🏞｜general-with-images

[feral pebble]

Interesting. I've done much inpainting (and never had good luck when I did).. probably should watch some videos

Comfy UI looks neat. I'm not sure how much I feel up to fiddling with nodes, but the power is appealing

And the efficiency

[sterile kiln]

2h i'm working on

[wild sorrel]

you need inpainting model for inptinting, I suggested outpainting tho...it's adding new thing to your image, expanding it to the sides

[oak osprey]

where did you get this photo of me

[sterile kiln]

It is Gerard Depardieu, a famous french actor

[weak sage]

guys you gotta try genning with 8 steps of 2m sde karras the new model, its so fast and good

[oak osprey]

that sampler is broken

if you go look at #sdxl you'll see the devs mention they're using it to break the bot's outputs

[weak sage]

doesnt look broken to me

[oak osprey]

it generates incorrect images, which you probably can't tell in your result there

try DDIM

[proud dagger]

Funnily enough - you chose exactly the perfect image to hide the issues lol

the key issue with broken samplers is weird colorations and splotchy oversaturations, which is... exactly the intended content here

[weak sage]

xD

[wispy nest]

one more day and we can play with XL again lol

[split rover]

I mean the bot and clipdrop are open for the moment if you are itching to use it haha, but yea it should be going up for api and platform tomorrow

[spiral cipher]

Come to the Dark Side! We got Nuts!!!

[wispy nest]

I thought it was off Clipdrown til tomorrow, didnt know it was back. noice.

[livid ermine]

Working on a next generation Starship replacement 🤣

That version there sucks, I'm starting over

[shut sinew]

Where does the flag come from

[oak osprey]

some prototypes

[paper quail]

Exams incoming...

[oak osprey]

AI cryptoid party

love the cool dude

[wet warren]

can i run a sd2.1 model with a 1.5 lora?

[cyan snow]

Dude, you just gave me an idea. I'll try to get mini Shrek, LOAB, and elder magic in the same photo

[cyan snow]

Yes

[wet warren]

ty

[oak osprey]

shrekuamole

[smoky oak]

Anybody know if there is a way to turn down the image preview res in SD for inpainting? Cause I have a 4k image loaded in there and it lags my whole UI when trying to scroll or move sliders

it doesn't lag for img2img, but it DOES for inpainting

[fleet goblet]

Icecream.

[smoky oak]

Wow that NAILED the sub surface soft surve texture

[fleet goblet]

ArtiusV21 is really good.

The fact I can generate 1280x1280 images is insane to me.

1280x1280 at 1.40/s.

[smoky oak]

Wow, thats artius v2?

[fleet goblet]

Yeah.

Takes me 28 seconds per image.

[smoky oak]

Color me impressed. Thought it was SDXL, but I did think it messed up the cone

but I don't even care, cause that icrecreame surface is so damng ood

Looks very very good

[oak osprey]

cryptoid hunting without negatives is wild, man

[smoky oak]

Is it a hunt if they are so obvious? lmao

[oak osprey]

no idea what that is, looks like a human vase and i don't even know

[oak osprey]

@smoky oak when i showed you this a while back you were super impressed with it too, that's Artius

native high res is pretty good on there

[smoky oak]

I wasn't impressed with it, cause of the duplicates, but it is cool how well it can not kill details it seems

[oak osprey]

the background you were

[smoky oak]

though I remember it still having the crunchy 2.1 look when I tested

[oak osprey]

maybe, were you trying 768x768? they look weird

[weak sage]

Does the CFG number affect stuff like loras and hypernets

[nova hound]

Can you create allah pic

[livid ermine]

I think it was a kid who drew it

And the prime minister saw it and said "This will be our new flag!"

IIRC

[shut sinew]

I meant like on the image lol

[wet warren]

is it normal that it takes so much time to switch to a new model

[fleet goblet]

Yes. It should only take a couple seconds.

[wet warren]

oh it is already done

this might be a stupid question but how can i toggle loras on?

[wet warren]

i read online that it should be included in the prompt but i saw youtubers having it in their ui

[oak osprey]

got it

[fleet goblet]

Negative prompt: Disfigured, Mutated, Multiple Arms, Legs, Fingers, Toes, Limbs, Ugly, Multiple beaks, Multiple eyes/lenses, Duplicates.
Steps: 20 | Sampler: DPM++ 2M SDE Karras | CFG scale: 6 | Seed: 1270141109 | Size: 1280x1280 | Model hash: 4ee4331304 | Model: artiusV21_v21 | Clip skip: 1 | Version: 4867daf | Parser: Full parser```

[fleet goblet]

I don't trust it.

[wispy nest]

It's not when something one think should work but didn't, it's when something one thought wouldn't work…instead did work. :P
I learned something new today by the way ;P

[oak osprey]

me too. i can now make realistic hands

[wispy nest]

well, at least I'm past that as well, it was such a rough road from 1.4->1.5->2.0->2.1->1.5 :P

[cyan snow]

whoa, is this a new AI crypted

[oak osprey]

that's houseparty shrek

[cyan snow]

also, this is smile cube, the oposite of LOAB

smile cube like water

[oak osprey]

that layout of people cracks me up, how could they even pile up like that

[cyan snow]

this is a perfect example of something 2.1 makes

[oak osprey]

well good news bucko, it is

[cyan snow]

i knew it, i don't know how to explain it, but you can immediately tell when an image was generated with a model based on 2.1

[oak osprey]

because you can tell it's the best we'll ever be at AI images

[wispy nest]

I know a couple of ways, but I believe we all know which they are in the end ;P

[oak osprey]

i literally prompted for an AI cryptoid housewarming party

what do you expect

[cyan snow]

idk man, i bet if i would prompt it with 1.5 or SDXL, it will be entirely different

[oak osprey]

probably. i bet if i went to my photo model it would be better too lol

this WIP checkpoint has some contrast issue i'm working out

[cyan snow]

2.1 is kinda sometimes in the "uncanny valley"

that's the best explanation i can think of

like, with 1.5 models you can often fool people that the image isn't made by AI, but that's much harder to do with 2.1

[oak osprey]

looks pretty real to me

[wet warren]

is there a way to fix the fingers?

[cyan snow]

hey bro, nice settings you got there =]

also yes, loras and hires. fix

[wispy nest]

1.5 has a lot more painting/digital art styles. Haven't seen 2.1 have much to do with anything but photo stuff. And when it comes to photos then 1.5 also have a lot more specific styled realism than 2.1.

2.1 has stuff which is better than 1.5 that 1.5 doesn't have, but I don't think a lot of people enjoy that "style" so it's seldom even thought of.

[wet warren]

what lora? i am using a character lora

[oak osprey]

2.1 can make very real looking people idk what y'all talking about

[wet warren]

reminds me of the "name one thing in this photo"

[cyan snow]

that is uncanny, i can still immediately tell this is 2.1

[oak osprey]

@cyan snow really? because it's an SDXL image

[cyan snow]

WHAT

no way

[oak osprey]

it's a prompt, check out my cousin larry in his dirty room, he is a damn mess

i got it from Emad's twitter

[wet warren]

best upscaler?

[wispy nest]

no image the ai creates is so real that someone will never note something weird about it. I mostly think about how many wants to create X type of style in the first place. That's why 1.5 stuff is so much more popular, it's because it has everything people want from an ai. Read into it how you'd want, but it's not about what a single person want, but what most want. :P

[cyan snow]

i usually use 4x_uniscale_soft. that upscaler is decent for pretty much anything

[wet warren]

hmm ok i will try it

wait where is it

[wispy nest]

of those I'd say it's ESRGAN_4x for detail, and Latent for the rest as long as you have high extra noise

[wet warren]

i don't see it in the chat

[cyan snow]

it's a custom upscaler from the upscaler wiki

[wet warren]

ty

[cyan snow]

but yeah, also ESRGSN is also pretty good

i personally uniscale_soft

never disappointed me

[wet warren]

fuck

[cyan snow]

ESRGSN moment

[wet warren]

i went overboard with the upscaling lol

[wispy nest]

No problem, but I'd also say that there's no one answer. I recommend you to try them all and get a feel which might be better at what. They are all useful as long as you know how they, or what they, make :D

[cyan snow]

i usually set it to 1.5 upscale and 0.55-0.6 denoising strength

that should do the trick for weird fingers

[wispy nest]

there's a extension named ultimate-upscale-for-automatic1111 and it can do wonders, if you learn it. Something I've not, but people at least say it's really good! :D

[wet warren]

wow

[wet warren]

you generate 512x512 and upscale it to 1024x1024?

[wet warren]

still some weird fingers but it is still majestic

[oak osprey]

does the native 1024x1024 not give it away @cyan snow ?

[wet warren]

i love ai

[solid gate]

any idea how to solve this? I get ERROR when I activate the "only masked" on inpaint

[cyan snow]

hell no, i do 1024x768 than upscale it to wallpaper. i can do this because i optimized my inference and i have a 4070ti

[wet warren]

i will try downloading it ty

[cyan snow]

yeah, me too. i never tried it before

[wet warren]

i am using a 8gb rtx 2060 super ;-;

[wispy nest]

if you use the sampler DDIM then try another one, if not, try another one as well, and if that doesn't work, then I got no clue from that image alone, sadly :(

[wet warren]

btw where can i find the denoising strenght option

[solid gate]

what is DDIM?

[cyan snow]

then you should ask someone with a similar GPU, i know the best settings for 40xx cards, i have no idea how to get most out of a 2060 super

[wet warren]

how many hire steps?

[wispy nest]

the sampler for the images. Sadly, if you don't know that part, it's really hard to help. :(

[cyan snow]

you can see what i use in the image i just sent

[wet warren]

what does hires mean btw?

i just started using SD yesterday lol so i am kinda new

[solid gate]

oh yeyeye im go to see

[oak osprey]

what is in the training set that it can make these messy areas LMAO

[wispy nest]

it creates the first "pass" of the image at one size, then upscale it to another larger size with a upscaler
it's like a combo of txt2img and then img2img

[solid gate]

damn no work, F

[wet warren]

fuck it generated two raidens

[wet warren]

Ei and Makoto maybe

hmm maybe the generator knows genshin lore

[oak osprey]

hires is one of the greek gods. like hermes

[cyan snow]

damn, anyway, the best advice i can give you is when you get a new GPU, try to get the latest gen CUDA GPU if you ever buy another one

[oak osprey]

get an H100

[wet warren]

woah what a nice tip of course if i had the money for a new gpu i would buy it

[cyan snow]

bro they are new, don't confuse them

[oak osprey]

:|

[wet warren]

i should try adding 1girl to the prompt

[oak osprey]

i don't think that does anything

[cyan snow]

yes, and add 2 girls to the negative

[cyan snow]

it does, trust me, i know

[oak osprey]

most models it does not, and just need better aspect bucket training

[wet warren]

it kinda looks cool tho

i saved it anyways

[wispy nest]

if you use a anime checkpoint trained on one with the list of tags, then adding solo will probably be enough as a normal prompt

[cyan snow]

isn't that something that's actually in that game?

i remember there being 2 girls like that in genshin

[wet warren]

i am using meinu unreal

[wet warren]

yes ei and makoto

makoto is dead tho

died like 500 years ago

you are probably thinking of ei and yae miko

[cyan snow]

no no, there was the chick that lives in her own world, and the one who is her puppet or something. they both looked the same

[wet warren]

man i love ai

[wispy nest]

don't know which one that is, but most anime styled checkpoints use the best quality tagging from my knwledge :P

[wet warren]

raiden ei and raiden shogun? that would probably not make sense because they share the same body

[wet warren]

ALMOST five fingers which is good enough

now i will watermark it with my name

[wispy nest]

highres fix is a thing I use to…fix fingers, it's not a 100% sure thing, but better than not using it :P

[wet warren]

lora:raidenshogun1-000009:1, raidenshogundef, upper body, smile, blush, outdoors, day, simple background, blue sky, short hair, sky, temple, looking at viewer, stairs, mountain, moody lighting, facing viewer, 1girl
Negative prompt: lowres, ((bad anatomy)), ((bad hands)), missing finger, extra digits, fewer digits, blurry, ((mutated hands and fingers)), (poorly drawn face), ((mutation)), ((deformed face)), (ugly), ((bad proportions)), ((extra limbs)), extra face, (double head), (extra head), ((extra feet)), monster, logo, cropped, worst quality, jpeg, humpbacked, long body, long neck, ((jpeg artifacts)), deleted, old, oldest, ((censored)), ((bad aesthetic)), (mosaic censoring, bar censor, blur censor), 2girls
Steps: 28, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 2409588786, Size: 1024x1024, Model hash: d1c96253ed, Model: meinaunreal_v3, Denoising strength: 0.55, Hires upscale: 1.5, Hires steps: 27, Hires upscaler: ESRGAN_4x, Lora hashes: "raidenshogun1-000009: 8c2e28fdf17c", Version: v1.3.2

[wet warren]

i use hires

[cyan snow]

yeah, if without it AI fucks up 40% of the time, with hires it does like 20%

[wet warren]

OH hires means high res

[cyan snow]

yes

[wet warren]

i am getting smarter everyday

[cyan snow]

bro, we are the complete oposites with AI, we make the most different kind of art

[wet warren]

this is sdxl?

[cyan snow]

nope, the 1.5 model i made

[wet warren]

because my ass 8 gb vram card will never do something like that

[wet warren]

oh goodjob

[cyan snow]

i bet it can, with lower settings

[wispy nest]

it's not the vram, it's how well you've learned how to use the ai, you'll get there in time don't worry! :D

[wet warren]

i am starting to think to switch from --medvram to --lowvram

[cyan snow]

yeah man, imagine what model i will make when SDXL releases

[oak osprey]

none because your system won't be capable

[cyan snow]

WHAT

[wet warren]

i didn't include ass shot in the prompt but the ai can read my mind... :)

[cyan snow]

what makes you say that

[wet warren]

i am getting annoyed by the fingers

should i lower this?

[cyan snow]

yeah, a little bit

like to 35

[wispy nest]

everyone writes their prompts different, mostly because of different checkpoints used, but also because of the style one wants, opinions etc.
I'd write the entire prompt like raiden girl, solo, portrait, blush, outdoors, day, short hair, buddism temple, pov, step by step, Fiji mountain myself for example, with no negative prompts at all. I add negative words when I notice a consistant error/thing I don't want after the fact I get them.

But I also probably don't want to create the same thing as you so what I showed might be no good either :P

[wet warren]

one with 16 hires steps

[wet warren]

i have to add the kew word of the lora

[cyan snow]

like, even the base un finished SDXL0.9 can compete with this, so imagine what it would be if i do the same thing i did to 1.5 to SDXL

[wispy nest]

not if you include it by calling it with the <lora:> thing

[wet warren]

no i have to include both

anyways it won't really affect the fingers that much

[wispy nest]

well, I never do and it seem to work, but I haven't used loras all that much yet so I might be missing some details around them :)

[wet warren]

it is a character lora soo

[cyan snow]

you can also stack loras, you can use lora:add_detail and lora:good_hands and whatever lora is for that character and it shouldn't effect the speed

[cyan snow]

also, character loras are usually less good than character embeddings

[wet warren]

what does character embeddings mean

[cyan snow]

it's a different kind of model, one that usually does specific characters very well

[wet warren]

where to find them

[wispy nest]

no idea if this is raiden, the only raiden I know of is the highlander one from the first mortal komabt movie :P

[cyan snow]

civitai.com

then filter to textural inversion

it will show you embeddings

[wet warren]

hmm

nobody did a raiden embeddings

[cyan snow]

https://tenor.com/view/pepe-lonely-sadge-sad-pepe-lonely-pepe-gif-24375295

[cyan snow]

then stick with the loras =/

it's basically the same thing, exept loras can sometimes effect speed

[sweet sandal]

Why are you interested by this concept?

[cyan snow]

idk man, it's pretty cool

[sweet sandal]

Yes, I know you think that. I implied that when I stated you were interested. I asked why.

[cyan snow]

why not? it's just a prompt i like

[oak osprey]

way to low-key imply you hate it

[sweet sandal]

That's a passive-aggressive reflex of yours. I'm asking him.

[sweet sandal]

Yes. Ask yourself and answer: Why?

[cyan snow]

i can create whatever i want and there is nothing you can do about it, that's the whole point of stable diffusion. i don't need to take instructions from people, i can make whatever kind of art i want, i don't judge anyone here.

[sweet sandal]

You are reading what I'm saying utterly incorrectly. There is no judgement. Tell me why.

[wispy nest]

high clip skip can be weird :P

[wispy nest]

well, back to secure normal clip skip instead :P

[jovial umbra]

hi whats the diff between Clipdrop and Dreams websites?

[wispy nest]

Dreamstudio doesnt have 0.9 til tomorrow (i think)

so right now clipdrop has the newer model

[jovial umbra]

ok thnx so SD owns 2 websites?

[jovial umbra]

I just want to support the team so are both Dreams and Clipdrop owned by them?

[oak osprey]

@jovial umbra yes sir

[jovial umbra]

ok thnx

[cyan snow]

is that what i just made but in a comics style?

[oak osprey]

yea

it's what my pseudo-journey-v2 model biases to when i don't prompt for a specific style

[cyan snow]

oh, that's weird. so it's a comics model? the prompt i used doesn't say anything about a cartoony-ish style

[oak osprey]

its just a weird model lmao

it can do a lot

it has biases toward Midjourney 5.1 style, that's all

[cyan snow]

i can give you a copy of mine if you want to generate that kind of image

[oak osprey]

oh i just like riffing on concepts

[cyan snow]

if that's what MJ would make then there is no need for SDXL, my model is also meant to be almost identical to MJ.

[oak osprey]

well it's a lot of the 2.1 weights. when i do the MJ prompts on that model, they look totally different

it's kind of sad because i wanted more of certain stuff out of it, but it's not represented often enough. like i can't make a kid version of Keith Richards, but there's a fair number of images of him in the training dataset

[oak osprey]

i didn't want to damage the text encoder, so, it just didn't "over-learn"

[wispy nest]

and I just merged a little too much toward yellow! >:I
Gah, will fix that tomorrow after me beauty sleep :D

[cyan snow]

yeah, overcooking a model is never a good idea

[oak osprey]

i tried harder and i accidentally beat that SDXL image

[cyan snow]

eh, but does it beat what my model makes?

[oak osprey]

probably not 😛 you have that concept down to a science

i am but a casual visitor to your domain

[livid cedar]

Should my stable diffusion look like this? I git pulled and it looks like ass now

[livid cedar]

can't even read model name

[oak osprey]

they come out very cute on my model

[cyan snow]

dude, you think it's possible to merge a 2.1 model and a 1.5 model

[oak osprey]

nay

you can generate images for each other but that's about it

to train on i mean

[cyan snow]

idk, model checkpoints are basically huge math equations with billions of parameters, if you can combine them on a specific formula, it might be theoretically possible

[oak osprey]

that's just sounding like training with extra steps

[cyan snow]

like, keep the architecture of one, then multiply they're data on eachother

[oak osprey]

[cyan snow]

i feel like it's mathematically possible tho. i'll look into it later

[livid cedar]

Is my previously posted image how the automatic1111 gui is supposed to look now, or have I done something wrong

It looks kinda ugly now

[cyan snow]

#🤝｜tech-support

[livid cedar]

Sure thing

[weak sage]

On one hand, AI letsyou do crazy stuff like this , but on the other, it doesn't know better when a ask it to draw a girl holding a candle, it draws a girl with a candle lit on top of her face

definitely good for creativity tho

[cyan snow]

-_-

[weak sage]

nice hands, did you just gen that?

[cyan snow]

yep

[oak osprey]

wanna buy some candles, man?

just got the fixed contrast ckpt uploaded and running too

[cyan snow]

and that's just a prompt with 3 words, if you are more illustrative with your prompts you'll get exactly what you prompted

[weak sage]

no i get stuff like this xD

[cyan snow]

idk, works fine for me =/

[weak sage]

what model u usin

[cyan snow]

1.5 finetuned model i made

[oak osprey]

looks like she's a mousepad

[cyan snow]

what does that mean?

i never heard someone describe a face as a mousepad XD

[cyan snow]

oh no

why does it look like it's in pain

[cyan snow]

i just prompted: anime holding a candle

xD

[oak osprey]

didn't take much fine-tuning with offset noise for darkness to come back

[weak sage]

do you guys ever make a bunch of images at like 4 steps, pick the good seeds, and redo those seeds at like 12 or 24 steps

[weak sage]

prompts for a firefly
gets some anthro furry stuff i shall not post
xD

[oak osprey]

i finally trained myself into the AI so i can put me in fun places

[weak sage]

hopefully funner places are coming soon

[oak osprey]

he doesn't seem to want to leave

[smoky oak]

Anybody have any ideas on where to post/warn people about a scummy Etsy based AI art employer?

[oak osprey]

well assuming it won't get you banned, on etsy

[smoky oak]

Good point, though I'm not really trying to warn customers as much as I am people who are signing up to work with them, such as myself

[oak osprey]

well that's why i thought you would do it tho lmao im confused now

but you can't rate buyers? maybe i dont get whats happening actually

[smoky oak]

They are scummy people with sweatshop wages who took advantage of me, and are currently trying to take advantage of other people.

I want to warn people in the stable diffusion community about signing up with them, because they are not nearly as good as they may seem on the surface

[oak osprey]

they sound like someone else

[oak osprey]

probably reddit tbh

[smoky oak]

That was one of my ideas as well, I just wanna make sure that its well known that they are taking advantage of their artists

[livid ermine]

Sorry what did you use here? Looks very cinematographic as if taken straight out of a movie

Damn man you got some nice work

I have to do a bunch of stuff in post-processing to achieve somewhat of a cinematographic look (ie. www.dannyhiele.com) like writing scripts like https://gist.github.com/basicfeatures/a083eb1ac1602f583d4993c3aca65358 and various color graders/VSCO presets

[oak osprey]

that's my checkpoint i just finished today

and, thank you 🥹

that earlier version wasn't the final one

[livid ermine]

Can I send a PM?

[oak osprey]

go for it, but i am busy

might take a while

[livid ermine]

No probs

[livid ermine]

Your message could not be delivered. This is usually because you don't share a server with the recipient or the recipient is only accepting direct messages from friends. You can see the full list of reasons here: https://support.discord.com/hc/en-us/articles/360060145013

[oak osprey]

🤷i can't send one either lol

[livid ermine]

Very odd

Try adding me as a friend

Prolly your Do Not Disturb thingie

[oak osprey]

just love this model so much. this is without LoRA or TIs

[smoky oak]

Yeah, so that shitty comission compant banned me and with held my pay, luckily I have all of our chat logs and stuff, so fcuk them

Their name is WattleArtCreations, and they scammed me, underpaid me, and also purposely tried to feed me wrong information to try and upsale the customer into forcing them to have to buy more revisions

When I demanded contact with the client to know they were not feeding me false info, they banned me for "unreasonable claims"

I worked nearly 20 hours on a single art piece of which I was making sweatshop wages on cause I was passionate about the client and character, cause the client was very nice, and now I have been screwed out of my pay, and they have been screwed out of my work

[oak osprey]

man, that blows. i really wish there were better ways of holding people like that accountable

[smoky oak]

Same here, it is so shitty of them

[sterile temple]

have you set up your own store @smoky oak to work directly with clients?

[smoky oak]

I have plans to, but etsy has been giving me account issues, so I will just have to make a new account after trying to salvage mine for well over a month

I suppose I can put this image up for what I am capable of doing

@median wren

[oak osprey]

don't let it get to you, and i mean that in the best way, i don't want some one stupid person to change you or how you trust others, you are fine the way you are and anything that disrupts that would suck

[smoky oak]

#1100170312106127410

Thats where you use them scifi

[smoky oak]

I'm not used to you being so nice!

[sterile temple]

I just watched a video last night (turned out to be click bait) about someone trying to set up a ai art store on etsy. they said new accounts get blocked by default and you have to wait weeks to appeal it

[median wren]

Right I figured out where to run it but I was looking for a guide like with the styles and things

[oak osprey]

i really thought about my words that time

[median wren]

All new stores or ai stores? I haven’t set mine up yet to use but I created it recently i guess That was good

[smoky oak]

which is why I wanted to use my old account, its 5 years old, but apparently Iit got hacked over covid and somebody opened a fake shop and scammed somebody, but they keep putting me in the "process to reclaim it" then they just don't

[oak osprey]

but i've been through that too, and it's just, it can really affect your outlook. and i don't want to brush it off. but you're stronger than those folks and they needed you, not the other way around

[smoky oak]

I am already putting stuff out on reddit about it, I have the screenshots from our convos

I really REALLY just wanna get back in touch with the client, cause they were so nice

[sterile temple]

I got the impression all new stores. Might be a way for them to 'manage' all the spam stores by just making people appeal and wait a while

[oak osprey]

yeah focus on that part, silver lining

[smoky oak]

oh, they are built in, if you need any more info, I am down to help

[median wren]

Glad i finished the setup portion then at least

[oak osprey]

can you still do what they were needing?

you were having issues with the training

[median wren]

cool I just wondered what the options were

[smoky oak]

I have 0 contact with the client, cause they were all power trippy about intermediating everything

[oak osprey]

ohhhhhHHHHHH

red flag right there

[smoky oak]

oh, very different thing, thats self employment LoRA training

[median wren]

like I assume there are only specific styles built in

[smoky oak]

this was an art comission

[oak osprey]

you're so many-faceted

[smoky oak]

oops lmao

AHA

SCIFI

it was the slash dream

thats why it blocked it lmao

[oak osprey]

yeah they do that so the channels don't get flooded with that

[smoky oak]

yeah, for sure

[oak osprey]

🫡

[smoky oak]

we were trying to figure out why when I was showing them, SDXL lmao

I assume they were in the wrong channel haha

[sterile temple]

I asked about the styles prompt in the bot, and the answer I got is it's basically just adding the word to the prompt. But i guess it's there to get people to try the different ones they know work.

[smoky oak]

I try to be, I do all sorts of creative things, but the one time I find something where I make a little money (even if it is pathetically little, AND They take 60% of it for doing jack shit), I was pretty desperate to get some money, and now I have been screwed out of $65 or whatever it was

[smoky oak]

yeah, and there is always a negative as well, its a default one that we can't see

[median wren]

Wait what do you mean

[smoky oak]

if its outside of the bot channels, its blocked

[median wren]

Ok so the styles pop up? I guess I didn’t realize that

[smoky oak]

yeah

[median wren]

Oh

[sterile temple]

you can tell they are constantly tweaking the settings. last night I was playing around with it and the quality suddenly dropped to really ugly images 😄

[median wren]

Maybe I used dm by mistake

[smoky oak]

See! I noticed that as well!

I was so confused why they were so bad all of a sudden

[sterile temple]

they need a thumbs down button lol

[median wren]

Thanks I didn’t see it before

[smoky oak]

RIGHT

[smoky oak]

No worries, glad to be of help

[median wren]

Why do i keep getting its voting time and it won’t work. I thought you just had to vote on your own but i voted a few other peoples now but it still keeps saying that

[sterile temple]

there's a showdown every 30 minutes where they block image gens for a little while and encourage people to vote in the #1100484581037195384 channel

[median wren]

Thanks. How long does it last

[sterile temple]

couple of minutes?

[oak osprey]

it takes so long to wait

[median wren]

Ok cool

[sterile temple]

you can scroll down to the bottom of the channel to see the countdown, the bot should be available again shortly after it hits 0

[wispy nest]

Hello, this is image generated in SD (txt2img, controlnet segmentation). Now I need just to change the pendant to different one. Is there a way to accomplish this via img2img, please? Thanks.

[sterile temple]

try inpainting using an inpainting model and describe the light you want?

[cunning geode]

i compared my models photorealism vs. that of deliberate, dreamshaper, and realistiv vision
mine is arguably worse, but also doesnt have any noise fix applied and i would argue my model is more diverse reagrding the output, e.g. the clothing, faces, backgrounds, etc

the grid doesnt show the negative prompt but it was:
anime, cartoon, digital art, cgi, render, 3d, drawing, sketch, instagram, pastel, dada, zombie, ugly, surreal, text, watermark, abstract, old, fat, jpeg, black and white, vintage, amateur, film grain, evil, damaged, concept, unfinished, model, cover, clay, figure, toy, pixelated, bad, inexperienced, illogical, random, oversaturated, overexposed, rough, fake, unrealistic, sloppy, artificial, low budget, unprofessional, cropped, out of frame, low-quality, poorly drawn, deformed, bad proportions, malformed, imperfect, unnatural, extra, rushed, weird

[sterile temple]

resolution, sampler, steps, cfg?

[cunning geode]

full generation parameters:

medium-shot photo of a woman, shot in 4k high-quality with a Fujifilm X-T3 camera with natural lighting and f1.6 bokeh applied
Negative prompt: anime, cartoon, digital art, cgi, render, 3d, drawing, sketch, instagram, pastel, dada, zombie, ugly, surreal, text, watermark, abstract, old, fat, jpeg, black and white, vintage, amateur, film grain, evil, damaged, concept, unfinished, model, cover, clay, figure, toy, pixelated, bad, inexperienced, illogical, random, oversaturated, overexposed, rough, fake, unrealistic, sloppy, artificial, low budget, unprofessional, cropped, out of frame, low-quality, poorly drawn, deformed, bad proportions, malformed, imperfect, unnatural, extra, rushed, weird
Steps: 35, Sampler: Euler, CFG scale: 7, Seed: 525484821, Size: 512x512, Model hash: fd7ce58be7, Model: st-AI-le_st-AI-le_v1.0, Denoising strength: 0.4, Hires resize: 1024x1024, Hires upscaler: R-ESRGAN 4x+

i guess i could have done that from the start lol

[oak osprey]

i achieved better photorealism than realistic vision using 2.1

[cunning geode]

@sterile templethat 2.1?

[sterile temple]

no it's a 1.5 mix I've been playing around with for a couple of months. photorealistic pretty women is not my main goal though 😄

more creative arty stuff as the goal

[cunning geode]

yeah its not my main goal either
my model combines multiple very distinct styles, photorealism is only one thing it can do

[oak osprey]

i was wondering why the skin was so excessively smooth

[oak osprey]

@smoky oak i'm so stoked The Darkness has returned

[oak osprey]

ai cryptoid party?

[jovial umbra]

Launch day of No Mans Sky

[oak osprey]

wat mine do

[jovial umbra]

cant wait for open source release of new model in July

[oak osprey]

apparently that is midjourneys latest beta

looks as good as SDXL

[wispy nest]

"good" can mean a lot of things. It's hard to judge something without all the details I'd say. For me, someone liking their art is good. but that doesn't mean I think the art is good, just that it's good that people get what they want :P

[oak osprey]

https://huggingface.co/spaces/allknowingroger/Image-Models-Test10

apparently my flex-base model is now running here and can be tested on this space 😄

damn. my pseudo-journey-v2 model is there too. it looks soo good

[wispy nest]

humble as always? ;P

[oak osprey]

that test uses no negative prompt

you tell me which img you prefer 😛

epiCRealism does a really stellar job too

[wispy nest]

for me, it would be the second to last one, the one at the bottom, three from the left

[oak osprey]

that one to me looks like they hyper-trained the model on porn

it doesn't follow the prompt very well

[wispy nest]

you asked which one I prefer, and I answered :3

[oak osprey]

the black squares are NSFW images

[wispy nest]

if it's about perfection, then the three black ones are perfect because they don't have any errors in it! ;P

[oak osprey]

their error is that they don't follow the prompt at all

it's just asking for a white woman in white clothes on white background

the anime ones are wow, broken looking lmao the studio one is nice but that's not a white background

[wispy nest]

the black ones might be because you didn't tell it to toggle the lights in the room? 🤔🤣

[oak osprey]

or put her clothes on 😛

[wispy nest]

but you said the one you liked the most was because it followed the prompt, but isn't the background gray and not white? :O

[oak osprey]

holy cow what the burned model hell

that's hilarious, it made her disheveled as hell

epiCRealism ballin' result again

this prompt is more generous to each model

except pastel-mix which still looks like trash

[wispy nest]

was my first result on the random model I had loaded :P

[oak osprey]

nice! if it can also do a completely darkened room then that model is likely pretty good if it has a balanced training set too

[wispy nest]

judge out on that ;P

[oak osprey]

looks like it has an averaging issue

oh well, not every model needs to make dark images 🙂

[wispy nest]

I've only seen loras be able to get good (in my eyes) images so far. Might be because it's hard to find many images taken that's all dark and not have some sort of light source

[cunning geode]

i got this random action pose and decided i like it so much i gotta inpaint it and fix some stuff

[oak osprey]

it's because of the noise issues with the training schedule

[sterile kiln]

Sorry to repost,
but i'm very proud of it

[wispy nest]

as I don't understand what that mean, I'll take your word for it :P

[oak osprey]

the reason most models have issues with brightness is because "inference sampling is incongruent to training sampling."

in plain english, when you are training, you take a sample and iteratively add noise to it. the last timestep is the noisiest.

oppositely, during inference, you take a sample of noise from the data distribution and you iteratively remove noise from the image.

these are literally opposite processes. training is a forward process and inference is a reverse process.

the proposed issue states that the inference process should match the training process. it should start at the last ("trailing") timestep and work "backwards" to the first timestep ("leading").

this means that the inference process picks up where the training process left off, with the beginning state of inference being conceptually similar to the ending state of training.

i'm sure that still doesn't quite help explain it but at least you can take any single part that confuses you and beg GPT to break it down further 😄

[wispy nest]

"plain english" :P
It reminded me of this simple example of how to explain something to someone who don't understand the technical stuff

[oak osprey]

[wispy nest]

it's so simple! everyone knows this ;P

[oak osprey]

😮 someone made it so my model can be used directly on the model card page

that's gonna suck if it doesn't use negative prompts lmfao

[wispy nest]

I'm too shy for uploading my weird mergers because people will probably laugh! and then laugh again when they look at me 😶‍🌫️

maybe they do so when I say stuff like, "It's true, mardi gras, is really just a color!"

[oak osprey]

Absolutely, your explanation is solid and clearly outlines the concepts. Let's try to simplify this further:

Let's imagine you're in art class. You're given a beautiful, clear picture of a landscape. Your teacher asks you to add layers of paint splashes on top of it, each layer making the original picture more and more difficult to see. After several layers, it almost looks like a mess, nothing like the original picture. This is like the "training" process in the model. You start with the original image (the 'sample') and add more and more noise (like paint splashes), until you end with something very noisy.

GPT4 is trying to help me explain it to a high school student

Now, imagine you're given a similar noisy, messy painting, but this time, your task is to gradually remove the layers of paint splashes until you uncover the beautiful landscape beneath. This is like the "inference" process. You start with a noisy sample and gradually take away the noise, aiming to reveal something that looks like the original picture.

The problem is, sometimes the picture you reveal doesn't look like the beautiful landscape it should. It's because the process of uncovering the picture isn't matching the process of covering it up perfectly.

The proposed solution is to make sure the uncovering process starts exactly where the covering up process left off. That way, every step in revealing the picture directly corresponds to a step taken while covering it up, increasing the chance of revealing a beautiful landscape.

In short, the proposal is to make sure that the process of revealing the picture (inference) matches perfectly with the process of covering it up (training). By doing so, we improve our chances of creating better and more accurate images.

i asked GPT4 "but this sounds like counting 6 eggs from left-to-right vs 6 eggs from right-to-left. why would "sampling timesteps in reverse" be an improvement?

That's a good point, and your analogy holds in many situations. However, when dealing with diffusion models and similar types of generative models, the process is not symmetric, and thus the direction matters significantly. This is due to the inherent nature of the data transformation that happens in these models.

During the training phase of a diffusion model (the forward process), the model starts with real data samples and progressively adds noise until it transforms the data into a predefined noise distribution (often Gaussian). So, each forward timestep brings the data closer to a noisy state.

In contrast, during the generation phase (the reverse process), the model starts with a sample from that noise distribution and tries to reverse the transformations applied during the forward process. It takes a noisy sample and step-by-step reduces the noise, aiming to generate a sample similar to the real data it was trained on.

The critical part here is that the transformations applied at each timestep are not symmetric; the changes made to the data during the forward process can't simply be undone by just "counting backward". Instead, the model has to learn a separate reverse process to go from the noisy state back to the original data.

For instance, if the model has learned that at timestep 3 during the forward process, a particular type of noise is added that blurs the image, the reverse process at timestep 3 needs to be able to undo that blurring. But the reverse process can't simply "remove" the blur; it has to generate details that look like they might have been there before the blurring occurred, based on the current state of the image and what it has learned during training.

So, when the suggestion is to "start from the last timestep", it ensures that the model begins the reverse process from a state that is equivalent to the one at the end of the forward process. That way, each step in the reverse process corresponds directly to a step in the forward process, which is important for maintaining consistency between training and inference.

[wispy nest]

:)
I like nr 3 and nr 4 the most ;P

[oak osprey]

starting at the basics takes too long

start at the end and work your way backwards 😄

basically start sampling your learning from the last timestep

[wispy nest]

b-but that's even longer! 😵‍💫😵

[oak osprey]

NO because you have a shortcut to the things you don't understand

the pyramid of knowledge is widest at its base, and there is so much to learn

if you go from the top down, you find out exactly what you need to know

[wispy nest]

hehe, this is the reason I'm so simple minded, well except being simple minded, I just want to give/recieve the quick answer to something. Like, if someone asked about how to get better hands for the billionth time (yes, that was very passive aggressive, sorry :( ) then I'd rather say "it's luck" than, becoming the real life example of this masterpiece video:
https://www.youtube.com/watch?v=aW2LvQUcwqc
:P

[oak osprey]

but for me i just give all the info and answer questions about deficiencies in understanding. i won't know all the answers and those we find out together

[wispy nest]

I always got answers, might be good answers, but they are still an answer :3

[oak osprey]

trying to give foundational knowledge to someone whose background i don't know, is impossible. a lot of the time you end up discussing shit they already know, and in the process, boring them...

[wispy nest]

might not be*

[cunning geode]

i took a shot of that prompt in my model
right is without negative prompt, left is with
it decided to ignore "inside cafe" tho lol, prob because you also have "storm clouds" in there

[oak osprey]

using textual inversions or LoRA?

[wispy nest]

those shades she wears in the top right is killing me! hahaha, it's so good! :D

[oak osprey]

same

[cunning geode]

my own model, a 4gb checkpoint (or 2gb pruned) https://civitai.com/models/97393/st-ai-le

autumn portrait of a russian young woman inside cafe, knitted fox-hat, fog, close-up, storm clouds, wearing red sunglasses, red shirt, (artstation, deviantart)
Negative prompt: anime, cartoon, instagram, pastel, dada, zombie, ugly, surreal, text, watermark, abstract, old, fat, jpeg, black and white, vintage, amateur, film grain, evil, damaged, concept, unfinished, model, cover, clay, figure, toy, pixelated, bad, inexperienced, illogical, random, oversaturated, overexposed, rough, fake, unrealistic, sloppy, artificial, low budget, unprofessional, cropped, out of frame, low-quality, poorly drawn, deformed, bad proportions, malformed, imperfect, unnatural, extra, rushed, weird
Steps: 35, Sampler: Euler, CFG scale: 7, Seed: 525484821, Size: 512x512, Model hash: fd7ce58be7, Model: st-AI-le_st-AI-le_v1.0, Denoising strength: 0.4, Hires resize: 1024x1024, Hires upscaler: R-ESRGAN 4x+

[oak osprey]

@cunning geode nice, in 2 months i spent $25oo to train my 2.1 model 🙂

[cunning geode]

right is same just without the negative prompt

[wispy ether]

Spam

[cunning geode]

its just the experimentation with captioning, datasets, hyperparameters, etc that took so much money

[oak osprey]

your goal to have a flexible model is noble

[cunning geode]

and its still not perfect 😢

but good enough

[oak osprey]

i hated all the 1.5 and 2.1 fine-tunes that end up nuking a lot of the base knowledge

well you can't do what you've done without damaging some of it. but it's a trade-off. and by "what you've done" i mean, training it so far that you don't need negatives

[cunning geode]

yeah i tried to preserve that, even merged some of the 1.5 text encoder back in
you still gotta emphasize morgan freeman by like 1 or 2 () if you wanna show him on a surfboard, but at least it works

[oak osprey]

did you freeze any layers of CLIP? it doesn't have very many, and it is easy to screw it up as a result.

[cunning geode]

depends on the prompt, i noticed that the photos look a lot worse without the negative prompt

[oak osprey]

well yeah. but they're not truly incoherent like stable-diffusion-1-5 or 2-1

[cunning geode]

no freezing

[oak osprey]

my Pseudo Journey v2 model is the same. when i run negative-free tests it has the BEST outputs of any 2.1 model

but it gets left behind when you start using negatives

[cunning geode]

is that the one with zero terminal snr

[oak osprey]

all of my models have zero terminal SNR but only pseudo-flex-base managed to keep it stable

[cunning geode]

yeah i saw that, and experimented a ton with zero frequency noise and terminal snr, as well as v prediction on 1.5, tons of testing, in the end i found that nothing worked without making the quality worse

so v1.0 has no noise fix for now

[oak osprey]

interesting

[wispy nest]

autumn, portrait of a slavic woman, coffee cafe, knitted fox ears beanie, depth of field foggy, storm clouds, red-tinted sunglasses, warm red shirt, artstation, deviantart
Would be how I'd write it, but I'd also not write it like that if I wanted to give a cherry picked image to show how awesome I am.

grumble, why it not make fox ears >:I

[oak osprey]

i had to kick the model in the face at the very end with 1 epoch of SNR gamma and 1 epoch of offset noise

otherwise it wouldn't generalise contrast

[cunning geode]

my model includes anthros and such, hence its probably better with the fox ears

[oak osprey]

it's actually supposed to be her holding fox ears in her hands, from the original image i lifted the prompt from

they used prompt editing stuff i removed tho

[wispy nest]

nah, it's probably because you used evil black magic or voodoo (not evil voodoo, just the normal kind) ;P

[cunning geode]

after having just deleted 32gb of 3 months worth of test samples, i am glad i can finally rest some and just generate cool images with my model

[oak osprey]

yep it was like a sigh of relief to delete 480GB of training data and test samples

think i sighed so hard i shit myself a bit

[wispy nest]

what? create images instead of critiquing someone elses images, and/or the tech behind INSERT_VERSION_HERE against INSERT_OTHER_VERSION_HERE?! :P

[oak osprey]

it's frustrating seeing amazing images being made in here and i'm like halfway stuck between two checkpoints and every model i have access to is TRASH

[cunning geode]

ive never actually used other peoples models lol, only ever my own test models

[oak osprey]

i started out with Deliberate and pixel upscaled 512x512 images to pretend we had high-res and then i learnt about controlnet tiles and then i figured out 2.1 fine-tunes are superior to 1.5 fine-tunes and set out on a quest to create a 2.1 multi-aspect contrast-fixed base model for others / myself to fine-tune from

that all began about 3 months ago 😄

i think the best part was writing my own fine-tuning scripts and trying to implement any missing pieces that i want from A1111 using Diffusers directly

[cunning geode]

oh yeah i dabbled with 2.1 too, but all i ever got was trash even with radically different settings, e.g. TE layer freezing etc

[wispy nest]

for me, 2.1 just didn't give me images I wanted. 1.5 was a lot easier, and because I'm lazy, I don't always finish what I've sta

[cunning geode]

whats that one famous american oil painter called with all the landscapes? he also drew the one with the falling roman empire
not thomas kinkade

[oak osprey]

contrast comparison of my 2.1 f/t vs 2.1-v and realism-engine

[cunning geode]

ah thomas cole

[oak osprey]

and multi-aspect support comparison. my model is a base resolution of 1024x1024

[cunning geode]

yeah my first published model was 1024 res, but i realised it probably doesnt increase quality enough to be worth the time and effort

so i stuck with 512 now

[oak osprey]

to reflect upon your assertions about 2.1, it is really hard to train because of the amount of noise remaining in the text encoder and the unet, but you can't train both together

you have to first fine-tune the text encoder, and then the unet (or the other way around) but fine-tuning both together makes the weights too much of a moving target and convergence is like the last thing it's going to do

this is a 2.1-v model fine-tune. the base model res of 2.1-v is 768x768, and the latent space internally is 96x96 as opposed to 64x64 for 1.5's 512x512 model

[wispy nest]

took me way to long to remember, not remember, and then look up the artist named JAMES GURNEY to say that I liked his artstyle a lot, and even the style the ai can create using it :D
But yeah, I like it, might not be the same for anyone else hehe

[oak osprey]

2.1 loves the art of Bob Eggleton and Darryl Eggleton but you have to put one or the other's names into the negative prompt, so that it doesn't blend their art style

[wispy nest]

Darryn Eggleton?

[oak osprey]

sure, him

[wispy nest]

typos in the negatices can have fun results :P

and no, I didn't make that typo above on purpose…

[cunning geode]

i just realized that sometimes including artist names can really help with my model
(((oil painting))) by william adolphe bouguereau and thomas cole of korra with lh hairstyle as a superhero with cape vs. (((oil painting))) of korra with lh hairstyle as a superhero with cape

[wispy nest]

yeah, artist names are one of the best things in my opinion, it's so sad to see people wanting them gone :(

[oak osprey]

well i am one of those that want em gone 🙂

[wispy nest]

it might even be the best thing for the ai in my eyes as well

[oak osprey]

i don't think adding the artist improved the style much, but it improved coherence? same seed

[wispy nest]

I write my prompts by what they do, not what I want them to do
That is, I learned that prompt x creates one thing, and prompt y creates something else. So if I have to write a ice cream colored hat to get the correct color I want, then so be it :P

[wispy nest]

heh, didn't know 2.1 also had that detail on the forehead :P

[cunning geode]

fuck is this good

[oak osprey]

forgot the training has kept going and after another 400 steps it's losing its dark contrast again that shit is so unstable

this kid finally figured out how to open the pack of Marlboros and it turns out it's a lollypop

@cunning geode concert crowds are one of my favourite unsolveable AI gore prompts

[iron field]

Mary Blair Lord of the Rings

[cunning geode]

whats the prompt

[oak osprey]

A cyborg playing a futuristic instrument in a crowded concert

HRA_hyperrealism_art is great without negatives, damn

[cunning geode]

with some proper manual editing by an actual artist, this shot could easily pass for real in some cheap isekai anime

[novel ravine]

https://civitai.com/models/97799

[cunning geode]

https://tenor.com/view/perfection-michael-fassbender-perfectionist-perfect-gif-7910211

[oak osprey]

how A1111 produces images

fuck that thing sucks ass

it has a broken tiling implementation too

[cunning geode]

i generate all my stuff with A1111 and never faced such issues

[oak osprey]

i'm using a 2.1-v model and nothing in A1111 is really optimized for these, it seems like they're begrudingly supported rather than something they really want to work properly

a lot of issues people face with 2.1 is just due to this broken shit in A1111

[wispy nest]

weird that no one noticed, wonder if it might be because no one uses it as to be noticed? ;)

[oak osprey]

well, no one uses it because it's been noticed and won't be fixed. the people responsible for fixing it, do not care about 2.1. my assumption for why that is, they want to make booba and it won't

[wispy nest]

if people are creating things they want, then it's good :D

[novel ravine]

get religion VAE

[oak osprey]

VAE is included in 2.1 and in fact you have to add --no-half-vae to sdwebui-user.sh so it will even produce images without "NaN tensor produced in VAE"

this thing is broken

[novel ravine]

i am using sd 2.1 768 right now with no extra arguements

just --xformers

[oak osprey]

mine is trained using a new noise schedule from the "schedulers are flawed" paper, and requires sampling from the trailing timestep first - which, A1111's implementation of DDIM doesn't do. i don't think there is a single sampler in A1111 that works correctly

[wispy nest]

if the results from people I've seen are using a broken sampler, then I don't want it fixed :D

[oak osprey]

euler A gets the closest but looks all blown-out and hyper contrasted because it doesn't apply CFG properly

[oak osprey]

this mindset, to me, as a developer, is really frustrating. just to let you know 🥹

[novel ravine]

`waltbrba lora:HeisenbergDogu2.1:0.72 ,in a restaurant, (pizza on plate:1.25), , solo, portrait shot, 8k uhd, dslr, high quality, film grain, Fujifilm XT3
Negative prompt: (closeup:1.2), nfixernext nrealfixer Neg_Facelift768
Steps: 40, Sampler: DPM++ 2M Karras, CFG scale: 5, Seed: 2739604252, Size: 512x768, Model hash: b346aa1648, Model: rmadaMergeSD21768_v70, Denoising strength: 0.2, Hires upscale: 1.5, Hires steps: 30, Hires upscaler: 8x_NMKD-Faces_160000_G, Lora hashes: "HeisenbergDogu2.1: 533dae38bd9c", Version: v1.3.2

Used embeddings: nfixernext [046b], nrealfixer [ff26], Neg_Facelift768 [c1dc]
Time taken: 29.63sTorch active/reserved: 5895/7370 MiB, Sys VRAM: 8192/8192 MiB (100.0%)`

[wispy nest]

as a developer, it's not to me :3

[oak osprey]

i didn't expect DPM++ 2M would work so well

but the paper for DPM starts out with the assumption that t=T, so, it implements the congruent sampling schedule

[novel ravine]

it IS the best sampler

[oak osprey]

probably for A1111 yeah

[wispy nest]

I just click on the combination of numbers and letters I find most funny when picking the sampler, most of the time :P
other than the ones that gets those blue artifacts

[novel ravine]

DPM++2M SDE Karras is also good if not as good, but produces monstrosities every once in a while

[oak osprey]

DPM++ SDE is not good

[novel ravine]

yeah hit n miss

[oak osprey]

DPM++ 2M SDE is whackadilly too.

[iron field]

I mainly use this one, what's the difference between it and just 2M?

[wispy nest]

sometimes it's also not the hammer, but who wields it ;P

[novel ravine]

https://civitai.com/models/95270/kim-jong-diffusion

many previews were made with dpm 2m sde karras here

[oak osprey]

i get that you're trying to be insulting in a clever way but look at the 'correct' results for the same prompt and tell me that the others seem to be functioning properly or that it's a prompt issue. it's NOT an issue with the prompt. these samplers will not do contrast correctly, no matter what you prompt

technically, this follows the positive prompt of a smiling woman, i guess, but it doesn't have any coherence to it, so, DPM fast = trash

DPM Adaptive produces the same shitty noisy images that 2.1 is known for

[novel ravine]

not all dpm samplers are made equally 😉

aight time for walter v2 training

[wispy nest]

I'm mostly trying not to sound mean to try and show that when one person gets results they don't like, while everyone else gets results they want. Then it might not be because of what's used, but who's using it. Insulting someone is far from my goal. I don't show bad results and tell that this or that is objectively bad, while at the same time knowing that most people who will read it are the ones who think their results with it is not objectively bad.

I like to show what I like, and results that are good to try and lift up people and show the sky, and not the easier path of bringing something down toward the dirt :)

[oak osprey]

the first four Karras samplers all either don't clean up all the noise, or they don't do contrast correctly, or the closest it gets is by hyper-smearing everything

i'm not saying "they're bad because i don't like the output", i'm telling you they are technically incorrect in their implementation

the DPM++ 2M Karras sampler does contrast properly but not denoising though it seems to be less hyper-smooth/smeared than DPM++ 2S a Karras.

[wispy nest]

sadly, I don't believe you on that. Sorry, not trying to be mean, just being honest. Saying something is this or that and make people agree is very different from showing something and making someone understand it in the way they need it to understand.

[oak osprey]

more poorly executed contrast from DPM++ SDE Karras

[oak osprey]

i'm not even the only dum dum that believes this. the stability devs did the a/b testing on the bot to showcase how bad some of these samplers are

i think their favourite bad sampler is LMS

[wispy nest]

which one of the devs?

[oak osprey]

it was literally the whole team testing that, but mcmonkey and joe penna were the ones talking about it on here

[wispy nest]

I'll see if I can ask them that as long as I know who to ask first :D

[oak osprey]

#🌠｜show-and-tell message

DPM++ 2M SDE Karras is pretty bad lmao is this on the surface of the sun?

i think most disappointing is A1111's version of DDIM because the Diffusers implementation of it works and the A1111 version doesn't know what the hell contrast means

and what is PLMS even for? 😛

[wispy nest]

thanks. :D
LMS is the one I agree with, that's the one I meant with what I said earlier with those blue artifacts. But I didn't think a dev could be that wrong to say the other one was broken. Maybe need to ask them to clarify about that later.

[oak osprey]

so DPM++ 2M and DPM++ 2M Karras are the two best samplers that have the latest and greatest science inside them

i hated DPM++ 2M SDE. it is total crap

but hey if you like their outputs, i said earlier not every image needs to be bright or dark, sometimes it just works in the middle, which is what most of these broken samplers end up enforcing

[wispy nest]

DPM++ 2M SDE doesn't seem broken to me at least. I don't get any "breaking" artifacts at all. But I didn't test more than 10 or so image generations, but that should happen all the time. Might be a 2.1 issue if so

[oak osprey]

well broken is in a technical sense. remember we're making images out of noise and it's a miracle it makes sense at the outcome at all

i picked a worst-case prompt that showcases exactly how bad samplers can be

it's honestly not even the prompt, it's the new noise schedule in my model that has a much wider range of contrast it can express, but these samplers end up like, doing the visual equivalent of a high-pass cutoff (audio thing)

[wispy nest]

I'll set my tests to DPM++ 2M SDE for a couple days and see what happens. I read "broken" as "a thing that doesn't work"
Not "broken" as in "it might do something wrong sometimes if you wait"
Because then the entire art ai would be broken :P

[oak osprey]

DPM++ 2M SDE reliably does the wrong thing on a v-prediction model with a zero terminal SNR noise schedule

maybe v-prediction is trash but it has so many training benefits

i'd rather just repair the samplers. and when they are 'correct', they have way nicer outputs, way more often

[wispy nest]

why don't you repair it?

[oak osprey]

i'm ignorant as to how it's done

the simplest samplers are fixed first

i could maybe do DDIM, because i reviewed that change for Diffusers, but i haven't really looked at it yet. that code is scary

[novel ravine]

v2 looking better I guess

[wispy nest]

I've never seen your coding skills, or code, so I'll take your word for it :P
The ai part of art is scary enough to me :P

[oak osprey]

https://github.com/bghira?tab=repositories

go nuts and see how bad my code style can be when there's no grouchy team lead to complain at me

[mental cipher]

i upscaled my image 4 times the size from 512x768, but in my outputs its under extras not under a size like all my other outputs i have filed, how do i know what the demission's are for the upscaled size, i know this is a newbie question but could someone help 😅

[novel ravine]

controlnet?

[oak osprey]

@wispy nest also the original terminal SNR paper's authors should have provided an example like this, which demonstrates the difference between 0.7 rescaled guidance and 0.0 rescaled guidance.

the 0.0 one is darker. and 0.0 also allows images to go WAY brighter.

the interesting part to me is that a rescale of 0.7 with a 'correct' sampler resembles the images you get from the 'broken' samplers in A1111

with a broken sampler and my model i was having to push the CFG over 9 or 12.0 in order to get true dark images. this added a bunch of noise and garbage to the image.

with the fixed DDIM, i can set CFG to 5.0 and get true blacks or rescale it by 0.3-0.7 and have a middle-of-the-road type image.

[mental cipher]

no, i generated the original images at 512x768, then went to extras and upscaled x4, but when i look at the png info to see the new upsclaed size, it just reads x4, how do i know the demensions after something has been upscaled?

[novel ravine]

can I tell u how I upscale stuff?

[mental cipher]

yes please

[novel ravine]

generate 512x768 with DPM 2M karras at high step count usually 40-50
if there are too much deformities like fucked up eyes etc I send it to inpaint,
after a satisfied inpaint job I send it to img2img,
0.3 denoise for little changes, 0.4/45 for some more enhancements
and use the same image in controlnet, tile preprocessor, controlnet is important check,
and I use ultimate SD upscale with scale from image anywhere 2x to 4x

done

and yeah- 4x ultrasharp for upscaling

[cyan snow]

or just get hiresfix xD

[novel ravine]

nah controlnet is good,
high res fix is for lazy ppl

or who just want to get job done

doctor heisenberg

[cyan snow]

it does almost everything you mentioned on it's own

it can be used with controlnet

[novel ravine]

you forgot a tiny detail where 8gb cards can't do above 2x in high res fix alone

u need tiling son

[cyan snow]

eh, skill issue =/

i have a 4070ti, so i guess i can't argue with that

[novel ravine]

money issue

[mental cipher]

i used high res fix yesterday lol, still learning controlnet, but i would love to really get comfortable with this type of workflow

[wispy nest]

I checked some of the code on your github, but I don't write pyhon stuff so I couldn't understand most stuff that wasn't json stuff but I sent it to a couple friends who knows more about python. But they haven't checked yet, or rather written me back :P

[oak osprey]

hires fix is interesting i like it

it works really well with the flex-base model

[wispy nest]

I'm nothing without highres fix! if that'll vanish, so will I! 😱🫣😵

[mental cipher]

512x768 upscaled 4x would be 2048x3072
is it as simple as just multiply it by 4? 😅 sorry im such a newb

[cyan snow]

not exactly, but yeah, you can say that

the main advantage of using hiresfix is that it reduces AI fingers and AI conjoined twins rate

if without it there is the AI syndromes 16% of the time, then with it it's 7% of the time

[novel ravine]

walter white - crack dealer

[novel ravine]

yes simple as that

how much vram do u have?

[oak osprey]

first time using it today for me

[cyan snow]

there is an upscale wiki with custom upscalers that are made specifically for hires

i recommend the ones on there

[oak osprey]

i'd rather just not rely on hires fix at all 😛

[wispy nest]

I used it before, and then after the whole drama around when auto1111's webui changed how it worked. I was one of them who was really, REALLY angry from it. makes my blood boil even thinking of it still...even when it's actually quite nice now even with the change :P

[oak osprey]

it certainly made it easier to clean up an image with a very basic prompt

buuuut i think less of images that are made with hires fix

[cyan snow]

it reduces the rate of AI syndromes by a lot

[oak osprey]

i don't need that with this model

[cyan snow]

i use it in almost everything i make

[oak osprey]

no hires fix, same prompt, same native res that hires fix made earlier, and still just one subject

theres a bit of residual noise but that's likely just a bug in the samplers still

[cyan snow]

the image you sent before and said you used hiresfix was kinda better

[oak osprey]

i agree, but it's a very basic prompt too

could improve it with either method

the hires fix will maintain most of the image though

the 1920x1080 native gens still have some issue but according to stability devs that's a limit of the transformers inside the older unet architecture

[cyan snow]

that's with hiresfix?

weird

[oak osprey]

without

native gens

[cyan snow]

oh, that makes sense

i bet if you used hires fix with 0.6 denoising fingers might be fine

[oak osprey]

it does 1920x1080 without dupes sometimes, without hires fix. but it's not as reliable and apparently starts to 'lose integrity' across the scene as you can tell

still i'm impressed that experiment worked as well as it did

doesn't look like it would take much architecture changes to get reliable 1920x1080 out of it

[wispy nest]

here's how I most often use the highres fix for. That is, make images larger, and also fixing some details such as the typical smudge of a face from a distance. But as seen, it's never a perfect fix. But that's what the "generate" button is for, to make another image :P

[mental cipher]

8gs of vram 3070 graphics card

[oak osprey]

yeah the one thing i'd really love hires fix for is faces at a distance but i've managed to improve them to an extent i'm super happy with anyway. these men in this image would not even register as faces for GPFGAN to fix before, but now it does

[cyan snow]

yeah, i don't know how to use hires on that

[novel ravine]

512x768 with 2x hires is the most u can do then

just like me

[wispy nest]

but sadly, it can't fix every finger, not when the first image has them as if pointing a dagger :P

[cyan snow]

try increasing denoising strength or switching the upscaler

[wispy nest]

yeah, I know of those things, or at least I think I know :P
It can probably be fixed somehow, but I just wanted to show that it can't fix everything every time. Or I can't make it do that hehe

[oak osprey]

my pseudo-journey-v2 model with the same prompt/params

it's kinda promising actually, i bet the 1920x1080 gens all look pretty good if they have mostly landscapes or abstract items rather than human subjects whose shoulders shouldn't morph randomly

[cunning geode]

@oak osprey the original rescale extension for a1111 is broken and wrongly implemented

this is a fixed version https://github.com/AMorporkian/sd-dynamic-thresholding-rcfg

[oak osprey]

i use no extensions at all

[cunning geode]

you need an extension
auto1111 doesnt have native reshale support

[oak osprey]

yeaaah i'm not going to install an extension, my point of testing is to make sure that the model will work out of the box for A1111 users

[oak osprey]

they should implement it built-in. no need to have it as an extension, since it's a core piece of the diffusion process

[cunning geode]

well it just wont ever. auto1111 has no native rescale support snd ztsnr only works correctly with rescale (altho it never did for me, it did for others). and the original extension is borked, so the one i linked is the only true working one.

people also install extensions for controlnet and such

they will manage

[oak osprey]

the negative reviews on a model don't stop coming in just because there's some obscure extension available

[wispy nest]

wait, there's a extension to change the highres fix?! I just use the one without downloading something...I think, maybe I forgot it if it's super old? :O

[oak osprey]

damn, SDXL is on a 12 billion image dataset?

[wispy nest]

this one is without any highres fix, i.e., made "just as" with settings, no loras, etc. But I forgot to change so it's using the very broken DPM++ 2M SDE ;P

now to get me some ice cream before it melts from my computer crying flames from creating a image at that size!

[cyan snow]

also, this is pretty pog

[oak osprey]

anything is possible with enough fine-tuning!

[cyan snow]

they also said that txt2video is about to be a real thing

[oak osprey]

since i've seen what i've seen, i withhold judgement on how fine-tunable something is. people told me 2.1 was harder than 1.5, and i guess they just never really tested a serious cross-section of prompts on their 1.5 finetune.. for me, 2.1 is easier. for kaibioinfo, 2.1 is easier. if everyone tells me SDXL is hard to fine-tune i'm more likely to assume it's the same as 2.1

but no one at stability has remarked on fine-tuning the text encoder(s)

[cyan snow]

there will be specialized finetuning tools

[cunning geode]

no

rescale is something else

[cyan snow]

trust me, i WILL make an SDXL version of my model, and it will be insane

[oak osprey]

those tools are unlikely to focus on fine-tuning the text encoder. it isn't something stability does

the unet can be tweaked but you can't gain cross-concept coherence with just the unet training, in my experience

the text encoder is what learns to generalise that stuff for the most part, if you want some dipshit celebrity riding a camel, it needs to know how to associate those concepts

[cyan snow]

idk what to tell you, whatever i did to 1.5 worked, that's all i know

[oak osprey]

he's in your model. he sees the filth

[cyan snow]

bruh wtf

[mental cipher]

ok so if i upscaled my image 4x already then is that all i can do with my image or can i enhance it more with inpaint now? or would it be better to work off of my 512x768 images?

[cyan snow]

better to inpaint then upscale

it's easier for the AI to mess with less pixels, in fact- it will be times 4 easier for it

[oak osprey]

some days, right guys?

[cyan snow]

i made these mario blocks again

[oak osprey]

i do and don't want to know what's in SDXL / 2.1's training sets that they can make these images

[cyan snow]

dude, i hate these things

[oak osprey]

they're the maori heads from easter island

almost called it easter egg island

[weak sage]

cracker dealer, you mean? he is in the soup aisle

[weak sage]

you need to realize that hires fix isnt magic, its just automatic upscaling and img2img

img2img has a weird effect of bringing out random details in a piece, im not sure i like that

[oak osprey]

i've literally written my own hires fix implementation. it's not "just" upscaling and img2img. that's a really misleading statement

[weak sage]

oh really you writing the code bro?

but really what am I missing out on there, it gens the image, then runs a chosen upscaler, then runs the img2img with the selected denoise strength, no?

[cyan snow]

i'm not sure, but what i AM sure about is that it reduces AI obstructions rate.

[oak osprey]

it ties into the diffusion process andhas a lot of control over which stage the noise is removed at

[cyan snow]

yeah, that sounds right

[weak sage]

now that's interesting

[cyan snow]

that's why it's slower than the upscale tab

[oak osprey]

it's also why no other tool has it

[cyan snow]

yep, i'd say hiresfix is probably one of the most crucial A1111 extensions

[weak sage]

i never use it

[cyan snow]

it's useless if you have low vram

[oak osprey]

i would have agreed before i spent all this time fixing zero-shot wide aspect / portrait gens in 2.1

[weak sage]

plenty of ram to load oblivion textures

[oak osprey]

now that i have a model that works straight up in Diffusers as i expect the models to, hires fix is merely a "nice to have"

[cyan snow]

yeah, but i'm not sure if it's even relevant once SDXL1.0 releases

[oak osprey]

it is, mcmonkey said as much

[weak sage]

wahts sdxl anyway

[oak osprey]

sdxl can't natively gen images (yet) beyond 1 megapixel, which i assume they did to keep the attention layers focused on the task at hand

after ive seen what happens when you train 2.1 on >1024x1024, i don't blame them. mcmonkey has been showing some 512x512 gens from sdxl that look "fine"

[cyan snow]

new architecture, pretty much

[oak osprey]

the SD2.1 768 model can't make 512x512 without looking broken because its attention layers are so conditioned on the detail it expects to exist in a >512x512 image. when it tries to make one, it expresses details that the model can't cope with and these come across as artifacting

i assume they included even 256x256 training data in the SDXL pool

[weak sage]

Do they resize training images both up and down?

[oak osprey]

not just new but like, bleeding edge experimental. so bleeding edge they might end up cutting off a huge chunk of it before 1.0 lmao

i have no idea whether they crop, down/up sample, or flip / rotate images for SDXL

[weak sage]

I think they didn't used to flip images at all

[cyan snow]

they said there will be a near perfect base model, so that's what's going to happen

[cunning geode]

extensions?

[oak osprey]

"perfect" and "base model" are so difficult concepts to combine, @cyan snow

[weak sage]

I was doing a text thing, and if I flipped the text element left to right it couldn't map it as well into characters

[cyan snow]

still it's damn closer to that than 1.5 is

[oak osprey]

@cyan snow the base model will know a lot and hopefully the data distribution of its weights are relatively even! this means no unintentional overfitting. but the even distribution of weights, seems to make prompting more complicated

it has so much knowledge that you need to know how to manoeuvre the weights via your prompt. the current implementations we've gotten to play with for prompt attention emphasis are pretty powerful but it's almost like they're too strong, as, little shifts in weighting have outsized impact

[weak sage]

4-6 fingers instead of 2-8 per hand? ill take it!

[oak osprey]

as i understand it, this is just how a base model is. if you fine-tune it for a specialty then it becomes more biased toward that form of output, and obtaining it becomes easier

so you look at midjourney and see it bias toward colourful hyper-contrasted explosive artwork and then obtaining real images from it is like searching for a needle in a haystack

[cyan snow]

nah, i found out why you couldn't get to the same level of detail with SDXL as my model, the discord bot and clipdrop uses lower settings, the images showcased on their website were generated with normal settings, i am sure that after it releases i will make a model that has better detail then my 1.5 one.

[oak osprey]

bro i'm talking about prompt coherence, not image details

it's a personal experience. some people's natural prompting style is like 1.5 loves, some people think how midjourney does, and others do well with 2.x prompting where a bunch of 1.5 users would be like "how are you doing that?"

when i try prompting SDXL, my initial results are amazing! i love them. they're aesthetic by default. but improving it to do what i want beyond that isn't easy

this will be easier with fine-tunes

[cyan snow]

that's also the case with any current base model tho =/

[oak osprey]

but you'll have to balance the fine-tune in pursuit of obtaining those results, and not destroying the ability to do other stuff.

@cyan snow but those models only have one text encoder to bump the weights of

text encoder fine-tuning is hard

it's harder in a different way than unet training is imo

[cyan snow]

but possible

[oak osprey]

every bit of fine-tuning you do with a text encoder, damages it

fine-tuning is the art of careful destruction

[cyan snow]

trust me, if it will be needed, the community will find a way, exactly like with 1.5

[weak sage]

It would be cool to semi-automate prompt refinement, so you could generate two complementary prompts related to the original and have the user select the one they prefer, or such

[oak osprey]

when it comes to finetuning the text encoder we don't even know the vram requirements. we know that it will be quite high, the models are huge. you can save on some vram requirements by freezing most of the encoder's layers. but it only helps so much

and yeah, "people will still tune it" but those people are few and far between

look at how many 2.1 tunes there are

[cyan snow]

that's because 2.1 was a fail, SDXL isn't

i bet after XL releases there might be more SDXL finetunes than 1.5 finetunes

[oak osprey]

it's funny but a handful of the SAI devs working on SDXL love 2.1, and its text encoder was selected as one reason, due to that success they hoped would continue with the larger OpenCLIP

that's a bunch of terrible english, i'm sorry

OpenCLIP worked great, and bigger OpenCLIP is better, is all i'm saying 😄

[weak sage]

your english is fine, you know other languages too?

[oak osprey]

thodee see hindi

i just read that sentence back though and it felt like r/sentencegore

if the person who wrote it is like "what the fuck lol" it's bad

[cyan snow]

but yeah, this bit of information makes me real happy.

[weak sage]

you from india? thats cool i just know a little spanish, im so grateful i can pronounce foreign stuff thats been romanized now

[oak osprey]

nope, i am canadian

[wispy nest]

phew, took way too long, but I learned that I hate python code :P

[weak sage]

curly braces ftw

[oak osprey]

https://www.youtube.com/shorts/Xe2MbMxuUuY
how i speak spanish

[cyan snow]

also, this was during the show and tell about an hour ago

[weak sage]

no yeah this is what happens if you DONT learn how to prnounce romanized text

[oak osprey]

the disappointing part to me is the implicit statement that 1.5 finetunes have more detail than SDXL base

[weak sage]

vowel massacre

[cyan snow]

we already tested that, remember?

[oak osprey]

i don't necessarily believe it though, that's just how Emad sees it

everything is personal and subjective, and i have terrible luck with 1.5 finetunes. remember, you have to eliminate hires fix from your mind. this is all zero-shot image gen

[weak sage]

do an empirical experiment if you disagree and it matters

[wispy nest]

so it's not very broken aye? 😉

[oak osprey]

in my world, i have no hires fix. everything is zero shot, and today is the first time i've used textual inversion

[wispy nest]

in my world, I have a good memory, I'm rich, and handsome, and have a good memory! :D

[oak osprey]

residual self image, Neo

it's how you view yourself in The Matrix

[wispy nest]

I'd probably be one of the nameless people in one of the booth when the real hero escapes. I'd probably think they'd also be crazy and out of their mind before having my 4231 refill of stale coffee...mmm coffee

[oak osprey]

you seen the movie 1408?

[wispy nest]

no, never heard of it, is it like brazil, or how it was spelled?

[oak osprey]

omgggg it's not at all like that

it's a really good John Cusack film by Stephen King

it has Samuel L. Jackson too

highly recommended viewing especially since you like AI art so much. it's not AI art, but, you'll see what i mean

[wispy nest]

last movie I saw was the first deadpool movie, so if it's after that, then I've not seen it by default :P

[oak osprey]

i think this is from 2007

holy cow, it is

oh, another one from 2009 same AI art-esque feel, Mr. Nobody but some people can't stomach Jared Leto anymore so, viewer beware i guess

[wispy nest]

hope I remember that, gonna be without much internet time for a while. And trusting my memory, I know I've already forgotten it! :P

[oak osprey]

put it in your phone notes 😛

i use Keep on my phone for random shitty lists

[wispy nest]

I...I didn't actually think about that one, I never use my phone for anything other than calling someone or sending a sms.

I hope people will have fine tuned that new sdxl when I return, one day! :P

[solid gate]

What program do you recommend to animate an image like this?

[weak sage]

idk but that's pretty hot

[cunning geode]

im following a trail right now
i always thought i need 0.4 denoising strength at highres fix to not get clones
but i think i can even do 0.6 without problems

[cunning geode]

0.6 vs. 0.4

[cunning geode]

nvm
this prompt here shoes well how the model deterioates if you go above 0.4 denoising strength
0.4 vs. 0.5 vs. 0.6

hm now im wondering if 0.3 might not even be better than 0.4

ah fuck i was using the wrong model lmao
but yeah seems like there isnt a big difference between 0.3 and 0.4
so i was already using the optimal value

[sterile temple]

I noticed the adetailer extension got pretty good results, and they have 0.4 denoise and 28 steps. So I started using that too

[cunning geode]

coherency of wings and tails in this model is definitely a ton worse than in my previous test models, but its also much more diverse and flexible so ill let it slide for now
definitely will try to improve that tho

[oak osprey]

opentalker

[cunning geode]

actually, it seems to do a lot better when i ((emphasize)) the wings and tail
still far from where i want it, but better

[solid gate]

im go to see!

[oak osprey]

https://civitai.com/models/97808?modelVersionId=104575

thanks, i've added that to the model page on civitai

[sterile temple]

downloading...

[oak osprey]

see the notes that you need the cfg rescaling extension, and perhaps to use DDIM or DPM++ 2M

[sterile temple]

weird, when I click download it opens a new tab and nothing happens

[oak osprey]

without the cfg rescaling extension, DPM++ 2M worked alright but there was definitely contrast ceiling left on the table

the safetensors file is also on huggingface

[sterile temple]

will it work in comfyui?

[oak osprey]

yes sirrrr

though i can't tell you HOW it would be done

😄

are you a ComfyExpert?

[sterile temple]

only used it a little bit 😄

[oak osprey]

compared to my experience with it that makes you the intellectual juggernaut of the two of us

[sterile temple]

I'm thinking of using it more. It can be tiresome having to fiddle with all the config in A1111. It takes a bit more time to set something up, but you can make workflows A1111 can't do and then repeat them easily

like coding

[oak osprey]

man so the a1111 config fiddly bits are the same even for power users?

i bet there's an extension to fix that but i didn't bother looking