🏞｜general-with-images

[dry crow]

good 😄

did you removed --medvram too?

[sterile kiln]

i will try again embedding

[sterile kiln]

no

should i?

[dry crow]

yea its not needed for 8gb cards

[sterile kiln]

Hmmm

[dry crow]

i have a GTX1080 8GB and can train embeds xD

[sterile kiln]

Without it, on SD 2.1, had crashes, unsuable

[dry crow]

only using xformers

[sterile kiln]

Yeah you're right, i added Xformers after

maybe it works like this yeah

[dry crow]

for sure

[sterile kiln]

Oh, just without --nohalf, atm no crashes

20/3000 steps, still no crash 😮

[dry crow]

nice

[sterile kiln]

dude

7200 MB used, 40/3000 steps

[dry crow]

looks good xD

[sterile kiln]

thank you so much man

[dry crow]

no problem 🙂

[sterile kiln]

months and months, trying to get it works

[dry crow]

months ^^ keep an eye at #🤝｜tech-support there you found help pretty quick 😄

here an image for this channel xD we follow the rules

[sterile kiln]

x)

@dry crow Is there a way to get it more faster?

[dry crow]

only if you go like 10 / 1 but the vram says no xD

also it depends on the training steps

the stuff with the 0.005 ...

[sterile kiln]

0.05:10, 0.02:20, 0.01:60, 0.005:200, 0.002:500, 0.001:3000, 0.0005

[dry crow]

yea, idk how to adjust them xD

sry

[sterile kiln]

np 🙂
I have to be patient (4 hours)

[strange jungle]

Man playing soccer. Why can't I get detailed faces in long shots?

Here is a real image of a man playing soccer. His face is clear. Why can'

can't I get this in SD: https://images.unsplash.com/photo-1517466787929-bc90951d0974?ixlib=rb-4.0.3&ixid=MnwxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8&auto=format&fit=crop&w=686&q=8

[sterile kiln]

Use Stable.Art plugin, to change details like face, its perfect !

[strange jungle]

is that amodel?

[strange jungle]

Oh, facenook pluging?

I'd have to do image by image...I'd like to do it in SD. However, I will try this

thanks

[sterile kiln]

https://www.youtube.com/watch?v=19ZNiw6sgBw

[strange jungle]

thanks

[sterile kiln]

this thing is completely amazing, I use it all the time

[strange jungle]

OK, I will try it

[sterile kiln]

I posted it above, each detail is done step by step, 1 by one

[analog goblet]

any of you around and willing to help me some with textual inversion / embeddings ?

[mystic marlin]

At first I was eh on the pink colored sclera for the eyes, but...it kinda works.

[waxen bramble]

Jaina Proudmoore?

[sterile kiln]

Yes x)

[sterile kiln]

😂

[eager reef]

I have been messing around with dreamlike diffusion model. It's pretty good, but I can't get landscapes as high quality as this. Any prompts that could fix this?

Example:

[sterile kiln]

guys, look how its amazing to add a shit google layer and inpaint him with Stable.Art... Awesome results, the original without retouching with the plugin :

[hybrid lava]

Added some animation functionality to imaginairy. This was using instruct pix2pix with "make it christmas" https://twitter.com/bryced8/status/1619516001967808513?s=20

[sudden wraith]

Superhero in orange pink costume

[rich flicker]

I just notice that it's quite difficult to let characters to look at other directions,
I tried to generate an image about a person reading book
but instead the person is holding a book while looking at the camera.
I tried prompts like read book, look at book, face at book in positive and in negative i tried face at camera, and look at viewer, sadly it didn't work

Anyone here have solution to this problem?

[grizzled sage]

might be the model you're using. i can't reproduce that behavior

2.1 is very consistent with what you want. simple prompt.
photograph portrait of a character, reading a book
neg: jpeg low quality, compression artifacts, low resolution

kind of consistent here too

go figure robots won't get it right. there are some limitations as you get more specific i suppose too.

[rich flicker]

hmmm I see so sd 1.5 and 1.4 are not able to make some specific commands

funny that it can't do something that is quite mundane and common.

[rich flicker]

could there be a reason why it can't follow such commands?

as it's quite strange that sd can't understand simple task like reading book.

[grizzled sage]

an embed may work. i don't play with 1.5 much these days, but its the same data it's trained on i'm rpetty sure. just with different filters and techniques to it. so it should know the concept still and that's where i think an embedding could help

[rich flicker]

oh embedding? any embedding you may suggest? I only know the bad prompt and bad image embeddings.

[rich flicker]

I guess sd do understand what reading books means, but the head seems to keep trying to look at the camera, I do wonder what makes it keep focusing on the camera despite I have command it not to in the negative prompt.

[grizzled sage]

I don't know of one no, but I bet it could be easily trained

20 good reading images maybe

These are more denoising targets defined with clip that it's trying to hit, not a command it'll follow. Ymmv

[rich flicker]

I see, i guess I better start training my own textual embedding

[grizzled sage]

thats the spirit

[strange jungle]

Cant get Photoshop to connect to server. Followed that video to install files etc.

Had to add --api in the ui launcher. It is in webui-user.bat, but not running from there

[strange jungle]

To do this, you selected the eye area, generated new images?

[strange jungle]

When the face is blurry, it generates, for example, blurry eyes

[sterile kiln]

What version do you have? YOu need recent version of Photoshop.
Often people use cracked versions of Photoshop, with a firewall rule to block any network connection with Photoshop. Obviously if this is the case this rule should be removed.
To do this, you selected the eye area, generated new images?
No, use Inpaint tool, results will depend of you prompts (& negative), models, and yeah, if you play with denoising, on what there is under you selection.

[strange jungle]

Yes, also erasing around the edges of the new layer helps

[strange jungle]

24.1.0

[sterile kiln]

Often people use cracked versions of Photoshop, with a firewall rule to block any network connection with Photoshop. Obviously if this is the case this rule should be removed.

[strange jungle]

This image is very clear compared to what I get with full body images.

[sterile kiln]

This is the original, with this kind of website, you can have EXIF data, with prompts, models, presets... etc : https://jimpl.com/

I used Protogen 2.2

[strange jungle]

Right, but it's not a full body image, action image.

I get very clear portraits, but in an action shot, the face is blurry

[sterile kiln]

On Photoshop, You can simply enlarge your working area (and therefore the size of the image), and then, with the Inpainting try to continue the body

[strange jungle]

vs this from the internet

https://www.sciencephoto.com/media/674805/view/men-playing-football

maybe it's the resolution...mine is 512x512. My card is weak, so it can't handle large sizes

[sterile kiln]

And in any case, the img2img allows you to take an existing image of a character who has the kind of posture you want, then with the appropriate prompt and settings, have the right posture with a new image

[strange jungle]

with a blurry face

that's what I get from that image of the two guys playing soccer. With that image, their faces have details. With the image I generate, it's like a minimalist painting

[sterile kiln]

man playing football, action shot Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 1124112417, Size: 512x512, Model hash: 67abd65708, Model: elldrethsLucidMix_v10, Denoising strength: 0.75, Mask blur: 4

Dude, its very minimalist prompt, steps, CFG scale, resolution, its normal

As I told you above, when you see beautiful images made by people, download them, upload them to a site (or other), to consult the EXIF ​​data and see the settings that people used to understand what which makes them so good.

[strange jungle]

I do...it shows in the PNG tab ...or is exif more info?

[sterile kiln]

Use the link above, upload a picture, see at bottom of the page

[strange jungle]

Ah I see...all the meta data

I will work with this now, thanks

No I got this result.

much better. He looks like a kid I tutor: 6'7", 320, when he changes a tire he doesn't use a jack. He lifts the car by hand.

[sterile kiln]

you should use another model, increase steps at 40, cfg scale at 10, develop background details, character details, increase resolution

[vital axle]

Man talking to a women in comic or manga

#Man talking to a women in comic or manga

#1014939219904450590 Man talking to a women in comic or manga

[waxen bramble]

Me feels like my TI needs some optimisation. 😄

No but really I'm trying poses for the first time.

Harder than I thought

Like few things I have noticed is hard to train TI for. Faces with parts of it covered or missing. And poses where body parts are missing or covered.

[sterile kiln]

3 Chars from War III

[novel ravine]

1st test result of my henry embed

[sterile kiln]

Sylvanas

[novel ravine]

dont lookup the prompt

[hardy robin]

what did you mean? I'm not good at English

[novel ravine]

haha nothing

[ripe cedar]

Are there many examples of sequential art with AI? I would have thought that with trained models and depth2img, consistency could be had to make a web comics but I haven't seen that to be the case for most part.

[tranquil kindle]

Ferrari

[hardy robin]

could i get banned for this?

I have a lot of content like that 🙂

[ripe cedar]

Most do these days...

[cunning geode]

from a 2 months old overtrained model of mine
a taste of whats to come in my new model

[slim scaffold]

/spoilers never gonne give you up

[cursive lance]

this is the best imagen

[cursive lance]

thx

[wispy nest]

this drive is trash, absolutely useless

[tranquil kindle]

I like the calculator in the background.

[pseudo aspen]

Okay...

That hair lol

[slender drum]

Hey everyone,
I'd like to ask for help, if anyone has advices regarding my current issues !
I've been training a character embedding and so far.. well, the results are not so bad

The issue is that I absolutely can't use this character in any scene describing an action, or with another character. It won't fight against a wolf, won't play the banjo with a monk, won't do anything. Any prompt I submit is a rather good portrait/pose of said character, but that's quite all.

Did anyone face this at some point ?

A few details :

Character embedding trained using TI.
12 sample images, full body poses/portraits from different angles
Initialization text : young girl black armor purple hair
N° of vectors per token : 6
Unchanged LR
Batch Size : 2
Gradient accumulation steps : 6
Current training steps : 500
Prompt Template : unchanged style_filewords.txt
Captioning using BLIP, removing anything that is the character, leaving only the scenery, background color, objects in the image, etc.
Model : Protogen Eclipse Advanced

[strange jungle]

can you compost it with other things?

Also, have you tried depth maps?

[slender drum]

No, I must say I'm quite ignorant about what they are ! I'll read about them 🙂

[broken oracle]

What does Hypernetwork Strength do? It seems pretty self-explanatory, but even at max it seems to have a negligible effect on the images being drawn.

[broken oracle]

@grizzled sage I changed the live preview settings to this. Now it no longer gives me an ETA.

[grizzled sage]

eta may show up after it gets a running avg. sometimes i've seen it take a bit for me too. you've got it set to update every 1sec.

[broken oracle]

Progressbar set to update every 1000 milliseconds

[grizzled sage]

1000ms is 1s. but every update it checks to see if it can figure an avg. When i run huge batch jobs, its often over a minute before it gives me an eta

[remote prairie]

Dat thic Lara.

[broken oracle]

@grizzled sage Suddenly started getting this error message with after several successful renders.

[faint fjord]

imran khan

[broken oracle]

nvm Flow. I think I figured it out. I was trying to use a Lora file as a model.

[tawdry wraith]

@celest mountain @strange jungle

[strange jungle]

the bottom image there's something wring with the hand on the right

It has an almost extra finger tucked behind the other hand

[tawdry wraith]

there's something wrong with a lot of things in SD gens, we just don't scrutinize as much

[strange jungle]

all three are images of only hands. It gets much worse much more often when a person is shown

I scrutinize

[grizzled sage]

yeah thats another nan issue dealing with half and full precision iirc. cool to see it was solved

[whole river]

my lora weight has no impact on the image at all...

[grizzled sage]

i've come across it while researching mine own. I think i'm dealing withh a psu issue because mine's about 5 years old now, annd has gone through a few "hottest year on record" heatwaves

[broken oracle]

are you using the right keywords?

[grizzled sage]

and it's a DELL 1000w psu so like, 5 years is good for it i think

[broken oracle]

Also how do you had Lora weight?

[whole river]

I used the filename as the keyword
Which parameter is the one I should use as prompt xd?

[broken oracle]

what tab is this under?

[whole river]

txt2img, under "Script"

I used the X/Y plot feature

I'm using the lora extension

[broken oracle]

lora extension is built in, but added net weight doesn't seem to be there

[whole river]

there's an additional extension

https://github.com/kohya-ss/sd-webui-additional-networks

this allows me to use lora not as a model, but as a "helper"

[cursive yacht]

Looks like someone forgot to turn on hires fix 😉 https://www.reddit.com/r/CrappyDesign/comments/10p9de5/i_think_we_found_jack_skellingtons_sister_or

[sinful kite]

https://giphy.com/gifs/reaction-please-chile-26DNiumgTwJi73VYY

[vague oasis]

@graceful escarp yo amjad above is sharing 🌽

[graceful escarp]

I hath been summoned.

[graceful escarp]

Dealt with. Thanks for letting me know.

[waxen bramble]

@strange jungle here is an example of a hand I used to train palm up: Yes I did it on photoshop. Took about 15 minutes to paint.

[strange jungle]

that's a mangled hand

[waxen bramble]

Well make a better one.

It was fucking enough to train the pose

And I had to make 5 variants

[strange jungle]

did it make a good hand in that pose?

[waxen bramble]

Yes... and 4 other images I don't have on my laptop

I made them all like that

[strange jungle]

well,,,then good enough

[waxen bramble]

So yeah... How about less art criticism. Since it needs to be good enough to fool the AI

It doesn't need to win a competition.

[strange jungle]

OK

[waxen bramble]

Also keep in mind, the fingers are slightly hooked up

Not spread sraight

Not visible in this but the 4 others it is clearer

Also... Learning to draw like this ain't hard... It takes at most 30 days practice like 20 minutes a day and giving a fuck. I am confident anyone can learn to draw at this level. And I am not that good anymore since haven't drawn for 4 years on the account of engineering studies eating my time.

And I used my own left hand for reference.

[static tusk]

together we can do it....

[wispy comet]

lol

[delicate oar]

¯_(ツ)_/¯

[cursive lance]

XD

they are chicken feet xd

[cunning geode]

i am very excited to test my 4000 image model soon
it includes a gigantic fuckton of images of people holding things or doing other things with their hands, all properly captioned
i am very curious if its going to improve hand generation or if it was a gigantic waste of time and effort lol

[hasty nova]

Oh that sounds super awesome and like a lot of work. Even if it doesn't work it could still be an interesting model!

Would love to see some images

[cunning geode]

me too, unfortunately 4000 images dont caption themselves

[hasty nova]

did you use something like BLIP as a base and go through and edit, or was it all by hand?

[cunning geode]

all by hand

[hasty nova]

dang

[cunning geode]

i dont use scrapers or blip or other such automatic tools, i prefer doing everything by hand for best quality and control over the end result. i hand search for images on various websites, hand download them, hand edit them (removing watermarks, cropping, etc), hand caption them

[hasty nova]

How long was the dataset in the making for? And how many steps/time did it take to train?

[cunning geode]

i think its been 2 weeks now
as i said before i havent trained it yet as i am still captioning it. i am like ifk 25% or so done with the captioning

[hasty nova]

Oh wow

Would love to see it when it's done

It would be interesting to do a smaller scale one as an embedding so people can use it with other models faster

[cunning geode]

its like 1500 photographs, 1000 legend of korra images, rest digital art and a few other specific artstyles like darkest dungeon or north korean animation
around 100 winged people images, 30 superhero costumes with capes, 100 hybrids, 30 or so giantess/growth images, 130 or so transformation images

and some other stuff

[hasty nova]

oo

Well I look forward to trying it out@

[cunning geode]

this is from a very old (2 months) and highly overtrained test model testing hybrids and such

cherrypicked tho

[hasty nova]

Inbteresting

[cunning geode]

i aimed to be finished with the captioning by wednesday but itll probably take until saturday and then the training will take like 3 and a half days

[hasty nova]

@dense haven here is an image I found explaining how it works pretty well, hopefully it answers some of your questions from #💬｜general-chat

[royal thicket]

QQQ: what Auto1111 img2img option do I use to fill in this image using the shapes as a comp guide?

[dry crow]

@strange jungle

[strange jungle]

Yes, I can get certain images with one or two good hands. But examine that image closely

Her right hand has a kind of double knuckle on the finger by the thumb, and there is no thumb.

[dry crow]

Yea its not perfect but not ugly

I have better images on pc

[strange jungle]

As I mentioned, even when folks post a pic of a good hand, pick an image to show good hands, there's a mangled hand at least half the time.

[dry crow]

You cant get 100% xD

Like with words on tshirts you cant even get 50%

[strange jungle]

The thing is, with 50%, it's more like 85%, of images with mangled hands, the chances of getting a good hand when other elements are good - those chances are slim

[dry crow]

I struggled a lot with hands before but merging + upscale fixed the most stuff

[strange jungle]

Upscale when generating?

the only upscale I do is post because my video card can't handle upscale when generating

[dry crow]

Yea upscale while generating changes the image, this cant be done afterwards with the normal upscale

[strange jungle]

right

Man holding bat. I removed "gloves" from negative. So it makes a guy in the field with a reversed left glove. Thumb on wrong side. I told it full angle, mentioned cleats/shoes. It gave me a portrait

#33 for the Rasgacs. He has a three fingered right hand, perfect for throwing a forkball. I guess he got hit in the eye, maybe with a come backer.

[dry crow]

Haha xD i see

[wispy nest]

Dream , capitã Marvel na praia usando maio azul

[dry crow]

I want to try this later

[wispy nest]

Cadê a imagem

Ué

[strange jungle]

here's the Martad's slugging 3B. He has two right thumbs gripping like a pincer. His left hand his flat, so he's trying to hide it.

At least these baseball players are coming out like regular people. All the soccer players I did looked like dwarves.

[dry crow]

Would suggest to add more negative tags, like Mutation, deformed, multiple limbs, etc

[strange jungle]

Cedric Smith from the England R's. He has a flipper hand, but he makes up with it by using the force to float the baseball in front of him.

I see what's happening here. SD couldn't handle the perspective of the two hands on the bat. It tried, so it merged some elements of the right hand on the bat with the left hand.

I actually found one soccer image SD must have had in training. It results in a warped face on many soccer action shots

One of the Awers best players. Despite a melted right hand and lengthened left fingers (he was the victim of a taffy pulling accident), he is known for fielding the balls he himself hits.

Every player looks something like Aaron Judge.

#41 for the Rochester Qadvins. He has three thumbs on his right hand - one shaped like a baseball. He paints that thumb the color of a baseball.

This is with a merged with protogen model, two finger/hand fix TI's, and a hand fix VAE

That's all we get before image #13, when the Doorknob Man begins appearing.

Im going to try it with just protogen to see if it makes the guy a dwarf. Maybe I can generate some Hobbit images - the dwarven band

not one of the baseball images is free of mangled hands

[dry crow]

Im back, Did you test it without the vae and 2 embeds ?

[strange jungle]

I did

That didn't work, which is why I tried them one by one, then in combinations

I will post some with protogen alone, none of the TI's or VAE

[dry crow]

Okay but then you need more negative tags

[strange jungle]

Tomorrow, or I guess today, I will run 100 of each thing, and record the % with mangled and unmangled hands

What negative tags?

When I do "bad hands" in negative, for example, it hides the hands

embeds them in sand, in a nearby object, makes a balloon in front of the hands, anything to hide them

[dry crow]

Try with this pls:

deformed, blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra_limb, ugly, poorly drawn hands, no color, weird colors, censored, deformed glasses, lowres, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, (((worst quality))), (((low quality))), (((normal quality))), jpeg artifacts, signature, watermark, username, low resolution, bad anatomy, siamese twins,

[strange jungle]

siamese twins?

[dry crow]

Against duplicate persons

[strange jungle]

why are you against that?

I'm siamese twins

[dry crow]

For real ? xD

[strange jungle]

no, I'm kidding

[dry crow]

Its just a fix that i dont get two heads above each other and stuff like that

[strange jungle]

I wonder if "conjoined twins" would also help

[dry crow]

Never tried, but maybe

[strange jungle]

my mother had a twin that didn't develop. Had to be removed

[wispy nest]

I can't make images in the bot, it doesn't appear in the text writing bar

[dry crow]

Bots are down: #1047610792226340935

[strange jungle]

the bots are gone

we dont know why. To make an image, give a try, to run diffusion on your system

[wispy nest]

Again

[strange jungle]

First image with protogen alone, all those negatives: six fingered left hand, right hand appears to have two pincer thumbs. This guy plays for the NU M's.

27 for the Sicily Manwhoks - he has good hands

#40 for the Norway Hanes. He has six fingers on the right hand.

He uses a giant golf tee as a bat

[dry crow]

But 5 fingers

[strange jungle]

No, 6

[dry crow]

Oh the little one

[strange jungle]

see the ungloved extra finger pincering with the gloved finger at the top of the hand?

[dry crow]

Nice one

[strange jungle]

So far, 2 out of 5 are face shots only

One way SD responds to negative tags about hands is to hide the hands

[dry crow]

You can use cowboyshot to get more body

[strange jungle]

#3 from the Bcagea - that's an asteroid that orbits the sun. He has three fingers on the right hand, five and a thumb on the left, with some merged fingers.

Its orbital frequency is 3,547 years, so road trips there tend to be very long.

[dry crow]

His shoes wtf

[strange jungle]

everyone is now looking like Bobby Bonilla

[dry crow]

Who is this xD

[strange jungle]

Now got four more face only images

He's a baseball player who, for three teams at least, led a cheating scheme

[dry crow]

Ah okay

[strange jungle]

One more image before I have to stop it

[dry crow]

Try to add Belt

[strange jungle]

Belt?

[dry crow]

Yea

Belt

Like i said yesterday with highheels tag + blue eyes you get mostly full body

[strange jungle]

is that a model?

[dry crow]

No just the tag

[strange jungle]

belt to the positive or negative?

[dry crow]

Positive

[strange jungle]

It better not make him look like Brandon belt

[dry crow]

Haha xD

[strange jungle]

I will send some of these images to some of the players

James Harden didn't think he could shoot well with six fingers on his left hand

43 of the Banies. He has a tiny right thumb. It's Jonathan Loáisiga

The Baekels are represented. The thumb is malformed - half finger, half thumb.

Ah, the Cleveland Fujaits. 5 fingers plus the thumb on the left hand.

[pseudo aspen]

Once they get over with the hands issue the AI will be nearly perfect

[strange jungle]

The Pinac Blakies. An extra knuckle on the right hand, three fingers and thumb all extra long on the left.

[strange jungle]

Hardly. It can't do interaction between people, show multiple people well - in perspective.

[pseudo aspen]

That's why I said "nearly"

[strange jungle]

For characters at least, once it gets over hands and feet, images of a single person standing will be good.

[pseudo aspen]

I can only agree

[strange jungle]

or sitting, I guess, maybe not action shots

Many of that last run were face only, despite mentioning things like full shot, wide angle, the cleats in the prompt

SD hides hands and feet when given prompts not to have bad hands and feet

Sometime I'll get a person standing...in a tall corn field, so you can only see the upper torso

Or standing...up to his chest in sand

or buried in the earth up to the chest

or in water up to the chest

I think I'll start a gofund to raise money for a hands bounty. Maybe we can get $10,000 as a bounty for whoever makes something that can produce normal hands and feet 100% of time, unless something different is specified

[cursive lance]

what genre would it be?

[analog goblet]

abstract art

[cursive lance]

o thx

[analog goblet]

i like it

[pseudo aspen]

Actually did you try Protogen model?

https://huggingface.co/darkstorm2150/Protogen_x3.4_Official_Release

[strange jungle]

Yes. It didn't change the hand results except more often hide the hands

[pseudo aspen]

Very clever...

[strange jungle]

I tried 3.4, 2.2, and 5.8

And with protogen, eventually it makes everyone have dwarf proportions, short legs

even if I switch models, it keeps happening until I close and re-start SD

Someone showed me some negative prompts that might help

[pseudo aspen]

Something like "short legs", "short arms" in negative prompt didn't help I guess

[strange jungle]

not once this starts

It seems the negative list the other person gave me is helping a little, but I am doing image of kids playing soccer. They all look like bodybuilders with short legs

the prompt includes "athletic pose, playing soccer, athlete" - it's taken athlete to be a muscle bound dwarf

so usually I put "beefcake" in the negatives to get rid of that. But once protogen starts this, it won't stop

[pseudo aspen]

Hmm

[strange jungle]

I just moved "beefcake" to the front of negaitves...thought I did before. Now we'll see

maybe musclebound, bodybuilder in negatives would help

I want regular looking people - people who play soccer aren't generally musclebound monsters

[pseudo aspen]

I understand lol

[strange jungle]

still musclebound monsters

[strange jungle]

Ah, perhaps the seed

[dry crow]

I need to try this later!

[deft dune]

A bunch of images from me creating my daily image entries. Still very new to prompting so hard for me to get good figures. https://www.youtube.com/watch?v=xUfScnTbxPk

[wispy nest]

pretty nice!

[wanton ore]

thanks! more here https://www.instagram.com/neofuturoai/

only SD of course

[wispy nest]

nice loads of creative stuff in there !

[carmine hollow]

People protest against bees

[wispy nest]

Lmao END BEES

thoughts?! Trying to add ~~more ~~cybernetics

shame the hands went crazy on this one

[wispy nest]

dope!

[hasty nova]

From an embedding I am currently making/testing

[hasty nova]

Same seed, different point in training, what is better?

[wispy nest]

I'd say they are both great images so it comes down more to objective opinion of the viewer, I guess whichever one looks like what you wanted

I always ask myself "which would I want to print out"

[hasty nova]

I see

[dry crow]

Good image but it seems you missing a vae

[hasty nova]

The embedding is probably better at portraits, like my old one, but can add more of a neon "style" to an image, like my old one adds a "ghost" style to images

[wispy nest]

vae should be built in but when i made the custom mix I did use a vae AND a safetensor with an embedded vae so sometimes the result gets a little messed up but not much, that is using an embedding

I could try renaming the vae to the same as my custom mix and see if the same input changes

also @dry crow do you know if it works to use two embeddings? that is what I was trying in that image

[dry crow]

Yea maybe anything or orange mix vae or wd1.4 would help

[dry crow]

Sure you can use multiple embeddings in one image

[carmine hollow]

Greta thunberg protesting against herself

[hasty nova]

you can combine embeddings, yes, but depending on their vectors/token, it might not work as well if that number is high due to using so many of your tokens

[carmine hollow]

Two "lizards" kissing

[wispy nest]

when using embedding can I prompt it to nudge only a certain amoiunt like (embedding:0.4)

[carmine hollow]

Beardie bodybuilder

[hasty nova]

I'm not talking about that, tokens are these things. The more you have, the less each token is represented. Embeddings use a certain amount of tokens defined in training, so having too many will make the image less representative of the prompt without you realizing I think.

[wispy nest]

OUHHHHHHHHHHHHH

going to try putting in the orangemix vae, just not sure I have enough vram since the vae is 800mb and my mix is 7.5gb

[dry crow]

@hasty nova, @wispy nest on Automatic 1111 webui max token is disables so it shouldnt matter

[hasty nova]

it is disabled, yes, but that still means that more tokens = less coherency to prompt

[dry crow]

Thats right

[wispy nest]

so then I guess at that point it's time to mess with cfg scale to get a desired output

[dry crow]

You can also try to do:
prompt + embed words BREAK prompt + embed
You would split the prompt into two fields of 75 tokens

[wispy nest]

is it possible to embed this vae into my mix or should i not worry about it? Just wondered if it might remove any unused / messed up vae from my merges

[dry crow]

You mean Bake in the vae?

I Wouldnt

[wispy nest]

ahh so basically the mix i'm using is dream with grape at 0.3 then moistv2 at 0.3

[dry crow]

It messed up a merge model of mine if its not for the specific model. But when you just select it on top it will work

[wispy nest]

and i embedded grape vae

[hasty nova]

portraits, I think I am going to go with the 250 step one. (both from 250) A whole extra 4750 steps wasted, I guess.

[wispy nest]

I might remake the mix without embedding then just play around til I find a good vae

[dry crow]

Thats the best Option, then copy and rename the best vae for auto apply

[wispy nest]

thanks, I can see the preview showing an awesome image then freaking out on higher steps at the moment so im guessing my mix is just a bit borked

[hasty nova]

:)

looks like an embedding cover image!

[dry crow]

Looks good

[hasty nova]

ty :)

Welp time to start training the actual embedding I have been trying to train. I haven't been able to due to bugs.

[wispy nest]

what do you thinik to these settings? I'm trying to keep it a bit general but bias towards anime, nsfw possible but not focus ( its grapefruit anime nsfw )

is 0.3 a good value to start with the mix?

[dry crow]

Yea 0.25-0.3 is good to start

[wispy nest]

heres what I was getting with the likely broken mix last night

so im excited to see if i can push it way more with the vae

[dry crow]

Colors will be better if you find the right vae but image composition is very nice

[wispy nest]

that was after filtering in adobe, before the colours were:

[dry crow]

Ahh

[wispy nest]

either way I'm going to just remake it without embeddings so i can try orangemix one etc

[dry crow]

And this one was unedited?

[wispy nest]

yea!

[dry crow]

Okay

[wispy nest]

thats where I tried to use two embeddings

[dry crow]

Yea try orange mix or direct AnythingV4.5 (orange mix included)

[wispy nest]

youre a legend thank you I'll get anything 4.5

works with 1.5 based mixes ?

[dry crow]

Sure its also base 1.5

Only WaifuDiffusion 1.4 is 2.1 atm

[wispy nest]

for some reason on huggingface I don't see the usual .pt format vae for anything 4.5

[dry crow]

Its using 4.0

[wispy nest]

https://huggingface.co/andite/anything-v4.0/blob/main/anything-v4.0.vae.pt this oine ?

[carmine hollow]

Cheesbus

[wispy nest]

doesn't look best but decent enough

originally by DALL-E 2 with Playground AI inpainting and some paint.net edits

still needs some work

[wispy nest]

bro you were completely right about the vae / embedded vae being a bit screwed up, got a bunch more coherant detail now

[wispy nest]

I think its really neat

thanks!

[wispy nest]

as a little extra, this is supposed to be showing sort of status of my current relationship I have

I was creating it with that in mind as a base

I'm basically just prototyping ideas to help me write a scifi book and have some frames of reference ^_^

also just a computer geek so

I get ya no worry :)

@dry crow the mix using anything4 vae

[hasty nova]

https://discord.com/channels/1002292111942635562/1070000581877440542 Released the embedding :)

[dry crow]

Very nice!

Glad it worked

[wispy nest]

spicy hot mate this is pumping out some great stuff now

[carmine hollow]

No to veggies

[hasty nova]

hmmm... have you ever wanted to add a neon-y style to your images?

[dry crow]

Which vae are you using rn?

[wispy nest]

anything v4

[wispy nest]

yes?

I love cyberpunk / futurism

[hasty nova]

Well then! I would love to see what people can make with my embeddings, and I just released a new one if you are willing to try it out. I need it to be tested on as many models as possible to see what works

[wispy nest]

absolutely

[hasty nova]

https://discord.com/channels/1002292111942635562/1070000581877440542 link here :)

[dry crow]

Okay you should try and compare to wd1.4 vae

If you only used 30% anime model it may bet better than Anythingv4

[wispy nest]

thanks I will do

[wispy nest]

rendering it now and my god I think youve done an amazing embedding here...

[hasty nova]

Really? Thanks!

[wispy nest]

ive used the same seed same settings so you can compare

will be done in 50 seconds

[hasty nova]

I think it can make people look a little funky, but other than that it works pretty well

I don't think it's as good as my ghostly diffusion embedding, which you can also try out on my civitai page if interested

[wispy nest]

in this case it's actually completely fixed visual problems with my character

[hasty nova]

👀

[wispy nest]

i had to double check it wasnt a random seed lol

[hasty nova]

Oh did it change it that much 😆

[hasty nova]

Oh, wasn't expecting that

Wow

Nice

[wispy nest]

that is a long prompt too ill PM it you

[hasty nova]

Ah that's why it didn't make it very neon-y if the embed was at the end. But it still worked some, I guess

Looks great!

[wispy nest]

OH youre meant to put them at the starT?!

and yeah either way that image is just, damn.

[hasty nova]

depends on how much you want the image to be affected by it. Words at the start have more influence on the image.

[wispy nest]

I've been generating this stuff for ages and I've learned so much in the last few days , how did I not know that

[hasty nova]

mine with that prompt

imma put the embed at the start and see

[wispy nest]

Looks awesome

what diffuser is that?

[sturdy latch]

Flying wheelchair

[hasty nova]

Forgot to set the seed to yours 😩

[wispy nest]

honestly that seed looks better anyway

[hasty nova]

another random seed, but with neon at the front

[wispy nest]

at this point I'm getting almost always a good image, the odd hand thats messed up or wonky eye and thats about it really..

its crazy!

[hasty nova]

yeha, nowadays face restore makes images worse

[wispy nest]

I stopped using it like after a week haha

it kept putting human eyes in anime style was scary

[hasty nova]

yeah it's not great for anime

[paper hinge]

https://cdn.discordapp.com/attachments/1049275551568896000/1070005316411994274/output.png

[wispy nest]

thats awesome!

[paper hinge]

its not even mj

[hasty nova]

alright I am trying it now with your seed and correct cfg scale

[paper hinge]

https://cdn.discordapp.com/attachments/1049275551568896000/1069744470322577448/output.png

[wispy nest]

I dont get the hype with midjourney , I do get having an online playform is extremely good for people who dont have the hardware to run SD though

[wispy nest]

@hasty nova im cooking you up something epic with the promp i was playing with before

[wispy nest]

so cuuute!

I love rabbits

[hasty nova]

well here is pretty much your exact settings.

[wispy nest]

Woah thats drastically different? Is that with Heun ?

[hasty nova]

yep

Heun, 120 steps, 10 cfg, same prompt/seed

[hasty nova]

different model tho

[hasty nova]

and embed is at the front

[wispy nest]

H O L Y

M O T H E R

[hasty nova]

👀

Wow!

[carmine hollow]

Bees in sweaters

[hasty nova]

Now try your prompt with the ghost embedding...

[wispy nest]

@hasty nova my hear actuially skipped a beat when that rendered tbh that is cool as HELL

[hasty nova]

it's super cool

[wispy nest]

Ive got so many tabs open in my browser lmao can I have the link to your profile again please

[hasty nova]

actually, I will do it for you

https://civitai.com/user/junglerally_

wow I have 7 followers 😆

[wispy nest]

8 now 😉

[hasty nova]

There are two versions of the ghost embedding, I would go with 4k fbecause of the size of your prompt

[wispy nest]

I'll try both , this prompt is small tho ill dm it you

[hasty nova]

oh ok

[hasty nova]

Same as this but with ghst embed (slightly nsfw, so spoilered jic)

not very ghosty due to the large prompt I think

[wispy nest]

so compare with the last one on the model im using this is WILDLY different

thats with 3k

[hasty nova]

same settings and prompt as your last image?

[wispy nest]

yeah!

[hasty nova]

very tiny person

[wispy nest]

could just be that specific seed

trying 4k then ill try a random seed

[hasty nova]

👍

4k doesn't change too much, but it changes some things

well, I can';t test any more due to me going on to train another embedding

This one is not ghost, but vampires, so lots of blood probably

I alread y have everytyhing ready, just got to hit train

[wispy nest]

for me it did change a lot maybe due to input length

looks awesome too

[hasty nova]

still a tiny person down there

[wispy nest]

yeah haha not sure why but I;m going to do a random one now

[hasty nova]

👍

[wanton ore]

mjv4 level with protogen infinity

[paper hinge]

https://cdn.discordapp.com/attachments/1049275551568896000/1070008613743108096/output.png

https://cdn.discordapp.com/attachments/1049275551568896000/1068161820201848872/output.png

[wispy nest]

haha love the foxes dude

[carmine hollow]

Im dissapointed she is not breaking them

[whole river]

I trained a .pt file with dreabooth lora
but it doesnt do anything to the generations

[wispy nest]

@dry crow ( recoloured in post ) getting there with the mix, think I will take grape out because its difficult to make them sfw enough

[dry crow]

looks good !

[wispy nest]

thank you mate

[half crystal]

Hi, I'm wondering if anyone thinks it is possible to train a model for a style transfer from/to the styles in this image. For instance from A to B or from A or B to C.
Complete novice here. If anyone knows someone that does this kind of thing for a fee I would love to get in touch. Feel free to DM.
Thanks!!

[wispy nest]

thats epic

[dry crow]

Thanks! But sd didnt liked it as dark as i wanted so it added these lamps xD

[paper hinge]

https://cdn.discordapp.com/attachments/1049275551568896000/1070111458635030548/output.png

[hasty nova]

Anyone interested in some samples from my latest embedding that I have yet to publish (still in testing)?

[hasty nova]

https://discord.com/channels/1002292111942635562/1070124391213572266 Here is the embedding

[wispy nest]

check it out guys its really good

[static tusk]

nice one @hasty nova

[hasty nova]

Wow, that's super cool looking! And thank you!

The only thing the embed it bad at is teeth, but a lot of sd is bad at teeth so

[static tusk]

yup teeth are like hands. pepehands

[hasty nova]

But vampires probably don't have the best teeth anyway

[hasty nova]

punk vampire 🧛

those teeth are pretty good too

[hasty nova]

Oo

What model are you using?

[static tusk]

good results dude. well done

I have a custom merge of 3 models

[hasty nova]

I see

[hasty nova]

Thank you so much! Now I need to move on to improving my "neon" embed I made this morning

I think I am gonna add some different images to it, and see how that works

After I make a reddit post with my vampires first of course

[hasty nova]

oo

Some very deep eyeshadow

[hasty nova]

https://www.reddit.com/r/StableDiffusion/comments/10qeptn/vampire_diffusion/ Post created!

[static tusk]

[hasty nova]

I just updated my neon embedding!

https://civitai.com/models/6016/neon-diffusion

[fossil solstice]

portrait male frontal 25 years disney colombia style physical features of hyper realistic, super detailed, 8k, 3D with suit victorian era--stylize 1000

[daring glade]

#1047759610737610784

After his baptism, as Jesus came up out of the water, the heavens were opened and he saw the Spirit of God descending like a dove and settling on him.

[split finch]

Portrait of a black male with a sword on horseback

[strange jungle]

Prompt: "Grok Kill Og With Rock"

Caveman Grok kill Caveman Og with rock

Og wasn't dead. He invented the knife, and now he wants revenge.

[strange jungle]

Elder Zog find out Grok and Og fighting. Elder Zog not happy.

[austere jay]

Seed : 75154 | Scale : 17.07 | Steps : 40 | Img Width : 576 | Img Height : 704 | Negative Prompt : cartoon, 3d, ugly eyes, 2 people, deformed iris, deformed eyes, bad eyes, (disfigured), (bad art), (deformed), (poorly drawn), (extra limbs), (close up), strange colours, blurry, boring, sketch, lackluster, face portrait, signature, letters, watermark, grayscale, moody lighting | model_version : custom_realisticVisionV12_v12 photo of one 29 year old woman, pale skin, homeless in new york city, upper body, dark hair, detailed skin, detailed eyes, realistic eyes, 20 megapixel, canon eos r3, detailed skin, detailed face

[wispy nest]

input: booba head

[strange cliff]

I think I've just invented new pokemon

[edgy girder]

Имперский штурмовик

[wispy nest]

Hello o/ I'm trying to get some architecture from a long distance, something like in this picture i'm linking.

problem is all i get is closeups of houses and such

i do my tries with SD 2.1 checkpoint

any prompt inputs that could help?

[dry crow]

maybe far away and wide view

scenery

[wispy nest]

TYVM

[dry crow]

np 🙂

@ripe cedar @wispy nest Vae comparison on my merged model

[wispy nest]

@dry crow dude, the quality

[dry crow]

awesome!

[wispy nest]

removed cos a little nsfw

[dry crow]

xD

[wispy nest]

I like the way anything looks

it's not hard to saturate the colours of an output afterwards and in fact i think the saturation on the far right is too much!

[dry crow]

yes i think so too, it sure depends on the model you pick

her is an older compare from me:

here anything is not good xD

[ripe cedar]

VAEs only work with the ckpts they come with?

[dry crow]

no

with every model but like you see in the image above int wont work good with every model

so use the ones that come with the model or the creator recomends

[wispy nest]

150 steps

[ripe cedar]

My brother is in my room with covid so I can't access my desktop to try stuff. I need to read up a lot on all these AI things

[wispy nest]

frankly, in both cases i prefer the far right picture. i prefer sharp and colorful original, to which i can always add fog, dept of field, or whatever else, including desaturation where needed

@dry crow

[dry crow]

yea colorful is good! my fav is wd1.4 vae 😄 the good thing with vaes is you can experiment with them

[wispy nest]

thicc

see I'm the opposite just because I like to take them into Adobe Lightroom afterwards

i usually take them to clip studio paint, then to photoshop

CS paint is great for smearing and mixing colors. it works much better than mixing RL colored chalks, for instance. very smooth. helps a lot when making composits from multiple works. PS is just for final detail sharpening or blurring

[dry crow]

<- hears expensive image editor names, "i like Paint.Net" 🙂

[wispy nest]

I use paint.net for basically anything that I don't need Photoshop for lol

have done for years

[dry crow]

😄 yea its my one and only allround editor

but i might get me a lifetime license of Affinity Photo

[wispy nest]

CS paint costed me like 50 US$ about 5 yr ago. In general it is a very cheap professional 2d painting software compared to the rest. it also has very good vector brushes etc

very powerful software for that money

[dry crow]

okay thx didnt know it, i should look how it compares to affinity photo

[wispy nest]

it was known in old times as "Manga Studio", today as "Clip Studio Paint" - check some vids on ppl using it on YT

[ripe cedar]

I have CSP but really disappointed when they decided not to implement AI features because of public backlash when they announced they wanted to release AI features.

PS already has AI stuff

[strange cliff]

whaaa?

[wispy nest]

CSP does have some really smart features. borderline AI if not AI. Autocolor of a sketch for instance is damn advanced stuff and super useful. They will jump on to AI train when everyone else does anyway kek

[dry crow]

@wispy nest #outpaint ez 😛

[wispy nest]

kek

[wispy nest]

mines actually gotten quite close but it's deffo not 100%

hang on just trying it agin and having a look

so close but so far

😄

[dry crow]

ouh not bad

[wispy nest]

I'm just using my own mix for the outpainting but I hear theres a specific 1.5 outpainting model ?

perhaps worth a try idk ?

damn, i'll have to get new Windows. Tried to stick with Win7 as long as i could, but damn SD can't use my GPU on this OS, so i use CPU only. RTX 3060 with 12 Gb of VRAM plugged in lol

[idle crater]

/drem promt: logo a

How it's working?

[wispy nest]

these are not made by the discord bot, the discord bot is not online

@dry crow try with this https://huggingface.co/runwayml/stable-diffusion-inpainting

apparently it works great for outpaining

I wonder if I took my model then merged this and try that?

as my model inpainting

[dry crow]

wow i fully forgot that we need to use the inpaint model, i have it already xD

[wispy nest]

thats cool

[woeful pond]

thanks

[wispy nest]

run it again! I'm downloading the SD inpainting now im really hopeful

if that checkpoint can allow some poor mans outpaining then I'm in

magic ]fill tool in photoshop can probaly fix the rest

also another note with the outpainting, how I got the reult i did was to remove any prompts defining the charcater and just reuse the ones that define the background

[sinful river]

hello how would I go about generating art based on some of my own ?

[dry crow]

@wispy nest ohh i didnt switched back to poor mans now, its on mk2 and i accedently picked 2.0 inpaint instead of 1.5

in 1 min we know more xD

then i test 1.5 with ppor man

[wispy nest]

im using outpainting mk 2 aswell ye

oh i thought poor mans was just an earlier version of outpaint mk2

[dry crow]

yea it was the first

different type of outpaint i think

but maybe mk2 beats it now

[dry crow]

you would need to train a model, its called dreambooth, or you can use Lora or Embeddings

[wispy nest]

i could be wrong but I think embeddings are likely easier place to start

[sinful river]

thank you for guiding me. Dreambooth is also open source like SD ?

[dry crow]

@wispy nest thats it !! SD 2.0 512 inpaint ema with mk2

[dry crow]

yes its a tool/extension for SD

[sinful river]

Then I dont have to pay ?

[wispy nest]

O LORD WE BE LEARNING TODAY BOI

[sinful river]

i paid for mid journey but im not getting what im after

[dry crow]

if you have a good GPU you can use SD local for free

[wispy nest]

now my next question if i don't get as good result with the outpainting is, would merging with the inpainting1.5checkpoint work / produce better result

[sinful river]

yes i have a 3070

[wispy nest]

mid journey is over-rated imo it takes a lot more learning to do offline but you also have a greater control

[sinful river]

any good tutorial on how to get started with embeddings please ?

[strange cliff]

I only have 3060

[wispy nest]

3060ti here lol

[dry crow]

sure here:
https://www.youtube.com/watch?v=2ityl_dNRNw

[wispy nest]

@dry crow so, deffo can see the default sd 1.5 isnt producing anywhere near as good results

but it IS more coherant

[sinful river]

great so this is called an embedding . I will dive right in . Thank you 🙏🏼

[wispy nest]

so now I gosh darned gotta know if merging works, pray for me

I mean what weight would I even put?!

[dry crow]

you need to have SD local on your PC to do that

[strange cliff]

wait 3070 have only 8gb of vram right?

[dry crow]

For embed 8gb should be enough

[sinful river]

yes i will see a video on how to get that done first , then the video you linked to

[strange cliff]

just asking cuz curious lol

[sinful river]

sorry i have a 3060 12GB

[dry crow]

the same guy made a video for install pls look at his video its the most updated:
https://www.youtube.com/watch?v=VXEyhM3Djqg

[wispy nest]

making style embeds is where I'm going to start looking at next, I'll share them with community when I do

[strange cliff]

ah same then

also I'm curious why isn't there anyone make anime gore model yet?

[sinful river]

thats wonderful. If you were to name one wizard on this forum who could replicate the art I have (12 pieces so far) who would that be on these forums.

[wispy nest]

@hasty nova released a vampire one that produces gore as far as I can tell

[dry crow]

@wispy nest poor mans with 2.0 inpaint

[strange cliff]

oh? link pls

[hasty nova]

I got you ;)

[sinful river]

like this

[hasty nova]

https://civitai.com/models/6079/vampire-diffusion

[strange cliff]

thank you

[hasty nova]

👍

[wispy nest]

recon mk2 is better?

[strange cliff]

here's the vampire I made before

[hasty nova]

Hey, I updated the neon embedding last night

[dry crow]

@sinful river look at this model, i think it would fit very well:
https://huggingface.co/Fictiverse/Stable_Diffusion_PaperCut_Model

[wispy nest]

OH DAAAMN YOU DID ! I didint want to hassle you so I didin't message

[hasty nova]

Ah it's all good

[wispy nest]

been working outpainting :L

do you think it will work if I merge ( no idea what weight) the sd1.5 outpainting with my custom mix so that the style is more consistant?

or am I getting the idea wrong

[sinful river]

its not the same style as mine but I'll try it any way 🤞🏼

[dry crow]

@wispy nest i dont know if its possible bevcause inpaint, depth and 2x are models that are like pix2pix model, different than normal

[hasty nova]

Let me know if the embedding works for you

would love to see what you make

[strange cliff]

👍 I will

[wispy nest]

on the civitai discord a user posted a guide on how it can be done bro

[wispy nest]

nice one

[hasty nova]

would anyone be interested in an embedding on merpeople?

[strange cliff]

this's what I made with your embed

it also allowed for really close recreation

only difference between the 2 is bat and wings

[wispy nest]

looks great

[strange cliff]

thanks

[hasty nova]

ooo I like the purple

I haven't really tested my embeds with anime models so it's good to know that it works well

[tough wharf]

its enabled / imagine

?

[hasty nova]

#1047610792226340935

no

[cyan snow]

WHY ARE THERE MULTIPLE ONE WORLD TRADE CENTERS?

[wispy nest]

these look nice, offline model with outpainting?

[hasty nova]

I'm trying some new settings for my next embedding, and I am running it on my laptop so I can do things on my pc, so it will take quite a long time. I am probably going to stop it at only a few hundred steps.

But it's a merfolk embedding

I hope it works

[strange cliff]

[cyan snow]

i used photoreal2.0 for the model

[cyan snow]

also this is done on my amd card, which wasn't possible on windows until last month

[strange cliff]

that moment when they use 2 eyepatch

[keen juniper]

/ dream prompt rainbow over the moon ultra realistic

[wispy nest]

your stuff is so good mate, I might look at that model for merging into mine its just already a mix of 5 models so I'm in too deep and my heads nearly exploding

I am in love with this mix but it tends to draw veeery squiggly building lines

which tbh is kinda goofy but at the same time to me I like the way it looks

[cyan snow]

im on amd, its probably impossible to have the same model as me, its converted from CKPT, but i used this https://huggingface.co/dreamlike-art/dreamlike-photoreal-2.0/tree/main

[dry crow]

good model indeed

dreamlike v1 is also cool

[cyan snow]

only for abstruct stuff

[wispy nest]

shame they use the modified CreativeML OpenRAIL-M license, with all the outrage around ai art I'm worried about mixing stuff that isnt just the defauly creativeML OpenRAIL

[cyan snow]

the real art is creativity, the painting brush is the pc

[wispy nest]

I agree I jus thope to use some as part of a web comic or something at some point so yeah, I dont know if using part of some ones mix with a modified liscence might make all my work basically worthless or illegal at some point

the sad reality of the situation :/

[dry crow]

they cant prove you used specific models in a image

[wispy nest]

yeah true, recon I'll have to redo the inpainting model though?

[dry crow]

why?

[wispy nest]

if i go ahead and mix in that to my main model ?

I would then have to redo the inpainting mix or?

[dry crow]

yea thats true

[wispy nest]

screw it im in

:L

I'll take what I have and drop OFT idk 0.2 or 0.3 mix? any guesses? I just want something that will help sharpen up the buildings thats it

dmp++ 2s a Karras at 60 steps using the inpainting mix holy shot nvm

[strange cliff]

I wanted to make normal soldier but AI say "you're making USSR soldier today"

[cyan snow]

did you add photoreal? the buildings still look slightly off

[wispy nest]

nope! and also I was meaning to show it to some one whos in the other chat bvut rendered that using an inpainting model, just trying with 0.3 photreal added now!

[wispy nest]

hmmmmm interesting 60 steps is the sweetspot for SDE and Euler I'd say

I know a lot of people want quick images but to me like if I get more better ones its faster than generating loads of bad ones

[dry crow]

60 was my go to with Berrymix last year

[wispy nest]

ahh ye olde berrymix

GUYS IM DYING INSIDE BECAUSE THE 0.3 HAS FIXED MY BUILDINGS I LOVE YOU

I "think" let it pass all the steps

[cyan snow]

i use 70 steps in most my images

[wispy nest]

thats so much better

[dry crow]

reminds me a bit of Astral Chain

[wispy nest]

😄 ❤️

and yeah I was going to say @dry crow

adding 0.3 weight of photoreal , has not changed the filesize by a single kb

some one tell that to the laywers attacking ai gen

see if they can explain how its possible

[dry crow]

hahaha xD

[wispy nest]

BUT BUT HOW YOU STOLE MORE IMAGES

NO KAREN YOU DONT GET IT

I'm quite tempted to release this mix but I wouldnt want to do it without checking the full conditions of the modified liscence for the photoreal

early testing has indicated at this point its really versatile

[hasty nova]

Which is best?

Based on these I am thinking 1900

But it doedn't have as much 'tail' even though that's what sd always struggles with anyway

[static tusk]

on the second pic the hair on 1900 is faded

[hasty nova]

looks like the hair is worst in 1800 to me

And I still think 1900 has the most coherency. Now it is time to test 1850 and 1950 as well, to reallky figure out which is best.

[static tusk]

agree

[hasty nova]

First two have more color

but 1950 is the best composition I think

Least messy

this is a hard decision

1850 filled the whole frame

1950 has a weird face on the tail

1900 most mermaid-y

[static tusk]

yup 1900 looking better

[tranquil kindle]

How do you get 1850, 1900, and 1950?

[hasty nova]

I save an embedding every 50 steps

[tranquil kindle]

Ok cool, I like all

[hasty nova]

1900 is best still

[hasty nova]

mans got an 8 pack

[tranquil kindle]

I like the 1900.

[hasty nova]

Same. I think I will release the 1900 one.

Has the best coherency.

And best details.

[tranquil kindle]

Yep

[hasty nova]

ALrighty, time to start working on a civitai page!

I think I will call it "merfolk diffusion"

Keeping with the theme I have on my page of <word> Diffusion

[wispy nest]

1900 between both images

[wispy nest]

1900 again

[hasty nova]

yeah that's what I am going with

[wispy nest]

great job again mate

[hasty nova]

Thkanks! It's not my best, but it works well enough

SD still sucks at mermaid tails so it doesn't work that well

[hasty nova]

(slight nsfw) good enough for a cover image on civitai?

[analog goblet]

seems to fit 🙂

[hasty nova]

cool image but the hands are funky

[hasty nova]

this one has some really cool detail=s

very funky

would have been cool without the creature thing

https://civitai.com/models/6242/merfolk-diffusion Tis done!

[pale cipher]

https://giphy.com/gifs/flambo-flambothedog-the-dog-fBRG99AeYpmT2kXgM0

[wispy nest]

amazing job.

[hasty nova]

Ty!

[ripe dock]

Jesus Cristo, religious iconography, holy, illuminated manuscript, intricate, highly detailed, gold leaf, biblical scenes, digital painting, artstation, illuminated, vibrant, illustration, art by greg rutkowski and alphonse mucha
Negative Prompt:

secular, mundane, modern, dull, plain, lacking detail

[primal tree]

Dog

[fleet solstice]

imagen que transmita paz

[cerulean spade]

#✍🏼｜rules-and-tos

[fleet solstice]

hands holding a heart with a heavenly background and doves around

#1014939219904450590 hands holding a heart with a heavenly background and doves around

[valid verge]

yo what is the difference between inpainting and finetuning

[sharp bronze]

Hey guys! Is there a model where I can a similar "concept" art look like Midjourney? It's hard to get similar results. Here are 2 examples from Midjourney:

[tawdry acorn]

Hello all!

I meed some help regarding Stable Diffusion video generation.

As there are so many models and UIs available, can someone guide me which is the best and convenient to create a series of images (coherent or linear) and then convert it into a video.

[dry crow]

Automatic1111 Webui with the Deforum extension is what you need

[tawdry acorn]

Thanks a lot. Shall give it a try. Can you attach a link for tutorial to install the same

[dry crow]

Np sure this is for installing the Webui:
https://m.youtube.com/watch?v=VXEyhM3Djqg

@tawdry acorn this is for Deforum:
https://m.youtube.com/watch?v=Zvo86Y7hgfY
He has also tutorials how to use it.

[tawdry acorn]

Amazing

Thanks a ton! Really helpful

[fast sleet]

Yo everyone! We've made an Adventure Time x Ghibli Studio video 😄

https://www.youtube.com/watch?v=-McJYKLr71U&lc=Ugx0O8SWx3yHyENt2NJ4AaABAg

[novel ravine]

lamborghini x Pickup

[strange cliff]

now I want this

[hearty wasp]

Trying to find my style, hard to find the right one haha

[cyan snow]

https://imgur.com/a/ts5J0Xn

[steel yarrow]

it tries so hard, bless it's little heart

[sterile kiln]

Galadriel

[sterile kiln]

Aragorn

[sterile kiln]

Gimli

[steel yarrow]

you gotta name the fat ones too

[tame forge]

no offense but male mermaids look gay

[jolly sigil]

Mind blowing

❤️

[analog goblet]

this's like a female red haired demon hitler ? 🙂

[mighty marsh]

working on training a model to the style of wes anderson's grand budapest hotel -- pretty happy with training so far (these are training images, not using the embedding yet)

[tame forge]

for rent: $2500/mo, no pets, must make 3x rent, $1500 deposit required, utilities not included

[analog goblet]

kinda 🙂

[sterile kiln]

Elrond

[lament ridge]

Hey guys look at these awesome images I generated

[strange cliff]

double spine skeleton cuz why not

[spring wave]

can anyone suggest me any model that can make art like vagabond?

or should i try embeddings?

or train lora

[vestal vale]

The photo of hills

[cunning geode]

samples from my current work in progress model

https://www.reddit.com/r/StableDiffusion/comments/10sf9nf/some_samples_from_my_current_wip_model_dealing see also here for more info

(the spoiler is for people with arachnophobia)

[fervent rune]

A LoRA model I'm working on

[south heath]

why colorless? I've tried all the models, orangmix, anyv4, anytwam1.1, all the anime models I mentioned, they all come out colorless, what do you think went wrong?

[storm sundial]

VAE

[south heath]

vae model?

[storm sundial]

VAE as in go get the anything v3 VAE from the anything v3 huggingface, its in its downloads

[south heath]

So,can't use v4.5 or other models that don't have Vae

?

[storm sundial]

you can, vaes can be applied at anytime to any model