🍥｜anime

[desert island]

and muted 😁

[lyric zealot]

wait what, did i get muted

[brittle rampart]

Maybe a big round and incomprehensible fish (is it even a fish)?

[green temple]

egg

[brittle rampart]

The dragon will be angry!

[desert island]

ohhh this is a vibe needs some clean up, but I like it.

[green temple]

is this itadori yuji behind her ? ( the red-haired girl )

[desert island]

🤷‍♂️ kind of looks like the same sort of anime that scene would be in 😄

[green temple]

more eggs

[desert island]

Love this one. Gorgeous! 👏👏👏

[brittle rampart]

some strawberries and sleep...

Well... that sounds kind of ambiguous...

🤣

[desert island]

There's always just that one more... 😄

...I'm such an idiot I read that as ambitious...

[brittle rampart]

All! after dessert, definitely go to bed!

[dawn jasper]

Let's hang out sometime.

[finite herald]

this is so good i wanna make it my pfp

[nova remnant]

Fruit wants to steal mikus

[lyric zealot]

wait

did that guy got banned?

[supple raptor]

all mikus inherently belong to fruit so she can't steal them

@finite herald

[finite herald]

Promoted

[unique mortar]

wallpaper 1920x1080

[desert island]

by the way was not muted 😁

[desert island]

bad actor on the server...pretty sure got the 🔨 .

[near elbow]

back in action (from a DM conversation, too good not to post)

[unique mortar]

I create a terrifying pic. It is not about the graveyard environment. The terrifying thing is all about the girl. Look~ she has 3 legs~ 🤣 🤣 🤣 🤣 🤣

[brittle rampart]

she just came by to get some spare parts.

and here is the "1+1" promotion. two for the price of one.

[near elbow]

why just three? 🙂

[desert island]

holycrap 👍

[bitter latch]

no, holychibi

[sterile comet]

Springtime Blossom

Aw, this is such a cute anime girl! Her hair is a beautiful light blue, styled in two pigtails with adorable little bows. Her eyes are a striking, bright blue, and her skin is a soft, rosy hue. She has a sweet, innocent expression, and her cheeks are a little flushed. She's wearing a lovely traditional Japanese-style kimono. It's a mix of greens and blues, with delicate floral patterns. The details on the kimono are really lovely, and the obi (belt) is tied in a pretty knot. It seems like there are little white flowers scattered around her, adding a touch of springtime charm. The whole image has a soft, dreamy, and magical aura. Meow! So precious!

Neko AI

[bitter latch]

not sure what shes mad about

[green temple]

"aw, this is such a cute anime girl" 🤣

[near elbow]

"anime tsundere evil"

[desert island]

Okay I'm done with this prompt. 😈 onto something completely different...

[topaz juniper]

much good cat content today

[rough oxide]

ionoooooo

la+ darknesss too but i don't have any la+ saved right now xD

[desert island]

Finally got the weights right... sucker punch incoming... j/k 😄

[zinc flax]

She's cute

[lyric zealot]

https://cdn.discordapp.com/attachments/1280389884745482313/1291889294843052095/nai_generated_125331697867816961.png?ex=6701bd2a&is=67006baa&hm=cbaf51371399e4ec9fc3f56c04e818db283aaaa361e5b6985ed04a60c275f8a0&

[nova remnant]

Time for baguette

@finite herald baguette

[nova remnant]

Smol baguette @green temple

[nova remnant]

@brittle rampart baguette!

[mint blade]

ur adjusting weights? how does that work? is that makine ur own lora or....

also lots of bread today lol

[desert island]

no just prompt engineering... like "(baguette:1.3)" the ":1.3" part states that the weight or importance of the subject is greater. as for the bread... @nova remnant said it was baguette time. 😆

[mint blade]

aahhhh ok. got it thank you..... also lol

[nova remnant]

(baguette:1.9) or nothing

[mint blade]

can i ask how u all get such a consistent OC? was there some sort of training on a dataset or are u just extremely detailed with prompts?

[nova remnant]

I've seen a video like this once

[desert island]

In comfy UI i use an IP adapter.

[lyric zealot]

Sometimes it’s because of Lora, sometimes you just have to describe your character, and stick with it

[mint blade]

hmmm ok. noted noted.

[mint blade]

?

[desert island]

yup

[nova remnant]

Same model and artist tags/style loras will generate basically the same thing

Anime models aren't creative

Unless you want one singular little detail here and there

[steady grail]

ipadapter is incredible, but other than costantly burning down my gens, it even burned down my patience, because the slow gens

[lyric zealot]

Your weight might be too high

[mint blade]

alright. 🧑‍🎓

[steady grail]

yeah, i know, ipadapter is very heavy and need a countinue touch on the settings

[nova remnant]

Nice way to say fat

[mint blade]

ngl ive never even tried a LORA yet. ive been using the illustrious model, and idk if its ok to use loras with dirivatives of models. lol

after all arent the loras designed to fine tune model weights?

[steady grail]

he fat shamed me

[lyric zealot]

https://cdn.discordapp.com/attachments/1067638175478071307/1288208397849722880/mods_get_him.gif

[vital raptor]

I just use the same prompt for the character and then add clothes, location, etc to the end of it

[nova remnant]

You need loras specifically for illustrious

[lyric zealot]

Don’t believe @nova remnant

[mint blade]

haha

[steady grail]

i'm split in two tonight i'm finetuning a new very promising mergin workflow

but i need to load 3 different models while testing it

[mint blade]

ill try it

[steady grail]

and everything became so slow

[desert island]

CR Model Merge Stack is the node I use.

[lyric zealot]

make sure those models are on your ssd for faster loading

[steady grail]

i'm working in transferring styles from a model to another, by just transferring particular blocks, purifying and merging them

[vital raptor]

all this bread is making me hungry

[steady grail]

i'll try to move and load them, ty

@lyric zealot

[green temple]

ne too

[desert island]

same... now I gotta make a damn sandwich.

[vital raptor]

i always forget there are people out there still using non-ssd storage

so bizarre

[lyric zealot]

yup yup, make sure when you are always loading and unloading, put it in your fastest drive, it will drastically decrease the amount of time taken

[lyric zealot]

https://tenor.com/view/poor-no-money-gif-24226168

[mint blade]

wow great hands lol

[vital raptor]

if you're poor, don't waste your money on crap hardware 😄

[steady grail]

this particular approach is prone to gen mutations, so i'm looking to set things better to stabilize it, but maintaining the model creative enough

for now, when it not generate mosters, the results are very promising

[lyric zealot]

using nai

[lyric zealot]

yup, be like me, just put some commonly used models/lora at your ssd

[steady grail]

my objective is to make animerge cuter >.<

and for all that is good in this world, i will make it cuter >.<

[desert island]

for me I just had to play with the ratios until I got good consistency. Ratio's for both the model the clip mind you... I currently have 3 models merged... using Animagine, Dreamshaper, and SDXL... although now that I'm looking at it again... I cheated with dreamshaper...

[vital raptor]

ssd are my slow drives, I don't buy anything but nvme now

[steady grail]

ehehe

[lyric zealot]

https://cdn.discordapp.com/attachments/1067638175478071307/1266853046546337915/NeuroPoor.gif

[steady grail]

nvme ?

[lyric zealot]

the faster "SSD"

[steady grail]

i shoul look for them and move my merging workflow on them

[vital raptor]

they take a different connector in the PC

NVME is direct pci-e connections, not sata

[steady grail]

100€ 1tb is affordable

[vital raptor]

most new motherboards only have 1 or 2 nvme slots, if any

[steady grail]

oh....

that's a problem

[vital raptor]

I had to specifically buy one with 4 slots when i built this computer

[steady grail]

they are already in use

[vital raptor]

they're specific slots only for nvme, not the normal pci-e slots

[steady grail]

i must have confused them, let me look what they looks like

[vital raptor]

that reminds me though, i still have an extra 1tb nvme sitting around not being used, i should do something with it

[steady grail]

found it

near the cpu and ram

only one?

[desert island]

https://tenor.com/view/rucoy-highlander-bees-knees-gif-13863171

[desert island]

Highlander? Anyone? ... smh... I'm old... 😭

[steady grail]

i'll buy one, i like the idea to dedicating genning on an hard drive alone

[vital raptor]

depends on the board, but most pcs only have 1 or 2

depending on how old it is, it might not have any

[steady grail]

if that's the one, i can find only one of them

[vital raptor]

just make sure you actually have one, adding another ssd might be easier/cheaper

and depending on the workload and speed of the rest of the computer, it might be a small speed difference in reality

[vital raptor]

i use nvme, but i also have a 13900K cpu, fast ddr5 ram and a 4090, everything is fast

in my servers and other computers I just use normal ssd because it doesn't really matter that much for normal use

[desert island]

too much bread?

[lyric zealot]

yup, her tummy hurts

[supple raptor]

just dropping this neat little free tool here for anyone who might not be aware of it https://github.com/Sanster/IOPaint
before, after

[nova remnant]

Why use this instead of inpaint

[vital raptor]

I imagine if you use comfy it would be more useful, comfy doesn't have built-in inpainting as far as i know

swarm and a1111/forge have inpainting though

[supple raptor]

because inpaint is not very good at removing a thing that is already there since it uses the existing image as the starting noise

[vital raptor]

i remove stuff all the time though, that's usually what i use it for

[supple raptor]

also this doesn't need a prompt, just draw the mask over the area you want to clean up and ~1 second later it's done

[vital raptor]

maybe the UI is a little easier

[supple raptor]

here's an example of one that took me about 5 seconds to do

and it's light enough on resources that i can do that while generating another image with sdxl at the same time with 12gb vram.

[lyric zealot]

so nai declutter with slightly extra step at no cost

sounds good to me

[supple raptor]

pretty much. it has other features and other eraser models available too that i haven't tried yet. i'm just using the default Lama model

for anyone interested in training it's especially nice since you can run it in batch mode through CLI. watermarks baked into loras with <100 image datasets is pretty unacceptable with this tool existing

[vital raptor]

i guess if you're mainly editing images you didnt generate, or if you have a use for the cli part of it, it could be pretty good, for someone that uses a UI that already has inpainting in it, I couldn't find any features mentioned that I don't already have

I'm usually just editing or removing stuff from images I just generated, and for that the built in inpaint works really well

the inpainting in swarm is also a bit unintuitive, not as easy to figure out as the a1111 inpainting, so if that has an easier to follow UI then that could be a bonus as well

[supple raptor]

i think you're misunderstanding what it is. it does have inpainting, but that's not the main feature.

[vital raptor]

i mean, i'm just going off their github

if there's more to it, they should maybe say so

[supple raptor]

It's running a very small model that is specifically trained to intelligently remove the thing in the masked area. It's not just just using your SD model to inpaint the area.

i guess technically that is a type of inpainting, but it isn't really the same as just inpainting in webui/swarm

[vital raptor]

it says it supports various AI models, including pretty general inpainting models based on dreamshaper, realistic vision, etc

has segment anything, ersgan, face restoreation, removeBG..

[supple raptor]

yes, it does support those models if you wanted to do prompted normal sd style inpainting. the eraser models are what makes it worth using though

[vital raptor]

all i'm saying, is that for my use case, inpainting in the webuis already does all this, it may not be as user friendly or streamlined or whatever, but i don't see anything listed on their github that i can't already do with the existing built-in tools

if there's more, then ok, but i'm not going to install it to find out 😛

[supple raptor]

here, maybe trying it without installing will explain it better than I am. keep in mind the HF space is very slow. on a 3080 12gb they complete in 1.1-1.3s per mask that you draw. https://huggingface.co/spaces/Sanster/iopaint-lama

and it can just be run alongside your webui of choice and for most people can even be inferenced on both simultaneously without slowdowns in either.

[vital raptor]

it's definitely easy to use

[supple raptor]

and a hell of a lot faster, no need to bother with a prompt, no need to switch models/loras to keep the style consistent, and it is far better at actually removing things than inpainting with an image gen model is.

[vital raptor]

removed all the little white spots and her tail, it smuged her hair and sleeve and left a bit of a 'blob' where the tail was

obviously can't tell if it's faster since the huggingface space is not fast, but i dunno

[supple raptor]

yeah, hf space is slow af

[vital raptor]

i don't need a prompt or same model in a1111/swarm either

[supple raptor]

here's the last several runs that i did. the couple of long ones near the top were when i was messing around with masking a huge section of the image to see what would happen (don't do this)

[vital raptor]

same model and prompt gives best results if you're trying to fix something, but for erasing, you just leave the prompt empty and put it on random seed

[supple raptor]

and likely let it run several times until you get one that actually worked. i've done a lot of impainting

[vital raptor]

in the case of the images i posted above, the blurry smudge left from the tail, i'd have to try again as well, it's not perfect with 1 try either

it's definitely easier/faster to use, especially if you don't use inpainting enough to be familiar with the settings to quickly remove something

[vital raptor]

the parts it over blurs and smudges badly (even if by accident) can't seem to be repaired, so that's a downside for me, but overall it could be really useful for some people

[rain tree]

this is so good wth

[mint blade]

oh thanks. im still super new to this

im learning from the others in here ^^

[rain tree]

i want to start a youtube so ive been looking for SD artists

yall with artistic talents r wild to me ㅠㅜ im jealous

[mint blade]

you can do it too. i only have been doing it a few days. i really mean it when i say i dont know what im doing xD

[rain tree]

i have boxer hands.. hands are very shakey XD trust me ive tried getting into art

[mint blade]

im just generating it

[rain tree]

wait these are ai generated

[mint blade]

yeah

[rain tree]

i didnt know that haha

thats still so cool tho

[vital raptor]

what did you think 'SD artists' were if not ai generation?

[lyric zealot]

i draw using computer code

[vital raptor]

    redo();
} else {
    echo "Yay!";
}```

😄

[rough oxide]

i can not confirm or deny on wut other people post, but the stuff i post is definitely ai-generated entirely.

.....but yea like 99% of the stuff posted in this channel is ai xD

[lyric zealot]

my code

[rough oxide]

i wish the cardfight vanguard video game had suisei... but they didn't include any of the collab cards.

i wanted to play a suisei&miku deck

new one probly won't have quintessential quintuplet cards either T_T

[lyric zealot]

Imma pretend i know what game that is

https://tenor.com/view/pop-team-epic-i-dont-know-i-dont-understand-i-dont-get-it-nod-gif-25890418

[rough oxide]

one of the most popular tcgs.... in japan.

same creators as weiss schwarz

that's the one that's more well known in the west

so that's why they could do hololive collab for cfv too most likely xD

(but in jp cfv >>> weiss, i think only shadowverse from them is maybe more popular globally?)

[lyric zealot]

this is me retweeting

[rough oxide]

...weiss is cool cuz u can get cards signed by the voice actors

so u could get say, a signed hoshimachi card by hoshimachi

....my friend just sold their signed ame card for $500

;_;

[lyric zealot]

[rough oxide]

value went way up on this thing xD

imo it'd be worth way more when ame actually makes a comeback/returns one day

mayhaps.

but oh well

500 bux is 500 bux

[lyric zealot]

understandable for her, since she's not gonna stream anymore right?

[rough oxide]

not necessarilyyyy?

it isn't a banishment

she can return any time

for now it's a temporary ceasation of all livestream activities is all

[lyric zealot]

i mean, she's kinda like graduating

yup yup

that's what i meant

[rough oxide]

that is tru

but she's gonna be in a cover song this month

don't quote me on that

so like she still appears in projects

[lyric zealot]

she will still appear on concert and such iirc

[rough oxide]

yeh

[vital raptor]

i think the only hololive stuff i actually enjoyed was ame and gura playing minecraft back in the day, those two were hilarious together

[lyric zealot]

i'm gonna be honest, i don't really watch hololive stuff at all. Just love suisei singing, and her that insane stream where she plays tetris while singing live

[vital raptor]

i pretty much only watch clips from some vtubers, none of them are hololive, sometimes I'll watch gura clips, but it seems to be pretty rare lately

[lyric zealot]

same

i only really watch neuro/evil now

[vital raptor]

then again, i really only watch clips from any streamer, i'm not big on sitting through a 4+ hour livestream

[lyric zealot]

[vital raptor]

i've been watching a lot of neuro stuff the last few days

[lyric zealot]

https://media.discordapp.net/attachments/1173360037020827728/1173364285515251752/ezgif.com-gif-maker_3.gif

[vital raptor]

i found out about neuro because I was basically looking to make the same thing, just without the streaming 😄

[lyric zealot]

the only hard part of all of these, would really be to know when to get input/responses

[vital raptor]

after way too much time wasted trying to figure it out, i decided I need to write a bunch of code to make it work the way i want, and I don't know python very well so I put it off

[lyric zealot]

i tried something similar, like a computer assistance. But i only made it so that it always feed the llm what my computer can see, and it's just spamming the llm with unnecessary infomation when it's not needed.

[vital raptor]

considering where a lot of the software available sits at the moment, the hard part is the hardware to run it 😄

[lyric zealot]

you really need a way to get the threshold of that

depends, i mean, if you want to run everything local, then yeah

but i was just testing with free api from gemini

[vital raptor]

yeah, for me it would have to be 100% local since i'd be tying it into my home automation stuff, which is also all local

internet outages are bad enough, but when they affect your lights and door locks it's really dumb

[lyric zealot]

i mean, if you are not afraid of data going to training, you can still make api calls for that

true true

you will wanna do something with function calling then

[vital raptor]

my first smart lights (phillips hue) were really nice and easy to use, but it sends every request through their internet servers before doing the thing, so turning on a light took over 20 seconds sometimes and I have pretty good internet

like, push the button, wait 20 seconds, then light turns on

moved everything to home assistant running on my server and got rid of the hubs, now it's all instant

[lyric zealot]

hmm, it depends on how that smart lights are setup i guess. If you connect it locally, then it should never need to leave your local network

[vital raptor]

yeah, it's just the hubs/apps that most systems use by default, they all use internet traffic and slow down

even alexa/google can be really slow

having it all local just makes everything run smoother, so having my AI assistant be local makes sense

just takes more hardware to run the llm and everything

[lyric zealot]

i guess because it's easier to pre-setup their hardware to auto connect to their server, than to get the user to setup properly using local network

[mint blade]

something is going very very wrong with my hands.... 😦

[lyric zealot]

send full pic with metadata

[supple raptor]

connecting it all through their servers gives them an excuse for why your data needs to be sent to them for harvesting

[lyric zealot]

yup, that's also one other reason why they did that

all the more so for home camera and security system

[mint blade]

im trying to figure out how to img to img to let it fix things up

this is the img i was trying to rerender

[lyric zealot]

im not good enough at reading comfyui metadata

[vital raptor]

currently have a 3080TI running a 13B llm, a 40 core VM on my server running TTS model with a 1050TI helping a little, turns out the 1050TI is mostly useless for AI stuff, even small models because the architecture is so old

[lyric zealot]

if you just wanna fix the hands, then do inpainting instead of full img2img

[mint blade]

ok ill try that

[lyric zealot]

i heard the new whisper model might be good

[supple raptor]

yeah, avoid the 16xx series too

[lyric zealot]

not sure if it's new tbh

[lyric zealot]

what is your noise for inpainting/img2img?

[mint blade]

pffft.... ummm. the denoise? i had 0.3

[vital raptor]

yeah, been looking to find a used lower end 30 or 40 series that can fix in my server, but I'm probably going to have to find a reasonable datacenter card since it's an actual rack server

[mint blade]

bc i actually want it to change, right?

[lyric zealot]

you will want it to be higher, since the base image hands are bad, you want it to be higher to make it noisy, so that it can generate on them again

low noise is for upscaling (most of the time)

[vital raptor]

was thinking something like a used p100 or something, but i dunno what the real life performance is

[lyric zealot]

like hires fix and such

[mint blade]

oh ok ur right smol, i had it backwards. ok

[lyric zealot]

[mint blade]

but ill also try this impaint tool. i just gotta figure out if it goes b4 my prompts or after

[lyric zealot]

since you intend to upgrade to 5000 series, why not just wait, then once upgraded, get your used 4090 for that?

[vital raptor]

the 4090 won't fit in the server

at least, not that one

[lyric zealot]

ahh, the node? That one im not sure, i don't really use comfy so you will have to figure it out, or get @supple raptor's help

[lyric zealot]

that's what she said

[supple raptor]

maybe with a riser so you can lay it flat?

[vital raptor]

https://a.co/d/8KVUBZE something like this fits

[supple raptor]

not sure how deep your case is

[lyric zealot]

i would assume it's after your generation, since you need to generate it first, then you can inpaint it

[vital raptor]

it already has a riser, the problem is the power cables from the top

also my particular 4090 is watercooled, so it also has a radiator and tubes to deal with

[supple raptor]

just go linus's route and install a pool in your back yard with pipes under it that you can use as a radiator then create a loop for the whole server

[lyric zealot]

actually you know what, just to get rid of your problem, i will help you. just ship it to me

[vital raptor]

I have a dell r730xd server, it can fit up to 2 slot wide card, already on a riser for horizontal mount

powersupplies would probably struggle with a normal 4090 as well

[lyric zealot]

too much talking, here have a suisei for now

[vital raptor]

i could put it in my secondary server with the 3080TI as long as there's no real CPU resources needed, since that's just a old 4 core system

but the dell server is 56 cores with 128gb ram (expandable to 3TB)

[supple raptor]

you installed Swarm right?

[mint blade]

i did not. lol. i stuck with comfyui for now to learn xD

but i prob should have like u said

[supple raptor]

ah, inpainting in comfy is well.... i just don't do it.

[mint blade]

haha

[supple raptor]

i load up swarm if i need to inpaint and wasn't already in it

[mint blade]

ok. i was searching for a mask tool i can just, interact with (paint), but i cant find one

[supple raptor]

i know i've seen one somewhere. let me take a look and see if i can find it.

i found the one i was thinking of, but i'm trying to figure out how to actually use it. it stretches my image to a square aspect ratio when i drop it into the node https://github.com/Lerc/canvas_tab

if you search "canvas" in comfy manager it will show up.

[mint blade]

oh ok

[lyric zealot]

i salute you for using comfyui as your first ui/step into ai

[mint blade]

haha thanks. its thanks to nubby really. i basically used his workflow to learn, and he helped me a lot

[lyric zealot]

i do agree you will learn a lot more than just normal users like me

[mint blade]

thats assuming i dont get stuck xD

ngl i was hoping this would let me edit the laoded image

[supple raptor]

ok, i figured it out. so when you click that edit button, or most of the time when you drag an image into it, it will open the actual editing tab. with that open, you want to set the Incoming Images dropdown to Create New Image (top right) and delete any existing images that are on the canvas (red circle top left of each image in the little tab. With the canvas empty, then drag your image into the node and it will load the image so you can edit it. For inpainting, you'll want to have the mask layer selected and draw over the areas that you want to inpaint

you can also edit the loaded image by adding other layers between the base and mask layer, although the editing tools are kind of light, so it might be better to do your editing in something like Krita or PS first if it isn't a very simple edit.

[mint blade]

ok. im comprehending this

[supple raptor]

with the send image to editor tab node, i believe the image should be present in the editor once you've run it and you can ensure that specific image is the one that is selected by clicking the green tab hanging off the top right side of the image in the editor tab.

now that i know how this thing works, i might occasionally inpaint in comfy XD

[mint blade]

wait @supple raptor, u can paint and edit the mask right here in comfyui. is this the same as "inpainting"?

also i could never get it to actually recieve the image i was trying to send that web editor. it would just open blank images.

[green temple]

can i run flux with 6gb vram ? 🥲

on forge atleast

[desert island]

Okay normally when I come back to the channel I like to read back through so I can catch up... but man that's a lot of reading. 😄

[lyric zealot]

if you have that discord summaries stuff, it's great for catching up

[supple raptor]

maybe... it will be painfully slow though. it's like 1.5 minutes on my 3080 12gb

i think that was with q5 gguf models iirc. lower quants would be helpful with less vram, but i'd still expect minutes per gen.

if you're looking for a new model to try have you tried out Illustrious or any of the finetunes/merges of it yet? https://civitai.com/models/795765/illustrious-xl

[green temple]

n4 is good for low vram they say

[green temple]

no

[green temple]

ok i`ll try it 👍

[supple raptor]

i think n4 was a little faster for me, but i had trouble getting loras to work and loras are basically required for flux anime.

[green temple]

oh

[green temple]

np

[supple raptor]

illustrious is really good. it's a little undertrained, but still easily the best open sourced sdxl model for anime. it has even better character knowledge than NAI v3 and broader, but less accurate artist knowledge. styles tend to be messier, but some can look really good. nearly everything i've posted for the last couple weeks has been from illustrious asside from a couple of old NAI gens that i reposted.

[green temple]

i don`t really use artist tags much

[supple raptor]

adding this lora between 1.0-1.5 weight also helps to stabilize it a bit https://civitai.com/models/510261?modelVersionId=567119

[supple raptor]

you'll probably want to with illustrious. the base style doesn't look that good. you can look through the gallery on civitai to find some artist combinations that you like and just stick to those.

[desert island]

this makes more sense now... I couldn't for the life of me understand why people liked it... but this helps. Thanks. 🫂 I was just playing around with the base and was like... ewww gross... now I know. 😄

[green temple]

ok 👍

[lyric zealot]

[green temple]

oh , rimuru

?

[lyric zealot]

yup, doesn't really look like her him, but yeah

all of my stuff should have metadata ||as long as you have the stealth pnginfo extension installed||

[green temple]

i dont have it 🥲

[lyric zealot]

[green temple]

but prompt well be somewhere in the regular pnginfo right ? or not

cause ur using comfy

[lyric zealot]

I'm using forge. It will probably not be detected using normal way of getting pnginfo

but if you are using my bot for that, then you should be able to do it

[supple raptor]

here's a few styles that i like on illustrious. (sorry, comfy metadata)

[lyric zealot]

I'm not sure if you have ever used my bot tho

[lyric zealot]

oh, even the first one is illustrious

[supple raptor]

1girlyesyes

[green temple]

np , thx 😍

[supple raptor]

i have a bunch in my gallery on civitai as well. just not some of my most recent ones.

[lyric zealot]

https://tenor.com/view/steal-this-is-mine-run-gif-15431338

[lyric zealot]

wait, this is for illustrious?

[supple raptor]

not specifically, no. but it works. i use it on basically all of my gens with illustrious now.

[desert island]

this was fun... just seeing if I could do it with my current model/s 😁 .

[lyric zealot]

the slime is

[desert island]

I'm working on trying to get an iron throne image... so far this has been the best but it's still not right.

[green temple]

this is with animagine

[desert island]

/emote throws his hands up!
"I give up!" 😭

[zinc flax]

Rip bot no more super secret server to gen super secret nsfw stuff

Was about to make a comparison with nai and illust lol

[desert island]

lol. Source material mashups are fun... 😁

[lyric zealot]

w-wait, im talking about my another bot, not the ||nai|| one

[zinc flax]

A

Oh yeah the illust bot so true

[lyric zealot]

you can still use the very secret bot on that server or dm

[lyric zealot]

nah, i meant the bot that can view metadata

[mental marten]

#🍥｜anime message```yaml
(ningenmame:0.95), (ciloranko:0.95), (shosho_lwlw:0.95), (tianliang_duohe_fangdongye, as109:0.9), (maccha (mochancc):0.9),

best quality, masterpiece, absurdres,

1girl, solo, hatsune miku, | |, blue hair, field, single bare shoulder,
Negative prompt: bad anatomy, bad hands, error, missing fingers, extra digit, fewer digits, (worst quality, low quality:1.1), jpeg artifacts, signature, watermark, username, sketch
Steps: 24, Sampler: Euler a, CFG scale: 7, Seed: 149707138, Size: 768x1024, Model hash: 3e15ba0038, Model: Illustrious-XL-v0.1, VAE hash: 66f78ad136, VAE: sdxl_vae-fp16fix-blessed.safetensors, Denoising strength: 0.4, Clip skip: 2, Hires upscale: 1.5, Hires steps: 10, Hires upscaler: 4x_fatal_Anime_500000_G, Version: f0.0.17v1.8.0rc-latest-269-gef35383b```

[zinc flax]

Oh wait I didn't think of that latter solution

[lyric zealot]

this bot ^

[lyric zealot]

or you can dm me your server id, i can whitelist it next time

[zinc flax]

All good, thankies

[zinc flax]

There we go

Was just trying to make her lol

Wish illust could do clean white bg like nai

[lyric zealot]

start genning using img2img starting with a full white image

might work not sure

[zinc flax]

Life hack

[zinc flax]

Illust version

[lyric zealot]

that's pretty good

[desert island]

finally... ughh

[green temple]

first illust image 🥲

[zinc flax]

Nai at home is real

[green temple]

🤣🤣🤣🤣🤣

[desert island]

the nai flux model?

[zinc flax]

I was talking about illust

[desert island]

hmmmm.... let me go search again...

[near elbow]

flux schnell should be ok - just google for "flux 6gb vram" and come across hits like this one https://stable-diffusion-art.com/flux-forge/

[near elbow]

why? flux schnell unchained does anime nicely, what would you need a lora for?

[green temple]

thx

[near elbow]

flux schnell unchained with "game of thrones, iron throne made of swords, a cute anime princess sitting on the throne holding her blue pet blob slime" - first try, no postprocessing

[desert island]

it's not Rinmuru though does flux know the character? I could test but then I have to load up schnell 😁

[near elbow]

TBH i have no clue about all the details of manga and anime lore and characters - i would simply describe her key attributes and say it's her 🙂

[zinc flax]

Don't think flux know any anime

[near elbow]

"game of thrones, iron throne made of swords, (((Rinmuru))) sitting on the throne holding her blue pet blob slime"
is that the right one? then flux knows her

[zinc flax]

I might be completely wrong

It's just anime style without any character knowledge

[desert island]

No Rinmaru from "That time I got reincarnated at a slime..." It's fine... 😄

[desert island]

I think you're right that's what I remember from my time with flux as well.

don't get me wrong I love me some flux model, but it's not a one size fits all and for now I like the control I have with the SDXL based stuff. I load it up mostly for alot of graphic design work.

[supple raptor]

animagine is bad at touhou. i'm surprised that came out as well as it did. flandre is especially cursed.

[zinc flax]

Sdxl baker try not to bork a model challenge

Quite impossible

[green temple]

yes flandre is very cursed 😂

[supple raptor]

it can only do basically 2 styles. it's either instagram style or that fairly mid anime screencap style.

[supple raptor]

"oh you said flandre so that means you want crystals EVERYWHERE right?"

[green temple]

😂

[supple raptor]

the people training animagine were very against adding any touhou to the dataset in the beginning and only did it after tons of complaints. it makes me kind of wonder if they trained them badly on purpose because the model is quite good at the 3-4 gacha games that the team actually wanted to train.

[near elbow]

well, depends on what you want - if you need character whatevercooldude from the series anotherobscureanime, then you're out of luck with flux (maybe). but if you want anime style images that look quite nice, that's perfect i think 🙂

[zinc flax]

All that hassle just for animagine to be undercooked

[near elbow]

same seed photographic

same seed expressionist

if you just want very specific anime characters, flux is nothing for you
if you want the whole world of AI image generation, welcome 🙂

[zinc flax]

Shame it can't do nsfw

[near elbow]

i beg to differ 😉 - look at the expressionist version above (and i can generate a ton more, but this here is an anime channel)

[supple raptor]

the issue is that there is no flexibility in style on flux. for general concepts, it's outstanding. the styles it can do get boring fast imo though. personally, i would also want more character knowledge in a model that i was going to use full time but that isn't the main reason i'm not using it.

[desert island]

not crapping on flux 😄. I ❤️ flux too.

[near elbow]

guess what the "unchained" part of flux unchained does? 😇

[green temple]

ayaka with your artist tags 👍

[zinc flax]

Sure ig lol

[desert island]

I tried...

[unique mortar]

Some Halloween Wallpaper 1920x1080

[desert island]

I tried some more ... 😡 ... and I'm done...

[supple raptor]

are you going for flan? try starting with this prompt and adjust to get what you want. 1girl, flandre scarlet, solo, crystal wings, blonde hair, one side up, mob cap, red shirt, puffy short sleeves, red skirt, frilled shirt collar, yellow ascot, wrist cuffs, laughing, covering mouth, pointing at viewer, cowboy shot, best quality, amazing quality, very aesthetic

[desert island]

sorry don't know the character just used some reference image I pulled off of google but yeah that's the character I was going for.

[supple raptor]

that's ok. marisa is the best touhou anyways.

[desert island]

not falling for this trap again... I'll go down a rabbit hole trying to make this character... 😄

[supple raptor]

1girl, solo, kirisame marisa, touhou, blonde hair, long hair, braided sidelock, yellow eyes, witch hat, hat bow, white bow, black tabard, white turtleneck sweater, neck ribbon, red ribbon, black skirt, cowboy shot

i have made a possibly unhealthy number of marisa gens

[desert island]

https://tenor.com/view/shame-game-of-thrones-gif-11381550

[supple raptor]

there is no shame in making too many marisa

[desert island]

I have a similar affliction with me making mountains of sprites for silly tavern.

[green temple]

same with ayaka 🌚

[vital raptor]

I don't gen specific characters besides my own OCs, but when talking about unhealthy numbers of gens.. let's just say they pile up pretty quick when it only takes ~3s each 😄

[supple raptor]

i've deleted far more than i have. just the last large purge i did was 500k images. i have no clue how many images i've genned at this point but i know it's in the millions.

i also came here in a time when it was pretty normal to gen 100+ images in a batch in the hopes to just get 1 that was worth posting though.

[vital raptor]

yeah, i'm really picky about what I keep, maybe 1 or 2 for every few hundred at best, and i go through and purge older images every couple months

[near elbow]

just for fun, wide screen

[desert island]

dark times...

[vital raptor]

plus pretty much anything i gen for 'bed time' ends up getting deleted 😏

[supple raptor]

wd1.2 was dark times indeed

[vital raptor]

i estimate i probably gen 3-5K a day though, more if I'm after something specific

[desert island]

/emote: shudders and rocks back and forth

[supple raptor]

once upon a time this was a really good gen

[vital raptor]

i actually had a fun idea the other day to setup my 3080TI to just use a llm to create random prompts within guidelines, like 'cute anime girls' or whatever, and then just have it feed itself its own images through some kind of auto caption system and make it try to improve it's own prompts so it makes images it things are good, and then see what it makes 😄

[desert island]

I wish I still had some of my old stuff

[vital raptor]

which reminds me, i need to figure out how to make sillytavern use the swarm api..

[desert island]

I started something like this in silly tavern but they just got worse overtime... I need to go back to working on my q learning ai project at some point.

[vital raptor]

or maybe see if I can get sillytavern to use the comfyui backend that swarm is running

not sure how that works

[supple raptor]

i think swarm's api is simpler to use than comfy

[vital raptor]

probably, i think i have to write the extension for sillytavern though, or at least modify an existing one

[desert island]

getting it to use comfy on it's own? You can edit the code and have it so that it can trigger it's own actions kind of like a shitty agent. That's how I set mine up anyways.

[supple raptor]

swarm api docs are here if you haven't found them https://github.com/mcmonkeyprojects/SwarmUI/tree/master/docs/APIRoutes

you can always also bug mcmonkey in the SwarmUI server there's something undocumented.

[vital raptor]

i also need to stop being lazy and make an extension (or PR) for swarm

the 'upscale' button/menu item annoys me, i want one that uses the 'refiner'

[supple raptor]

I really wish someone with UX experience would step up and make a proper mobile layout for Swarm. I'd probably use it full time if it had that. As it stands, I just use Comfy when I'm at my PC and load Forge when I want to gen from mobile.

[vital raptor]

like, phone mobile, or tablet?

[supple raptor]

phone mobile

it's completely unusable currently

[vital raptor]

yeah, ive noticed

if i ever get around to digging through the code and figuring things out, i'll put that on my list of things to fix 😄

[supple raptor]

over the last few months swarm has also become considerably slower than using straight comfy for some reason. takes about 30% longer per gen despite using the exact same comfy install with the same venv for both.

speed difference is the same whether using generate tab with swarm's logic to build the workflow or just loading my normal comfy workflow in the comfy tab.

[vital raptor]

i havent noticed any slow down, that's weird

[supple raptor]

i did a fresh install of both comfy and swarm thinking maybe something just got screwed in my venv or something, but no change.

[vital raptor]

to be fair, i don't use comfy by itself, so I can't say if swarm is slower or not, but swarm hasn't changed speed for me since i started using it

[supple raptor]

only logical thing i can think is that swarm is using a different memory preset when it launches the comfy backend that's slower with my configuration.

[desert island]

is there a front end for comfy?
example: If I wanted to make my comfy more modular and setup triggers to execute workflows kind of like function calling for workflows. My workflows can becoming pretty sprawling sometimes... but that's because I want it to do different things at different times and I don't like to have to jump between workflows for different jobs.

Further explaination: like when I create a still image and then I want to animate it but I don't always want to one-shot it so I have to disable certain sections and then enable other sections at different times depending on what I want to do.

[supple raptor]

SwarmUI is the best frontend available for Comfy. https://github.com/mcmonkeyprojects/SwarmUI

there's another one out there somewhere, but it was always super janky and i don't think it's gotten any better.

[desert island]

I'll take a look to see if it has the functionality I want.

[steady grail]

without all the jank, krita is pretty good

[supple raptor]

tbh, there's a pretty high likelihood that you can just do what you want in the generate tab and never have to touch the comfy noodles.

[vital raptor]

imo, at the moment, swarm is the best frontend period, the comfy workflow being available is a bonus

but, i don't use a lot of extensions or anything either, so i'm biased for the simplicity side

[desert island]

not unless it has florence and other nodes in there standard. Plus I use alot of custom files to setup loops and iterations based off of specific project needs.

[supple raptor]

i see...

comfy

good news is that you can still make your schizo comfy workflow and load that into the generate tab, you just lose the extra <> functions that swarm has by doing that

[desert island]

wow okay... did I say something to piss you off?

[supple raptor]

no, sorry. that wasn't intended to be hostile, tone was lost through text. "schizo" is just a common way to describe the massive comfy workflows in the touhou server where i spend most of my time.

[vital raptor]

also a term used in general AI servers/channels for massive prompts.. probably also places with people from the touhou server 😄

[supple raptor]

it's not a bad thing unless you have to try to untangle the noodles in someone else's big workflow

[desert island]

ahh okay, sorry was unfamiliar with that way of using that abbrv. didn't even know about touhou until today :smile.

[supple raptor]

all good. i realized how hostile it could have sounded to someone unfamiliar as soon as you said something

[desert island]

how is virtual cable management a thing! 💢

[desert island]

😄 thanks for clarifying and helping me understand. My social circle consists of basically this discord channel, my wife, and work emails.

[supple raptor]

i know that feels

[vital raptor]

thats like me, without a wife or work emails 😄

although, i drive uber full time, so i actually talk to people all day, probably more social interaction with random strangers than most

[steady grail]

big works flows, reminds me something 🙂

[supple raptor]

@desert island and @vital raptor if you like ai assisted prompting this is something you might want to take a look at. idk if it's been shared in this server yet. basically you feed it a basic prompt that can be danbooru tags, natural language or both and it gives you a prompt that can be danbooru tags, natural language or both. https://github.com/KohakuBlueleaf/z-tipo-extension

[steady grail]

cute

how big is the llm model?

[desert island]

I don't use ai assisted prompting persay but I'm always down to experiment with and try new ai related stuff. I'll take a look at it tomorrow, thanks 👍

[vital raptor]

might be fun to play with, haven't tried using custom nodes in swarm

[supple raptor]

only 400m param, but trained specifically for this purpose so it does a shockingly good job at the task. it's intentionally smol to keep the requirements low since you're intended to be using it alongside a much larger image gen model.

[steady grail]

wow

that's very good

[vital raptor]

restarts the comfy backend

[supple raptor]

for especially low resource systems he even provides quantized 8bit/6bit gguf versions

[vital raptor]

ERROR: Failed building wheel for llama-cpp-python
that's probably not a good sign

[supple raptor]

iirc it will still run, but will run on cpu which isn't painfully slow typically, but might be for you since you're used to 3 second gens

[steady grail]

this is a sign you are not chosen by the llama model

this model is only for the worthy

like the thor's hammer

[vital raptor]

apparently it's missing the compiler for cmake

[steady grail]

little girl, that's a flute, not an axe

[brittle rampart]

I'm sure she still doesn't know how to play the flute “correctly”... 😁

[steady grail]

i can imagine her try to cut a tree with the flute....

[brittle rampart]

@supple raptor I wanted to ask - is this the result of a pure prompt or with modifications?

I mean improvements in i2i

[supple raptor]

pure prompt, it only passed through upscaling, hires fix, and face detailer.

[steady grail]

Captain Harlock lora?

[supple raptor]

no extra i2i or manual edits

[steady grail]

the resemblance is pretty spitting

[brittle rampart]

cool result!

[supple raptor]

it's using this lora on illustrious https://civitai.com/models/822028/league-of-legends-lol-splash-art-or-illustrious-xl-01-or-style-lora

[steady grail]

tonight i had another idea for the workflow .

[brittle rampart]

thanks for the hint. quite interesting

[steady grail]

cute lora

[brittle rampart]

yep

[steady grail]

now it's mine lol

wow, cute

i can't wait to test the new test on the workflow

i hope it works as intended

if it does, it should fix the mutations problem

damn, the difference from my old hard drive and the ssd in something like 10 20 times slower

i didn't think the difference was so much

[nova remnant]

[subtle kindle]

gremlin

[brittle rampart]

not true.

[nova remnant]

reddit

[nova remnant]

lmao

I think he was just trying to generate something

[brittle rampart]

Ha! I know.

[nova remnant]

@brittle rampart is this a gremlin ?

[brittle rampart]

https://tenor.com/view/range-gif-22199229

[nova remnant]

@green temple smol gremlin

I don't think nai has much of an idea what a gremlin should be

I think neither do I

[brittle rampart]

Tagger - gave out such tags when describing Gremlen's photo no humans, pokemon (creature), solo, brown eyes, realistic, full body, standing, blue background, smile, animal focus, parted lips

🤣 just Pokemon!

[nova remnant]

I got this

[brittle rampart]

🤣

[brittle rampart]

almost similar to the original

[nova remnant]

[brittle rampart]

You can't drink so much alcohol... 🤣

[nova remnant]

https://tenor.com/view/cat-gif-5215455

[ivory pebble]

stare

[brittle rampart]

https://tenor.com/view/kanna-kannapog-gif-20442777

[lyric zealot]

read

[brittle rampart]

translate 😭

[nova remnant]

I fixed it with declutter

👍

[vital raptor]

😄

[odd fern]

https://picsart.teamhydra.dev/picsart-536e29a2-d8fa-4b4a-bc22-eee0f57ed39e.png

[lyric zealot]

if only i understand

[lyric zealot]

stoopid

[lyric zealot]

declutter it again

[lyric zealot]

a better job

[rough oxide]

still nai? these are good.

[nova remnant]

[nova remnant]

alex only uses nai

I think

probably

don't remember using anything else

[nova remnant]

lick

[nova remnant]

grabby

[lyric zealot]

https://cdn.discordapp.com/attachments/1067638175478071307/1288208397849722880/mods_get_him.gif

[green temple]

+!

+1

[lyric zealot]

[rough oxide]

hard to tell w/o metadata ;_;

[nova remnant]

you're against me too

im innocent

[green temple]

no

[nova remnant]

well

if you want something like nai

there is illustrious

its basically like NAI

it was trained mainly on anime stuff

you're against me too @brittle rampart

[brittle rampart]

we love you. each in his own way😁

[lyric zealot]

https://tenor.com/view/cat-no-nonono-noooo-cat-no-gif-13597132752790194338

[green temple]

no

[ivory pebble]

[nova remnant]

[ivory pebble]

Yup

[ivory pebble]

Tried it for a bit but all the images are.. very low quality for some reason or blurry. The linework on the images is kinda messed up and the backgrounds are odd

Could just be a skill issue on my end tho 😭

Only messed with it for a couple hours so I'm probably not using it correctly

[nova remnant]

nubby has been using it a lot

I think

just wait for his messages to finish cooking

like 10 minutes

[supple raptor]

you definitely have to find the right artist combinations to use for illustrious and it doesn't really like natural language.

[ivory pebble]

Still a fun model tho, it recognizes a lot of characters and we FINALLY have a model that has artist tags

[rough oxide]

slight skill issue. but also v0.1 is understandably underbaked as well.

better base than pony though for loras

[ivory pebble]

Yeah I was basically using the same prompts from NAI but I was tweaking the weights a bit

[ivory pebble]

Yup, that's why it's exciting

[supple raptor]

it's also quite picky with samplers in my experience. dpmpp_2s_ancestral with ddim_uniform scheduler has come out the best for me usually

[ivory pebble]

I see, it's honestly very impressive considering it's underbaked

[ivory pebble]

Oh wow, I only use euler a lmao 😭

[supple raptor]

but yeah, it's got even better character knowledge than NAI, and the range of artists it knows seems to be very similar to NAI, though often a bit less accurate, especially on styles that have very clean lines. some of the sketchier artists look even better on illustrious than NAI imo.

i've successfully gotten tags with less than 100 images on danbooru to work with illustrious. nai's cutoff typically seems to be closer to 500+ before a tag works at all, with some exceptions.

[ivory pebble]

Yeah it does know a lot more characters, and it can actually generate multiple subjects even better than NAI at times so it's quite impressive

[nova remnant]

will they update it

[supple raptor]

that's a little unclear at the moment. it was the plan before someone leaked it while in early access and emailed the ceo directly.

[ivory pebble]

The image composition is also a big factor with models like illustrious and NAI compared to Pony. Pony might be more consistent with poses but it just looks very bland/boring

[supple raptor]

the email was not about the leak, it was some sort of business proposition or something like that.

[supple raptor]

yep, and i expect pony v7 to be even worse to use with the grouped art style thing that astralite wants to do.

[rough oxide]

whatever the email was about it threw a wrench into plans. dunno what's happening with the future of the models ;_; but worse case scenario nai is planning to update to a v4 at least

but no idea when that's out

everything is soon™️

[nova remnant]

2 weeks

[rough oxide]

kwsk

did kurumuz say anything about it?

[supple raptor]

i'd like to be proven wrong about pony v7 because more good models is a good thing for everyone though

[rough oxide]

i don't even like pony v6 (except for loras)

i HATE how good pony is for loras

it's an abyssmal base model though

u can't even get a leek in pony

[rough oxide]

cuz they nuked old training & such xD

[nova remnant]

someone just said its bad above

[rough oxide]

nai you can at least get some real world concepts still

[nova remnant]

I can't trust anything

[ivory pebble]

To be honest I'm hoping it doesn't fail cause the more models we have the better but at the same time like... when you have such a big dataset mixed with anime/cartoons/3d stuff I don't know if it's gonna be good

[nova remnant]

I can't trust her either

[rough oxide]

i recommend konoko's pony merge

for loras

[nova remnant]

im 4gb, what do you think you're recommending it to me for

[rough oxide]

4gb's more than enuff for sdxl isn't it

¯_(ツ)_/¯

[nova remnant]

no

also on amd its a no no

and a 6 years old amd card, no no no

[rough oxide]

o yeh

trying to find one of my old posts for knk's ponymerge

[ivory pebble]

That's mainly cause v6 had only 500k anime images in the dataset, it will probably recognize what a leek is in v7 with a bigger dataset but who knows honestly 😭

[rough oxide]

https://civitai.green/posts/4878994

a lot more on main civit site for that post

xD

[supple raptor]

nah, i'm sure we'll start seeing preview images from Tabloid before launch like we got before NAID3.

[rough oxide]

but i mean if nai works for the char it can do it just as good as pony w/ lora most of the time

[rough oxide]

yeh, i'm guessing it's still a couple months down the line maybe.

[lyric zealot]

isn't pony v7 also on a totally different base model?

[rough oxide]

well

it takes them like half a month with their hardware once they have the dataset all curated & all that right?

[ivory pebble]

Ye the base model is called auraflow or something like that

[rough oxide]

maybe november

....but maybe my math sux

[nova remnant]

wait v7 is going to be on aura

wait

is v6.9 a thing ?

[rough oxide]

i'm assuming like 75000 hours on 256 h100s still for what was naiv3

[supple raptor]

yeah, i think he's planning to use AuraFlow

[ivory pebble]

I don't think so, I think astralite is going straight for v7 on auraflow, but either way the next Pony model is gonna be on auraflow

[nova remnant]

astra was gonna do a v6.9 as an update

I kinda wanted to see that

hopefully is gud

[lyric zealot]

from what i heard, auraflow is better at following prompts, and the licensing is better than sd3

[rough oxide]

i thought they fixed sd3 licensing

after outcry

[nova remnant]

yes but

[rough oxide]

but as a consoomer i dun pay attention to the licenses.

[lyric zealot]

i think they did a community poll about whether to just move forward with pony v7 or spend more time with pony v6.9

[nova remnant]

theres still the newer sd3 stuff

[ivory pebble]

Apparently they're training their own base model so that's gonna take some time as well, they won't be using SDXL or Flux or anything like that

[rough oxide]

i'm just kinda sadge everything is soon™️ for the past like 3 months xD

flux was a nice shadow drop tho

[lyric zealot]

they fixed it, but it's still not as good as auraflow is what i heard

[rough oxide]

ohhh. ok then i'd probly move it back to december for expectations. no doubt it'll probly be the best thing available again lol

[supple raptor]

isn't auraflow also the one that's trained on a shit load of ai outputs?

[lyric zealot]

is it? No idea about that

[nova remnant]

https://tenor.com/view/objection-court-judge-object-lol-gif-4833328

lmao

[ivory pebble]

Yeah I think it's trained on Ideogram images

[rough oxide]

i really dislike when ai is trained on ai images

[nova remnant]

nothing wrong with that

literally

they are all images

[rough oxide]

if it's like entirely ai images

[nova remnant]

theres no "real" stuff coming out of AI anyway

[rough oxide]

this is why i have no fayth in indie projects xD

big data bettr.

[ivory pebble]

To be fair I'm not sure how well they're gonna do this time cause making a base model is a lot harder than just finetuning an already existing one but we shall see

[nova remnant]

I just want some godamn AIs that help people make games

so I can play more games

[lyric zealot]

and run it on your 4gb vram?

[rough oxide]

and in just like 5 months illustrious seems more SOTA than NAI (on the current version we dun have access to ;_;) but again they have corpo backing

[nova remnant]

imagine AIs optimizing games

kinda of how nvidia optimized factories

lmao

then I can run them

[nova remnant]

pony > nai

[rough oxide]

only when pony is using a lora for a char u can't make in nai

otherwise nai>>>all currently available stuff to us

making use of vibe transfer is a huge part of it too

[ivory pebble]

I feel like it would be a bad idea to train a model using AI images cause don't they have a bunch of artifacts? Not to mention all the anatomy errors for characters

[rough oxide]

yessss.

[supple raptor]

we already have that to some extent. using AI assisted coding is a lot more practical than using AI for art assets currently

[rough oxide]

glad you can explain it better than i can

[nova remnant]

just select better images

[ivory pebble]

I'll be honest the only reason why I have NAI above Pony is simply because of artist tags, a pure Anime dataset, better innate character knowledge and obviously better colors/vibe transfer (that's a lot of reasons ngl)

[supple raptor]

there's probably no problem with using well curated AI data as a relatively small portion of the full dataset, but actually doing that curation at a large scale is very difficult with current methods.

[nova remnant]

vibe is fun

I have some chinese manga art on it

and it makes angy girls

:p

[rough oxide]

being able to get this style of drawing just by a small strength vibe transfer is gr8

i suddenly realized my "random prompt" had "restrained" in it

xD

[ivory pebble]

Ye most of my images are made with vibe transfer, that Deku image I posted earlier was made with like 7 vibe transfer images lmao

[supple raptor]

yeah, vibe transfer is pretty easily the biggest advantage that nai has over opensource especially now that we have illustrious

[rough oxide]

i cap out at like 4

cuz then it starts making artifacts

with 7 do u have to drop info extracted?

[ivory pebble]

The Inpainting model as well

[supple raptor]

yep, i was just typing that

[lyric zealot]

[rough oxide]

do u has the metadata for it? or do you wanna keep it secret? xD i kinda wanna see how u tuned the vibe transfer to use that many

[lyric zealot]

i think even with metadata, you don't get the vibe transfer images, right? Unless you meant how the settings work for vibe transfer on such amount of images

[ivory pebble]

I'll be honest I have no idea, I just move the slider left or right and pray that it works lmao 😭

[rough oxide]

not the image but it includes slider info

[supple raptor]

i've tried to use some ipadapters to get something that works even remotely similarly to vibe transfer and the only conclusion i could come to is that the current ipadapters we have for sdxl are very bad.

[rough oxide]

like info extracted/ref strength

that's what i'm interested in

[lyric zealot]

ahhh, okay

[supple raptor]

i've done a few images with a lot of vibe transfer images and generally, the more of them you add, the lower you want to set the sliders. i don't have specific examples though.

[ivory pebble]

I deleted the vibe transfer images and the original image sadly, but I basically just go to safebooru and find a bunch of images that have good backgrounds/interesting colors. Sometimes I even add a couple of images that have the manga/comic style to it cause when I add "motion lines" to the prompt it can Impact the image a lot

[rough oxide]

nuuu it's not the vibe transfer images themselves

i'm interested in the metadata that shows how u used the sliders

[ivory pebble]

Ohh I see

[rough oxide]

cuz i have never gotten one to work well over 4 vt images

[ivory pebble]

That's strange, are you using NAI on your phone or pc?

[rough oxide]

pc

[ivory pebble]

I don't think it matters tho

[rough oxide]

at 4 images it starts getting noise/artifacts

[ivory pebble]

Yeah I did it on pc as well, you can use more than 4

[rough oxide]

no i mean... the actual outputs lol

that's why i wanna see metadata T_T

[ivory pebble]

Let me see if I have another one hold on

[nova remnant]

vibe transfer

[ivory pebble]

Not sure if this is what you're looking for but this is all I got

[rough oxide]

ye that works

this was the bit i was interested in

[nova remnant]

[rough oxide]

so you tune information extracted down a lot then?

[ivory pebble]

I'd let her slice me into a million pieces

Yeah pretty much

[nova remnant]

I do 0.2/0.5 everything

[rough oxide]

i need to play around more with information extracted cuz i usually leave that full

[nova remnant]

it works its fine whatever no thinking brain empty

[rough oxide]

i do like 0.2-0.4 for strengths

[ivory pebble]

If it's too high up then it will end up looking like the vibe transfer image and I don't really want that unless I'm looking for a specific pose

[rough oxide]

that's usually the strength, but yea if ur trying to stray too far from src image

then the info extracted might do that too

[ivory pebble]

Not sure, but you can mess around with it and see what works best, that's what I do

[rough oxide]

'tis good insight, thanks.

[ivory pebble]

At the end of the day all of it comes down to RNG

[rough oxide]

true

[ivory pebble]

And a lot of Inpainting 😭

[rough oxide]

lol yea

i find myself inpainting less n less these days tho

the new karras sampler is rly good with bodies

[ivory pebble]

It depends on what you're looking for honestly, I mainly just inpaint the hands or the feet

[rough oxide]

btw do u ever get rate limited?

i think one day i must've just genned too many in a short period cuz it rate limited me til the next day xD

[ivory pebble]

Yeah, I just refresh the page and it works again

[rough oxide]

tried refresh too, i must've just done 2 many cuz i used the api in a game

[ivory pebble]

THE NEXT DAY!?

That's insane 😭

[rough oxide]

not sure if you can get rate limited normally thru normal use

[nova remnant]

date with a cute girl

[rough oxide]

the game basically genned like 360 pics in 15 mins tho xD

but sometimes i can gen a lot in a small period of time, idk if there's some kinda limit i should be looking out for just using normally

or if i just triggered it cuz of the game

[lyric zealot]

Let me update the vibe transfer images allowed to 100 images, then you do one gen on that

[ivory pebble]

I think you can but it usually fixes itself once you refresh the page

[lyric zealot]

[rough oxide]

u gen a lot like all day too? xD

[nova remnant]

put 100 images at 1/1 strengths and see what happens

[ivory pebble]

Hate to break it to you but SHE'S NOT REAL!!

[nova remnant]

what a relief, imagine dating someone

[lyric zealot]

[lyric zealot]

SHE’S REAL TO HIM

[ivory pebble]

Depends, sometimes I just keep rerolling cause I'm never satisfied with the images I get. They look but I always want something more fancy 😭

[rough oxide]

if it's truly unlimited i shuld stop being so stingy with friends and gen some of their requests too LOL but whenever i hit rate limits i start wanting to just wanna min-max my gens

ye i hit the rate limit while manually rerolling an alter saber thing i was trying

[nova remnant]

I have spammed so many images

[rough oxide]

probly rerolled it 40+ times but

[nova remnant]

[rough oxide]

earlier in that day i used the api game

[ivory pebble]

Based

[rough oxide]

so dunno if that was part of it

[lyric zealot]

How do you even get rate limited, are you on opus?

[ivory pebble]

I've genned so many Sabers I forgot which one's which 😭

[lyric zealot]

W-what’s the api game?

[rough oxide]

these new tricks turned my rin tohsaka into a mermaid and idk how

xD

[rough oxide]

AI Roguelite

u can attach NAI to it but it seems like a bad idea cuz it gens so many

triggers the rate limit i think

but only after an extended-ish period

everything in the game is generated from the text scenes (it also uses NAI text api)

to the items, npcs, enemies, etc

[ivory pebble]

I'm not exactly sure if we're the same page but isn't the rate limit when you get that message when it says that this session is blocked or something like that? I usually just refresh and it fixes it

[rough oxide]

i used to even just use it just to get random gens based on the situations going on cuz it saved em all in a folder

sometimes there'd be a surprisingly good gen

from all the prose

[ivory pebble]

This is the first time I hear about that, pretty neat