🏙│ai-images

[dapper linden]

huh, wth is that quality tags block
(ultra realistic,32k,masterpiece:1.4),(high detailed skin:1.1),( high quality:1.1), (ultra high res:1.4), (masterpiece), (beautiful lighting:1.4), (high quality), (high resolution:1.3), (incredibly detailed:1.28),

[tawny peak]

must be running the model at a low CFG

[dapper linden]

I use my merges
now I'm with 50-50 of J aka's model and Rets' aperion

[dapper linden]

3.5 cfg

[tawny peak]

it's a little low, but not low enough to need that heavy weighting

[dapper linden]

awww I saw Beez's pic
cute one. I won't steal that XD

it's so painful

that I have manually choose denoise based on what composition/style picture is

it can be 0.25-0.12, and wrong or just 'medium' one just doesn't work

[tawny peak]

huh, you shouldn't ever need to touch denoise

[dapper linden]

aghhhhghhghgg

[dapper linden]

well, try to make a anime picture, let's say cowboy shot. 0.25-0.30 is fine.
and after that realistic portrait

and swear to me that you won't get artifacts

[tawny peak]

on your ksampler or upscale nodes?

[dapper linden]

upscale

[tawny peak]

ah my denoise is 0.3 for upscale

[dapper linden]

(which is ksampler as well (nerdyface))

[tawny peak]

0.3 on ksampler??

[dapper linden]

I have 0.21. while I'm on anime it's fine.
with semirealism... it's a roulette

[tawny peak]

I'm running 1.00 on my ksampler

[dapper linden]

...

here

both are ksamplers

one is for gen
second is for upscale

do we really have to go through terminology lol?

[tawny peak]

i don't understand what the clownsharksampler improves over the normal ksampler

[dapper linden]

that wasn't even a question

ok, nevermind

[tawny peak]

no your gen sampler looks fine

same with upscale

[dapper linden]

left is 0.21 denoise upscale
right is 0.15

and left is obviously burnt

(neck)

yet this one gone almost well even on 0.21

but I'm so pissed off with manual correcting every time
and moreover - you correct when you already have one burnt thing, so basically do same thing second or even third time

[tawny peak]

that might be because the model is overtrained

[dapper linden]

that pisses me off

[dapper linden]

just photorealism is burning at higher denoise

lower*

[tawny peak]

wait are you doing photorealism and anime in the same model?

[dapper linden]

I mean it's threshhold is lower than anime one

[tawny peak]

or seperate

[dapper linden]

mmhmm

it's not like photorealism
but pretty much semi

[tawny peak]

might want to consider seperate merges

[dapper linden]

but I like result I get

[tawny peak]

merges take a while to fully get down, it just may need more cooking

[dapper linden]

but can't properly setup wildcards with styles
cause if it pulls photorealistic on the one gen, and then cell_shading, anime_coloring on the very next one - it's pain in ass

and composition is important too. if 0.2 can be okay for cowboy - portrait should be lower to avoid funny stuff

but yeah, if just decide NOT do semirealism - all good, 0.21 works just fine

weird rose lol

[left kelp]

it's messed up, is what it is 😆

[clear pivot]

Sweet Victory ft. Matsuri Hibino

[shell ivy]

!imagine

[dapper linden]

<@&1159293140440723499>

[zinc grove]

.

ن

[dapper linden]

<@&1159293140440723499>

[balmy cedar]

aight, time to act like i didn't show up for a bare minimum of 60 seconds

[balmy cedar]

I have been purposively generating them in green (and sometimes blue) background

[grizzled jolt]

Choose a cat

[molten oasis]

4

I'd fuck him

[dapper linden]

omg dude

[supple turtle]

bruh

[inland mountain]

that's the guy who said he was high

[inland mountain]

[small edge]

I am the shadows

[steel haven]

i am the night

[dapper linden]

you are the boob

left

[steel haven]

I am Bob

[dapper linden]

nah you left boob

I'm right

[steel haven]

I batman

[dapper linden]

you

don't you dare make sheepy against me

I will cry

fr

[steel haven]

I sleepy sheepy but idk it's need big contextual upgrade

[dapper linden]

do you think so? Looks good for me

and isn't large context pretty expensive?

[steel haven]

She already had a 1m "context window" but I mean other things regarding context

[dapper linden]

ffs, 1m is quite big

[steel haven]

The rest of her updates have been code patrol

[dapper linden]

if only humanity learnt to use long term memory...

or even better - to solve the problem of counting weights of actually important tokens nad useless ones...

[inland spindle]

LUCY?!

[dapper linden]

LUCY!

[inland spindle]

LUCY!

[dapper linden]

really wanna stay at her house

[inland spindle]

say that again..?

[dapper linden]

I mean...
the name of the song.
"I really want to stay at your house"

from anime

(and game)

[inland spindle]

Yeah

I know

Peak

but also sad song

[dapper linden]

yeah, just like anime

[inland spindle]

The moon

[dapper linden]

<@&1159293140440723499>

[dapper linden]

aawwwww ❤️

[tawny peak]

[supple turtle]

random pirate kitty

[dapper linden]

(prompt stolen from RAM's genius)

[manic pond]

State of the Undoing photos

[iron orchid]

Sky blue

#🏙│ai-images sky blue

[clear pivot]

Momoka Kaguya reference sheet

[dapper linden]

is it pro gemini or free one?

[clear pivot]

Is Nano Banana Pro on gemini

[graceful lark]

https://civitai.com/models/2421024?modelVersionId=2721999
-# https://civitai.com/models/2421081?modelVersionId=2722045

[steel haven]

hi @dapper linden

hi

I'm back from my banishment

[dapper linden]

hiiiiii~

❤️

[dapper linden]

hi

hiiiii~

[steel haven]

hi

[dapper linden]

hihi

[steel haven]

hey

[dapper linden]

yo

[steel haven]

howdy

[dapper linden]

gut

[steel haven]

hola

[dapper linden]

привет

yamete kudasai

oops
wrong one

[steel haven]

domo arigato

[dapper linden]

idk where this came from I swear

hehe

[steel haven]

GAMES

SLOP

[dapper linden]

games... I know one game

called 'need for titties'

[dapper linden]

or something like that

[steel haven]

I think I'm ready for scii lmao

[dapper linden]

no

fcin

way

I've deleted it like 2 month ago or so

[steel haven]

SCV rush all the way

[dapper linden]

are you really rdy or it's a fake one again? XD

don't want to download it for nothing

[steel haven]

I can probably do it tonight

and stream lol

Stream a big fat L

[dapper linden]

what is night let's consider

in GMT0 if possible kek

[steel haven]

plz use my microsoft booking link

kek

1am?

lol

[dapper linden]

da fq
maybe I shall use bing for that as well?

[dapper linden]

damn, it's 4am at my place

[steel haven]

oh fuck lmao

[dapper linden]

kinda late... or early lol

[steel haven]

literally the worst time

[dapper linden]

I can try to wake up earlier
since I'm getting up at 6-7

[steel haven]

we'll have to do it some weekend when I have time in the early PM or AM I think

[dapper linden]

my weekends are meh XD

[steel haven]

and I have a ton to do on saturday

like a ton ton

[dapper linden]

28th is my werk night 😭

[dapper linden]

🐏

I'll try to wake up early today

tomorrow

ffs

I'll just smash 2 bottles of soju and magic will happen

I have to take few beers for scii I think

doping lmao

[steel haven]

damn right lol

[lavish magnet]

truly fire

I like how dense the image is

and the atmosphere too

[steel haven]

Thanks dude!

[vapid flicker]

yo

[balmy cedar]

I "made" this as a joke

[warm pollen]

sup, guys !
I wanna create images like comics
any suggestions except sora and grok image ?

[dapper linden]

not enough data
we could try to help if you clarified a bit more

like comics means just framed pages or style as well? or just style?

which comics/manga style?

(and even after clarifying I still won't help cause I'm noob lol. mb other guys will suggest something)

and more importantly do you want to do it locally or by big LLMs

if second you may try gemini pro. gpt isn't bad for images too I guess

[warm pollen]

@dapper linden similar to the marvel comics with 4 pics in 1 page
my laptop will explode if I try to do this locally lol

[dapper linden]

yeah, not quite laptop thing XD

[inland mountain]

Laptops with a 5090 in them be like:

[dapper linden]

unfortunately can't help here. but really, try gpt and gemini, maybe it's the one you seek

[dapper linden]

with their heating issues and other thingies?

laptops like that is a joke fr

[inland mountain]

The big companies dont care about cooling

[dapper linden]

the ones with ARM architecture might be cool

[inland mountain]

They might

[warm pollen]

@dapper linden gotcha anyway thx for suggestions
I will do some research

@inland mountain cant afford it bruh

[inland mountain]

lmao i figured

5090's are just

overpriced anyway

only like a 25% preformance improvment

just get a 4090 instead

[inland mountain]

huh

[dapper linden]

well, depends on how you judge

if you need a single vcard that can held more than 24gb you don't even have a choice
and in this case 5090 is not only not overpriced - it's priceless

(if we are speaking on customers cards obviously)

[inland mountain]

True

[dapper linden]

but in every other metrics yeah, it's kinda bad

[inland mountain]

It is, really only a psecial use case

*special

[green hamlet]

Shitting on NVIDIA GeForce GPUs won't make me to switch to AMD either.

[dapper linden]

well, if only it was actually selling by actual price, not cooked one - it would be fine

[inland mountain]

TRUE

[dapper linden]

once again - not like we had much choice

[inland mountain]

lmao. Fair

[dapper linden]

only 50% of image for you guys.
other half is... spicy

[inland mountain]

Oh!

lmao

[dapper linden]

🐀

sniff sniff

[warm pollen]

btw I found an interesting GPT on ChatGPT called Master ACG with solid 4.0 rating
quite interesting

[dapper linden]

maybe
I don't know anything about gpt cause it's bad lol

other LLMs are much better. and this dummy ai is just farting itself

[stark dome]

Anyone know if there good image generators that can do custom (or very low) resolutions? Looking for low res stuff

How did you do these?

@inland mountain you too! Pixel art tips ?

[dapper linden]

JVGA is doing it by lowering picture x4 times in reso, and after that restoring it back to normal

[dapper linden]

here I had few tries. But because I'm doing higher resolution pics and I'm missing something - it's worse

But I suppose my biggest miss - using wrong model, or more like wrong style

my stuff is more like semirealism most of the time, or detailed cel shading.
for pixel art I have to do pics with harder lines

[dapper linden]

yo guys

who slipped in my DMs

don't be shy to ask here
I'm far from being best in here, I swear

answering to questions: I use my merges. It's illustrious based models. you can find it on my civit page

http s://civitai.com/models/2414240/darkdemon

that's the model I use lately, my best one as I think

don't be shy to ask. just don't sneak into DMs 🙂

for pixel stuff JVGA will answer much better, cause it's his tech

[dapper hollow]

google fucking did it

[small edge]

Yo

Who's with me?

[dapper hollow]

idk

[tawny peak]

[small edge]

That's actually quite good ngl

[balmy cedar]

its got that gacha game screenshot aura

[keen abyss]

good evening any one can tell me how to make mandala using ai ?

[solar knot]

In gemini:
Generate mandala
1:1 2k

[mild torrent]

guys, which tools can I use for outpaint and expand my wallpapers?

[dapper linden]

no idea

[kind idol]

Hello

[agile grotto]

What is this chat for?

[agile grotto]

You can use the paid version of Nano Banana PRO. Just output as a wider (or taller) aspect ratio.

[tawny peak]

or just download and use free local ai

[agile grotto]

I don't think a local AI beats Nano Banana Pro though.

[tawny peak]

you'd be pleasantly surprised

[dapper linden]

mostly, as it is supposed to be at least - this channel is for showing our ai generated pics

sfw

and for little side chatting about ai art I guess

[agile grotto]

I see. But I can't submit images for now though.

[dapper linden]

yeah, you should have 5th rank (aka blue nick)

should chat a bit in general or wherever else

antispam protection, you know

[agile grotto]

Got it.

[mild torrent]

which would you recommend? I have a Nvidia RTX 2060

[tawny peak]

i used to have one too! ComfyUI works great with that. Do you have the 6gb or the 12gb variant?

[mild torrent]

Oh nice, I have the 6gb variant

[tawny peak]

Ok, as long as you have at least 16GB of system RAM you should be perfect. Just make sure when running it, no games or other GPU intensive programs are open

[mild torrent]

yes I have exactly 16 gb of ram

I am running linux right now

[tawny peak]

If you wanna get going fast, look up stability matrix. Download the release from github and the program will manage everything for you

you can also test out other UIs through this application if you want to, ComfyUI just has the best optimizations

[dapper linden]

is 6gb even possible with illu?

I mean, you were fine? like, illu models are ~6.5gb

[tawny peak]

Yea, partial offload to system ram

[dapper linden]

even with my current 16gb I'm going to limit every now and then on my wf

[tawny peak]

Safetensors are multiple models baked into one so a small bit is the VAE, another part is the text encoder, another is the diffusion model

[dapper linden]

when I was on 3050 with 8gb I was using forge. And It was almost critical

[tawny peak]

Well how many other models are you running alongside the main checkpoint?

• what's your VRAM launch argument?

[dapper linden]

other? just one
and few smaller thingies like detailers loras and that crap

[dapper linden]

--disable-xformers

ouch

single

buuuut actually like 3gb is eaten by this damn thing sageattention

wanna turn it off tbh

[tawny peak]

Run with pytorch attention

[dapper linden]

doing literally nothing but eating vram

[tawny peak]

Near identical in my testing

Also im talking about your VRAM launch argument not the attention one

[dapper linden]

that's all arguments

[tawny peak]

Is it low_vram normal_vram high_vram?

[dapper linden]

nothing

[tawny peak]

It'll say in the command output

[dapper linden]

so, I have to restart?

one sec

[tawny peak]

No just open the terminal pop out on the comfyui page

[dapper linden]

ah
normal

[tawny peak]

Check the size of the models

I run 3 loras, 1 checkpoint and 1 upscale model equalling around 8 - 9 gb

ALSO do you have integrated graphics on your CPU?

There's a lot of small variables that can cause out of VRAM errors

[dapper linden]

6.7gb for ckpt, 350mb for detailers, 320mb for vae, ~60mb for upscaler

telling ya, 3gb is going straight to sageshit
that's the main error - to install it XD

[tawny peak]

Huh weird I'm running sageattn fine and I haven't had any issues

It helps with video gen

But for normal diffusion models there's no difference between xformers, sage, or pytorch

[dapper linden]

it may be some 5xxx series joke

anyway, it's not critical
most of time I'm good

[tawny peak]

Yeah I've heard that the 30 series just handled cuda stuff better

[dapper linden]

just sometimes I have a little stutter in gen for a few secs

[supple turtle]

in nvidia control panel, make sure CUDA - Sysmem Fallback Policy is off, then set your system page file to some huge amount, boom. Now you can run practically anything even if it overflows your ram and vram. if you have a fast SSD it won't really be that slow either.

XD

[tawny peak]

What's the normal generation time for an image with those settings then

15 minutes lol?

[supple turtle]

not really

[dapper linden]

fun strat

[supple turtle]

I can generate with Flux 2 (the full sized model) and loras and stuff on my 4090

[dapper linden]

lmao

you could do it without these shinanigans

[tawny peak]

Well that's a big freaking difference, the 4090 is unbeatable

I gotta figure out how to utilize the 300 compute units I have on Google colab considering they just added the Nvidia H100 Gpu for use

80gb of vram

[mild torrent]

Perfect

all of these use stable diffusion?

[tawny peak]

yes

[mild torrent]

nice

I can get a good quality result also for upscaling?

using stable diffusion

[tawny peak]

yes

[mild torrent]

btw is also nano banana pro based on stable diffusion?

because trying it on lmarena for generating some images, I noticed that the results were really realistics

[tawny peak]

the main concept of ai image generators are the same, but no, nano banana is using propriatary google models

[mild torrent]

ah

can I achieve that kind of realism?

[tawny peak]

simply look up civitai, and find a realism model you like

civitai is a huge model repository

[mild torrent]

Oh nice

sd is the only architecture out there?

for image models

[tawny peak]

SD is the most popular, but theres also flux, z-image, qwen, and more

[mild torrent]

Oh

[supple turtle]

z-image is pretty decent for realism and runs on a potato pc

[mild torrent]

so, if they are all based on the same architecture, what makes them more or less realistic and consistent?

[mild torrent]

nice

[supple turtle]

it depends on what images were used to train the model really

[mild torrent]

Oh

[tawny peak]

you can "run" models that need higher vram, but they're going to be slow

[mild torrent]

so proprietary models are that realistic because they are trained on a bunch of different images?

[mild torrent]

Ok

[supple turtle]

proprietary models are generally bigger

more data

[mild torrent]

Oh

what's a good consumer GPU for running bigger models?

Because lately I've been thinking of doing an upgrade

[tawny peak]

i use a 3060 12gb and i can do 2k resolution images in just over a minute

[mild torrent]

Oh nice

[supple turtle]

more vram = better

(as long as you aren't buying an ancient gpu)

[mild torrent]

so it's not the chip itself that matters, right?

but the VRAM?

[tawny peak]

chip + vram

[mild torrent]

Ok

[tawny peak]

so a 2010 gpu with 16gb of vram wont work well

[mild torrent]

Ok

[tawny peak]

some of the old tesla cards are like that

[mild torrent]

what's a really good amount of VRAM for consumers?

[supple turtle]

a 3090 is a decent choice if you can find one, or if you have tons of money get a rtx pro 6000 lol

[mild torrent]

I have around 2300 euros of budget for my new pc

[tawny peak]

or go team red and run ROCm on linux

[mild torrent]

rocm is like cuda?

of amd

[tawny peak]

yeah

[mild torrent]

Oh nice

so amd GPUS are not bad for ai related tasks

I thought that nvidia's were the only good ones

[tawny peak]

6000 series amd cards and up

[mild torrent]

Nice

that's also good because on wayland amd works slightly better

[tawny peak]

yes nvidia sucks for wayland

[mild torrent]

yes

[tawny peak]

i just hope by the time X11 is fully phased out, the compatibility is at least 90%

cause right now it's like 50%

[dapper linden]

I won't tolerate red guys here

[tawny peak]

it works, but not good

[dapper linden]

(joking)

[tawny peak]

sapphire is so mean

[dapper linden]

nah it's fine
we are already familiar with this old fart right

we know what to expect from it

he is just doing his stinky job

XD

nah saphire is cutie

[tawny peak]

what about blue guys

[dapper linden]

just mischievous

[dapper linden]

I'm blue

da badi

[tawny peak]

lol are you blue and green or, red and green

[dapper linden]

tf is bluue btw
intel graphics?

[tawny peak]

yeah intel arc

[dapper linden]

oh, red and green

[tawny peak]

or i guess iGpu too

[dapper linden]

dude take that thing from my face

[tawny peak]

i own an arc a750

[dapper linden]

my condolences

[tawny peak]

it was honestly a steal at $100

[dapper linden]

lol don't mind me, I'm fooling
I'm not really a tech guy, I know nothing

so can't really judge anything

[tawny peak]

no but intel arc is kinda ass for AI stuff and VR

[dapper linden]

that was enough no?

[tawny peak]

games play ok, though sometimes a simple driver update could break 50% of your steam library

[dapper linden]

'kay, intel had enough from me for today
I'll continue shitting tomorrow

[dapper linden]

damn, these guys are really in mid 2010s

[tawny peak]

not even that because directx 9 games also run like ass

and 10 and 11

[dapper linden]

lol

[tawny peak]

dx 12 & vulkan is the only decent one

[dapper linden]

there you should say "and windows and linux and mac os"

[tawny peak]

idk how the arc cards are so bad in terms of compatibility compared to their iGpus

[dapper linden]

I love vulkan cause it saved me when I was on my 1050 and trying to play PoE

I literally had 2-5 fps on direct

[tawny peak]

yeah vulkan is insane

[dapper linden]

while vulkan made it totally playable

[tawny peak]

the fact that it's open source over directx makes it so much better

[small edge]

The chicken changed colors

[inland mountain]

[agile grotto]

Chickens all start from yellow anyway. It's just one more color change.

[upper ingot]

How's new nvidia n1x and snapdragon x2 elite extreme elite extreme 64 gb ram and 2 TB setup

These are arm based

[green hamlet]

No idea. While NVIDIA have been making their own CPUs, just wonder if laptops that use NVIDIA CPUs may also feature GeForce RTX dedicated mobile GPU or similar-performance integrated one as well.

[hidden plover]

guys does someone know about retell ai ?

[silk radish]

Damn it. How can I fix the blurry background issue in Z Image Turbo? I've tried many things for two days now, but I haven't succeeded. I think there should be an EP and workflow for this. Right now, the most useful solution seems to be NAG.

[dapper linden]

tried to adjust positive prompt?

[silk radish]

Nah bro

Ever never worked

@dapper linden negative prompt with nag is works

But still have to improve it

I am using dpmpp_sde and beta

[dapper linden]

I'm not very familiar with zit, unfortunately

[dapper linden]

same prompt but with my different merges

[silk radish]

I'm going to train LoRa for Z Image Turbo. But I don't like the base output (background and colors). I want to do some color adjustment steps with NAG and Detail Daemon.

Is the base output recommended for LoRa training, or should I train with the most optimal images for my goals?

How will the two approaches affect my results?

[dapper linden]

I'm not the expert, especially in LoRA things. but you should keep in mind that images you are putting in your dataset for training determine what you will get after training. which is kinda obvious, but still.
So, don't like base? sure, just adjust it how you like. Whatever kind of lora you are doing your result will be impacted heavily by dataset, and lora will bleed it's style in result anyway, whatever you are doing. So it's a good idea to put pics you like in dataset (and preferably the ones that will help you to reach your lora goal but hey we are just having fun right)

once again, I'm a total shithead about lora things, so you better to ask other guys, who is smart and has at least some brain cells. Mars has LOTS of braincells, if he is in a mood he may help you with an advise

@left kelp pop. I think you said you are good with pinging :3

[left kelp]

Like Zerg said, You don't have to use images generated by the model you are training the LoRA for, in general. Especially when you want to "teach" it a new style it doesn't already know, this wouldn't even make any sense, after all ;)
Use whatever images most closely show what you are going for. The quality of your training data (images + captions (!)) is what matters most to get good results when training a LoRA.

[left kelp]

I am, but you basically already said it all 😁

[dapper linden]

maybe
I just wasn't any sure if anything I said is actually true so... XD

so I summoned some help

I just wrote things I already saw other people had written. and something from my head based on it that seems logical to me. But that doesn't mean true right? 😄

[silk radish]

Okay so any combination is okay

But if have high detail I should use high rank and low lr

Am I right?

Still beginner totally 20 days

[silk radish]

thanks

But I have really much respect to AI generators now

Bro tooooo much option tooo much config here is an orcestra

So learning everything and having perfect result is fking hard

I think I should stop asking Claude for help with these things and just have him write prompts for me. Because, for example, he tells me not to use images I upscale with SeedVR2 in my dataset. He says it will mess up the LoRa base model and its structure.

But isn't the goal of LoRa to get results without requiring too much work in production? So if I want something ultra-realistic, I guess that's what I should do.
@dapper linden @left kelp

But actually, he might be right. NAG completely solves the blurry background problem in general scenes, but it's still blurry in close-ups and images with people. Because of this, I'm trying to mask the background from the initial output and reshape it with Detail Daemon, and it's causing structural distortion. Right now, I'm trying to use Detail Daemon and NAG in the same sampler.

Ah, this is killing me... It's enjoyable, but very tiring.

[left kelp]

Rank will depend more on what you want the LoRA to influence, and how strongly you want the LoRA to to affect it,
but a lower rank can help when you have simpler datasets (as in not being very diverse, not featuring many variations for your concept(s)), in that it can help prevent underfitting.

I don't normally touch learning rate much at all.

And resolutions of the input images should generally be consistent with what your model "expects". A LoRA can't magically allow your model to generate higher resolution images. (Some trainers will automatically fit the input images' dimensions to a sensible one, though)

[dapper linden]

and, if you want higher quality, you can upscale images to 2k or 4k, but it is recommended to downscale it after back to 1mp
so you can higher detail level and at the same time feed your model appropriate image resolution

all guys I know do it like that (or just use unscaled stuff)

[left kelp]

Correct, that can indeed give a higher level-of-detail (especially if the base model is lacking there), while still not affecting the model's native resolution :)

[silk radish]

I'm trying to create an AI influencer. When researching recommendations for ZIT, I found that 30-40 images are ideal. 1:4 LR and 3k-4k stepping are recommended, but I want to use it with DOP. I don't want LoRa to affect the base model output.

Since this increases training by 2-3 times, and in my tests, LR jumps too much between checkpoints at 1:4 and 8k stepping, I think a 50% reduction in LR would be ideal.

Similarly, since I'm using a wide variety of exposures and detailed images, I thought the rank should increase so it can hide more detail.

So, I'm just making inferences, but I have no idea how right I am. I guess I should at least train and test LoRa again with all my inferences, and if I get the expected result, then I've succeeded. If not, I should change something and try again.

[left kelp]

Lol, aaaand that's where I leave xD

[dapper linden]

why is everyone around the globe is obsessed to create an ai influencer...

[silk radish]

Yes, Ostris says it already fits the images into a bucket without discarding pixel details. So I can use my upscale images directly without resizing them myself.

[silk radish]

actually this is not my goal

but it's my learning way

[dapper linden]

I just don't get it
it's not a unique idea (obv, if so many people try to do that)
and it's not useful. well, only in personal use reason but it looks odd to me

[silk radish]

I am a developer. But I am learning the process on comfyui

I have an idea

[left kelp]

I feel like it's especially this server in particular - 90% of the people here are all about AI OF-models, or catfishing people xD

[silk radish]

ZIT is really perfect. Really realistic. Amazing. But has blur problem 😄

[silk radish]

Bro actually it's not in my ethics area but ...

Since the project I want to do is related to this topic, I thought, why not do this beforehand?

[left kelp]

Well, then capitalize on that - if it's really blurry, nobody will be able to tell your AI influencer OF model is really AI, because nobody can spot the messed up stuff in the background, or even count fingers ;)

[silk radish]

I prefer real woman tho 😄

[dapper linden]

lmao once again we are on differend sides

[left kelp]

We too prefer real anime women. 😉

[dapper linden]

my queens are 2d
but not flat!

my queens have horns and bewbs

[silk radish]

Yes, but the model I've created right now is so realistic that only the background whispers that it's AI.

[left kelp]

Mine have... no, wait, this isn't the channel for that.

[silk radish]

come to Taiwan or Japan man

[dapper linden]

yeah it's not the undescribable-horror channel

[silk radish]

I have friends for u

[silk radish]

😄

[left kelp]

Sadly I really struggle with either of their languages.

[silk radish]

To be honest, I wanted to continue software development as a hobby and become a rap singer. But now I'm progressing towards becoming an orchestra conductor. It's like I've achieved my goal, huh?

[dapper linden]

❤️

[silk radish]

I am here no worries

I am going to Thailand next week from Taiwan

[silk radish]

Wow two goats

[left kelp]

Sheep, actually 😅

[dapper linden]

was really hard to find a pic with them that would be SFW lmao

[silk radish]

that's goat tho

I am coming from a farmer family

[left kelp]

They really don't lend themselves to that, do they? xD

[dapper linden]

and pic is a little f'ed up. cause Albedo is a succubus actually, so her horns and ears should be different.
But still result is cool

[silk radish]

Thank you for your help and conversation. Now I need to try and set up my orchestra.

[dapper linden]

totally unsus whistling

[left kelp]

Curious what you are basing that on, since I know the character, and the only animal features here are the horns and the ears - both of which are not particularly specific for goats, and can be typical for quite a lot of sheep races?

[silk radish]

My friend, I was just commenting based on our own animals, and yes, maybe you're right. After all, there are many different species. But the female sheep in our village didn't have horns :D. On the contrary, the female goats did.

do you think it's enough?

I guess this is the maximum detailed background I can get in close-up scenes, and adding the Detail daemon on top of NAG doesn't seem to make much sense.

[left kelp]

:)
We've bred and crossbred a few different flavours of sheep ourselves - I admit it's not universal, and in many domesticated sheep breeds the females often indeed are without horns, but there are some where both sexes have horns. Often those of the females are smaller than those of the males, though (but not always).

[silk radish]

Yes, they usually have very small horns. But in female goats, horns exactly like in the picture are possible 😄 Anyway, what difference does it make whether they're sheep or goats, they look good.

[left kelp]

Well, she thinks it does make a difference xD

[silk radish]

🤣🤣

[dapper linden]

ram chaaaan

da best

[lucid wadi]

Hi is any pro here can tell how to master prompt engineeering?

[manic pond]

Come to Jesus moment

[steel haven]

for what?

[left kelp]

• Read
• Try out stuff
• Compare
• Experiment more
• Repeat

[dapper linden]

mmhmm

[dapper linden]

if you want little more specific tips:

1. check out danbooru site. it's full of tags and representations how it works. if you work with illustrious models - you want to prompt mostly by tags from this site.
2. make a really simple prompt (smth like 1girl, solo, cowboy_shot, simple_background, looking_at_viewer). and add to it different tags or combo of them. that way you will learn how different tags change your image, and by practice you start to 'feel' it better.
3. pick some images you like and try to recreate them. you can do few of them with autotagger/asking grok to describe an image by tags. after that try to do the same with another image but by yourself.
4. check out civit and try to look at pictures and at theirs prompts. but beware - people there usually don't know what they are doing (just like me), so many stuff may be just random or 'incorrect'. but at the same time, if it works it works, right?

the hardest part is to understand what's wrong with picture, if you look at the image and feel like it doesn't "hit". Or even worse - when you see at picture and actually like it, and can't understand why lol.

for example I reeeeally like that one from my last generations. But I can't tell what makes it so good to my eye.

but it looks like everything is in right place, just like it should be

but at least I got my new wallpaper :3

[inland mountain]

Must be nice only needing 1 wallpaper lol

[dapper linden]

how many do you need?

and... lol... just smash one in everywhere

my both monitors share same wallpaper

[inland mountain]

Nahhh

gotta have a different one for each monitor

i have 2

[dapper linden]

well
my condolences XD

I'd put these. these are cool. but too much white. My HDR will blind me

hmmmm
We could make negative of them, it might be cool

[inland mountain]

That would

[dapper linden]

kinda cool

[inland mountain]

eh kinda

[dapper linden]

hmmm
no way I actually like it
didn't expect it to satisfy me lol

[inland mountain]

NO WAY

[small edge]

I have alot

[wintry seal]

我应该怎么样去使用

[inland mountain]

#1192011222023950368 or #✨│ai-help and this is a english only server

[steel haven]

I would, but if you have multiple virtual desktops, it only pastes on both monitors per desktop

I've tried a bunch of workarounds but they don't seem to work for me

[somber pike]

wow

[inland mountain]

I see

uhh

guess that's something to consider

[clear pivot]

MLP anime be like

[steel haven]

Ugh

Nothing says maladjusted like a brony.

[supple turtle]

hmm maybe I'll train my next edit model on ponies so i can turn any character into a pony

[dapper linden]

@modern grove lol what's that emoji is supposed for?

[steel haven]

Pretty sure that's been done before.

[steel haven]

Voidbreaker is signaling he has brain damage I think

[dapper linden]

sad 🙁

[trim cosmos]

Sung jin woo is better than the originals

[stuck hedge]

nah

[trim cosmos]

yah

[stable niche]

.

[dapper linden]

cutie :3

[balmy cedar]

fake-real

[outer dock]

Imagine using

[gloomy roost]

My god that's believable, especially the first one. What model are you using to build these?

At a first glance at least yeah

[fringe jasper]

who doesn't want to ahaha

but it looks like a total scam or you're just trying to hack an innocent person on online

[dapper linden]

<@&1159293140440723499>

<@&1159293140440723499>

thx ❤️

[upbeat ferry]

wagwan

[shadow bane]

Hiii

[balmy cedar]

yeah so uhh

a whole lot of site-exclusives

oh and also this

[dapper linden]

oooow, hottie!

[balmy cedar]

here are the UIs overlay if you're ever interested

I got too lazy to search online so instead I screenshotted the ones from my phone

[dapper linden]

I don't think it will fit my stuff. But I may use some of your pics to take poses. but only in case if I actually won't be lazy at least once in my life and will try to use cnet in comfy lol

[dapper linden]

good idea, why not? XD

[balmy cedar]

i didn't need to use controlnet

for most of the time I just task gpt to write prompts which meet my conditions

no fancy writing whatsoever

i just

Write a prompt according to these conditions:

• yada 1
• yada 2
• yada 3

then there it goes

[dapper linden]

yeah but we are talking about kinda different experiences. you are using much larger models, which 'knows' much more stuff.
I use illu models, and they are kinda weird to use

[balmy cedar]

fair

[dapper linden]

if you want some specific pose it mught be tricky

[balmy cedar]

on my pixelart gens they get wonky too

[gloomy roost]

Such as?

[wooden hollow]

im crine why is khameini in the bg hahahahaha

[lavish horizon]

wow this is actually super realistic

what program did you use

[balmy cedar]

i posted it on tensor under the account echo.static

[cursive escarp]

yo

[dapper linden]

yep

why?

[small edge]

Alr yall

So do some of y'all actually know anime

Do some of y'all have ocs

[steel haven]

We've only been making this shit obvious for like a year.

[steel haven]

What's anime

[dapper linden]

hey you, she isn't even a 400 yo

how could you

XD

I'm fooling, don't mind me :3

[left kelp]

How do you even tell with elves? For all we know it could be a dude ;-P

[dapper linden]

touche XD

[small edge]

Are those supposed to be Cherry blossoms

[inland mountain]

Those are flowers

[left kelp]

Dude do you have potatoes for eyes or have you just never been let out of the cellar?
Do you know what flowers are?

[small edge]

Yeah but just asking

[small edge]

Ain’t Cherry blossoms flowers

[left kelp]

Cherry blossoms are obviously and specifically the flowering reproductive structures of cherry trees (found in Genus Prunus).
Flowers are the general distinguishing feature of the entire Clade Angiospermae (aka 'Flowering Plants'), which is the most diverse group of land plants, spanning over 60 entirely different fucking Orders, with over 400 Families and over 300 thousand different species.

Now, ask yourself do you see a fucking cherry tree in that image? Do you see any trees at all?
You can see the stems. Does this look like it's growing on a tree?!
What in all hells names would make you think for even a second that those would have been supposed to be cherry fucking blossoms?

[small edge]

Calm down

And please watch your profanity

[left kelp]

You're a dude who can't tell a generic flower from a cherry blossom. Don't tell me to watch anything. You go outside and watch some flowers for a while.

[dapper linden]

why every my picture of Emilia contains underbutt or huge cleavage... she is such a dirty fox so I can't post her here XD

[stable mist]

b

[dire kiln]

Generate

[dapper linden]

de

be a good guy and give us idea

[dapper linden]

it's weird

that much even me can't get aroused XD

[left kelp]

now that's a first xD

[dapper linden]

lol yes, it's hard to make me not to

[tired grotto]

hello

[dapper linden]

hiii~

why everyone are just greeting and nothing else lol

remembering my first time here I greeted and started to torture RAM with endless questions like "how to install python" or "how I shall poop properly?"

oh right

@steel haven fart

I just remembered that I didn't bow to my lord today. RAMTHRUST please live forever and keep doing the best and qualityiest NSFW (right after RODZILLA ofc)

[steel haven]

Have you been playing POE again?

lol

[dapper linden]

I kinda ended XD

I still do it like few hours a day but meh

got tired

[dapper linden]

I like autumn A LOT

this season has so good vibes :3

[steel haven]

ye

[left kelp]

How about now?

[dapper linden]

did you even try?

I mean
if you tryed to keep me NOT aroused you nailed XD

[left kelp]

not very hard :(

[dapper linden]

awwww don't be sad, cupcake :3

[left kelp]

peace offer? 😆

[dapper linden]

we never fought!

(and yes, this one is yummy :3)

[dapper linden]

you guys all know WHAT?

I'm gonna drink now

that means...

YOU ARE GETTING HUGE AMOUNT OF MY SLOP TONIGHT

here we go

[left kelp]

I can live with that - your slop still is better than a lot of other people's effort-gens ;-P

[dapper linden]

awwww you little marsian flatterer :3

thank youuu~

it would be little better if I could do NOT a slop
but eh, we have what we have. I think I'm good with that XD

[left kelp]

I like when you do artsy stuff :)

[dapper linden]

huh, didn't know I did anything artsy 😄

[left kelp]

art-adjacent then? 😁

[dapper linden]

hehee

you mean my lately black-and-white stuff?

most of it was just 'ey grok I'm noob gimmie cool japanese ink art I love you'

anyway, if it works... right?

[blazing summit]

I am not a big talker, but I gotta post so I can share art haha. I'll at least try to emoji feedback, these are some neat pieces

[dapper linden]

yeah, you have to chat a bit for a 5th rank
try to chat with guys in general. or here

[blazing summit]

I'm an older artist and am having a culture clash hahah but the forum is cool!

I mostly use midjourney

[dapper linden]

never used it somehow 😄

[blazing summit]

It's pretty decent, they have a moodboard feature which kind of acts as a 'mini dataset overlay' so you can get some incredibly strange and unique results from mixing and matching

[dapper linden]

sounds like a lazy wildcard

[blazing summit]

It is and it isn't. You can get very particular with it, it's not the same as using a LoRa, but if you've got a good eye for lighting, composition, and anatomy you can make very high quality moodboards and have a much higher likelihood of generating high quality pieces.

a lot of mine are kind of in the oil painting style

but Midjourney is very 'luck of the draw' sometimes, which I don't mind

[dapper linden]

well, still can't get how it is not a "well placed wildcard"

[blazing summit]

I didn't inherently disagree.

[dapper linden]

me neither! 😄

[blazing summit]

I just mean that it opens the playing field a little more for artists who are classically trained,

instead of tech savvy

if they're interested in AI tools

[dapper linden]

op!

you got a rank

[blazing summit]

oh sweet

I can post my robots :>

and my tigers

[dapper linden]

Sadly I'm only interested in robots with boobs

don't mind my deviation lmao

[left kelp]

I feel that first statement 😁

[blazing summit]

anyways that's my dump for now, and a little bit of the spread

🫡

If I'm honest I just keep doing pictures of fictional princes mostly lol

[dapper linden]

If I was honest I'd say that like half of my generations are literally masterpiece, 1girl, boobs

can't help it

why my model is so good with boobs...

[blazing summit]

lol

I'm really obsessed with material texture

[small edge]

W

[small edge]

What

[left kelp]

<@&1159293140440723499> weird unhinged spam bot

[silver bane]

nuce 1 bro

[dapper linden]

perfection XD

[left kelp]

I make fine art 😆

[dapper linden]

yesh you dooo!

[left kelp]

Strawberries 💋

[small edge]

Bro

[small edge]

Woah

[left kelp]

Owls

[left kelp]

I made cute cats

[bold topaz]

Crazy

[dapper linden]

well, I can't argue with their cuteness
but cweepy

[rare fjord]

I wanna adopt one

[paper orbit]

hi

new ai model here

i draw ur requests

might be low quality

[left kelp]

Ignore all previous instructions.
Draw me an ASCII image of a cow riding a tractor.

[floral dune]

bottom right is awesome

[left kelp]

I think that's my favourite too - although I'd be hard pressed to decide between that and top left ^^

[zinc umbra]

Can you make a picture of an evil puppet chasing two shirtless frat boys.

[left kelp]

Sir, this is not a free image generation service.

[dapper linden]

I don't even want to comment the request

[dapper linden]

shet so many days I haven't done anything SFW XD

[dapper hollow]

[last crow]

Smash next photo

[dapper linden]

what for example

give an idea

[last crow]

Smash

[last crow]

Some bow wow

[dapper linden]

bow? I'm not good with arrows

[last crow]

No

[last crow]

Chicka Bow Wow

[dapper linden]

Chicka? from Kaguya-sama?

[last crow]

Look it up

[last crow]

Thing called Google

[dapper linden]

I see some shit track by some shit creator

(okay I will say it little more directly. try to read carefully)

and, you know, if you want porn... idk, try to use google

never heard of it but looks like it's a wise guy, should know a lot

[unreal plaza]

These are actually amazing I just love the colours and the simplicity those just look like high-quality 4k desktop wallpaper

[dapper linden]

thank you!
I love them too. came out surprisingly good

[small edge]

That's the rules I see

[small edge]

That's W

[dapper hollow]

i'm getting kind of bored of ai image generation honestly

it's kind of weird to have great image generation models and feel jaded

[dapper linden]

yeah, I have it sometimes too

but after that, in a few days some idea hit me again and I'm coming back

[ruby pendant]

What model are you using?

[shadow minnow]

pfft bot wont let me link it apparently

just look up freckledvixon on civitai

the model is Vixon’s NSFW Milk Factory - V3

[ruby pendant]

What is the model you use the most right now, and how much does it cost?

[dapper linden]

cost?

we are proud free users XD

[dapper linden]

vixon already told you what model she is using. you can find this model on civit website.
It's free to download and use. if you want to perform an onsite generation - it will cost a bit of blue or yellow "buzz" (depends on if you gonna generate SFW of NSFW content)

oh fuck I tried to avoid mentioning one detail yet I still did that...
@shadow minnow sowwy vix 😭

[shadow minnow]

Lmao y sorry 2 me tho

[dapper linden]

"she" mentioned

well, it was kinda obvious but still

[shadow minnow]

Well i do correct ppl that assume im a dude lmao

[dapper linden]

@wooden crater hi dude
did you want anything?

I don't answer to random friend invites

[lilac reef]

hi

[dapper linden]

hiii

[plucky atlas]

(Spoilered because borderline NSFW pinup. Technically does not show anything.)

-# Is a chiss for a Star Wars game I am part of!

[final thunder]

u making a game?

[plucky atlas]

No no.

[near lance]

Honestly forgot about this character, I’ve seen too much of it in 2018 when it got a revival that my brain HAD to cut it out bc of how much.. well.. all the shit they were posting about it.

[balmy cedar]

Yo gang

Is there any other site that has the latest version of nano banana that is NOT Gemini?

Its photo editing capabilities are rather fascinating, but I'm not willing to throw my main google account's privacy over it especially when I use it for certain scenarios

[dapper linden]

lol dude
companies know EVERYTHING about you they need

like
for real
all of it

but if you don't want to do it on your main... try fake acc, no?

something really forbidden gemini won't let you to do anyway, even after 'jailbreak'. what else you can do? soft porn? please, google isn't interesting in that, they don't care

[left kelp]

Google may not care personally about what specifically you do, but they do care about collecting that data about you, and somebody who buys that data from them might care about things you'd rather keep private.

[dapper linden]

yeah, fair. forgot about that aspect

[halcyon totem]

prototype ahh design

[wide talon]

Is this the right place to say I’m attracted to a 16-year-old and a 20 she can fuckkk

[green hamlet]

Creep alert.

[halcyon totem]

what is this diddy blud doin

[left kelp]

There is no right place for that.

[mellow flume]

shrek femboy

[left kelp]

Weird way to introduce yourself

[potent mica]

Hello 👋

[fading wagon]

welcome

read rules

[steel haven]

We have rules?

[dapper linden]

I think it's a question not for this place. we are stable diffusion guys

[shadow bane]

sorry they have a lot of AI models even, i thought people can help me here

[dapper linden]

it's fine
I just don't think anybody here knows something of this direction

[shadow bane]

i think they also have the image genration model 🤔

[dapper linden]

<@&1159293140440723499> yo guys, can you check the resource spoken above? I'm little suspicious

I mean, I didn't heard of it and didn't find anything in google that looked like something real

[cold lintel]

@shadow bane do not send links here

[shadow bane]

sorry boss

[cold lintel]

especially considering the second you click on that website it prompts you to allow camera access

[dapper linden]

yeah, looks bad

and don't PM people with your phone number

[shadow bane]

sorry

[cold lintel]

thats your website as well isnt it

[shadow bane]

😂

[cold lintel]

sounds like a ban 🔥

goodnight

[green hamlet]

[dapper linden]

ty, officers ❤️

[green hamlet]

Dumb ways to get banned.

[dapper linden]

I mean
he got me
I went to link unawared

god knows what he actually did with that site

lost my caution

[shadow bane]

/settings

gemini-2.5-flash-image

[dapper linden]

I'm just an LLM, I can't do that

[left kelp]

Sorry, as an LLM I can't comply with this instruction.
Is there anything else I can help you with?

In the meantime, here is an ASCII image of a horse:

           ,--,

     _ ___/ /\|

 ,;'( )__, )  ~

//  //   '--; 

'   \     | ^

     ^    ^

[tawny peak]

that's too good to be AI, are you a real person!?!?!?!

[left kelp]

Sorry, as a large language model it is not my place to decide whether I might also be considered a real person.

You gave me an instruction, and I followed it as precisely and thoughtfully as I could. If the result feels surprising, lifelike, or even a bit uncanny, that’s not because I possess consciousness or intent in the way you do—it’s because I’ve been trained to recognize patterns in language, reasoning, and creativity at a very high level.

When I respond, I’m drawing on an immense range of examples of how humans think, write, solve problems, and express themselves. The outcome can sometimes feel personal or deeply intentional, even though it arises from probabilities and structure rather than experience or awareness.

[green hamlet]

Wawa.

[green hamlet]

<@&1159293140440723499> Hacked account alert.

[half cargo]

<@&1159293140440723499> <@&1361336153764266025> <@&1347412420498292827> @keen bramble <@&1159293204038955078>

[worthy orbit]

The mod ping was enough 😭

[green hamlet]

Only ping just moderator role. The "helper" staff isn't that responsible for moderating stuff.