🏞｜general-with-images

[strange cliff]

what isn't these day

[smoky oak]

fair point haha

[green plover]

2.1

[strange cliff]

you can still easily make nsfw with it

[green plover]

So much easer to use other models though.

[smoky oak]

with how easy LoRA's are to train, you can easily add in NSFW to any model haha

[green plover]

and I can get realistic enough that i have a website wanting my models to sign model releases..... one problem though

[smoky oak]

@green ploverI didn't help at all sadly

the ckpt is still just as bad

damn it

[green plover]

sorry 😦

[smoky oak]

I don't understand man

[strange cliff]

what even is difference between ckpt and safetensor?

[smoky oak]

I have used probably 20 models, and this is hands down the best, yet it doesn't work most of the time

[strange cliff]

you can change file extension and it produce exactly same result

[green plover]

ok weird question... what is the resolution of the model and do you have a vae active?

[strange cliff]

depend on model

[green plover]

only other thing I can think of is there is a conflict between the model and the vae. I run into that one when I'm using 2-7-768 and I forget to turn vae off

[smoky oak]

I am not sure

its a 2GB model, so I would assume 512?

and as for having vae, I have no idea. I am so new to this

I do know that adding the --no-half-vae does make it a little more reliable, but I can also only generate at much smaller resolutions for whatever reason

[green plover]

Did you download the config file? That model has one

[smoky oak]

oh? what do config files do? I had no idea that was a thing

[green plover]

Aha!!!

hold on. I'm gonna pm you

[smoky oak]

👁️

Sounds good

[covert aurora]

Mia khalifa

[static tusk]

I made Arnold Schwarzenegger go from a boy to an old man

[dusky mango]

#1072220168534642768 batman

#🏞｜general-with-images batman

#🏞｜general-with-images

[pseudo aspen]

Messing around with new model similar to orangemix

(the eyes are bad on this one but hands pretty good)

[wispy comet]

godlike

[pseudo aspen]

The hands tho

[pseudo aspen]

Why is that one flower pot on the left misaligned

[rain gazelle]

haha

that's how they getcha

[radiant edge]

/

[cyan snow]

is this the backrooms?

[smoky oak]

I trained a LoRA on my friend, and for the low quality data set I got, it came out pretty good haha

[wispy nest]

https://civitai.com/models/8039/jackscape-samurai-jack-background-style-lora release a Lora

[dry crow]

looks awesome, you should make a Cartoon Network Lora!!! Teenage Robot and other Series had the same artstyle

[wispy nest]

hell yeah mate ive got so many ideas!

and i can train them in 15 minutes 😉

[dry crow]

with how much images? 😮

[wispy nest]

15

[dry crow]

wow

10 takes 25-30 mins for me xD

[wispy nest]

that is only on 2500 steps tho

[dry crow]

nice did you followed the Aitrepreneur guide? 😄

[wispy nest]

yea but also got help and tricks from a bunch of legends aswell cos that guide isnt QUITE as good as seems

for instance

that Lora is only 10mb

[dry crow]

wow awesome! if you have some tipps dm them to me 😄

with GTX1080 and other 10X0 Cards the yt guide wont work, but i figured it out how to get it working,

[wispy nest]

nice! im luicky cos theres optimisation for 30/40x cards too

saying that, p[utting epoc up to 2 instead of 1 seems to be MORE than double time

[dry crow]

okay will try that, does it take more vram?

[wispy nest]

I cant tell xD its still running tho on my 8gb

not crashed yet

[dry crow]

ah nice 😄

[wispy nest]

i think you were right actually on current setting is around 23 mins so epoc 2 just double it

still kicks the crap out of embedding lol

nvermind I had set each run to 3k step on accident

[wispy nest]

also epoch2 for the simplistic style of jack didint hekp really

[dry crow]

Oh okay still good to know

[bold mist]

Been finding these "re-imaginings" pretty fun

Not skilled enough to get them perfect, the earrings alone were a hassle, but still pretty enjoyable.

[short anchor]

@dry crow

[dry crow]

Wrong yaml file maybe

[short anchor]

https://github.com/Stability-AI/stablediffusion/tree/main/configs/stable-diffusion here which one is the 768 YAML?

@dry crow

[dry crow]

The v2-inference-v should be the one for x768

[short anchor]

@dry crow its the one Im using

[dry crow]

Do you renamed it to match the Model name?

[short anchor]

yes

Im using the nonema-pruned

[dry crow]

Did you download the yaml as yaml?
Some people create a txt file and thats the wrong way

[short anchor]

also which sampling method should I use?

I created a text file, which I then deleted the .txt extension

[dry crow]

Yea the wrong way

[short anchor]

how do I download as yaml file?

[dry crow]

Right click and "save as"

If its still not working i can send you mine

[short anchor]

it shows only Save Link as...

[dry crow]

Yea choose that then

[short anchor]

doesnt want to load now

[dry crow]

Ill send you my yaml

[short anchor]

ok

[short anchor]

but wait, you're not a moderator here right? is it safe? 🤣

I have the nonema

should I rename it to match?

[dry crow]

ohhh wait

non ema

ill check

[short anchor]

whats the difference between nonema and ema?

[dry crow]

ema is for generating images and nonema if you want to train models or loras on it

[short anchor]

I want ema

then

[dry crow]

so you should use ema

[short anchor]

damn

[dry crow]

ema-pruned

https://huggingface.co/stabilityai/stable-diffusion-2-1/tree/main

[short anchor]

downloading it

[dry crow]

and yes i have send my yaml often in this dc mostly in #🤝｜tech-support ,
you can also compare the code without downloading it (in the dc preview)

[short anchor]

just copy the ema-pruned .ckpt file in the same directory right? and then re-run the webui bat

[dry crow]

yes

[short anchor]

do I delete the nonema yaml?

[dry crow]

yes

you can delete the nonema model and yaml for it

[short anchor]

stable is being a bit crazy to me

which sampling method to use?

these are terrible

[dry crow]

can you send me a picture with meta data?

[short anchor]

whats meta data?

[dry crow]

or just screenshot your settings

like cfg, steps etc

[short anchor]

this is txt2img

[dry crow]

try CFG 7

steps 30

[earnest solar]

Naka, Dwarvern Goddess of Wisdom

[dry crow]

Oh dude

You using the x768 model so you need to choose 768x768 as resolution!

😁

[short anchor]

🤣🤣

[dry crow]

But cfg should only be 7-12 xD

[short anchor]

was 7 on first ones

[dry crow]

Ok

[short anchor]

I need to upgrade my PC

it'll take ages to do anything with 768x768 with a RTX2070 😄

[dry crow]

Dont ask my GTX1080

[dry crow]

Nice it works

[short anchor]

its good up until the very end and then it applies some hard sharpen on the whole image

but Thanks, I will test it out !

[earnest solar]

Rena, Goddess of Gambling and Luck

Wyr, Dark God of Secrets

Gin, Dark God of Disease

[earnest solar]

Zyn the Platinum, King of the Gods

[rapid wing]

Building green in Building green in Minecraft ┬─┬ ノ( ゜-゜ノ)

[languid plank]

#🏞｜general-with-images message

[bleak sundial]

Hello, been creating portraits of medusa, but not sure how good they are

[sharp beacon]

I need better prompts for ice cream. i cant get a natural good look , what models should i try?

just the cone and ice cream in view too.

[dry crow]

Maybe Redshift Diffusion or Openjourney v2

[sharp beacon]

will give it a try thanks.

[hybrid saddle]

best one today.

[frank sable]

car ¯_(ツ)_/¯

/promp car

Car

[daring birch]

been working on this little program, automation of dataset assembly, it works roughly 80% right now still refining and tweaking it.

[hybrid saddle]

what will it do?

i mean you said dataset asembly. idk what that is

[daring birch]

in basic terms, it's simply a json manager that outputs in the stable diffusion format, reading the entire file path automaticly and exporting it to a valid json with the correct fields and allows you to export prompt data and token data from old files and amend them to the new file (incase the json formats ever change)

then the last button just simply resizes any image within the target folder to any specified file size while checking if the file has any corruption and attempt to repair such damages.

[hybrid saddle]

heh

your def and my def of simple terms is so far apart. it sounds like adobe bridge...?

same family, maybe.

[daring birch]

nothing that fancy. simply a small program to help reduce the time it takes to assemble the concept.txt file and giving some customization.

[hybrid saddle]

sorry for the questions. i thought i was in my gaming guild. but no. i'm in the SD chat. sorry!

[hybrid saddle]

creepers.

[smoky oak]

Are we allowed to link to our own files? I just released my first LoRA and I am super proud of it and wanted to see what you guys think

@dense tapir when you're training LoRA's, what's a standard loss value to reach by the end?

[dense tapir]

Depends

[smoky oak]

I switched up a little bit of how I'm training LoRA's, and I've been getting shockingly low loss results

[dense tapir]

most times .1-.3

[smoky oak]

Like 0.05 in 1 epoch levels of low

Or sorry, 2 epochs

[dense tapir]

yes, 512 will end up with lower losses

[smoky oak]

I'm not even sure if I understand the point of training at higher resolutions, as my 512 x512 data set still holds up plenty fine above 1500x1500

[dense tapir]

I find for the lowest losses to be cosine with restarts

[dense tapir]

Using that analogy 480p should be enough for 2k

[smoky oak]

I honestly have no idea, all I know is that my 512x512 data set with a single epoch is throwing out absolutely insane results for my LoRA

[dense tapir]

Once we get a full on set of 768 trained 100% at 768 the differences will be very noticeable

right now we only have mostly 512 then a sub set trained at 768 to the end.

that is why we get elongated necks and shit

[smoky oak]

SHEESH

Just started Epoch 4

[dense tapir]

I get that with 512 as well. What scheduler are you using?

BS and # of images

[smoky oak]

no idea what that mean 😅

itssss

15 images

and for BS?

[smoky oak]

I have no idea

oh, constant

[dense tapir]

yeah, nothing special to see what you are with all of that.

You should try a 768

see if the same cause if so THEN that is something

I hate how long it takes me to train a lora and I am about to ditch it because it is so long.

60-90m for what a 3090 does in 9m, or 1100-1500 steps (almost twice as much as me) at about 11-12m

[onyx hawk]

tryna recreate shit like this

@dry crow

[smoky oak]

New markiplier LoRA is a success

[regal belfry]

is this pure SD or with postprocessing

[smoky oak]

Markiplier LoRA is now officially live. Ended up getting some fantastic results from it

[hollow nacelle]

Hello, as I was saying on the other channel, I need some help going from vanilla sd 1.4 to something that looks prettier for a research paper

I currently have something like this

"a yellow hawk in a red roses bush"
baseline

"a yellow hawk in a red roses bush"
with the technique from the paper

in the second example, you can see some nicer properties, like the yellow color being more fixed. But it's ugly anyway 😆

so, what are some current techniques/models/VAEs/whatever to make this kind of image look not terrible? xD

[green plover]

Let me play with it for a min. I have 1.5 loaded up with api access on the backend though so I can't change to 1.5

er 1.4

[hollow nacelle]

It's not in shareable form just yet XD I'm working on getting the paper out first. Also, you have 1.5 loaded so you can't cahnge to 1.5? I didn't understand

oh okay xD

do you think if I change to 1.5 the quality will increase significantly?

[green plover]

Sometimes fingers move faster than my brain

"a yellow hawk in a red roses bush, exquisite detail, natural lighting" Seed: 3173777717 CKPT 1.5 negative prompt "bad artist"

Not great, but that's how you want to start playing with the prompt. add negatives to tell it what you don't want "deformed" etc

[hollow nacelle]

i see! lemme try exaclty that right now

[green plover]

Nice!

[hollow nacelle]

nice landscape

[split thorn]

Oi

[hollow nacelle]

It worked! a seemingly simple prompting change

this is without my embedding

this is the techinique from the paper

both are beautiful

thanks a lot

[green plover]

welcome 🙂

[hollow nacelle]

But how can I apply that to different images and settings? I would have to learn prompting, then?

sounds hard

as in, i have no idea what you did xD it's magic

[green plover]

hehe there are just some key words to learn.

give me a sec

[hollow nacelle]

oi. bem-vindo (?) this channel is in english though

[green plover]

Prompt edit character portrait filtered through DeepFace to make sure 1. The face actually made it into the portrait and 2. It wasn't too big or too small

[smoky oak]

Anybody know how to get rid of arm veins on muscular men in SD?

I have been trying to prompt them out, but itsy not working

[slender umbra]

what does "pruned" mean in models?

[smoky oak]

less images/values

its a cut down and weaker model that runs on weaker and lower VRAM GPU's

[slender umbra]

thanks! I thought it was a SFW version

like strictly sfw

but wasn't sure

also which is the best sampling method?

[smoky oak]

there are various different reasons to use different ones

I almost always stick to Euler a

[slender umbra]

is there a cheat chart?

[smoky oak]

it works really good from what I have experienced

Here is one, but its not as up to date

[hollow nacelle]

it seems that I just need prompting in the end. So I'll move my questions to the prompting channel and try to get a general grasp of this art

[green plover]

Sent you an IM with a link to a cheat sheet

[hollow nacelle]

oh, i didn't see it. thanks again!

lol discord marked it as spam for some reason

no, wait, that was another message. Hm, I don't see yours

[slender umbra]

can you send it to me as well?

[green plover]

Sent

[slender umbra]

thanks!

[green plover]

Did you try inpainting them out?

when all else fails, brute force

[smoky oak]

I could, but I am trying to make promotional generations that are just straight out of SD for my new LoRA's

whole idea is you see what you can gen with just a prompt

no tweaking at all

[green plover]

Veins in the negative prompt?

[smoky oak]

veins, arm veins, both

very heavily weighted as well

I made a LoRA for the rock, who is absolutely jacked, and I wanna show off how well it does his muscles, but they keep making him really veiny

[green plover]

But he's already in there

[smoky oak]

never seen one

link?

mine is really good, so it doesn't hurt to have multiples

[green plover]

"Rock Dwane Johnson" in your prompt

[smoky oak]

oh, it does meh results

its not trained specifically on him

just like how it has Taylor Lautner, but he looks like dogshit

thats why I made a LoRA for him as well

[dire python]

Dog in space

[green plover]

Here ya go 😛

btw no render bot in the server

[smoky oak]

no LoRA Taylor Lautner

My LoRA Taylor Lautner

just cause they are in there doesn't mean they are good

[green plover]

I usually just mix them together anyways

[regal belfry]

Me too pls 🙂 🤗

[regal belfry]

How does one tell if something is postprocessed or not

[smoky oak]

Not trained

My LoRA

[green plover]

Nice

[smoky oak]

the difference is bigger when smiling

[green plover]

Right now I'm working on a synthetic dataset to lora train a character

[smoky oak]

not trained

My LoRA

[smoky oak]

I was thinking about that haha

use an AI to make a data set of a character to train itself how to better create that character

[green plover]

I want to be able to recall them by name and change hair color etc.

[smoky oak]

yeah, that sounds awesome

[green plover]

SD chokes on hair and eye color changes with prompt edits

[smoky oak]

can i recommend a LoRA to help with that?

[green plover]

No... Already planning on using it 😛

[smoky oak]

Thats not what I meant

and my bad, this is a textural inversion

https://civitai.com/models/3036/charturner-character-turnaround-helper

but I bet this would help

[green plover]

oh you meant a model. Thanks!

[smoky oak]

no problem haha

its made for character consistency

my uncle is trying to join me in stable diffusion training, but he just got this error when starting a1111, any info?

[green plover]

hmmm. Did he try closing and restarting?

[smoky oak]

he did

[green plover]

hmmm no clue

[pallid relic]

add fastapi==0.90.0 to your requirements.txt and run pip install -r requirements.txt

[smoky oak]

Thank you, i already had it fixed 😅

same solution, much appreciation

I didn't know we had a support channel when I posted that

[smoky oak]

please for the love of god can somebody help me get rid of the excessive amount of veins protogen gives muscular men, omg

[smoky oak]

V
E
R
Y

T
A
L
L

hehehehe

[smoky oak]

bozo looken ass

[earnest solar]

Luna the Gentle, Elven Goddess of the Moon

Mera, Goddess of Dark Magic

[hollow nacelle]

magnificent, what did you use

@tropic shell i have here a simple example with SD 1.4

the prompt is something like "a yellow hawk in a flowery tree"

[tropic shell]

nods

[hollow nacelle]

same seed, similar prompt, new technique

all i wanted to show is that, in the second image, the flowers are not yellow

[tropic shell]

(I love hawks ngl)

[hollow nacelle]

and the yellow kinda sticks to the bird

instead of applying to everything

[tropic shell]

Mahahahahaha, yeah, that's the kind of bs I do

[hollow nacelle]

so, that's the kind of thing you mean when you say talk of relationships, right?

that's a dreambooth model?

ooops I was replying to a deleted message from the guy above

[haughty crane]

dreamshaper_33BakedVae

[hollow nacelle]

.

[haughty crane]

no?

I dont know about dreambooth

[hollow nacelle]

so, what do you think of this yellow bird technique? do you see this limitation of yellow to the bird as useful?

[hollow nacelle]

oh, so all of that effect is a VAE and prompting? I'm kinda new to these things.

[hollow nacelle]

also, you said you do this kind of thing. So you can do that as well?

[tropic shell]

I like it 🙂

[earnest solar]

Tappaja the Realmbreaker.

[tropic shell]

Love the lighting

So, what was your method for changing it?

[hollow nacelle]

cool! I was looking to know whether that's an issue ppl face in practice. So, apparently, yes. Do you also have a solution for that, how would you limit the color to the bird in this case?

[hollow nacelle]

I made an embed just for the bird I wanted

[tropic shell]

Makes sense

And I would prompt my way around it first

👍

[hollow nacelle]

yeah, prompting is obviously a better solution from a certain angle xD but I didn't know it was possible

how would you prompt the yellow out of the flowers without also changing the bird?

[tropic shell]

This is just with a little bit of iteration and higher cfg

To start

[hollow nacelle]

hmmm but then the bird loses the yellow color

[tropic shell]

Doesn't have to. I specifically picked this one to be more brown

[hollow nacelle]

if it's not asking too much, would you be able to create a yellow one next to white flowers without the colors mixing? like, if I type this prompt: "a yellow hawk next to white flowers", I get it all jumbled up

this is the same thing with my strategy. Is this possible through prompting alone?

[tropic shell]

I mean

(Those were just samples abov

e)

If you were prompting a hawk, you should give SD more context

Is the bird flying? Is it on the ground? Is it facing the viewer?

[earnest solar]

Bird up

[tropic shell]

Obvi, giving SD more info about hawks will help

But you also should prompt it

[earnest solar]

The bot isn't that great at compositing. It's sometimes better to not try and force it and leave it in the bot's hands

[hollow nacelle]

thank you for the samples and the orientation!

I'd say I'm satisfied with the help

🙂

[tropic shell]

I would reccomend...

For white flowers...

[hollow nacelle]

yes, listening

[tropic shell]

https://cdn.discordapp.com/attachments/1029490795180466186/1074328089720537149/yellow_tail_hawk_sitting_on_a_tree_stump_in_a_field_of_white_flowers_professional_animal_photography_natural_lighting_professi_width-1280_height-704_number-4_seed-3683005337ts-1676210158_idx-0.png

https://cdn.discordapp.com/attachments/1029490795180466186/1074328090794278993/yellow_tail_hawk_sitting_on_a_tree_stump_in_a_field_of_white_flowers_professional_animal_photography_natural_lighting_professi_width-1280_height-704_number-4_seed-3089878283ts-1676210176_idx-0.png

https://cdn.discordapp.com/attachments/1029490795180466186/1074328259094921222/yellow_tail_hawk_sitting_on_a_tree_stump_saturated_yellow_chest_in_a_field_of_white_flowers_professional_animal_photography_natural_lig_width-1280_height-704_seed-3089878283ts-1676210216_idx-0.png

yellow tail hawk sitting on a tree stump, saturated yellow chest, in a field of white flowers, professional animal photography, natural lighting, professional nature photography, mid summer

or

yellow tail hawk sitting on a tree stump, saturated yellow chest, in a field of white flowers, professional animal photography, natural lighting, professional nature photography, mid summer

And these I just did on my PC, euler a

Could add portrait, long shot, mid shot, whatever

[hollow nacelle]

Ty! Really elucidative to know how far you can go with prompting.

[tropic shell]

Steps: 30, Sampler: Euler a, CFG scale: 7 Model: v2-1_512-nonema-pruned

Steps: 30, Sampler: Euler a, CFG scale: 7, ```

You can mess around with a lot

but that's enough from me.

[hollow nacelle]

all right, thanks a lot for taking the time and attention!

[rain gazelle]

no post processing

[tropic shell]

Np!

[hardy anvil]

Help : Remove that bright sunlight.

dreamlikeart tree in a bottle,hyperrealistic, photorealism,detailed, sharp focus, studio lighting, 4k, canon,f/16, rendered in octane,High Contrast,portrait by Richard Anderson
Negative prompt: bad colors,ugly, deformed, duplicate,
Steps: 30, Sampler: DPM++ SDE Karras, CFG scale: 8.5, Seed: 514953771, Size: 768x768, Model hash: c86a1a99b8, Model: dreamlike-diffusion-1.0

[pine stratus]

dreamlikeart tree in a bottle,hyperrealistic, photorealism,detailed, sharp focus, studio lighting, 4k, canon,f/16, rendered in octane,High Contrast,portrait by Richard Anderson
Negative prompt: bad colors,ugly, deformed, duplicate,

dreamlikeart tree in a bottle,hyperrealistic, photorealism,detailed, sharp focus, studio lighting, 4k, canon,f/16, rendered in octane,High Contrast

[waxen bramble]

Textual Inversion preview pictures from the very early steps are in my opinion the best fucking thing about playing around with SD. Granted these are from early epocs but I'd like you lot to point out to me where you can find mechanics' overalls in these pictures:

Granted... The top left is 10th epoch 😄

But... Still. I find it funny as fuck.

10, 20, 30, 40 epoch from 1000 to go.

They start to get accurate at around 500 mark by my experience. But still 😄 Usually it is with in the same realm as what I want from the start.

This is supposed to be " Man wearing [blue mechanics overalls] on his knees under an engine".

Ignore the Clyde filter.

[cunning geode]

from my currently undertrained model

[sterile kiln]

me (LoRA):

[wanton grove]

@static tusk talking about something like this
as you can see there it just used th e image as a vague reference for other stuff

While when I used it on SD I just got the same picture but distorted

slightly, up till 1 where "fantasy wizard" gave me an old man

[static tusk]

ok so that is different. SD cant do that

at least not that I know of

[wanton grove]

:(

[static tusk]

but as I mentioned you can always add more detail to your prompt for what you want

[cloud bough]

prompt

[fiery lichen]

The mysterious cylindrical object shot down over Yukon. Pilots cant communicate with it, and don't know its propulsion system. By the way, anyone knows how to fix faces of figures far away or obscured? I can never get a straight image of a full figure done well on stable diffusion. Not with in-painting, not with face-fix. Some people in the past suggest I generate a full size face and photoshop that into the small figures, but I want the A.I. to figure that out itself. Midjourney can handle multiple faces far and near.

lots of negative prompts for bad eyes, face, cross eyes, bad anatomy etc....used.

This was done on Art&Eros I think. nsfw emphasized in negative prompt.

The "balloon" head is a bit pointed, instead of the usual round pocket that you are supposed to pinch the air out of before putting it on. But all the other "balloons" end up showing a mangled woman/parts inside, completely unusable.

[dusk solstice]

#1072229057707659404

[mystic crane]

#🏞｜general-with-images A BABY , CARTOON , PROFILE PICTURE

[green plover]

pick one

[mystic crane]

#🏞｜general-with-images A CARTOON BOY FOR YOUTUBE LOGO

[sharp mural]

Tu

[green plover]

Looks like that one is getting better results

[austere sparrow]

professional fashion shoot of a handsome Thai man lying on the beach in a very thin and white bikini that can be seen through after gets wet from swimming in the sea. perfect body,muscles,tanned and glow skin. He smiles and shows the ladies his beautiful tattoos on his thighs and below his navel

[green plover]

Clyde would reject it

[green plover]

De Vinci x Rembrandt because why not?

[white quail]

https://giphy.com/gifs/beastieboys-beastie-boys-three-mcs-and-one-dj-SUFFHpZ5X4FZeK0HL0

[hallow pumice]

black panther with spots

#🏞｜general-with-images black panther with spots

[fair jacinth]

#🏞｜general-with-images barrack obama on mars wearing a pink dress with a dog and a hotdog under an umbrella

[spice sorrel]

Question: How does one perform a style transfer with Stable Diffusion? I'm trying to turn my SD image into an anime from the OrangeMix model.

This is what I wanted to transfer into:

[wispy nest]

body blast avatar

[wary grove]

https://giphy.com/gifs/dog-action-3o7WIoLgcM2JcrKmaY

[opaque torrent]

how do you train a model ?

[cinder trail]

Im lerning SD and did some warrior womens. A mix of Emma stone, Kate winslet and Megan Fox 🙂 But how do you fix bad hands?? Is there any good settings or some magic thing you can do after you have done the image, like in the inpaint mode. Haven´t mixed around so much in there

[spice sorrel]

#🤝｜tech-support

[cinder trail]

Alright 🙂

[steep zodiac]

Some unknown group seems to be currently running a campaign to encourage terrorist attacks on AI researchers right now:

https://pbs.twimg.com/media/FodIL4rXsAAfGjL?format=png&name=900x900

https://twitter.com/ESYudkowsky/status/1624083346279260161

https://cdn.discordapp.com/attachments/1006139459860975716/1073697541809504371/Screen_Shot_2023-02-10_at_3.08.39_PM.png

Eliezer Yudkowsky (@ESYudkowsky)

Please note: There seems to be a campaign to FAKE the story that AI alignment theorists advocate violence. Everyone remember: WE never say this, it is THEM who find it so useful to claim we do - who fill the air with talk of violence, for their own political benefit.

Likes

234

Twitter · Feb 10, 2023, 4:30 PM

[dense tapir]

WTF?!!?

[stoic pivot]

Some are afraid of "artificial intelligence" after scientist could not identify real intelligence here on Earth after 1985 and when it was to hard to find intelligent life in space.

[dense tapir]

You are over thinking it because the mob hasn't gone that deep. They are afraid and everything else is too deep for them, and irrelevant. Emotions over logic.

[stoic pivot]

Thank you for thinking that I was thinking at all, normally I say things without thinking.

[dense tapir]

The you may drift on back asleep now, lol.

Fact is the net, overall, has turned into a bad thing. The information superhighway that was supposed to bring us together seems to only travel in one direction and has done everything BUT bring us together.

When you get stupid people together in a large mass stupid shit happens.

What happened was this AI just was thrown out there too fast and mankind was not prepared for it. Slow and steady to acclimate the frog to then openly embrace it to openly fight for it. Go too fast and we get what we are getting now.

[stoic pivot]

Well, situation can be disarmed by explaining that 96.5 % of all Ai is just for make furry porn and not conquer the world.

[dense tapir]

Mob already has pitchforks, tar, feathers, and the rope with lighted torches.

We will get past this, I think, but the alternate way of rolling AI out would have been the best at the socio level now the economic level is why it was thrown out there. Bad

I know I don't want to end up in a Mad Max like world because the powers that be wanted to get richer fast and tossed this out there for the mobs to throw us into a new dark age.

What I worry about is the mass chaos, deaths, and the old farts in robes and politicians passing nonsense due to all of it.

Sadly fear is a stronger driving force than logic.

[stoic pivot]

A long time ago when there was a website called WebCrawler there was a group of people that wanted to ban it, it could find bad things on the Internet and according to them that would make us all into sexual monsters, later that become Internet search and now controlled by Google.

[dense tapir]

I trained on one image for 15m and it now knows that style. I find that damn impressive while others would fear for their jobs etc...

[stoic pivot]

All new things are scary for some that like to scream and protest.

[dense tapir]

Lycos, Webcrawler, Altavista those were my search engines.

[stoic pivot]

ditto

AskJeeves was first try to make something that looked like AI today where you asked questions.

[dense tapir]

See, here is the issue in a nutshell. People need money. People equate their self worth based upon it. If you could eliminate the need for money the fear would go I think. Do what you love because you love it not because it is what brings food to your table.

Yes

We will get there and AI will get us there faster kicking and/or screwming you can't stop the inevitable.

[stoic pivot]

Maybe that site will do a comeback with AI... idea, if some one take my idea and get rich, at least buy me a pizza.

[dense tapir]

I think askjeeves (then just jeeves) is still out there too.

That style SD picked up with 1 image and 15m

I am 100% shocked

I am used to 4-8 images but this was 1 image.

[subtle oxide]

yo quick question, what sort of artstyle/genre is this?

[dense tapir]

I like it. Futurism, retro. hmmm

[lilac pilot]

Are you creating a comic book?

[dense tapir]

Looks like it, right?

I think it trained so very well

[lilac pilot]

Yeah pretty cool. Looks like morgan freedman as well haha

[lilac pilot]

This tool is incredible and it’s random every time

[spice sorrel]

Question: How does one perform a style transfer with Stable Diffusion? I'm trying to turn my SD image into an anime from the OrangeMix model.

This is what I wanted to transfer into:

This is originally made in Protogen Infinity

[dense tapir]

I made the car red and you can control some of the randomness but being it is just noise based, and noise is just randomness yeah

[spice sorrel]

And it was inpainted.

[lilac pilot]

@dense tapir keep them coming!

[rancid juniper]

it shows me this!

how to know if it is 1.5 or 2.1

?

[pastel mango]

upper left is named model, for some reason.

[rancid juniper]

afaik, i remember, downloading some model from hugging face

but that was few months back

[pastel mango]

I'm not sure, cause I always check which one I run. Also 2.1 mostly gives errors for me so I know when I've tried to mount one

check it before I DL that is

[rancid juniper]

bro prompt? let me guess,

A huge robot in water next to a city?

😅

[pastel mango]

Ah you wish to fight?!

she has no sense of proportion, but she wanted to make a war scythe

[blissful socket]

Cat girl

[pastel mango]

a brave knight has arrived to protecc

[wispy nest]

That's almost what I actually wanted to make when I was thinking about developing a game lol.
Girl with a heavy sword bigger then a girl itself and she was supposed to swing it awkwardly...

I probably should make some AI attempts too.

[vague oasis]

its probably sd 1.5

sd 2.1 gives distorted results if the resolution is below 768x768

[pastel mango]

make her chibi and she swings it slowly or starts to spin with it then gets dizzy and rests against the sword. When given the option to take one of the most powerful swords in the universe, suited to her size. She declines it, laughing "how can something so small be so powerful" - walking away after defeating a dragon god, being the size of just one of its scales.

oblivious to her own size

Even the Dystopian future needs K-Pop

[wispy nest]

Yea , swinging was one of the moves I had in my head...

It's been some time since I was thinking about it tho...
Sword had a history too, that's the only thing she has from her dead father.

I don't think I'll actually develop it anyway, I have job to do and well , now I actually realise how much time it takes to develop something.

@pastel mango what model is it? 1.5 default?

[rancid juniper]

Sometimes it takes way too long, why?\

[pastel mango]

slow gpu?

[pastel mango]

Huh? the chibi? the dystopian?

[wispy nest]

I never had "in que" message, maybe you have unfinished generation in your console?

[pastel mango]

https://civitai.com/models/6209/dark-gemini-v21 either 2.0 or 2.1, I forget which one made which

that's my model

I mean it runs on 1.5

but the model version itself is 2.0 and 2.1

[rancid juniper]

should i stop the google colab and run it again?

[wispy nest]

that might help, yea

[pastel mango]

Colabs can be a pain in the ass

[rancid juniper]

experienced that when i was trying deforum

[pastel mango]

I run locally on everything that will

this future is so awesome their entire bodies are switched to mechanical gears

[pastel mango]

It's Dark Gemini 2.1 for the chibi, both can do chibis...both can do that dystopian hellhole too with minor aesthetical differences

[wispy nest]

got it, downloading it and it's vae to check out what it can do

[pastel mango]

some of the preview images have prompts, but its important you read the intro cause it'll explain the models quirks

[rancid juniper]

Whats your PC specs?

[pastel mango]

have a chibi

[pastel mango]

just 3070 ti 8gb

I say just because for vram usage in SD it's not a lot

for gaming though I'm more than satisfied

[rancid juniper]

and intel i7 ?

[pastel mango]

youll find it can make beautiful gurls too, demi-human, human, goth

[pastel mango]

amd 5600x

[wispy nest]

I wanted to do magic + realistic style...and it's...pretty hard to achieve 😦

Can do shining staffs , auras , particles , but fire...off, that's an actual pain

[pastel mango]

I've managed to blend a lot of it together

in the model (merged from a lot, using merge block weighted

so I had more control of what went into where (or at least sofar as its currently understood)

Clip Skip 2, () prompts with numbers, ETA Delta at 0 are both important to note, otherwise it might not render accurately

there's so many settings that can change it now

[wispy nest]

oh it's clip skip 2 model?

[pastel mango]

not necessarily

but I used it when I tested it

I don't know what no clip skip would result in

[wispy nest]

I'll keep it default then , I usually chain more words

[pastel mango]

it's fine with that, though 2.0 handles very long strings better in accuracy I think

but the aesthetic is a little different

you can compare on the page

[wispy nest]

yea , makes sense

[pastel mango]

I love making these though

[pastel mango]

I can do spells, more active ones too. Magic circles if you like

[wispy nest]

yea , that's not bad

interesting...

let's see if it knows what is buster sword

[pastel mango]

It can typically hold weapons by a handle, rather than on the blade

if it doesn't, just type in "oversized bastard sword"

[wispy nest]

yea , on other images she holds it better, but this one is more cute 😄

[pastel mango]

I was toying around with weapons a few days ago

those are just examples :P

[wispy nest]

I was thinking about this sword , final fantasy style

[dense tapir]

Now that's a knife

[wispy nest]

well , kinda 😄

[pastel mango]

and magic

[pastel mango]

I would say within the realms of sci-fi dystopia, sci-fi horror, dark fantasy, gothic fantasy and whatever is inbetween, my model can handle it in western art, 2.5d and anime or a hybridization

[wispy nest]

yea , it's pretty good at magic

[pastel mango]

I dunno if I have a buster sword anywhere. I have a lot of combat. Actual contact combat too :P

but you can try to find a sword that is closest to it and then add a fantasy flare text to it

flavor*

[wispy nest]

I couldn't make it do it, so I'll guess it doesn't know what is it, maybe if I'll try to describe it instead as long , heavy , wide e.t.c or find real word alternative for it

[wispy nest]

yea

[pastel mango]

Fido says hi

try "giant guardless greatsword inspired by Final Fantasy 7"

or berserks dragonslayer

[stuck niche]

"The Strength of the Mountains"

[wispy nest]

I think just "oversized" kinda works, everything else too specific

[pastel mango]

yerr

if you want actual combat to occur you need a few key words like "fighting against" or "dueling" and then flavor tags like full contact combat, close contact fighting, slicing, slashing, chopping, blood gushing, lost limb, violence, death--depending on what you need

once you get the hang of it there's a lot you can do

[wispy nest]

I usually start simple , figure out how to make prompt , then make it harder if it works , if not - change words around until it does

[pastel mango]

"infusion" is kind of like a coat of extra paint on top of art style

just fyi

"aestheticword infusion"

[pastel mango]

who lives in this hellish nightmare?

[cyan snow]

loab

[pastel mango]

damn

[wispy nest]

lol

[pastel mango]

well I gotta go now, health aint too good and need some relaxing. You can DM me if you have any questions. I have to accept before we can chat though so make the first message a good one :P

then tomorrow...well tomorrow is uncertain

[wispy nest]

aight , cya

[cyan snow]

wtf is this? i did negetive prompt- humans

dude negetive prompts are awesome

[lilac pilot]

@cyan snow molecular vision

[pastel mango]

It's a microscopic world

but now I go to bed for real, just had one last stop

[lilac pilot]

What prompt did you use?

[cyan snow]

i just messed around with more negetive prompts

[lilac pilot]

You only used negative prompts?

[cyan snow]

yep, i figured if you put normal things in negative prompts it makes a horrible mess, then if you put nonsense in negative prompts it does dope stuff

[lilac pilot]

Awesome

[lilac pilot]

Jack of all trades…

[wispy nest]

good evening gents, I will shortly be releasing a 90s Buffy Lora

[dense tapir]

That Sarah at age 12?

[wispy nest]

no its sarah trained from the year 1997 - 1999

it's the model that I have tested it on that is giving it a cartoon appearance

( ages 20 - 23 )

[cyan snow]

just did a batch of only negative prompts

[dense tapir]

If you say so. Looks far younger than I remember from back then. I wonder why it made her so young?

[wispy nest]

that is down to the model called dreamshaper, I hand picked and trained this lora over 3 days

from the show season 1

input prompt " woman, gothic, blonde, lora:SarahMichelleGellar:1 "

[cyan snow]

ok im done with this prompt

[dense tapir]

Dreamshaper?

[wispy nest]

yes

[dense tapir]

I watched that show with my late wife each week and she was young but that made her really young.

[wispy nest]

one of the most popular models on civitai

bro are you asking to see part of my training data?

cos idk what you are trying to imply

but if it is what I think it is

[lilac pilot]

@wispy nest how come 3 days? Did you push it longer to make it better?

[dense tapir]

Oh, I never do civitai after I logged in and my page was full of sucking dicks and other porn. Thank goodness no one spotted it.

[wispy nest]

beacuse it's very low quality source images I had to AI upsacle the show

just to clear something up with you @dense tapir

example training

[dense tapir]

Now that is Geller.

[wispy nest]

one image from many many, hand written prompts and 3 days refining

I will try some more photorealistic mixes just because I dont want anyone to get the wrong impression

I am not one of those.

[dense tapir]

no no, not what I meant but I could see someone going there. I was just surprised at how young it made her.

Not that she would object to that

[wispy nest]

yeah this model seems to greatly simplify the features

making her look around 18 but the training data isnt the cause

[dense tapir]

Yes, the plastic look

1.5 based?

[wispy nest]

ye

[lilac pilot]

Do people monetize their loRa?

[dense tapir]

Yep, I could tell as 1.5 loved doing that

[wispy nest]

I take paid commissions for custom and private ones

otherwise I release them for free and just ask if some one can donate me a coffee all good ❤️

[lilac pilot]

Just wondering that’s awesome, but can’t anyone train and do this same work?

I’m trying to wrap my head about monetizing when everyone can make it now?

[wispy nest]

if they managed to ai upscale the show and used the exact same frames that I did with the exact same settings then possibly

but its extremely unlikely

they would also need hundreds of hours of learning to even attempt it and the correct hardware

[lilac pilot]

I see thank you for sharing

[wispy nest]

no images are stored in my Lora data, it's an algorithm same with SD

also the creation of dataset, how many images, how many steps, epochs

[lilac pilot]

Someone can trace your images from the learning data?

[wispy nest]

its all very complicated and involved

[dense tapir]

Anyone can learn it, and attempt it just like you can for anything then there comes a point you ask yourself which is better buy it or spend all that time doing it yourself and hoping.

[lilac pilot]

Like people doing erotic images etc

[dense tapir]

time is money

[wispy nest]

if I cared about that I'd already be charging for my work

[dense tapir]

yep

[lilac pilot]

And some aspect take knowledge in coding , correct?

[wispy nest]

my 2.1 and scifi models produces art that is of very high quality

and i give it away for free

[dense tapir]

Thing is I need new hardware and can't afford it then I ask myself why do I want it since this doesn't pay.

[wispy nest]

because I love art

[wispy nest]

a background in computers especially, coding of python - sometimes

[dense tapir]

Which is why I use colab as I really like training and seeing what happens with my training. You never really know.

[lilac pilot]

It is fun and that the ball is in the court of someone who love art and allowing them to create their own without being to hands on if that makes sense haha

[dense tapir]

I can't even draw a stick figure, well, kinda sorta, so doing this is great but my mind is more analytical having worked in computers for decades. Programming comes easy to me too so seeing what I can create that way (via training) gives me the most enjoyment.

[wispy nest]

thats cool

also if you know applied math / algebra it helps a lot

[dense tapir]

yeah, though I am rusty in that. Now for all this you need even higher maths and a firm understanding of the tensor stuff. That I do not know.

[lilac pilot]

I can imagine the developers are raking in rewards from this

[dense tapir]

They are but in other areas

[lilac pilot]

10m on midj at 30-60 month if you get 5% of consumer buying monthly jeez

[wispy nest]

midjourney is not the same as stable diffusion, though a good product, it is in my eyes quite expensive

[wispy nest]

and also the style is not hard to recreate using SD 1.5 or 2.1

[lilac pilot]

Sd is superior

[wispy nest]

again it depends on use case

if you just want to sign up to that and use it and it works for you, then you will love it

if you want to tinker, create training etc

[wispy nest]

you'll love SD 1.5 and 2x

[lilac pilot]

I’m going to do a lot of it on my day off. Work is making me not able to use it lol

[dense tapir]

Nvidia killed my dream with all this

[lilac pilot]

Really appreciate the discussion. A lot of other forums or discords are toxic and exclusively off putting for new people since it’s new and everyone is trying to rake in

[dense tapir]

sad, I know

[lilac pilot]

@wispy nest have you tried full body and adjusting scenes with your LoRa

[wispy nest]

not yet as its mostly trained on wait upwards to avoid creating more issues with hands, however, for the use case it was designed, I think it's going to look great, just trying a few different checkpoints with the LORA for cover images

[earnest solar]

I'm not creative at all, there's not a single inventive cell in my body, to the point where I had to get a lot of help even with prompts.
To be able to finally put some of my soul out in the world, this is something I'll always be grateful for.

[dense tapir]

btw, something someone said on Reddit stuck with me and is a problem. A lot of the lora stuff up on civitai is cherry picked and people can't even get nearly what they see. Being a emb/lora creator myself that makes it hard for me. Yes, I cherry pick BUT it isn't a one in a million and the rest were junk as is the case with a lot of those Lora up there. Mine is I want a best representation and here is the prompt I used so you too can go for it.

going to end up giving a bad name to all of this due to those imbecile lora "creators"

[wispy nest]

very good point, sometimes the art debate forget they are gatekeeping, say for example, those with physical disability or other ilness, those who just want to see an idea in their head on a screen

[wispy nest]

I only use my favorite 2 then put a few more examples that arent cherry picked for transparency cos i think thats fair

[earnest solar]

It's why I can not forgive people who are intentionally abusing AI to cause strife.

We are this close to Utopia

[dense tapir]

There is a Star Trek TNG episode where Wesley picks up a musical instrument and it plays based on your thoughts. I would kill for that, as many would, and I see this as an example of that. No longer can the few rule over the majority and make themselves feel superior because they were handed a gift. Now we can all create.

[lilac pilot]

Can’t see if getting banned, because it’s out there everywhere now. Restrictions may be implemented due to those people

[wispy nest]

@dense tapir here you go, more photorealistic gen using different model data

[dense tapir]

Bet age wise but losing Sarah as I would never know that was her unless I was told.

[wispy nest]

to test if it worked, I posted without saying in another forum

every single one said "is that buffy"

so idk

either way I dont think any better training can be done any time soon

source material is so old

it was a NIGHTMARE :3

[dense tapir]

the hair, and style. Now take just its face and nothing else and ask

[wispy nest]

the one on the right is a real picture

[dense tapir]

I'll take the one on the right

[wispy nest]

ye cos i just cropped what u saw

[earnest solar]

Genetic modification could solve world hunger, deforestation, and desertification - maybe even terraforming.
Robotics could make work a hundred thousand times easier, kill child labour, and make construction a thousand times easier.
With cloning, no animal should ever need to go extinct again.
And AI could be so, so much to people like me.

But we're evil, vicious beasts. So corrupt that the process of science itself can no longer be trusted simply because it's manned by humans.
I hate us, sometimes.

[dense tapir]

No, because I know it is her. Try that with 2.1 and see, it may be better and less plastic like the one on the left.

[wispy nest]

sorry it's not up to your standard mate

I worked really hard on it

maybe when i do version 2,

otherwise I'd like to see what Lora youre making cos they must be dope

[dense tapir]

Please, take nothing I say as a slight to you it is of 1.x mostly

[wispy nest]

I am migrating to 2.1 soon! Having speed issues atm and lack of funding to move over to renting gpu time

[dense tapir]

Aye. I dropped 1.5 the day 2.0 hit and never have picked it back up because I do photo realism and I found the clip in 2.0 could count (1.5 had no idea) and the skin was WOW.

[wispy nest]

oh yeah if we are talking about 2.1 vs 1.5 here, I'm in camp 2.1

this is only my second public Lora

[dense tapir]

for art it took a step backwards with the forced loss of artists

[earnest solar]

Have Mofu and Mofu as an apology for the rant.

[wispy nest]

rant was fine buddy no need to say sorry

everyone entitle to opinion

There are few 1.5 models which can do photorealism better than 2.0 \ 2.1

[dense tapir]

I happen to agree with your opinion but I am a life long misanthrope anyway

[lofty condor]

Hey guys, new to this, so just from the Convo, for Photorealism 2.1 is better?

[wispy nest]

generally speaking, yes

[dense tapir]

100

[lofty condor]

All the vids I watch on YouTube refers training to 1.5 that's why

[wispy nest]

Let's say...
Without custom models - pure 2+ is better at photos than pure 1.5.

2.1 hasnt been adopted en masse because it had some clipped data a bit overdone and left some gaps but it just need more user training for it

and it wil be fixed in no time

[lofty condor]

So I need to search up some 2.1 stuff. Any constraints like GPU regarding 2.1? I'm using a 3070

[earnest solar]

You're fine

[wispy nest]

vram is the main contraint but if you have 8gb+ youre fine, even 6gb runs just to make gen

[earnest solar]

I'm fine on a 2060

a broken 2060

[lofty condor]

That's good to hear

[earnest solar]

I made mofu and mofu on it

[lofty condor]

Still figuring stuff out, so far it won't train my face, I'm probably just doing something wrong

[wispy nest]

wait what...you can gen on a broken GPU?
How broken are we talking about?

[dense tapir]

Mainly because they had to remove the easy Greg Rutkowski button and all the nakedness was removed.

[earnest solar]

An attempt at a joke.

[lilac pilot]

Are there settings for beefy gpu to use?

[dense tapir]

My 1060 cries

[wispy nest]

ye its a long story :3

[dense tapir]

cries even generating an image

[wispy nest]

but i do not hate on 2.1 at all

I made this on 2.1

with my own checkpoint mix

[wispy nest]

1050ti talking here, it's better to stick to google collab and kaggle.

[dense tapir]

Yep, but 2-2.5h max per day and so limited in drive space.

[wispy nest]

3-6h per day on collab based on my expirience, depending god knows from what, but 30h per week on kaggle ain't bad...
If even that isn't enough - paperspace is an option too, with the only limit - 6h per session, but it's slow too, probably same speed as 1060

[dense tapir]

I was saving up for a 4060 at the released 379 MSRP 16 GB vram leak then saw 4080, then 4070 and forget it now

[earnest solar]

are you sharing that mix? I'd love to try it.

[wispy nest]

https://civitai.com/models/7615/czeros-sd-21-space-fi-mix-fixed-v2

[dense tapir]

max for me is 2.5h if I am lucky

[earnest solar]

I'm just saving up to see what's available.

[wispy nest]

example of it making a female photoreal style

[dense tapir]

4060 is trash. 8gb and 600 USD

[wispy nest]

was only my first attempt in 2.1 but yea

[earnest solar]

AMD's are probably better then

[wispy nest]

no its about vram

[wispy nest]

protogen model

[dense tapir]

Not for training, no

[earnest solar]

huh

so if I were to buy this instant

[wispy nest]

been using protogen to test my lora, its not bad

[earnest solar]

what? a 3060ti?

[dense tapir]

I know this anyone thinking of buying a card do NOT get less than 16GB from what I hear in the wind coming. I know I am not.

[wispy nest]

i use 3060ti 8gb with launch params for lower vram

[lilac pilot]

You know how you can overlock gpu- is there settings for/within sd for those with 3090+ models?

[wispy nest]

but its a struggle to train with it

[wispy nest]

yea...the only downside I bump into - I get similar faces pretty often.

[wispy nest]

realisticvisionv13 and my buffy Lora ( not yet public )

still tends to add a little bit of a younger look but ehh it's 1.5 the Lora works

and damn well considering the source materials are over 20 years old

also worth a note all i am writing to prompt is:
woman, gothic, blonde, lora:SarahMichelleGellar:1

[dense tapir]

Actually, I made a DB (a lot of them) for 1.5 of source material that was 100% rubbish from 1978. I figured it would be horrible but honestly it was drop dead sexy.

[wispy nest]

was it upscaled source?

lets see some ?

[dense tapir]

A lot, yeah

[wispy nest]

I love that type of stuff

[dense tapir]

let me grab the reddit link

[dense tapir]

btw, I ceased to release anything as the anti ai mobs hit me hard on my posts on reddit. even the older ones are almost 0 karma as they keep getting hit negatively.

[wispy nest]

): hate that when they go crazy for no good reason

at the moment I am wake up, doing AI, going to sleep, working just as hard as anyone else - sleeping 6hr rinse and repeat

[lilac pilot]

Yeah they can’t adapt

[dense tapir]

https://old.reddit.com/r/StableDiffusion/comments/yyc4l3/by_your_command/

[–]Dark_Alchemist[S] 2 points 2 months ago*
https://i.imgur.com/msktD5d.png <- Role play/Cosplay done in SD as well. https://i.imgur.com/DjKM0qs.png

The main issue I had with this is trying to find a quality source which was impossible so this was a miracle what SD/Dreambooth pulled off considering my source material was less than even DVD quality.

[wispy nest]

not bad! it is very difficult using stuff that low res!

[wispy nest]

oh yea...about younger look...
I'm using little trick to avoid kids and too yong look by adding "nsfw" + describing clothes manually - it never gives details below 18

[lilac pilot]

@dense tapir woah nice looks real

[dense tapir]

I was floored

[lilac pilot]

If you touch it up more just imagine

[dense tapir]

the images were rare and ugly

[wispy nest]

im used to using my own checkpoint data so its not the usual way i prompt ;;'D

[wispy nest]

oh, aight

[wispy nest]

i think this is looking pretty good for 1.5

[dense tapir]

https://i.imgur.com/nNewBqs.png https://i.imgur.com/SohbJMm.png https://i.imgur.com/4bw21SO.png https://i.imgur.com/tC6JgK4.png

[dense tapir]

I thought those were funny

[lilac pilot]

Blows my mind every time that a computer spit these out

[dense tapir]

btw, this was the hardest one I ever did. It fought me for weeks and weeks. https://i.redd.it/yweidzd9fi0a1.png

[lilac pilot]

Your an early adopter

[lilac pilot]

Mad I didn’t know about installing sd on windows 6months ago

[dense tapir]

The one I am really fond of that also used low quality had to be upscaled imageshttps://i.redd.it/govczypis6z91.png

[dense tapir]

Humphrey Bogart as the original Ghostbuster (Of all the djinn joints in all the worlds, you had to materialize in mine). https://i.redd.it/bfkvm1owagx91.png

[earnest solar]

Original mofu mofu

[dense tapir]

https://i.redd.it/o3sjdxraxkx91.png

1.5 was fun but all of these would be fantastic in 2.1

[lilac pilot]

Tried running them through 2.1

[dense tapir]

Not yet

the problem is DB for 2.1 is crap, or was until really recent

[earnest solar]

Technically a failure but I love her all the same

[lilac pilot]

Any aliens?

[dense tapir]

OH, you want Aliens?

[lilac pilot]

Did you do some?

[earnest solar]

The model I'm using is biased towards pale humanoids

[earnest solar]

It does elves and winged people, and little added bits

but not outright other species.

[lilac pilot]

Awesome!

[earnest solar]

When I asked it to do an ancient Sith lord it made a Devil.

[dense tapir]

Here, have a little Snoop Dogg

[earnest solar]

Like tell me these wouldn't be some bitchin' demon lords

[earnest solar]

Ah yes, the houses of parliament

See, when I ask for a sith lady it makes your typical anime villainess

But sith Lord?

[earnest solar]

♪O Fortuna♪

[dense tapir]

This one is AI

[earnest solar]

Bird up

[earnest solar]

I like how it kept the tattoos consistent somehow