed mangled hands

without having to fix them in a successive step

[bitter latch]

40....

.. 80? yes lets

[valid flame]

I can't bother dude I just get a Lora

[valid flame]

It's pointless to go higher without increasing the resolution I heard

[bitter latch]

wow

[steady grail]

most of the times, i only use style lora, because lora makes things too average

[steady grail]

just adds more stuff

[valid flame]

I work on a compact let the thing do its thing workflow

[bitter latch]

changes some settings

[valid flame]

Tell me more cool prompt words like duo

[steady grail]

well, one of these guys that spend most of their time enginering, than prompting XD

[bitter latch]

and im just random

[steady grail]

try this prompt 1girl, 2000s_(style), angel, angel_wings, belt, blonde_hair, blue_eyes, blunt_bangs, blush, borrowed_character, clenched_hand, entity_(masa), feathered_wings, frills, grey_jacket, grey_skirt, impact_desu_entity_(vocaloid), jacket, kitnomaly, loafers, long_hair, long_sleeves, miniskirt, open_mouth, pleated_skirt, shirt, shoes, skirt, smile, solo, thighhighs, translated, vocaloid, white_thighhighs, white_wings, wings, zettai_ryouiki

[valid flame]

You want me to try this?

[steady grail]

yeah

let's see how it works on you

[valid flame]

K hold on

[bitter latch]

😄

[steady grail]

oh, so gorgeous

i love this surrear effect

[steady grail]

i don't understand if you are genning miku, or joker

[bitter latch]

... yes

Miku Joker

[steady grail]

[bitter latch]

removes the breaks

[valid flame]

Are you sure about the 2000s style I could get more high res without it I think

[steady grail]

it's an anime style

[valid flame]

My gojo finished cooking too

This is closet I ever got

Help me fix the fuse

[bitter latch]

tries the prompt from above

[steady grail]

things get fused when the steps stops in the middle of a transition, or when you gen with weird resolutions

[bitter latch]

hmm.. k

[valid flame]

Took me like a month to get that resolution and the others get saturated and stuff for some reason

[bitter latch]

adds some styles

[steady grail]

realism gives me chills as always

[valid flame]

Another one done

[bitter latch]

moves from the rpg paragon merge back to the anime NOMD merge :D

[steady grail]

try with 832 x 1216

[valid flame]

It's not working

[bitter latch]

yawns, ooh almost 4am

[steady grail]

eheheh

[bitter latch]

good thing I have 2 days in a row off

1girl white haired miku angel in skirt, action scene, simple background

ok wtf are those batons coming from?

[steady grail]

must be character stuff

[bitter latch]

grinz

waits

haha

ok lst one.. I think

done

hmm viggle is taking a while

worried look

heads to bed

ok that was creepy

[bitter latch]

finally

sleepz

[valid flame]

4 fuckin steps

Goated

[bitter latch]

150 steps.. goated?

cackles insane.. so tired.. ned.. sleep THUD

puts her to bed as well.. nite nite

[valid flame]

8 steps

[trim current]

my lora is working very well in comfy, happy

[trim current]

need to change the cooking times abit but pretty sastified with the first go

[unkempt cape]

Lol just figured out somth

[keen hollow]

cool sword pose

[trim current]

ya im enjoying this

[keen hollow]

I'm trying some new stuff... it's still a bit too abstract, trying to dial it in.

[keen hollow]

Man, I am terrible with this noob model

[green temple]

i think i`m the worst when it comes to noobai 🥲

[keen hollow]

it's a very "sketchy" style model, too, so you have to work with that

[green temple]

yeah

[keen hollow]

I'm also just better at natural language models.

I don't know characters or common anime/manga descriptors at all.

So these tag-based ones are tough other than very generic adjectives for me.

[green temple]

me too , i`m trying to find a style i like

the best i found so far is "anmi" but i got bored