🏞｜general-with-images

[glossy herald]

yes. so that you personalise the drawing itself to get a face that really talks to your client, it feels like him. so yeah, using their dedicated trained model seems essential here
Then you use this output in the txt2img controlnet for the color, the 1.2 step

I didn't include the mouth in my mask recommendation, because it's more part of the emote itself usually, and less part of the "person recognition"

[jagged crest]

Yeah but I'm automating this on youstick.fun so manually entering "blonde hair" in the prompt is out. Unless I first do some feature extraction from the images they upload. I'm not sure how big a model that would be wrt time and cost

[glossy herald]

training a small model like that, using collab, currently is about 10 minutes if you automate it. it will require something like 5 to 10 pictures of their face in different context if possible (you can put some of the quality responsability on them here)

then you can use that model to color, it should be all right in the 1.2, it should have learned those hair by heart so it should directly use that in that step, if you use the token for the client in the prompt

this will need quite some testing to completly automate for sure, in particular the denoising of controlnet, finding the right values and prompts here and there

but it should work quite well

[cyan snow]

bro wtf

[glossy herald]

big resolution duplication shenanigans

[cyan snow]

dude why is he so happy

[glossy herald]

(I'm going after this one : )

[wispy nest]

me and the boys

[glossy herald]

Cosplay time

nice hat ^^

[glossy herald]

(just making a new banhammer)

this one feels festive

[wispy nest]

That's interesting idea

[dry crow]

does somebody know why its not possible to outpaint upscaled pictures ?

[floral flicker]

Any tips to make an underwater scene, without visible water surface at the top of the image?
Prompts are beginning with deep sea photograph
I've tried in the negative prompt: sunlight, surface - image is now as dark as I want it but there is always the water surface showing up

[jagged crest]

Thanks a lot. Going to try this. I'll let you know how it goes

[fiery lichen]

I've seen a lora for deep sea scenes. Try that

[floral flicker]

Thanks for the tip, Ill see if I can find that one 👍🏻

[glossy gazelle]

#1072220168534642768 sunflower

how to generate image?

[main elk]

anyone knows how to solve this? Im using m1 macbook pro. i tried to use SD 2.1 but when i try to generate it shows this. I also added the config file in the model folder. v2-1_768-ema-pruned.ckpt this is the model and v2-1_768-ema-pruned.yaml is the config file. Can anyone help me with this?

by the way.. with other models it works fine....

[boreal hare]

Tried new lada models and wow, they are great. I used img2img on one of my old generation and new one is so much better

[grizzled sage]

got that heavy metal aesthetic \m/

[thorn salmon]

Hi all, this is my first post here.

Does anyone knows how to keep the same elephant (color, size, shape) but keep adding more to the picture?

is that possible, i tried img2img and locking the seed but no luck

[glossy herald]

Should be possible by using 2 more tools :
1/ inpaint, a tab inside img2img, lets you draw a mask on part of the picture you want to be modified, or not to be modified. This can let you preserve the elephant as is
2/ controlnet, an extension for SD. It can let you keep the current composition and details as most as possible

I'm trying it

[fiery lichen]

someone talked about how hard it is to get guns right in SD, especially 2.1 since they are sorta censored (sex and violence), not bad guns here.

[glossy herald]

here is an example, very stylized

[thorn salmon]

Thank you @glossy herald that is exactly what i was looking

[glossy herald]

prompt is

a purple and white elephant in a forest with a mountain in the background, intricate, very detailed, from

and add at the end the type of reference style you want

the first one was "studio Ghibli"

[thorn salmon]

ahh! perfect!

[glossy herald]

or you know, go photorealistic

[thorn salmon]

hahaha. i think i am going to give up on the idea lol

when masking it, it just places another elephant behind the mask

[glossy herald]

use the option "inpaint not masked"

by default, the mask covers what should be changed

here is the mask I drew

and here is the controlnet I used, to help keep composition outside of the elephant

last important parameter, denoising, I upped it to 0.85

the higher, the more changes will happen

[thorn salmon]

iam missing the canny model tho

[glossy herald]

yeah, the model is important

abd you may want to try and mask the husk too

the canny model is what will make it respect a lot more the original composition

https://civitai.com/models/9251/controlnet-pre-trained-models

here are the models

[thorn salmon]

Thanks!

[glossy herald]

no problem

download the ones you want (all ?) and put them in the folder where you installed automatic, in the extensions\sd-webui-controlnet\models folder

then click the blue arrow next to the dropdown of the UI and it should give you the models to choose from in it

[thorn salmon]

wow! that model helped a lot.

Even without masking the tusk lol

thanks a lot!

[glossy herald]

happy to see it work like you want 🙂

I saw a problem while testing

it seems to need to specify a lot of the colors

so the mountain going purple is mainly because it's not specified it seems

not sure, I had a same drift

have a happy editing 🙂

[languid ermine]

I trying deflickering, with controlnet on a custom style model I made, but I would like more consistency, any suggestions?

[glossy herald]

Checking on some frames how the controlnet preprocessing performs and fine tuning its setting could maybe help, pushing down the threshold sliders in canny mode for example.
Also adding a secondary controlnet, or even tertiary controlnet, will help keep for example the same depth or normal map, making it even less likely for the image to flicker

[languid ermine]

Mh, for now I'm working with a normal net, and a canny net, at a 1k Res, for both of them, I wonder if using depth may actually help

[glossy herald]

I feel like in the first half, the main problem was things coming in front and then going back, so maybe on that yes.

both normal and canny net have threshold in their preprocessors too. I didn't play with normal yet that much, but I saw that using the one in canny can change a lot of things

[languid ermine]

Yes, I played around with both, to be fair yeah, maybe i should not use normal as primary, is kind of not super trustworthy for this kind of stuff

[glossy herald]

you can just run the preprocessor on some frames to see the normal map it would use, but yep, the main feature of your clip is movement and depth, the flat surfaces aren't the most prominent. I'm not sure if the depth model will manage to read it well though

[languid ermine]

The normal one seems to not do well for this,but the leres does wha better

[glossy herald]

from this

somewhat better than the movie did

Powerpuff girl

original :

getting there

got it

[rapid current]

COD zombies Primis crew but real

[tawdry badge]

hey guys can i run SD with this laptop?

[tame tartan]

A picture of beatmakers of various races battling it out.
Similar to the style of Akira Toriyama.

The image size is 16:9 landscape.

[vagrant shell]

Nope

[fiery lichen]

your bot is killing its foe with its rifle? with bayonet, and immediately hosting the funeral with white flowers bouquet in hand LOL.

[dense tapir]

Well, it came prepared.

flowers to lay on ya

[smoky oak]

ok so like, has anybody ever run into massive issues trying to do concept art for weapons?

none of the SD models I am using are working

[dense tapir]

originally scimitars, bows were almost a no go without a lucky hit

[smoky oak]

I am like 100 attempts in, and SD is just ruining everything

like the results are horrible, and I don't know why

[dense tapir]

which weapons?

[smoky oak]

battle axe, battle hammer, war hammers, anything like that

it has 0 clue

bro, it tries to make axes as guitars

[dense tapir]

well, that is their nickname

[smoky oak]

yeah but like, a guitar is not a "battleaxe"

[fiery lichen]

He is like the G1 megatron

[smoky oak]

ok, so the AI has no clue what bladed weapons are, bruh

Tomahawks, hatchets, axes, anything like that

no clue

[dense tapir]

probably need to grab a weapon model

[smoky oak]

looks like it doesn't know what swords are either OMG

[smoky oak]

I have even tried RPG diffusion and dungeons and diffusions, and nothing

[fiery lichen]

We talked about this before sytan SD is censored on weapons/violence (to some degree)

you need to make your own loras sytan

and name them something else

dont call them swords and axes

Like our last conversation, the AI wont draw missiles and bombs on my jets either

[smoky oak]

Seriously? That is ridiculous lmao

[smoky oak]

I never talked to you about this

[fiery lichen]

we had this conversation a few days ago

[smoky oak]

what?

do you have more info?

[fiery lichen]

or maybe it was someone else

[cyan snow]

That really depends on the model, VAE, and lora. I just merged a bunch of models I like and this one gets the prompts very accurately. You can probably make a lora for guns.

[fiery lichen]

but i talked about weapons (specifically swords, bows, axes) from someone here

[fiery lichen]

Some one did make a gun lora

and a tank (type 99? china)

[smoky oak]

interesting, i guess I am gonna have to train these relics into the AI

[fiery lichen]

my f14 tomcats carry a shit load of external fuel tanks but no missiles/bombs no matter the prompt

1 of the images carried something that looks like an external gunpod...half gunpod half fuel tank

that was as close to arming my F14 as I could get

[smoky oak]

I never knew that was against what they were doing. That is insane lmao

[fiery lichen]

people got lucky with all the nsfw stuff on sd 1.5

that genie came out of the bottle and it will never go back in now that it is circulating

[smoky oak]

I am using 1.5, but I guess even that has no info on weapons and stuff?

[fiery lichen]

I just know that SD 2.1 is definitely blocking sex and violence (guns/weapons)

occasionally, something does pass thru (but not completely), I've had half? a nipple show up (discolored/colored to match the skin tone and not protruding fully)

on a 2.1 generation image

Sukoi's with external fuel tanks, no missiles drawn. I think they were SD 1.5 images, not even the censored 2.1

F14/18 bastard child with extra wheels despite negative prompt to avoid them and no missiles/bombs

[smoky oak]

Their censorship is obnoxious. It quite honestly feels like its against the whole idea of a widely available free and open source AI

"use it for whatever you want, as long as we want it too"

[fiery lichen]

The bottom jet looks like it is carrying some ECM pods for electronic warfare, maybe it got that from some EA-18G Growler images

[dense tapir]

Yes. Thing is you can train to add to it but there is just too many weapons missing. You mean out of all the images scraped there were no weapons?

[smoky oak]

no, they chose to add none

[dense tapir]

Precisely

[smoky oak]

its quite literally censorship. Luckily its not hard to work around, but man, it would take a lot of effort

for now I am focusing war hammers and battle axes

[dense tapir]

it would, and does

[smoky oak]

@dense tapirDo you know if there is a way to merge LoRAs?

Like if I made a battle axe and battle hammer LoRA, could I blend them to get one LoRA that does both?

[dense tapir]

There is

they must have the same dim size

[smoky oak]

Dim size?

like 512x512?

[dense tapir]

dim size is also called rank

you can also make multi concept loras

[smoky oak]

Sorry, I guess I don't know that about LoRA's. Do you have any links to places where I could find out?

[dense tapir]

omg no. This shit is everywhere and why it has taken me months to get it all down since the start of the year.

[smoky oak]

It just requires more images and more tag words, right?

Like 10 images with "battleaxe"

and 10 images with "battlehammer"

So then they only trigger when I use one of the two?

[dense tapir]

kinda sorta

you need to caption each image

[smoky oak]

yeah, but you don't want the caption info to overlap

[dense tapir]

not sure on that part but best no.

oh, I have a link for ya

just remembered

[smoky oak]

If you caption one with "two handed weapon with a giant hammer head"

and the other as "Two handed weapon with a giant axe head"

then saying "a two handed wapon with a giant ---- head" will trigger both data sets, no?

[dense tapir]

Love me long long time tomorrow cause now I am too tired. https://rentry.org/lora_train

[oblique kestrel]

It is very mundane

[smoky oak]

thank you

[dense tapir]

It is a great little guide

[oblique kestrel]

Can I use this I am newbie

[dense tapir]

That is what it is really made for.

[smoky oak]

@dense tapirohhh, so its saying to keep them in different folders? So a folder for battle axe, and a folder for war hammer?

[dense tapir]

yes

as the site shows

each caption for each goes into its respective folder

[smoky oak]

@dense tapirWait, so this uses a completely different LoRA trainer to Kohya SS?

This LoRA doccumentation is going against everything you and I have been doing. Its saying use 150 images? lmao

[dense tapir]

It is for anime

they use 250 images, and tags, without blinking.

[smoky oak]

Ah, so anime benefits more because there is more of a style and less data to pull from other things

[dense tapir]

pretty much. A bit flat

[smoky oak]

like obviously if you train on a real person with their eyes always open, the AI would know how to close their eyes cause they are a person. But it wouldn't necessarily know how to do it with an anime character in a different style

[dense tapir]

Pretty much

[smoky oak]

alright, I get it now

[dense tapir]

you need more images in anime for everything from open eyes, closed eyes, mid point, frowning, happy, etc...

[smoky oak]

ok, so Kohya does have the same functionality to merge, cool

[dense tapir]

yep

[smoky oak]

does this rip out the style of a checkpoint and make it into a LoRA, or am I missing something?

[dense tapir]

thing I have not figured out is when you merge what then? I know merging embeddings I never could call either again.

you aren't missing anything

[smoky oak]

alright, awesome

so this merge lora, extract lora, model conversion, and auto captioning

[dense tapir]

don't think auto is perfect because it is far from that

[smoky oak]

Auto?

oh, yeah

I actually use it just to add prefix and made the files for me. I always set it to caption 0-0 cause they kinda scuk

[dense tapir]

well, you would be amazed at danbooru tags

even non anime

[smoky oak]

but I will have it add like "a photograph of subject name" at the start of the tile

or just "subject name"

[dense tapir]

well, promptword class

class being. Well, I honestly do not know. Maybe weapon

[smoky oak]

I wish i could look at the full setup of like a kick ass LoRA and just figure it all myself, cause there is so much random info on LoRAs

[dense tapir]

oh, wow yes. I am beginning to hate loras

I may have to go back to embeddings since I do styles

[smoky oak]

@dense tapirDo you have any documentation on how LoRA ranks work?

[dense tapir]

ranks are those dims I mentioned

[smoky oak]

I don't really know what DIMS are

[dense tapir]

don't need to

more dims the larger the file the more info it can hold BUT that isn't always a good thing

[smoky oak]

I kinda do, so I know what to set them to 😅

[dense tapir]

what card?

[smoky oak]

ohhhh, so like its the index? The smaller the less it holds on to?

[smoky oak]

GPU?

[dense tapir]

yes

[smoky oak]

3060ti

[dense tapir]

128 is good

I see some anime with dim size 8

[smoky oak]

I am not actually sure where to change DIM size

I see the resizer thing, but nothing pertaining to DIMs in the training settings

[dense tapir]

about 16k per dim so a 100 dim is about 164m

[smoky oak]

or does it always train at 1024 and you prune it down?

[dense tapir]

I don't use what you do so no idea what it is doing

[smoky oak]

hmmm

oh!

Network Rank

[dense tapir]

yes

[smoky oak]

its at 8 👁️

[dense tapir]

alpha make it the same as it

8 is LOL

[smoky oak]

its at 8, with an alpha of 1

[dense tapir]

yeah, SHITE

make both 128

[smoky oak]

Jeez, so you're telling me I can get better results with LoRA's? they already come out so good, i couldn't imagine haha

[dense tapir]

these things can go all the way to 768 but past that all kinds of errors happen I have been told

[smoky oak]

is this gonna affect VRAM or speed?

[dense tapir]

not really

[smoky oak]

ok, cool

just wanted to make sure cause I ride at like 7.7GB VRAM when training haha

[dense tapir]

yeah, I can't train locally so I get frustrated all the time

[smoky oak]

Aha, I see

Now I wonder about the mixed/save precision, Learning rates, encorder rates, unet rate, I don't know what any of these things mean 😅

[dense tapir]

you should go search and find out cause there is no one size (simple button) fits all.

[smoky oak]

very interesting. Here is hoping I can find some info haha

[dense tapir]

precision make those bf16 always

[smoky oak]

mine are set to fp16

[dense tapir]

bf

[smoky oak]

change it to BF?

[dense tapir]

yep

way better

[smoky oak]

alright, you got it man

[dense tapir]

I can't do bf or I would

damn colab can't do bf either

[smoky oak]

Any of this stand out to you?

like, anything I should change?

There is also this

oh wait, I don't want buckets, my images are precropped

[dense tapir]

yep

[smoky oak]

so, nay red flags I should change?

[dense tapir]

a 3060 might require the gradient checkpointing

[smoky oak]

isn't that to save VRAM?

[dense tapir]

yes

at the cost of speed

what BS are you doing?

[smoky oak]

I have had no issues with VRAM

BS?

[dense tapir]

batch size

[smoky oak]

as in the number of images? Sorry if I am missing something obvious

oh, my dumb ass

I use 2

uses about 7.7GB of my VRAM. 1 uses about 6.1

[dense tapir]

see

I run 14.7gb of 15gb on colab

you should be fine then

[smoky oak]

alright, cool

[smoky oak]

so, out of everything else, any red flags to you?

[dense tapir]

nope, try it and see.

[smoky oak]

I am gonna retrain a LoRA that didn't do too well to see if there are any major inprovements

[smoky oak]

sounds good :>

@dense tapirOk, my bad, I just realized

I didn't load the config file that AItrepeneur gave

[dense tapir]

I didn't like his config

[smoky oak]

that one ises, 2 batch, BF16, and 128/128 DIM/rank

any major changes?

[dense tapir]

just try it

Not like it will blow your house up

lol

[smoky oak]

true haha, just wanted to see if there were any major changes you would make, cause this one is his preset

[dense tapir]

worst case it has a high loss, or worst a NaN for loss.

[smoky oak]

alright, cool

[dense tapir]

the automatic1111 extension is a bit flaky though

[smoky oak]

I am using Kohya SS

[dense tapir]

yep, I know

[smoky oak]

Oh, I thought you meant something different cause this one is its own separate thing, not an extension

[dense tapir]

yep

there is an extension (or was) and it was crashing on people and all sorts of shit

[smoky oak]

ahhh, alright

so its not this one that has the problem?

[dense tapir]

the training extension I meant

no, that one is fine

[smoky oak]

ah, alright

[smoky oak]

@dense tapirThank you again for all of your help, I appreciate it greatly

[dense tapir]

You are most welcome. We are all in this LoRA together, but the sad thing is there are a few who gate keep the info we need.

[smoky oak]

Yeaahhhh, i try to share as much as I can. I have already recruited 3 friends/family members to running SD themselves, one of which is my uncle who is using it to conceptualize his characters he makes for VR chat commissions

[dense tapir]

Cool. My nephew is all up in this but his pc can't run it.

[smoky oak]

COME ON

gimmie 2

WOO

@dense tapirso I tried the 1e-5 speed that comes default with Kohya, and it gave me huge Loss (.2 after first epoch)

after going back to .0001, I am at .08 after the first epoc (at 100 steps instead of 150)

so 0.08 loss at 500 steps instead of 0.2 at 750

oh wait, I also changed the base model lmao

ok, maybe that was a big contributor lol

[dense tapir]

yeah, normally 1e-4 for unet and 5e-5 for te. Just remember what you see at the beginning should not be anything to judge by. A lot of this doesn't lower until almost done.

remember loss is it being spanked for getting its answer wrong so it will not do it again.

[smoky oak]

to compare, on the other it was 0.2 after epoch one, and 0.17 after epoch 2, now its .09, .071, .053, 0.335, and now its at .025 (thats where it is now)

I wanna see how much benefit each epoch gives

[dense tapir]

yeah, the LR is more inline. also I don't use those tbh

[smoky oak]

LR?

[dense tapir]

learning rate

[smoky oak]

ah

[dense tapir]

I also prefer constant with warmups

[smoky oak]

this?

[dense tapir]

only good if you use constant with warmups as the scheduler

[smoky oak]

like this?

[dense tapir]

mine is steps yours seems to be %

not that constant

[smoky oak]

oh

[dense tapir]

yep, that one

[smoky oak]

so what kind of difference does it make?

[dense tapir]

now warmup make 5 from 0

[dense tapir]

yes

[smoky oak]

ah shit

my last Epoch failed

[dense tapir]

failed?

[smoky oak]

Was messing around with overlocks, looks like I went too far lol

[smoky oak]

it returned anon zero exit value

first time its ever happened, and it happened about 60 seconds after I changed the OC lol

oh well, still got 5/6 epochs lol

[oblique kestrel]

Did you made this? From which api?

[dense tapir]

automatic1111

prompt - walking through the ocean

[glossy herald]

nice !

[smoky oak]

@dense tapirQuick question, Do you do your captions as like "woman, tall, pretty, dress"

or "a tall and pretty woman wearing a dress?"

[dense tapir]

first

[smoky oak]

ok

and if you had a subject specifically, how would you start it off, out of curiosity?

[dense tapir]

first is tagging (.txt) second is captioning (.caption)

[smoky oak]

wait what

[dense tapir]

don't get the question

[smoky oak]

tags and captions?

[dense tapir]

yep

[smoky oak]

WHAT

I have just been using .txt, thats all I have seen

ohhh, i see what you mean

my bad

the styles are captioning and tagging

[dense tapir]

people/subject/object too

[smoky oak]

so people subject object, do you have an example of what you mean?

[dense tapir]

well, class people would be a man, a woman, a child, etc...

[smoky oak]

so

"Woman, Ariana Grande, pink dress..."

and so on?

[dense tapir]

actually, from everything I have read, do not name the person

[smoky oak]

oh really?

[dense tapir]

yep

[smoky oak]

I wonder if thats why I am having problems with my current lORA

[dense tapir]

might be. Remember I haven't really tried (except once) to do a lora person

[smoky oak]

ah, I see

[dense tapir]

don't name the item so use generic terms

woman, man, boy

I tag so 1girl, 1bot

*boy

[smoky oak]

oh wait

I upped the LoRA strength a lot, and now its a lot closer

with the others, if I went over like .3, it would mess up

[dense tapir]

show me one of your training images (pvt is fine if needed)

[smoky oak]

this one is still getting some artifacts

[oblique nimbus]

I lost my link to SD anyone want to help me out?

[dense tapir]

https://github.com/AUTOMATIC1111/stable-diffusion-webui

[oblique nimbus]

thank you but I was meaning the one for the web broswer

[dry crow]

Edit your webui-user.bat and add --autolaunch behind Commandline_ARGS=

The URL is also:
https://127.0.0.1:7860

[smoky oak]

Anybody know how to install extensions while SD is in share/listen mode?

[oblique nimbus]

thank you but I dont understand how to do that

[smoky oak]

here

the file you click to launch

right click, edit with notepad

[dry crow]

You right click the file and edit it

[smoky oak]

Where it says COMMMANDLINE_ARGS=

you wanna add --autolaunch

this will automatically launch when you start it

shit-

That deleted my previous arguments

Nevermind, got them back

now to figure out how to add extensions while IP hosting

[dry crow]

Disable the hosting, add extension, enable hosting

[smoky oak]

Ah, thats gonna change my local IP/URL to access stable diffusion, but that does work

[charred crystal]

I'm trying to generate images for Thri-Keen (dungeon and dragons race) and I'm failing a bit. They are an insect race, but in human form. I've tried the [praying mantis|human] notation to try combining them, but it doesn't work. Is that because there's not enough overlap in the categories?

[smoky oak]

thought I would much prefer to just be able to do it when I am authorized with my login

[smoky oak]

there should be models that actually have info on that race

[oblique nimbus]

with the --

[smoky oak]

one moment

[smoky oak]

yes

--autolaunch

no spaces, just like that

then just save and start

[smoky oak]

There is a model trained on thri-keen

https://civitai.com/models/90/dungeons-and-diffusion-v3

[charred crystal]

This is the closest I've come. That's just a tiny insect with a big sword.. 😄

[charred crystal]

thanks!

[smoky oak]

no problem :p

I am a fellow DND diffuser haha

it does illithids very well

tortle too lol

[charred crystal]

already love what you can get with the standard network (not sure what I'm using)

[oblique nimbus]

I dont get what I'm doing wrong

I cant edit it

[smoky oak]

I found the right command, it was --enable-insecure-extension-access

its password protected, so its no worry

[oblique nimbus]

ok I have to sign in or up to make changes

[oblique nimbus]

really sorry to be a bother but where do I have to open the bat file? do I need to download everything and open it with notepad ++?

[smoky oak]

I would say yes

notepad++ is great

[charred crystal]

Having some trouble getting thrikreens to be generated with 4 arms. Anything I can do with the prompt or should I hope for a model update?

[smoky oak]

its what I use

[oblique nimbus]

I'm still trying to learn this programing stuff

[smoky oak]

all good, I am shit at it, I just am a very fast lerner lol

[oblique nimbus]

ok then how do I download it from GitHub I can never find the link in all this mess

oh, and I was wondering what is sytan sound design. do you make music with some kind of software or are you working on the software for making music itself?

[oblique nimbus]

new problem python is not launching

[oblique nimbus]

nope not working

I give up

[stoic pivot]

Good morning Internet. Today I test "Illuminati Diffusion v1.1" and I do not like it, it so dark and it feel very limited. But there may be something wrong with my Stable Diffusion and Automatic1111, I often get ERROR written in the UI ang I get a feeling it not been updated even I do Git Pull at start,

I do not like "Illuminati Diffusion v1.1
" but I think that is a model folk should download and test, it may be very creative.

[dry crow]

do you want to install Automatic1111 webui?, come to #🤝｜tech-support for help

[wispy nest]

weird gen , but I kinda like it

[dense haven]

I have been trying illuminati diffusion without much success in making dark scenes. I have tried putting lamp and light in the negative prompt, and also edit the textual inversions that was suggested by the page. Not sure if it's working

[stoic pivot]

I suspect "illuminati diffusion" is trained on contrast with much dark and light and little did tones, so you will get a dark room but then there will be some lights that take a bit of the focus from the image,

[wispy nest]

ship walding repaire

[grizzled sage]

does illuminati diffusion use the offset fix that i've seen going around? i guess base SD and all the training algorithms have a bias towards averaging the brightness of a pic to 0.5. Someone released some math that helped the training process learn to create images that have higher or lower than the average for general contrast

[wispy nest]

I feel like that's just over hyped nonsense...
They are claiming it's SD noise algorithm problem, but...model, just trained on other images or Lora fixes it, even tho it doesn't change algorithm itself?

Am I not understanding something or they are saying dumb things , using smart words?

[robust oar]

attempt at making a model based on the simcity 4 art style

[smoky oak]

Oh this is fire

fantastic work :>

[smoky oak]

Don't you just love when a LoRA just doesn't work no matter how hard you try lol

[stoic pivot]

My hamster always have same or similar pose and the composition stay similar with small variation when use Illuminati Diffusion v1.1 or SD v2.1, what do I do wrong?

[smoky oak]

man, no matter how I train this LoRA< its always bad

[stoic pivot]

@smoky oak Send it to some to test, for some reason I have trouble with that most LoRA look a bit bad and grainy, it may be some setting or so, I do not know, I have no luck with LoRA.

[smoky oak]

Most of my LoRA's come out damn good, but something is physically fucking with this LoRA. Its consistently giving her details she doesn't have across several models, and I don't get it

The biggest being that it keeps giving her a dot on her forehead thats not there in any of the data set images

I really just don't get it lmao

its insane how consistently it makes somebody who looks nothing like her

it only looks like her if I specify short hair, which she doesn't have in most of the data set images

[stoic pivot]

What is up with my Stable Diffusion and Automatic1111 today, after running "Illuminati Diffusion v1.1" I get a feeling that SD v21 and SD v1.5 is messed up, even same prompt under "Deliberate" look wonky.

[meager flare]

What am I doing wrong?

[cyan snow]

#🤝｜tech-support

[opal tangle]

https://civitai.com/models/14605/howls-moving-castle-interior-scenery-lora-ghibli-style new LoRA for scenery style of Howls Moving Castle

[crystal bloom]

Using ControlNet to generate images based on my sketch drawings.

[stoic pivot]

I may been writing the prompt, but LoRA made it to "art"?

[stoic pivot]

LoRA's are crazy, and I have no idea who this is, but to use that LoRA made everything a bit creepy, this is a combination of two images.

[floral cloak]

@dry crow Just a couple examples. They are very similar, but there are loads of detail changes.

Flick between the two

No matter how many I generate no two are alike

[dry crow]

That could be the Upscaler too

Thats causing this

[floral cloak]

I've tried different ones

disabling xformers fixes it

[dry crow]

Try without one

[floral cloak]

Same issue

[strange jungle]

is same seed expected to generate exact same image?

For me, it always has tiny changes

[floral cloak]

It's supposed to. Disabling xformers arg and generating multiple images produces the exact same thing every single time.

[dry crow]

Yea xformers can cause this

[strange jungle]

exact same?

[floral cloak]

Yep

[strange jungle]

hmmm

[floral cloak]

Down to the same byte

[strange jungle]

img2img or txt2img?

[floral cloak]

txt2img

Ive not tried img2img

[strange jungle]

You want exact because you're upscaling?

[floral cloak]

Exact because its the exact same parameters and prompt surely? Is it not a bug in the xformers library?

[dry crow]

I dont think its a bug

[floral cloak]

This is the --xformers arg removed from the bat file. All 3 here are separately generated images.

if xformers introduced randomness and eliminates reproducibility, I dont see why this wouldn't be a bug?

[strange jungle]

those are surely the same

[shadow ore]

yeah i can't tell any difference

[shadow ore]

i just checked it online

if i move the slider the images are the exact same, literally zero difference

not even a minor diff

[floral cloak]

That's the point, without xformers they are exactly the same image, with xformers they differ a lot

[shadow ore]

yes

that's what happens with xformers

it literally states that on the automatic1111 wiki

Xformers library is an optional way to speedup your image generation.

This optimization is only available for nvidia gpus, it speeds up image generation and lowers vram usage at the cost of producing non-deterministic results.

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Xformers

[floral cloak]

Ahh, I didnt see that page, I was only looking at the full list

Thanks @shadow ore

[cursive nacelle]

When AGI takes over

[cursive nacelle]

just sydney carrying on Satya's legacy

[glossy herald]

Spiro

can't manage to better the last one,

[hasty nova]

I haven't trained anything in a while and I would like to train something cool for 2.1, is there a specific style or concept someone wants?
(I do embeddings)

[atomic goblet]

I hate it in Deforum when I get what would be an awesome render using 200 steps and 50 scale but it’s too distorted because those are so high and if I lower them then the render is different and way less detailed

[hasty nova]

Im thinking about stylized low poly/game asset embedding, but I don't know if I should do that for 2.1 or 1.5

I guess why not both?

[austere sleet]

⬆️ControlNet + OffsetNoise + LoRA = Stable Diffusion 3.0!

Read my Reddit Post below to learn more!

https://www.reddit.com/r/StableDiffusion/comments/11fsdtq/controlnet_offsetnoise_lora_stable_diffusion_30/?utm_source=share&utm_medium=ios_app&utm_name=iossmf

[simple hamlet]

i think i fed it with too much image to train , gpu goes brrrr

[unkempt basin]

flowers in space

[green plover]

Generated

Scaled out the way it's supposed to be

I love it when a plan comes together. 16x9 ti works! 😄

[sturdy wyvern]

cute Asian girl, long hairstyles #🏞｜general-with-images

[sinful mirage]

https://imgur.com/du0EDPW lil better, dog was going schizo in the beginning tho unfortunately, i used canny and the hed CN models

o and yknow what, i used a random seed for each img :/

[glossy herald]

yep, that would explain it :/

[sinful mirage]

i wish the pose model didnt take so long, can be so useful 😦 like 2x longer than hed and canny combined

[sinful mirage]

https://imgur.com/TQzXHX7 @glossy herald any tips? I think i just gotta let it run overnight with openpose, has odd limbs a lot

[glossy herald]

first, go on and check the setting "Allow detectmap autosaving" next time, and you'll be able to keep all the openpose precontroled frames, and reduce drasticaly the next try's time
It would also help you detect what openpose detection was buggy, and fix it, by making the frame openpose preprocessed image yourself using openpose editor
then, is that using the same seed ? you could gain a lot of consistancy through that, on the trees and the background car for example, not sure.
I find it quite good to be honest already though

I really need to start doing one like that too

[dry crow]

Is this done with ControlNet and batch input ?

Or the animate extension ?

[sinful mirage]

ah ok thank you ill run it again, it is the same seed, 75 denoise tho. I used guess mode on both hed and canny, just had to put "watercolor, guy, brown pants, striped shirt" as the prompt, using the deliberatev2 model

[sinful mirage]

its done with img2img batch and controlnet. I find a video on pexels usually, i use "free crop video" to crop to a square (if ur card can handle it then there isnt a need), then i use ez gif to turn the video to frames. I take the first img into a regular img2img page and get some decent settings with controlnet, then go to batch, link the files and an output folder, then let them generate. Then I put it back into ezgif but the other way, jpg to gif 🙂 make sure to use the same frame settings for both or itll mess with the speed

it almost understood handstand lol, using canny, hed, and openpose

[dry crow]

Wow thanks for the explanation, never tried Animations yet, mabe a project for the weekend ^^

[sinful mirage]

its super simple once you do it a couple of times, that last clip i sent took maybe 45mins-1 hour on my 8gb gpu

[dry crow]

Nice i have 8gb ^^

[stoic pivot]

Negative prompt: Popular opinion. -I think Illuminati Diffusion v1.1 just now is one of the most boring models/checkpoints.

[sinful mirage]

dmn :/ https://imgur.com/mLVl6ZH

[hasty nova]

I tried it out too, the images it can make are cool but it's way too specific and repetitive for me, even with my embedding.

[sterile kiln]

GUys, do you know a way to save the parameters of an image generation? Because the thing at the top right (for me anyway), only saves the prompts, it doesn't save the resolution, the CFG scale, the seed, the steps... etc.

[boreal hare]

Do those 3 look like one person in 3 diffrent stages of his life?

[stoic pivot]

That was same for me, can do some good and folk should test it, but it so limited just now. We hope they develop it more.

[sinful mirage]

https://imgur.com/CyMhzaU ill have to work on post processing tmrw, been a while

and i think the CN depth leres model would help with the background

[opal tangle]

https://civitai.com/models/14605/howls-moving-castle-interior-scenery-lora-ghibli-style-v2

v2 release today

[golden sorrel]

mw2 type shii

[pallid ruin]

It looks cool

[livid dawn]

Looks dope

[cedar reef]

https://giphy.com/gifs/MobileLegends-jujutsu-kaisen-mlbb-jjk-DGsDLr9nyz2LkVgKFs

[grizzled sage]

lol nice experiments. openpose model isn't trained for a lot of poses i think a lot of refinement wiill go on there most definately. particularly hand stand type styles. i discovered the limit when i was trying to generate olympic break dancing shots

[golden sorrel]

frfr

[pseudo aspen]

Hmm... Coincidence? (1st is AI generated image, 2nd is a real photo)

[dense tapir]

:/

[golden sorrel]

mute from r6 but with camo

[glossy herald]

hey people

what's the trend tonight ?

[lilac pilot]

Trying to make doodlebob 3d

[glossy herald]

ooooooo

that would be fun

doodlebob

never knew of the concept before

[lilac pilot]

Blender+ SD is amazing

slightly failed I dont know why the top and bottom got cut

[glossy herald]

i'm starting from far away it seems

I need to discover blender soon it seems

more and more things benefit from that knowledge

[lilac pilot]

Have you seen the videos showing you can use blender to get depth, canny, and open pose to fix hands and feet?

[glossy herald]

yes I saw that

I did some 3D like 20 years ago, but never touched it again since

I need to just get back on that horse

controlnet creates a real use for me right now ^^

but I may start with real life video input, I wanted to try that too

such a beast that controlnet

[lilac pilot]

yeah, I want to get into using this for videos

[golden sorrel]

this nike swoosh inside a coastline

a city, with another swoosh

[lilac pilot]

If you could do that with the offset extension for night with car lights making a neon effect swosh

[golden sorrel]

like this?

[lilac pilot]

That looks cool, not sure why i put neon, but how the lights look on hyperlapse videos

see without too much blur in

[golden sorrel]

TELL ME THE PROMPT

CUZ I CANT

THINK OF ONE

[charred crystal]

earth genasi

[lilac pilot]

not sure Im a caveman level right now. put hyperlapse lights see what it does. a lot of us are experimenting

[golden sorrel]

this

[lilac pilot]

[boreal hare]

Tried updating one of my old generation with new artstyle mix, how does the style look

[smoky oak]

looks clownish

[boreal hare]

I know, but I mean details, I tried mixing art novou with pop art

and bit of impressionist at the top

[smoky oak]

looks good! its low res, but for that res, its clean

its high detail for the resolution, is what I mean

[boreal hare]

What about this one?

[smoky oak]

That one looks sick

yeah, thats a cool style for sure

I've been flipping between hot men and wallpapers

the duality of man lol

[boreal hare]

This is impressionism, mixed with slavic pagan realist art

don't ask how I got that mix of artstyles

[smoky oak]

alright haha

but it does look very cool man

I have been focusing on wallpapers like this

[boreal hare]

It was fix of this generation I made months ago, so I am currently fixing my old generation to better with new models

[smoky oak]

high res generations mixed with upscaling

[boreal hare]

nice

[smoky oak]

oh yeah, new one is wayyy better

[boreal hare]

Here is another fix

[smoky oak]

Oh I like that one as well (she kinda looks like a kid tho IMO haha)

phone wallpaper I made

[boreal hare]

I mean, on bunny girl there is nothing suppose to nsfw, it is just suppose to be cute painting

[smoky oak]

Yeah, I get it

she just has a very very child like face IMO

[boreal hare]

True, for some reason model I am using has that problem, I sometime has to specify adult so it doesnt look like a kid

[smoky oak]

Man, I am looking at what I used to be blown away with when generating with midjourney, and I can't believe how fast AI has grown

[boreal hare]

It is not even year

MJ, Dalle and Stable diffusion all came around july

[smoky oak]

This was from less than 6 months ago man

[boreal hare]

and it progress from just couple smudges that barly resembled person to this

[smoky oak]

Midjourney and its OBCESSION with teal and orange colors lol

[boreal hare]

I want to try with new model, try to again do futurist artstyle

[smoky oak]

to think that in less than 6 months we went from this

to this

is insane

[boreal hare]

wow

[smoky oak]

might I add one was $30 to run, the other was free lmao

I love stable diffusion man. I am genuinely passionate for the art of image synthesis

I wanna see if I can somehow use SD to generate musical visuals for my album I am working on

I mean, the entire album is around raw sound synthesis, and synesthetic/chromesthetic reactions to music

[boreal hare]

Remember early MJ faces

[smoky oak]

yeah haha

[boreal hare]

This was once taken as peak of face generation

[smoky oak]

yeaahhh haha

now we can make basically photo realistic faces consistently lol

This was my first megabatch

and every single one is good. I just find that insane

[boreal hare]

Yeah, even finger in Lada models became less of a isue

[smoky oak]

for being a single generation across 64 seeds, every single one is unique and properly captured the look of Lady gaga in just 4 seconds an image. Insane

[smoky oak]

I'd go as far as to say I get pretty reliably good hands

[boreal hare]

not alway. At now it is like 50 50. But it is so much better then 1 in 1000 chance of having good hand in what was

before

The first one is image I tried my best to generate for hours and I needed a lot of work to have bearly human anatomy, but now I used it on new model and it is wow

[smoky oak]

I wonder if there is a way to interface with SD to do frame by frame auditory visuals that respond to specific notes and chords. Like sample the sound at the section, and associate specific colors with specific chords/notes. That would be an incredible way to have synesthetic visuals

[boreal hare]

I think some of more complicaded LADA models are using that strategy

[smoky oak]

Only issue is I have no idea how to code or anything, so I have no idea how I would even try to bring that vision to life

[boreal hare]

wish you luck

[smoky oak]

I really wanna release my album along visuals that fit the message, but man thats gonna be soooo much work 😅

[boreal hare]

I had just finished with new models first of my colections of gods and goddesses

[smoky oak]

Just this single sound right here would probably produce some insane color responses just off of note association, if that was a thing

[boreal hare]

yeah

[smoky oak]

(here it is in the little micro demo for my album)

The whole thing is based around alternative synthesis and audio scapes/expressions

oh, I forgot my icon is animated lol

I thought I was tripping balls lol

[cursive lance]

nice song

[smoky oak]

Thanks!

i have 2 more demos (much older and lower quality, but hey)

A-rythmia is 3 concepts for a single song together

and the intro is the placeholder album intro (its LOUD at the start)

I am sure you can hear how much lower quality their sound design and mix downs are 😅

Truthfully, me getting into stable diffusion was cause I wanna make art for my album

[cursive lance]

[fringe pecan]

Pixar style, a super cute and happy white fairy bunny, pink headphone sports car, bright big eyes, wearing a small yellow flower skirt. Sweet smile shiny snowflakes fluffy, fluffy tails, fairy tales, cherry blossom forest background, flying cherry blossoms, bright colors, natural light, ultra Clear, ultra realistic, ultra fine, Illusion Engine Octane Rendering 8K, 3D, HD-g 2-v 4-ar 2:3

[hollow grotto]

how

[smoky oak]

how what?

I was just showing midjourney 6 months ago vs stable diffusion now

[hollow grotto]

did you make that pic

i just really like the style

and composition

[smoky oak]

I did, i used 7th layer anime v3-a

Was my first ultrawide generation heh

now I make more realistic things

like this

just made this one too

[mystic marlin]

Here, have a free GTM_v3 generated Mustang. Great model by the way.

[glossy herald]

@tired nymph charming Polish village, complete with colorful traditional houses, cobblestone streets, and a lively town square

trying now with another model

[tired nymph]

thank you :3

damn those are really great, ai is great

[smoky oak]

damn thats well done

decided to try something

256x256 to 1024x1024

368 to 1024

512 to 1024

768 to 1024

and finally, RAW 1024

same prompt, same seed

crazy how base res affects it

and I would say they all have their own strengths at 1024 final

tho speed is not one of them

[charred crystal]

how long does 1024x1024 take? takes forever on my 1070

[smoky oak]

256x256 took 4.33 seconds
368x368 took 4.39 seconds
512x512 took 5.12 seconds
768x768 took 9.34 seconds
1024x1024 took 18.02 seconds
256 to 1024 took 15.91 seconds
368 to 1024 took 18.56 seconds
512 to 1024 took 16.77 seconds
768 to 1024 took 20.54 seconds

not sure why 368 to 1024 took so much longer

weird

using 30 base steps, 20 upscale steps, CFG 8

[charred crystal]

with what nvidia card?

[smoky oak]

so yeah, 512 to 1024 seems like the best option

[smoky oak]

3060ti

oh, I am using euler A

[charred crystal]

strange, does not seem much faster than my ancient one

[smoky oak]

its more about VRAM speed than core speed

I can send the exact gen settings, if you wanna compare

oh, I am sure the model matters too tho

[charred crystal]

512 to 1024 took 1m16 so yeah it's a lot slower

[smoky oak]

oh, ok yeah haha

76s vs 16.77

but still, thats not too bad

my very high res gens take longer than that, granted at much higher res

[charred crystal]

Still only 8 GBs of video ram. Haven't really kept an eye on it. Is that standard?

[smoky oak]

I have 8GB VRAM as well

[charred crystal]

I mean, my 1070 also has 8 GB

[smoky oak]

xformers allows me to go way higher res

are you using xformers?

[charred crystal]

don't know what that is

[smoky oak]

you use a1111?

[charred crystal]

yes

[smoky oak]

let me do you a massive favor

xformers made my GPU way faster and I can go to absurd resolutions now

you know the bat file you launch?

you wanna edit it and add --xformers

after the COMMANDLINE_ARGS=

Make sure to add no space

you should generate faster, and be able to go way higher resolution

this 1280x540 to 2560x1080 just took 69.34 seconds

[glossy herald]

yep, xformers is a must-have option, just pure positives on vram and speed

[smoky oak]

before I had xformers, my highest res I could do was 1024x1024

[charred crystal]

only for upscaling or also for generic 512 exploration?

[smoky oak]

now I can go way past 1080p

[smoky oak]

everything

[glossy herald]

yep, it helps on everything

[charred crystal]

why is it not enabled by default 🤔

[smoky oak]

its faster, and uses less VRAM/allows using shared graphics RAM

[glossy herald]

it's just an upgrade on the cross attention thing

[smoky oak]

new tutorials tell you to enable it by default 😅

its fairly new

[glossy herald]

compability issues with some harware maybe ?

the base cross attention is maybe easier to run

[charred crystal]

I followed an old tutorial and was stuck on 1.5 for two days until I discovered 2.1 was out 😓

[smoky oak]

it does only work with NVIDIA, I know that

[charred crystal]

does SD work on amd cards as well then?

[smoky oak]

and it benefits different NVIDIA cards differently

for me it went WAY higher res and faster

[glossy herald]

both have their uses. right now, there is also the "controlnet" extension that does magical things, but only works on 1.X models

[smoky oak]

I use only 1.5 as of now. No reason for me to go to 2.x

[glossy herald]

like this is base

this is controlnet

[charred crystal]

Combining things (like earth and skin) seems to work MUCH better on 1.5. 2.1 seems to get stuck on real faces

[smoky oak]

2.1 is a mess TBH

[charred crystal]

should I try 2.0?

[smoky oak]

removed nudity, trained with incosistent resolutions, blegh

[smoky oak]

I use strictly 1.5 as of now

[charred crystal]

don't care for nudity

[smoky oak]

and don't think 1.5 can't give the same quality as 2.x, cause it very much can

[dry crow]

Only use 1.5 or 2.1

2.0 is just not good

[smoky oak]

All raw 1.5 results, no in painting or editing

[charred crystal]

anyone excited for 3.0?

[smoky oak]

the version matters much less than the quality of the model

realistic vision v1.3 on SD 1.5 will still blow basically all 2.x models out of the water, cause of how well it was trained

all images above are raw, unedited realistic vision results

I generated them myself :>

@charred crystalquestion, do you use just like.. SD 2.1 base release?

[charred crystal]

I'm liking the animal/human hybrids more 🙂

[smoky oak]

ahem

[mystic marlin]

Random McLaren incoming.

[charred crystal]

as in for D&D, not nsfw

[smoky oak]

Like these?

just an example

[charred crystal]

haha no

[smoky oak]

thats anything furry v4 on SD 1.5

used it to generate my character for DND

[mystic marlin]

Oh, here is one I made a few days ago, Tifa. No, this is not from the actual game either.

[smoky oak]

@charred crystalso really, do you just use basic SD 2.1?

no specialty models?

[charred crystal]

no, 1.5 works better for weird stuff like Earth-Genasis

I tried the D&D thing for 1.5 but then you get bog standard stuff

[smoky oak]

What are you just using the official release?

[charred crystal]

for now

[smoky oak]

Oof, the base versions are not very good haha

[mystic marlin]

I haven't even touched SD 2.0 or 2.1 after trying them the first time. 1.5 is just....better for me all around otherwise.

[smoky oak]

I've been using stable diffusion for less than 3 weeks, and I can honestly say that every version 1.5 model I have used has been better than every version 2.x model I've used

[charred crystal]

same

[mystic marlin]

Now that we have Noise Offset for better contrast and darker images, total freaking GAME CHANGER to where 1.5 can directly match to even surpass Midjourney.

[smoky oak]

Stable diffusion already beats mid journey in a lot of places, like accessibility, granularity, confidentiality, diversity, cure resolution / detail output on the high level, the ability to train on whatever you want, and way way more community support/new features

But now that we can reasonably expect weighted contrast sometime soon, I don't really see what the benefit of mid journey would be overstable diffusion anymore, other than the fact that you don't need a powerful computer to run it

[mystic marlin]

I mean, just the Auto webui alone with how much control we have is more than enough to kill it in my honest opinion, but yes absolutely.

[charred crystal]

haven't tried midjourney and now I won't 😉

[smoky oak]

Things like control net also just bitch slap mid journey haha

[mystic marlin]

ControlNet is next level shit for us.

[smoky oak]

I was one of mid journeys original supporters, I supported it for three or four months, and racked up about 15,000 generations, granted it was very early in the beginning, before stable diffusion was released, and then I left about a month and a half after stable diffusions release

[mystic marlin]

I paid for it once and after SD came out I ditched. lol

[smoky oak]

Mid journey was fun back then, before they decided to kill it by censoring the fuck out of it

[smoky oak]

I tried stable diffusion when it first released, but it was so fucking trash back then, that I just went back to make journey, before I couldn't afford to continue with mid journey and I ended up leaving image generation

Sorry, I'm using voice typing so it's a little off

[mystic marlin]

I don't support the views of the CEO of MJ regarding censorship at all....

[smoky oak]

Yeah, it's completely bullshit. I don't know if it's as bad as it used to be, but when they first started cracking down on censorship, it was game ending for me

It was so unbelievably censored that you could do barely anything at one point

[mystic marlin]

Hold on a second...I have just cooked up something amazing for you....

[smoky oak]

Wanted an image of a minor in any capacity? Against TOS, want anything related to bruises, looking rough, torn clothes, grungy styles, all against TOS

Anything referred to as blood red? Against TOS, short skirts? Against TOS, gym clothing? Against TOS. Men not wearing a shirt in any capacity, even if it makes logical sense like them being at the beach, or being professional weightlifters? Against TOS. It was fucking horrible lmao

I canceled my subscription shortly after

It was a lot worse than just that, but I don't remember any other examples off the top of my head

[mystic marlin]

Damn, I forgot how rough that actually was.

[smoky oak]

Though I will say that I think it is absolutely bullshit that stable diffusion also censors content

Like on one hand I get the message, but on the other, it's just as annoying to deal with as mid journey at times

[mystic marlin]

You talking about Dream Studio, yeah?

[smoky oak]

No, I'm talking about stable diffusion in general

There are significant parts of it that are censored, and while I can understand parts, some of it's just ridiculous

[mystic marlin]

Oooooh, you mean like what they did to 2.0 and 2.1?

[smoky oak]

Nope, even before that

Try and generate a sword with stable diffusion

It won't do it, because they removed any and all weapons out of the data set

A gun? Nope. An axe? Nope. A tomahawk? Nope.

Hell, they went so far as to remove the guns on military vehicles, as well as the missiles on military aircraft

My friend was going to pay me to generate some concepts for cyberpunk/steam punk battle axes and warhammers for a character that they're creating, and I spent probably a solid 2 hours trying to do it, and it just would not do it

[mystic marlin]

Oh...now I see what you mean. Yeah...you're right on that one again. Weapons are a no-go.

[smoky oak]

Like I can understand them censoring illegal stuff like inappropriate underage content, and stuff like that, but a sword?

[smoky oak]

That seems ridiculous, good luck generating any convincing looking knight lol

I was trying to generate buff lumberjack men cutting wood, and it would always remove the axe

Granted, you can add it back in by manually training it yourself, but who the hell wants to do that for all weapons?

[charred crystal]

hmm

[mystic marlin]

We can barely even get lightsabers to come out decently. lol

That alone is a one in a million.

[smoky oak]

Exactly, I feel like them censoring such a light content as a hand knife or a crossbow is just very much against the open source nature of stable diffusion

It becomes basically useless to generate any form of melee-based D&D character, unless you want to generate them without their signature weapon

[smoky oak]

Thank you for reminding me, I want to see if I can find a way to generate hot cyborg men lol

[mystic marlin]

Well, at least I got this image back in January based upon one of my favorite visual novels. lol

[smoky oak]

Oh wow, that looks dope

[young solar]

Think any of that will improve with later versions, or is just up to the community?

[smoky oak]

It's not likely, stable diffusion seems to be taking away more than they're giving

However, we should all know by now that it's the community that makes stable diffusion, not stable diffusion itself

So I'd likely say that as more people find out about stable diffusion, and higher and hardware becomes more accessible, we can expect the quality to go up just on community support alone

[mystic marlin]

Alright, this is what I've been waiting to put up here for the last few minutes. My last and favorite image of the night.

[smoky oak]

That looks very nice

[dry crow]

Thats my Car in Saints Row 3

[young solar]

I would hope that after some of the hype, attention, and controversy dies down, the official models become a bit more broad. I have certainly run into the mangle weapons.

[smoky oak]

They will for sure, just think about when stable diffusion 1.5 came out, and how much better stable fusion 1.5 is now, that wasn't because it's stable diffusion, that was because of the people in the community putting in the time to train the AI better and better and better

[young solar]

Frankly, anything that someone holds is hard to generate. I tried to get a janitor holding a broom, and only the smallest part of the handle would show up.

[mystic marlin]

The Huracan didn't even exist until 2014 funny enough. Might have been based on the Sesto Elemento which predates it while being similar. XD

[smoky oak]

Yeah, I'd have to say that my biggest complaint was stable diffusion is that it doesn't generate what you ask for, it just generates in the general idea

That's something that mid journey still excels in, it's text encoder and understanding of what you're asking is significantly better than stable diffusion

However, that comes at the expense of a ridiculous monthly payment, limited access, limited image rights, and supporting a company that frankly I don't agree with

So I'll stick with my higher quality stable diffusion generations, at the cost of them not looking exactly how I ask them to be lol

[mystic marlin]

Oh don't you worry...one day we'll wield enough power to make this damn AI give us what we want without question. Then we take the world.

[young solar]

Yeah, prompting with SD is an art in and of itself. I have noticed that sometimes it is fine generating something by itself, but not in combination with other things. Like, heaven forbid you want only a partial cyborg in a few different models. I had to use inpainting with a different prompt to get one.

[young solar]

But I am just a noob, so maybe that is just flaws in my prompts.

[smoky oak]

Haha, yeah

[smoky oak]

Oh yes, one of my weaknesses, hot cyborg men haha

Something I have struggled greatly with as well

It also doesn't help that there is very limited content of anatomically mimicked cybernetics

[young solar]

Yeah. I am only trying to get a specific limb each time, then match it with a particular outfit. I have been learning a lot about inpainting, img2img, and upscaling. Most of the original image is gone by the time I have something I like.

[mystic marlin]

Johnny Depp?

[smoky oak]

That's fair, I only use in painting for specific things

[dense tapir]

yes

[charred crystal]

It also doesn't help that there is very limited content of anatomically mimicked cybernetics
I wish it did not need to have a source of that. Just smash the two topics together pleas. This in the shape of that or with the texture of some such

[smoky oak]

I wonder if I could train a LoRA on that NFL robot mascot. He's so hot, and for what lmao

Sorry, I am just too into men

😩

You can't tell me the man that made this robot does not have an eye for the male body lmao

[charred crystal]

Missing a cybernetic member?

[smoky oak]

Pffft

I can dreM

*dream

[mystic marlin]

Ha, over in the anime channel people had a tendency to make the toned gals tonight. A lot of guys seem to be into ripped chicks these days.

[smoky oak]

These people know what they're doing

They're cyborg baiting as hard as Disney furry baited with Zootopia lmao

Zwierzogród

Not even sure what language that is

[charred crystal]

polish?

[smoky oak]

No idea

Just the fact that it says rod on the screen next to those tigers-

Lmao

[charred crystal]

it's polish for zootapia 😄

[smoky oak]

Ah!

The more you know

Still feels intentional :p

[charred crystal]

yes I'm definitely going to use this knowledge

[smoky oak]

Also, complete sidenote

I love how different the tiger designs are between sing and Zootopia

Both look really good, but they are very different in a unique way. Just props to the character designers

[smoky oak]

Disgusting lik

*lol

[glossy herald]

love that model ^^

https://huggingface.co/dallinmackay/Cats-Musical-diffusion

[charred crystal]

haha omg