🏞｜general-with-images

[oak osprey]

here's what it looks like when you keep training and keep training the text encoder. just keeps adding more and more details

[smoky oak]

Glad you are finding a way to get rid of the dreadful 2.1 noise

[oak osprey]

that's been training for a full week now

[smoky oak]

The 2.1 noise is the number 1 reason why I don't use it

[oak osprey]

my glasses are so bad i just can't really see it until you helped me look closely lmao

if it helps you understand 2.1 maybe all the researchers are blind

[smoky oak]

A 768x768 2.1 model looks worse than 512x512 in 1.5 at pixel level detail, which is why I hate it

[oak osprey]

mm yeah controlnet helps with it even on some of the worst models at both versions

today's SDXL drop still has me waiting for good textures

though "bed-head" meant "towel on head" to SDXL and my 2.1 fine-tune which to me is just

this comes from one of my ethnographic kodachrome prompts

[oak osprey]

on the surface this image is really impressive and i love the aesthetic. when i get nit-picky, his features look so unnatural when you zoom in. or maybe my beard just sucks lmao

[dense tapir]

I still can't get a Lora/Locon/Loha to train a style. It trains something but you have no idea what it will do when fully trained.

[oak osprey]

keep trying

don't give up

[dense tapir]

I tried 3 different techniques and they each did something different but none did the style I was after.

...or trained on

[smoky oak]

I feel like I could do a style now, if I wanted to

[dense tapir]

Well, want it naow, lol. I have tried everything but it is really weird cause the style isn't what I was after with reg or without. A final piece is missing for a real style. iow, a style that when used whatever you do is in the style you trained on from brush strokes to mimicing an artist.

akin to 1.5's slapping greg rutkowski in all the prompts.

[oak osprey]

man style is a weird one so i totally feel you on that. though i don't even know if GA can see my stuff anymore. Sytan, does stuff like encoder layers come into play at all with LoRA?

[smoky oak]

I know it's very much possible, as the anthro model I use comes with 30+ included individual style LoRA's to be used on other models as well

[smoky oak]

Honestly, not too sure

[oak osprey]

because i know it matters a lot for style transfer in OpenCLIP

[dense tapir]

Yes, but the info is just not out there. I guess my colab is dead for the next 24h as I can't connect to a gpu.

[smoky oak]

I'll experiment when I get back to the US

[oak osprey]

it's a tough cookie to crack because it's so destructive to train a new style into the layers that seem to be most responsible for it

it's easy to enhance existing ones

kodachrome 😄

[dense tapir]

This is why I hate colab because I get really involved and time is up for the next day. Colab is slow too. I mean CPU is slower than my 5600.

[oak osprey]

slow CPUs are actually a constant across pretty much the entire thing until you start paying a LOT more

[smoky oak]

CP-Ew

I shouldn't be talking, my CPU is molasses lol

[oak osprey]

aye we always hate what we can't have

[dense tapir]

I watched it cache at 140s per it on CPU while on my 5600 it was 10 times faster

[smoky oak]

I just want something that's not a massive bottleneck in every way 😅

[dense tapir]

Jaw dropping time

[oak osprey]

you seem to get pretty damn near the max it/sec on it. what does it bottleneck

[smoky oak]

Jesus Christ, 140 per it?

[dense tapir]

yes

[oak osprey]

it's google drive's fault probably

[smoky oak]

My friend does SD on a 2 core laptop, and it's like 40 seconds/it, but that's at 512x512

[dense tapir]

it was to be about 4h to cache all the images

[oak osprey]

that sounds ... broken

but it's google drive

it's going to be slow

[dense tapir]

when it builds kohya it takes 4-4.2 mins

each damn time I run it

So bad I added this

[smoky oak]

Yeah, that is insanely slow, wow

[oak osprey]

why even bother doing that at all, it sounds super frustrating

[smoky oak]

It takes like 20 seconds on my CPU, and my CPU is pitiful for modern standards lmao

[smoky oak]

All games I play, and almost all programs I use, even more so now that I am on a 3080

[dense tapir]

oh, the gpu is a t4 so blows mine away but the cpu is horribly slow

paying for colab doesn't make it faster either, and as you see paying for colab gives you no support either.

[smoky oak]

Makes me wonder how other pay for work cloud services fair in that way

Fare? IDK

[oak osprey]

glad today's new GPT model can still be jailbroken

[dense tapir]

Amazon is actually really nice, and damn sophisticated in comparison but free they never have any GPUs available so I could see that. I know their gpus are modern ones not Pascal based.

[oak osprey]

GA, have you tried Paperspace free tier

🦗 🦗

oh he'll probably never know unless you tell him that exists, Sytan

[dense tapir]

Now that even TheLastBen requires a paid service to train his with a lot of the SD training is pretty slow right now.

Only the ones with decent gpus at home, or crazy to pay, is doing it.

[smoky oak]

@dense tapir do you have pseudo blocked?

[dense tapir]

yes, after my forced vacation.

[smoky oak]

If so, he's trying to share important information

Ah, right, I remember that

[dense tapir]

No worries I can live without it

[smoky oak]

He asked if you have tried the free tier on paper space

I forgot they even had a free tier

[dense tapir]

I did, yes

Most times completely taken

PS free suffers the same as Amazon free just no available gpus

[oak osprey]

tell him he can use paperspace's CLI tool to snag one when it becomes available

oh well though if he's going to be grouchy about it i won't help him with it, lol

can probably use amazon cli tools to grab free ones too. that's likely why none are never available, because others are already doing that

[sterile temple]

If you had 2 or more GPUs with say, 6 or 8GB of vram, could you combine the memory? or would it still need every GPU to have a minimum amount of VRAM?

[oak osprey]

you can kind of do that, in theory, but i've never tried

they'd both need to use the same driver versions and all that of course

[dense tapir]

SD doesn't take advantage of that and another thing that they killed on the 4k series by eliminating the nvlink

[oak osprey]

beyond that, a GPU is a GPU, and accelerate distributes training using multiple processes

each training has its own blocks it works on aiui

[dense tapir]

reason a 3090 ti is more valuable than a 4090

except in SD but for ML/AI otherwise way more valuable

[oak osprey]

@sterile temple you can put the text encoder on one GPU and the unet on another, manually, and train them on separate devices. this is how some people fine-tune deepfloyd

[dense tapir]

the above without my lycoris

[oak osprey]

you can put certain layers of each model on their own GPU but that becomes more of a like, ... hmm.. cpu bottleneck, i guess.

oh whoa, Fulljourney is crazy

what the hell lmao

deepfaaaakes

aw its not the good kind

@smoky oak new NVIDIA JENSEN chip

[smoky oak]

Lmao

[dense tapir]

What does it mean when your lycoris requires more weight?

train it more?

[oak osprey]

SDXL seems to do this really interesting pattern with details

like his beard does this and i've seen freckles like that

perfectly spaced patterns

[dense tapir]

Needs 2-2.1 to work right

[oak osprey]

https://images-ext-1.discordapp.net/external/XxvkSxErkvzkxLDBjqwIhBYmL0sdRjk6HtRz2_oYVJk/https/tripleback.net/public/discord/1686705841.782628601c9d4b14941cfca7e45d986c0fe5e.png?width=1004&height=1004

the people of the stable foundation discord, ethnographic photography, 1998

@smoky oak which would you be

[smoky oak]

None of them, I wasn't alive in 1998

Lol

Imagine being from last century

HELL

LAST MILLENNIA

Old asses

[oak osprey]

yeah i imagine if you were that old you might know what "would" means

when i had a youtube channel i expected my demographic would be the usual like 18-35, right?

it ends up being geriatrics around 65-70

[smoky oak]

They feel at home with your natural language of stroke

You speak it so fluently <3

[oak osprey]

jordan peterson being confronted by his deep fried past selves

didn't expect that

[oak osprey]

no cappin fam thas strait bussin', YAAAAASSSSS

[smoky oak]

Damn, you're so out of touch you can't even fake what we sound like properly lol

[oak osprey]

wat, that was lit as fuck af

[smoky oak]

You're gonna give me a stroke

[oak osprey]

i learnt thanks to a kiddo at the museum the other day (oh god, i am old) that "af" means "an f-stop" and its an old photography term that was adapted by the youths

thats something you can probably appreciate on many levels. pop culture knowledge and photography as well as just a unique ability for me to connect across a generational gap

[dense tapir]

I just looked and almost all of my reddit subs are now gone. I wish Reddit would collapse.

[oak osprey]

on the upside you're in a bunch of good communities then

[smoky oak]

Yup

Good sign of a good community

[oak osprey]

he can't see what i said so he's just going to be confused

[smoky oak]

Right lmao

I am all for the blackout, I just wish they didn't broadcast how long it would be for, cause now reddit knows they just have to wait it out for 2 days, then it's back to normal pocket lining

[dense tapir]

Precisely

What in the fuck kind of protest is that?

[smoky oak]

Honestly screw reddit. They are shitty for doing those API abuses

Flat out unacceptable

[dense tapir]

Everyone is so greedy in 2023

[oak osprey]

stable diffusion is still free, LLaMA came out for free, but black and white thinking, sure

reddit has too much power

crazy how many people rely on that corporate trash

remember usenet? oh wait, too young

[smoky oak]

Makes me think of hughes net lol

[smoky oak]

Insane lol

[dense tapir]

Hughes net a 100% rip off

[oak osprey]

over 8000 steps of training the apprentice has become a man

[dense tapir]

Get Skylink for the same price and live free.

No caps

Honestly, I am surprised Hughes Net is still around now.

Less lag on Starlink due to LEOS vs Geo Stationary, and way faster up/down speeds with no caps. 100 per month

[oak osprey]

oh no, artifacts

@smoky oak i did this to someone just now

F for another A1111 install

[oak osprey]

they were on Diffusers 0.14

[smoky oak]

I have awoken, I am on my way to Mexico

[sterile temple]

have fun!

Dang.. that totally surprised me. A 1200x900 image and it produced a cropped close up 🤔

[smoky oak]

That's impressive haha

Gonna do 6k, and it's just gonna be an eyebrow lol

[sterile temple]

TIL the task manager chart for GPU can be changed to CUDA

[smoky oak]

Yeah, very very useful

Tho sometimes it dies lol

[sterile temple]

I couldn't do 6k, but 2048 is.. not sure what that is

I copied the prompt from somewhere, it has a lot of keywords, too many I think

removed about 2/3rds of the prompt and it zoomed out 😄 is that the trick to larger resolutions? just add more random keywords?

[smoky oak]

Doing terms like "detailed" makes them zoom into the areas specified

Like detailed hair, will make it get closer

[sterile temple]

detail is mentioned 4 times in the prompt

((fearsome dull skin noise)) <- not sure what this is supposed to be

lamp for a leg. nice

[smoky oak]

Camper cat

*vamper

[sterile temple]

monster phoenix raffles pizza, glowing pizza with blood <- magic prompt 😄

[smoky oak]

I'm in a silli goofi mood

I suppose that's to be expected when awaking at 3 am to catch a flight lol

[sterile temple]

have you heard shpongle - divine moments of truth? good song for when taking off

goes for 10 minutes

[smoky oak]

Hehehehe that's a fun name

"shpongle"

That's a good mood off of first impressions alone haha

Sounds like some shit I'd make up when cracked out on sleep deprivation

[sterile temple]

it has the slow bit at the start for taxiing on the runway, then it ramps up when the plane takes off, and bit of a marching band drum thing at the end to get you pumped for the food cart 😄

this is from before they had fruity loops

[smoky oak]

Ah yes, Fruity loops, my one true lord and savior

When I get back from Mexico, I desperately need to work on my new releases

[sterile temple]

https://tenor.com/view/thirteen-drop-track-drop-track-thirteen-th13teen-airtime-gif-25959661

[smoky oak]

Or well... First releases lol

Everybody keeps telling me I need to release stuff already, but I need the right idea for my first release

[smoky oak]

Gato

gato en el traje

[sterile temple]

Sytan - Count Catula

[smoky oak]

I really need to think of a concept to follow for my first song

I wanna set up my brand image from day one, so I have an idea to follow said concept

[sterile temple]

And then the fans would be like, "Why did he have to change the sound with the new album?" or "This album sounds just like the last one..."

[smoky oak]

I already have several concepts set up to release later, but I need one for now haha

[sterile temple]

it's corn!

i mean look at this thing

it has the juice

[smoky oak]

IT HAS THE JUICE

corn

CORN

CORN

[sterile temple]

7 элемент - Витас

bllbbb ah ah ah

[smoky oak]

Gorn

[sterile temple]

https://open.spotify.com/track/6JDhxzaxkq3bOPPFDBNTFp?si=b8d7589137094c6d

@smoky oak you know clown core?

[smoky oak]

Yes lol

[sterile temple]

good, just making sure 😄

[smoky oak]

I should use this week away as a brainstorming idea for my first releases

[sterile temple]

https://tenor.com/view/uks-unbreakable-kimmy-schmidt-unbreakable-unbreakable-k-titus-gif-24322981

I was trying to find a gif of titus pointing and nodding, but saw this one and had to post

[smoky oak]

I mean, that's how I did it just a year ago in Hawaii, and that's how I did my whole rebrand and figured out my goal for my brand image and everything

I just need to think of ways to execute it properly

[sterile temple]

you should do an album that loops, the last song goes into the first song

[smoky oak]

Way ahead of you on that

Already have plans for an EP that follows not exactly that idea, but it's a free flow EP, so the idea is playing it in whatever order you want

I ran the concept by my professional friends, and they all seemed to like it as a core concept

Also have an idea of a abstract song, but I feel that would be too weird to come out swinging with lol

[sterile temple]

is there such thing as microtonal electronic music?

[smoky oak]

I'm sure there is

[sterile temple]

https://www.youtube.com/watch?v=MJ3PK8wQ3E0&t

not as good as kglw

[smoky oak]

I desperately need to go back into IRL contact with my music peers

I miss my friendsssss

[sterile temple]

https://tenor.com/view/go-home-leave-golden-girls-just-go-you-have-to-go-now-gif-22287144

well go on then

[smoky oak]

They are legit the only reason I am still in music, cause they all hyped me up so much

[smoky oak]

It's reallyyy hard, we all live so far, and they are real professionals always busy

One of my friends just came home from his tour

[sterile temple]

@smoky oak brainstorming music while on holiday https://youtu.be/ydeV1_8pM4o?t=135

[naive wren]

hey, I am trying to create a QR with img2img and I am getting horrible results, is this even possible? To beauitify a QR with a logo instead of a prompt?

this tutorial they used img2img and get a good result, but if I try it with a logo results a disaster https://learn.thinkdiffusion.com/creating-qr-codes-with-controlnet/

[sterile temple]

I tried following a few tutorials and couldn't get anything like these images they show

[naive wren]

same

[sterile temple]

this is the best I could do after ages

[naive wren]

promt or providing image?

[sterile temple]

using a prompt

[naive wren]

yea, img2img is the one I am interested

assuming you give the image of a building and the QR, and then fusion them toguether

like in the tutorial I mentioned

[sterile temple]

another one with a prompt, but it was a fluke 😄

[smoky oak]

Think the antichrist, my old phone works in Mexico

*thank

[sterile temple]

touchdown!

[smoky oak]

Nope, haven't flown yet lol

We just walked across the boarder is all lol

[sterile temple]

11... i should go to bed

walked?

[smoky oak]

Yeah, we live in southern california, so we drove all the way down to the boarder. We have a very cool airport here

The way it works is that you enter in america, cross the border through customs, then you walk to the other side where you fly out of Mexico. It allows you to take a domestic flight from New Mexico to Mexico rather than an international flight from US to Mexico, and it's WAYYYYY cheaper cause of it

[sterile temple]

interesting

Can't walk into another country from Australia 😄

Mexican foods is probably one of the lesser available cuisines you can get here, due to our proximity to Asia. Chinese, Thai, Vietnamese, Japanese foods are more popular than Mexican

anyways, I'm off to bed. have a great day!

[smoky oak]

[sterile temple]

looks like a flight to mexico would cost me about $2300USD

[smoky oak]

Jesus Christ, yeesh

It's $300 a person here 😅

[sterile temple]

Bali is our Mexico

can probably get flights and 5 nights in a nice hotel for about $300

👋

[smoky oak]

We're staying in a fancy resort cause my dad's side of the family is exuberantly wealthy, but our money can't be used in the UC cause they will take a massive tax to bring it in

And my abuelita is now trapped in Mexico cause the gov lied about her immigration status and got her permanently deported, so we have to refile for her residency, so she keeps flying us out to see her in the lavish resorts

[dry crow]

@naive wren @sterile temple img2img with controlnet tile resample can work:
Controlnet weight 0.9
Controlnet start between 0.23 and 0.3.
Controlnet end at 1

[naive wren]

Hey @dry crow would it work with a logo too?

for example Discord logo

[dry crow]

If you use a model or lora that is trained on the discord Logo sure

But for now it doesnt seem to be easy to get everytime the thing you want

I had a lot of failed qr codes before

[oak osprey]

lol civitai overloaded

wonder how they afford that site

[smoky oak]

Our plane almost left us

Omg

Our boarding was suppose to start at 8AM, they started it 35 minutes early-

Did the last call at 7:55, 5 minutes before we were even supposed to board

[oak osprey]

https://giphy.com/gifs/homealone-90s-home-alone-d2YWTOsVtuHgOHhC

[smoky oak]

Our plane is playing trance and acid while we are boarding lmao

[oak osprey]

thats house music

NOT trance

ugh

@smoky oak after killing someone's automatic1111 install last night by getting them to update it, i showed them how diffusers works and they said it's way less code, and switched over to it. it took about 6 minutes to do so lmao

[smoky oak]

I probably would, if I wasn't giving up so much stuff along with it

Maybe I'll take a look some other time

[oak osprey]

they develop a Discord bot that used sd-webui as a backend which i didn't know anyone had done yet

it's weird because they have like, inpainting built into it, and you can inpaint through a discord channel

so you paint a mask right there in the channel. how do i even do that lmao

[wispy nest]

you guys still use auto right? there hasnt been a hot new ui thats dropped has there?

[livid ermine]

Not very popular in Norway (who owns the Antarctic). But the downvotes were all from little jealous boys.

Granted this is going to look way better when it's done. Plus they're virgins, and I work with the world's top #1 architecture agency Snøhetta.

(or among the top #5 at least)

[oak osprey]

invokeAI is really good now

https://images-ext-2.discordapp.net/external/TLbIuS3kd9X2_PUYu98fRNdioDD9p8OSnRrAVTUT_MI/https/tripleback.net/public/discord/1686760520.850337369f01c4e4381a28c7b35f0223a2e39db.png?width=1004&height=1004

[wispy nest]

do you use it over auto?

[oak osprey]

i use Diffusers, the underlying library InvokeAI is built on, via python scripts/a websocket-http infrastructure i've built

uses a Discord bot as the main frontend for now

[wispy nest]

oh invoke ai has a sub cost

[oak osprey]

it's on github

https://invoke-ai.github.io/InvokeAI/#installation

[cyan snow]

the only disadvantage of InvokeAI is the lack of customizability and extensions. when it comes to capabilities the a1111 ui has way more features.

but in my opinion if InvokeAI could support A1111 extensions it will be the best ui, hands down.

the current best is probably stable studio, it can be run locally just as an interface and connect to a1111 as a backend.

[wispy nest]

Give me your best to throw on this

[wild sorrel]

Yea, a1111 has alot more features... invoke had last significant update pretty long ago, unless I'm missing something released recently.

It doesn't even have all a1111 samplers.
Yea, it has nice design, but it has 0 features a1111 can't do, while a1111 has lots of things invoke can't...

Do you really want to use webui which lacks functionality just for it's visuals?

it feels like devs just gave up on it....

[cyan snow]

no i don't use InvokeAI. i use the ui that Stability published and it's plugins to add all a1111 backend extensions.

basically i run a1111 as backend and Stability ui as interface.

[oak osprey]

A1111's extension support is implemented in a problematic way that isn't really possible to hook into with invokeAI

without having a LOT of complexity, in invokeAI itself

it can do extensions, but not the way sd-webui does them

i could be wrong, i'm just one developer, and i don't work with invokeAI people at all, so i don't know how they feel about this

[oak osprey]

stock photo surrealism

[still oak]

hello guys i was wonder if anyone know how to make stable image create 2 image 1 in the front the face of the character and 1 in the back whill it still the same character i apology if my english is bad

[oak osprey]

backside shots are hard, and you'd likely need to train a Dreambooth for that, and that's not easy. you can sort of approximate what you want, by finding a single seed and a strong prompt that invokes a particular looking person

[still oak]

okay thank you, i will try it

[oak osprey]

joey ramone

portrait of (joey ramone)1.4 from behind, the ramones, 1955, standing on the sidewalk, california, technicolor, UHD

portrait of (joey ramone)1.4, the ramones, 1955, standing on the sidewalk, california, technicolor, UHD

[oak osprey]

@smoky oak SDXL's latest lmao

[wispy nest]

looks like what I'd guess it'd look like. I believe that's the first image from that model I've seen :P

[oak osprey]

tis very zoomed in

[oak osprey]

@wispy nest ^

[wispy nest]

if that's their new model, then it still got a long way to go in my eyes, my blurry, blurry eyes :P

[oak osprey]

well there is supposedly some major improvements they've got internally with "control nets" that are not available on the bot

but to me this indicates that SDXL still suffers the fundamental core issues of SD 1.5 and 2.1

once the control nets are a part of the model's basic workflow, that is no longer the case in my eyes. but... that's not what i've got available to go by, right now

[wispy nest]

words are cheap, that's why I'll wait for it to be shown instead :P

[oak osprey]

in so many words that is what i said to them 😛

none of the demo images so far have showcased SDXL succeeding dramatically at these fundamental levels compared to what you can currently pull off with SD 2.1 or 1.5 and fine-tuning, let alone with ControlNet v1.1

i don't count the contrast fixes because those are super easy to pull off with either 1.5 or 2.1 as well with just a little bit of fine-tuning on a medicore dataset

it has no residual noise in the images, also something you can pull off with 2.1 or 1.5 but what's the point of a super deformed image with no artifacts. just trading off one form of artifacts with another

[smoky oak]

I'm not as pessimistic about SDXL, as it's going to clearly be an improvement. 3x the parameters, less VRAM use for higher res images, monumentally better TE, and results that are pretty damn good out of the box, with 0 fine-tuning

I think that's what everybody forgets, SDXL isn't finetuned. The 1.5 and 2.1 bases are like garbage to use, and if we got as good results out of them as we did, then SDXL will be light-years better

[oak osprey]

that's a logical fallacy though too, to look at things that have happened, and assume they will keep happening

[smoky oak]

In what way?

[oak osprey]

there's no fine-tuning your way out of fundamental issues

[smoky oak]

I did and an if and then, which isn't an assumption, it's a hope

[smoky oak]

And yet 1.5 and 2.1 finetuned aren't unusably bad like the bade models

The 1.5 and 2.1 base models have a hell of a lot worse issues that have been "finetuned" out

[wispy nest]

I believe I only made around 100,000 images using 1.4 and 1.5 default models before I learned of the first fine-tuned model :P

[smoky oak]

What wasted time haha

[oak osprey]

the reason i went with 2.1 is because it was amazing to use, despite the residual noise 😄

that's the kind of stuff i get from base 2.1

[smoky oak]

Looks decent

[oak osprey]

that's with the fixed DDIM

i haven't fine-tuned that at all

noise schedulers have such an outsized impact on this thing because of the way it already skips its final layer

i asked it for a mad magazine cover

alan alda's magazine cover has a bit of a "mule kick to the head" thing going on

[wispy nest]

there's still a lot of basic 1.5 images I love at least :D

[oak osprey]

my pfp is 1.5

😛

[wispy nest]

mine as well, one of the first images I made :D

[oak osprey]

some of the best celeb poop gens have been from 1.5

erm i mean that like youtube poop, not.. uh

[wispy nest]

1.5 also knew how to make fluff :D

[oak osprey]

hyperfluff

international vaping contest judges table

[smoky oak]

I am just a little confused, cause you point out these mostly minor issues with SDXL, and then praise 2.1 with massive and hideous deformities, I just don't get it

[oak osprey]

it's fundamental issues that remain unsolved, it's nothing new

it's really cool because they actually took a lot of care with their dataset this time, and it seems like they're not training on a frozen text encoder, but otherwise it's the same small internal representations

they hype it a LOT, meanwhile, no one likes 2.1, and thinks it's a totally broken and dogshit model

so, yes, i am a fan of the underdog, not the hype train

Bing image gen

[wispy nest]

I'm content with just making images I like :P

[smoky oak]

As it is now, 2.1 is that way, as I still have yet to see a version of it that isn't, but that doesn't mean it can't be fixed, but it just seems like a waste of time to work sideways rather than forwards, IMO

[oak osprey]

that's what sticking with 1.5 is

i'm just wondering what's the hold-up on releasing these 'awesome internal features' to the SDXL bot

i want to see it work better than anything has yet

[smoky oak]

As I have said for a long time, I will happily leave 1.5 when I find a 2.1 model worth adopting, but that certainly has not happened yet, so I wait ._.

[oak osprey]

i'm hoping my feedback is useful to them, that my observations cover issues they might not be aware of, and so on

there's very little wrong with that model. i asked how they know it's done, didn't really receive an answer. unless you care about people with freckles or acne or realistic beards, it's good to go for release already

[smoky oak]

I am excited for when we actually get our hands on it, and don't have to keep assuming things about it, cause that is the most annoying thing about all of this SDXL conversation, IMO

[oak osprey]

same

[wispy nest]

I have the same view on it as their fast two-step sampler that was just around the corner :P

[oak osprey]

look, i thought it was just under-training that led to the shitty small faces at a distance, so that's what i've been trying to fix for a long time. the Jack Black riding a bear in the swiss alps. his face looked like swiss cheese. but it turns out you can't solve that with training. and i've seen the same distortions in SDXL. like, the same exact ones. so i'm disappointed that it's not looking very ground-breaking yet

[smoky oak]

If the 50% traini g checkpoint images are anything to go off, then this model blows everything we have seen out of the water, but the gens from the SDXL bots are just not up there yet

Just frustrating, especially cause we are not able to mess around with the new variables and aspects of it reliably and repeatably

[oak osprey]

that checkpoint was from a month ago, it's been a lot further along now and if you look in #sdxl-feedback you'll see posts from people showing how much their prompts have changed for the better and for the worse since today's update

https://i.imgur.com/DloNK0w.jpeg

i'm just curious how 1.5 makes these tiny shreks all running along the shrek-hulk monster thing

[smoky oak]

I still don't think the images from the bots are nearly a good enough representation of what to expect when it's in hands

Very frustrating in various different ways

Oh, signal is a little shoddy here

[oak osprey]

so, not trying to frustrate you. it's just that i thought the new model would be a new architecture that looks to solve these problems

maybe it's just something they can't know until they try, and they thought it was too

cherry-picked images look amazing even from the bot

it's kind of like a philosophical debate, we're creating images of things that don't exist, so, what perfection even is, i do not know

[smoky oak]

It's not a new arch, I thought that was like... Super upfront

First time I ever talked to the devs they specifically said it's just suped up 2.1

It's just a test to get the good stuff figured out before they crack down and concentrate all of the improvements into 3.0. that's what I have been told for well over a month, probably 2 at this point

Granted, it has changed a lot since, but it's still what they said at the core

[oak osprey]

well souped-up 2.1 doesn't describe a new text encoder very well to me

the text encoder is a damn workhorse, it does most of the job

[smoky oak]

That is something that has changed since. I remember playing around with the SDXL beta the first day it came out over 2 months ago, and it handled very very differently

[oak osprey]

how can you just drop a new text encoder in

[smoky oak]

No idea, but that version was also trained on 2.1's dumpster fire of a data set, and it showed

[oak osprey]

hey that garbage gives us Rick Moranis as Harry Potter 🥹

and 2.1 got 2048x2048!

that's why it sucks at 512x512

[smoky oak]

I still prefer 512x base images to 768, cause the base res doesn't matyer if you're gonna high res fix it, and 512x is WAYYYYY more efficient for generating large batches for comp

[oak osprey]

ain't no one lookin' at 512x512 images, sweety 😛

[smoky oak]

Kinda the whole point of my message l

[oak osprey]

i gen 1024x1024

[smoky oak]

*lol

They are composition guides for high res fix

[wispy nest]

yeah, the numbers are probably around the same as people using 2.1 ;P

[oak osprey]

@wispy nest oh blue willow runs on 20 million discord servers and they exclusively use 2.1 with TIs and LoRAs

more accessible than 1.5, arguably

[wispy nest]

no idea what that is

[oak osprey]

it's a free image gen discord bot. though now they limit to 10 a day

sound familiar?

[wispy nest]

no, never heard about it :/

[oak osprey]

i mean the payment changes

midjourney did that

[wispy nest]

no, not that neither, never heard about it :P

[oak osprey]

[smoky oak]

Blue willow is MJ with a few extra chromosomes

[oak osprey]

hires fix sounds like a total scam lmfao how is it working for everyone but me

like i'm just not pyramid schemeing hard enough

[wispy nest]

can't be a conspiracy because that needs 2 or more people :P

[smoky oak]

It's clearly not, just based off.my results haha

[oak osprey]

mark zuckerberg again eh

@smoky oak your examples of input weren't tiny faces at a distance tho

[smoky oak]

Pretty sure your implementation of it just isn't right honestly, but the person who made it said that it was just pixel upscale into img2img, nothing fancy

[oak osprey]

i did a 1.5 based img2img and it kind of worked, it certainly made their faces less distorted, but by making them smeary and blurry

well mine does that too, a pixel upscale using Lancsoz or whoevers name it is, and then img2img through the long prompt weighting pipeline

[smoky oak]

Oh, that was one niche example, I was meaning in general, it works really damn good, especially for generating dozens of images for comp to pick from in like a minute, then pick the one you want and high res fix it, and you went up with a much better result than generating native 768x or 1024x, and it takes a lot less time

[oak osprey]

yea i agree it can do other stuff really well but i swear i've seen people get good results with distant faces before. but maybe not

[smoky oak]

I mean yeah, it's easy to do with even the most basic imainting and out-painting lol

[oak osprey]

they could be inpainting which if i have to do that, fuck it lmao

that's like making an image in an image

https://giphy.com/gifs/cat-infinity-XkDyB1jCZMxUs

[grizzled olive]

Every person I have ever generated that's not a close-up basically looks like they just got hit by lightning.

[oak osprey]

we're here for you

don't get me wrong, training can fix some of it, but i think this is the smallest a face can go before there's real issues lmao

[smoky oak]

Cause that doesn't have "real issues"

But I get what you mean

[oak osprey]

i mean just the faces, i wasn't doing myself any favours with this prompt asking it for "kelly hawk's tony honk's pro kelly slater 3 surfing playstation game box art"

[grizzled olive]

I would play Tony Slare KLKV

[oak osprey]

find me a model that can do people training brazilian jiujitsu as well as mine can

sure it isn't great but that's saying a lot about how bad the other models are. i can't even show them here

[wispy nest]

can I ask you to find a model that does something I want just as easy? ;P

[oak osprey]

yes

ask away

[wispy nest]

I was trying to get a style like this at the start of the year, but this was the limit I got to before all the details started to mess up too much. Can be seen here. But I'd like the same focus and detail as the eyes in the image :D

[oak osprey]

i don't think the technology is there yet

[wispy nest]

I got the eyes like I wanted, but the rest kinda, didn't turn into what I wanted :P

but the real question for every checkpoint: Can it make smol cat?

[oak osprey]

https://images-ext-2.discordapp.net/external/2gVIFvnOReIsGhKArdKiIivSOGHRTVAN0Lfhm-ecRDo/https/tripleback.net/public/discord/1686779193.8256648dbabbe235f2e06d350bd1e56642d1750.png?width=1004&height=1004

nailed it

[wispy nest]

thank you, I needed that laugh, but I didn't need my jaw locking for a second >:P

[oak osprey]

[wispy nest]

BUT, can 2.1 create constant images of boats/ships with figureheads based on prompts?
I've had MAJOR issues with that on every single model I've tried as it seems the ai has no idea what details on a ship is named as :P

[grizzled olive]

Me and the boys learning how to use controlnet

[oak osprey]

i dont even know what a figurehead is

does that have one?

this has an actual head

[wispy nest]

hehe, not really, it's those (often) mermaids in front of wooden ships :P

[oak osprey]

idk it keeps cranking things out

[wispy nest]

would fit in some sort of horror game if 2.1 only knew what darkness was as well :P

[oak osprey]

this is a good one

2.1 definitely knows darkness

[wispy nest]

I meant the black colored one, not the evil one ;P

or rather, night, as to not sound as weird about it :P

[oak osprey]

i don't even use offset noise for training

it's just better at it than 1.5

[oak osprey]

2.1 is ready to make your horror game.

[wispy nest]

I didn't even know where the edges of the images when copying it :P

But I will concede that using black and white is somewhat cheating :P

[oak osprey]

that's trained using offset noise. can it make very bright images?

currently looking at some tests of a 13,200 step unet with 4200 step text encoder

[wispy nest]

hmmm. Let's say it's a...draw? :P

[oak osprey]

he's so handsome when he vapes like that

you can smell the Strawberry Cheesequake vape juice from here

going to withhold judgement but so far it looks like the unet has swung back around into clarity land

i keep waiting for this one to become photoreal and i'm not sure it ever will completely, but holy cow is it better now than the SEGA Genesis 8bit it was originally

[wispy nest]

I never finished my mist/smoke/fog/etc embedding for 2.1, mostly because for some reason it started to create awsome borders instead :P

[oak osprey]

helo im robein

the over-trained unet with the older text encoder

[wispy ether]

Wtf

Bro sent nsfw

@moderator

Bruh

@sunn

FICL

@tropic shell

[jovial jungle]

Does this chat have more rules than #🍥｜anime or are you doing a le reddit bit?

[wispy ether]

https://tenor.com/view/luigi-dancing-gif-21641026

[jovial jungle]

oh a bit, don't scare me like that

[wispy ether]

[tropic shell]

If someone sent you something, then please send a ticket with the image if it's in DM or flag the image with report or ⚠️ , etc; thank you.

[wispy ether]

But i was lazy and randomly saw it

Didnt want to make a too big deal out of it

[wispy ether]

How do i know the ! Worked

Someone removed it

K

Idk

[wispy spindle]

Removed
Be aware of nsfw rules in server for future reference!

[jovial jungle]

k will do

[sterile temple]

good morning/evening 🌄🌃

[oak osprey]

M. Night Shyamalan

sorry, this one has no 'deformities'

[grizzled olive]

It's really good, i just imagined the first one with him in a really tall hat also

Is 5head an embedding?

[oak osprey]

lmao

[smoky oak]

It is so humid here in Mexico lmao

Also, don't mind me DRINKING

Hehehe, it's legal here lol

I'm such a ||not|| criminal

[sterile temple]

Adult drinking 😱

[smoky oak]

In the United States of assmerica, it would be illegal lol

[sterile temple]

such an odd mix of laws over there 😄

[smoky oak]

You can kill yourself for your country, but not smoke or have a beer, so fucked lmao

[sterile temple]

We all know substances are the root of all evil 😄

[dense tapir]

Been saying that since they raised the age to 21 from 18.

Hell, they want to give 16 yo the right to vote now but can't do anything else? come on.

@smoky oak Funny that one can go into the military, go off to some foreign war zone and be killed at 16 YET can't buy a smoke or a beer. Yeah, fucked up.

[smoky oak]

I strong agree, but I'm at least I am responsible with substances haha

I just drink for a little fun, and I never get drunk

Only in Mexico lol

[dense tapir]

I don't do either but I once did both. I got drunk too but I am a happy drunk and didn't like when I saw other drunks being mean. I was 19 at the time asking my friend why do they have to get so mean (meaning the others)?

As they say when drunk, orstoned your inhibitions are removed and you get to see the real person. I never forgot that so I do watch how someone is when either.

btw, it has been so quiet in the SD world with each day just getting quieter and quieter.

[oak osprey]

@smoky oak the fulljourney bot running my model now lmao

[dense tapir]

Oh, and reddit is still black so it has been over 2 days now I think.

[oak osprey]

not my prompt, just saw it randomly

i have no idea how they got a good elong mask

[smoky oak]

Oh wow, he almost looks human here

[dense tapir]

I see bitsandbytes 8bit is now on AMD ROCm.

[sterile temple]

An internal memo was leaked from reddit confirming they think this whole thing will blowover and they have no plans of changing anything

[smoky oak]

Good to know, time to take the reddits down for an unspecified amount of time and suffocate them out

[sterile temple]

I read an article, one guys 'conspiracy' theory you could call it. His theory is given the similar investors between reddit and openAI, this charging for access to reddit's data is a way for them to stop startup and opensource developers from being able to use reddit's data to train competing LLMs, since that is the biggest threat to chatGPT

[oak osprey]

aw fuck i thought it was a pretty realistic photo but if he resembles a human, it's not even close

i can't see their prompt so i can't manage to reproduce it as well but this seems accurate

[sterile temple]

giving himself a hug and pats on the back

[oak osprey]

it just "gets" him

cybetwuck

elon musk (infectious laughter) smiling, in a desert, germany, 1995, kodak, realistic, photography

ah yes the mountains of germany

in its famous desert

[smoky oak]

It's like how mark zuckerbergs meta avatar looks more human than him lol

[sterile temple]

Woot! Pay day! 💸

Totally forgot it was the 15th already 😅

[smoky oak]

Munie

[oak osprey]

oh dude now that you say it i can't unsee it

[smoky oak]

IT LOOKS MORE HUMAN THAN HIM LMAO

[oak osprey]

https://images-ext-1.discordapp.net/external/MtlyVGfwfFPVq8FlTNzDjfxxlTsK5e84rr3aZegBM2o/https/tripleback.net/public/discord/1686795381.2495325524cf9770b56d069a06009231a8eb80.png?width=1004&height=1004

so does this

[smoky oak]

Zark Muckerburg lol

[oak osprey]

https://images-ext-2.discordapp.net/external/5DMINHuCFqB9DV-9rQT2mUriEeko0iJgtPehAi9wWwk/https/tripleback.net/public/discord/1686795381.08801846141409b17317cc498be7489f0ab25d2.png?width=1004&height=1004

that came out like this. zark muckerberg

https://images-ext-1.discordapp.net/external/7bSGKgxK-Iw2vxbNQUALsfmiOSoN1Y1kDc_a2lwHiKo/https/tripleback.net/public/discord/1686795380.9888363593681250109ec40fb0a528e139561b7.png?width=1004&height=1004

octane render of mark zuckerberg

[smoky oak]

Lmfaooo

[oak osprey]

https://images-ext-2.discordapp.net/external/7tLcMGsWN4kIx66NA_n-cA1H0nRdJeY5UALRtHiSPvw/https/tripleback.net/public/discord/1686795512.82542063c5c89a9611000fb882f7ef4fde884c5.png?width=1004&height=1004

why is it so resistent to pixarification

i mean, why is he so resistent

whoops

@smoky oak 🥹 does mexico let you make AI images, if yes, can you pixar him plz

https://images-ext-2.discordapp.net/external/K_SdUtMfVfLvkPjAgX9a6p8KNQOTCWC2130Tut9aB04/https/tripleback.net/public/discord/1686795587.5289996f8ca6b7e0a99c84aaf2c255b4174ab01.png?width=1004&height=1004

you have to weight pixar ungodly high

[smoky oak]

I can access my PC in Mexico lol

[oak osprey]

im surprised that much pixar didn't set mark on fire

https://images-ext-1.discordapp.net/external/B0_wBSp4FN1R4XwrsvOBCsntFCwntfQdzDkvm8EGmhE/https/tripleback.net/public/discord/1686795617.32248a9881812868f6e4d763a9b187c8f41be.png?width=1004&height=1004

we're at CFG 9.5 now

https://images-ext-2.discordapp.net/external/Ax2K941mxG0vpCgyL246sXumxXuaO4pCb8iPhgCETTU/https/tripleback.net/public/discord/1686795692.970940898609b654220f1120f4a2c3e046ffd3f.png?width=1004&height=1004

fuck

it flipped over?

https://images-ext-1.discordapp.net/external/GbDBGPHRtbCIJOADx5txl5CvsRWpiRGdJKGmVrc8uWM/https/tripleback.net/public/discord/1686795779.8387895e2eaee3d5cbc4237e9b328a436215624.png?width=1004&height=1004

CFG 12 is getting weird man

https://images-ext-1.discordapp.net/external/TvA9OUQzpKaZMofsrQ49IKjEVuRoNPzWWSqnVBdQaDg/https/tripleback.net/public/discord/1686795820.1419344b6493b6705c91113673d29c7d4292240.png?width=1004&height=1004

bingo

[smoky oak]

Lmao, I did him with the furry model

[oak osprey]

oh nooo

put it on the dark web

[oak osprey]

https://images-ext-1.discordapp.net/external/sV3QIF4MlSzeKEsaWcwrBOKhdBCF8NeBk15MCgiPADU/https/tripleback.net/public/discord/1686796008.9916165d6fdcdced3574a9770aee3991aa004f2.png

https://images-ext-2.discordapp.net/external/F4rLVNGMUW9orCogByEveH4Wds0sRdEwFbVhFPZB0h8/https/tripleback.net/public/discord/1686796010.0907195d19ffe6ade11eeb0af28876e497b9efb.png?width=1004&height=1004

mark zuckerberg searching the dark web for a rocket to take him home

[smoky oak]

This furry model makes him actually look human, inronically enough lol

[oak osprey]

i still do not trust that face

https://images-ext-1.discordapp.net/external/rkbDpNGO9qsts0lZIzLSVDOg7vmxkgRE9PdeOda7TZM/https/tripleback.net/public/discord/1686796204.16796231c93597028825e3d40a09c0a1d623c7e.png

my friend got creative

[smoky oak]

I wanna do something cursed lmao

@oak osprey do I use the furry model to make him lewd faced? Lmfaooo

[oak osprey]

have him make the belle daphine face

[smoky oak]

I can do it lmfaoooo

[oak osprey]

https://images-ext-2.discordapp.net/external/WUMzMe2bOTPKlxDXV47n_PTDQEy22EbEGb4xl9uAjl0/https/tripleback.net/public/discord/1686796848.9103248aa3b01c0c4a85348c104821219f601b9.png?width=1004&height=1004

oh man oh man

why is this the first thing i ever successfully got .and() to work for

("mark zuckerberg", "as belle daphine").and()

i don't think it knows who belle daphine is

oh

i spelt it wrong

[smoky oak]

@oak osprey ok, I need to send in DM's, cause this is too NSFW for the server lmao

[oak osprey]

go for it

you nasty

[smoky oak]

Thoughts? Lmfaooo

I will so get banned if I share that, SO

[oak osprey]

yes

it's a fine product. te best product

[smoky oak]

Oh, my school had a Doinaois beanitois as well

[oak osprey]

the best lunch time meat spread

[smoky oak]

SAYS POOO, BENE IN POOO

[oak osprey]

is it too late to get in on the balenciaga jacket trend?

when you have to sneeze

@smoky oak try sneezing in slow-motion, uhd

[acoustic oracle]

Why isn't the lora appearing in addnet?

[oak osprey]

tourist trap

[oak osprey]

i can't believe how often people come here to generate nasty illegal stuff

makes me never want to run any kind of similar service

[sterile temple]

damn, $120 for Starfield, and 170 bucks for Starfield Premium Edition

and it's in the top sellers on Steam, 'cause Bethesda games have always been so reliable at launch 😄

[dense tapir]

I just spent a half hour training on colab and even though it showed saving nothing was saved.

[dense tapir]

No idea what colab is pulling as the next small attempt it saved but still took a really long time.

I just can't bring myself to pay 950+40s&h for a used GPU sight unseen even if a 3090ti.

[oak osprey]

night

[sterile temple]

1 month after giving up on getting fibre internet and switching to 5G, I get an email saying that I can now upgrade to fibre for free I'd have to pay an extra $50 to get a plan that is faster than 5G anyway

[wispy nest]

noice

[dense tapir]

Alright, my model works for style but lora/lycoris/loha fail

Using my style model

I will see if I can extract the style from it

1200 steps but I think 800 would have been enough

judging by this 1000 steps was the best loss

Oh, I forgot I didn't even call it.

Calling it really changed it

[split rover]

Come try something new in the bot channels above ;D

[dense tapir]

My new locon

[dense tapir]

Anyone know how a dylora is made or even wth it is? lol

[sterile temple]

never even heard of it, makes me think of dora the explorer for some reason 😄

[dense tapir]

Same, but it is for real

[sterile temple]

the purple thing is cute-ish 😄

[dense tapir]

Well, it tries, lol

[acoustic oracle]

It looks like a photograph, very good

[sterile temple]

thanks

[acoustic oracle]

what model did you use?

[sterile temple]

a mix i made

[acoustic oracle]

what did you use for the mix

[sterile temple]

it's been evolving over a couple of months, i add and subtract various models using the weighted block merge in the supermerge extension. some of the models I can remember using: cheese daddy, pastel mix, seek art mega, sd art complete edition, icbinpl, analog diffusion

[acoustic oracle]

ok

[cyan snow]

dude, i have been using euler_a up until now, DPM++2M Karras is WAY better

[sterile temple]

I change between those two

[wispy nest]

love it

[oak osprey]

lol. sdxl has american bias still

i wonder if normal looking people are in the dataset at all

just attractive influencers and celebs

[wispy nest]

wow damn

[sterile kiln]

Wow, nice

[smoky oak]

@oak osprey why are you like this? ._.

lol

[oak osprey]

pseudoterminalx, encapsulated in his own fears, all of the guilt, intertwined with, LOTS OF NOISE

escape imminent

[oak osprey]

@proud dagger film grain is possible to train into the unet by the way. this is both TE/unet, then just the TE, and then just the Unet, various combinations of the starting checkpoint with 13,200 steps at batch size 150

i haven't touched the VAE but you can improve small faces at a distance on 2.1-v with the (deep fried) negative

when your arthritis kicks in

@cyan snow lol

dead lmao

[torpid ridge]

Hey everyone, I'm a newbie here! I've been on the discord for a bit trying to look at old convos about how to gerenate a not-so-professional taken-on-an-iphone natural look using Automatic1111. If anyone has any tips or knows a good checkpoint model, please let me know!

[cyan snow]

?

oh, you also got to the point where your model turns most old people into witches and wizards?

[ripe cedar]

@smoky oak https://github.com/pkuliyi2015/sd-webui-stablesr have you tried this to upscale? It claims to be really good

[smoky oak]

I have not, I'll have to take a look sometime

[oak osprey]

yup

should call the model, elderjourney

[smoky oak]

@ripe cedar I'll give it a try, but they lost me when they boasted about 4k on under 12GB VRAM lol

You can do 4k on 4GB VRAM if you know what you are doing lol

Also, the lack of examples is mildly annoying

Show me what to expect before I sink that much time into it :/

They do attach examples, but not of their implementation

I don't see any benefit to major upscaling before giga-GAN, or a new diffusion method that is not size constrained and works as a refining process from edge to edge

It would be amazing if we could get our hands on a progressive refine diffusion method that is just a scanning addition, rather than a full content aware diffusion

Just as an alternative for some cool things

[acoustic oracle]

I used kawaice and a lora for eyes, any suggestions to improve?

[oak osprey]

i think they mean, in the same hour.

=p

[smoky oak]

A 3060 could do 4k in less than a minute likely

[oak osprey]

it takes 90 seconds for my laptop 3070 to run out of memory at the end

ahahahaha

[oak osprey]

toddler nascar drivers

that's not dr. phil as a hindustani toddler at all

[oak osprey]

@cyan snow what happens with ethnographic photography in that prompt?

[cyan snow]

let's find out, wait a minute

do i put this in one of the commas?

[oak osprey]

at the end

in my bot i can also do

("prompt piece one", "something else", "piece three").and(1.0, 1.2, 0.8)

and it'll blend those three prompt segments in the final output, offering more control. i don't know if Automatic1111 has this by default. if you have Compel 1.2.1, it should be an option.

[cyan snow]

didn't make a huge difference

ok, this one got fucked

i have no idea what went wrong here

[oak osprey]

interesting

[oak osprey]

really cool how the jar fucked up but not the internals lmao

[cyan snow]

i hate to say it, but that seems like it was a bad seed gatcha

[oak osprey]

well that's not a bad thing

sometimes the math just shits itself

[kind quartz]

try this seed 3809589491

[cyan snow]

huh

[kind quartz]

o.k. I just if not cursed seed. Got this

[cyan snow]

also, is it just me or if you use the same seed on the same model, but with a different prompt, the seed controls the angle? i did this expiriment once and did 2 landscape shots with different prompts but same seed, and the angle was very similar.

[kind quartz]

that i dont know. Will test tomorrow. Now i am going to watch sport 🙂
Good night!

[sterile kiln]

@cyan snow Where did you got this model??? mdjrny-v5.1

[oak osprey]

that's a trigger word, seemingly for a dreambooth model or inversion or lora that has the style attributes of mj5.1 data

[sterile kiln]

Nope,
Model hash: 35a537c88c, Model: mdjrny-v5.1

Its a checkpoint, a model..

[oak osprey]

probably a model merge he made locally

unlikely it's actually MJ 5.1

[sterile kiln]

That's why I ask him

[oak osprey]

i'm tellin ya lmao

[raven beacon]

hi

[cyan snow]

yeah i called it that because one of the models i merged was called mdjrny-v4, so yeah. it's an improved version of that model.

[cyan snow]

it's a model i trained then merged with a bunch of other models, making it produce results comparable to midjorney's.

[oak osprey]

that's so hot

[cyan snow]

i have never heard someone describe a landscape as hot

[oak osprey]

never been to Phoenix?

or Saskatchewan in summertime 😛

[cyan snow]

nah, i live in the backrooms, i don't know what a sun even is

[oak osprey]

https://prompthero.com/prompt/da045b2f949-openjourney-1-5-the-source-of-future-growth-dramatic-elaborate-emotive-metallic-baroque-and-rococo-styles-to-emphasise-death-as-a-transcendental-seamless

this prompt does not do that result lmfao

[mystic marlin]

First image of the day from me. Hope everyone is well.

[wispy nest]

LMS Karras 20 steps

[smoky oak]

Whuddup schnoodled goobers

[wispy nest]

think its even better than DDIM sampler

[oak osprey]

Karras ruins everything imo

[smoky oak]

Same here

[oak osprey]

a month ago i would have disagreed, but that's because the other samplers had implementation issues

now DDIM is fixed. and it kicks ass

[smoky oak]

Not sure why people stan it so hard, it's slow, needs more steps, and looks worse (it does have benefits in some 2.1 models, but not in the models people actually use :p)

[wispy nest]

also one of my biggest issue was character's face looking weird and if i used ((highly detailed face)) the image would turn out extreme close up

[smoky oak]

DDIM, my beloved

[dense tapir]

https://mods.reddithelp.com/hc/en-us/articles/16693988535309

[oak osprey]

because the sigmas are a huge factor in reducing noise

it overly blurs the images which is "aesthetic as fuck"

[wispy nest]

i should run a comparison test between DDIM and LMS Karras 🙂

[oak osprey]

the model makes a lot of difference to how those tests go

[wispy nest]

i can't change model for consistency

[oak osprey]

use base 2.1-v to see the true extent of the horror fest

[wispy nest]

im rendering two images with same settings and all but different samplers

[dense tapir]

Now that appears on all links I follow on Reddit.

[wispy nest]

this is what came out with DDIM 20 steps

one of those images are slightly weird

[wispy nest]

but I also see DDIM is faster in my console

[wispy nest]

are those from LoRA?

[cyan snow]

no

[wispy nest]

wish they were, so i could download them lol

[cyan snow]

oh, they are from the model i use. no need for lora

[wispy nest]

nice, ive been looking at some of the LoRA for anime but haven't found many that I liked

[cyan snow]

there are plenty of checkpoints that can do anime perfectly

[wispy nest]

yes but i wanted to retouch the background like what you have just posted

[cyan snow]

i think i understand what you mean, you want to make like anime landscapes? anime models aren't good at making landscapes usually.

[wispy nest]

yep, always wanted to make landscape that stood out

so far im using AOM3, Counterfeit .. but looking for some LoRA for the background

[cyan snow]

i think if you approach this in the oposite way, it will be possible. if you will use a model like the one i use then use a LoRA that does anime you should get good results.

[wispy nest]

that sounds interesting, i might try that 🙂 but which checkpoint are you using for those images?

[cyan snow]

long story. i calculated which combination of models can produce a model capable of making images comparable to the ones made with MJ. well, my experimentation seems to be a success. i haven't uploaded it yet, but i might in the future.

[wispy nest]

oh 🙂 sure if you post them some time later id like to try them out

and thanks, hope you actively work on them

[smoky oak]

Sigma males lol

[wispy nest]

I wonder about the face issue when rendering a full body view that always gets distorted, is that a technical lacking? cause with close up faces turn out fine.

[oak osprey]

lmao

[cyan snow]

oh no

what even caused that

[dense tapir]

It is called modern art

it is a style

Derrived from the pschyadelic art days.

[cyan snow]

that's what people on drugs see

[dense tapir]

LOL, this is funny

[dense tapir]

nfixer can be bad

[oak osprey]

https://www.twitch.tv/ai_spez LOL ai reddit ceo

[cyan snow]

bro wtf

[oak osprey]

@split rover i don't want to flood that channel with trash from MJ so here it is, an example of one of its highest quality images that makes me wonder why people care so much about it, other than that we cannot have it.

[smoky oak]

||sigma BALLLLSSSS||

[quartz ravine]

"highest quality": you should know that quality depends on the prompt

[oak osprey]

nope, i have ligma

[oak osprey]

the quality is the FID score that is calculated for this image

i'm not discussing this from some "user perspective" 🙂

https://wandb.ai/ayush-thakur/gan-evaluation/reports/How-to-Evaluate-GANs-using-Frechet-Inception-Distance-FID---Vmlldzo0MTAxOTI

[wispy nest]

Chong and Forsyth showed FID to be statistically biased, in the sense that their expected value over a finite data is not their true value. Also, because FID measured the Wasserstein distance towards the ground-truth distribution, it is inadequate for evaluating the quality of generators in domain adaptation setups, or in zero-shot generation. Finally, while FID is more consistent with human judgement than previously used inception score, there are cases where FID is inconsistent with human judgment (e.g. Figure 3,5 in Liu et al.).``` well, at least it seem to be some sort of standard still :P

[oak osprey]

oh it's not the best, or even great but it's all we have

[wispy nest]

isn't best defined as just that? As in, the best we got currently? :P

[oak osprey]

yes

you damn wizard of words

[wispy nest]

so if it's all we got, then isn't it also the best we got? :P

[oak osprey]

"it's always in the last place you look"

i want to smack anyone who says that ironically

[wispy nest]

yeah, I hate it, why can't I find my keys before the last place. Takes so darn long always >:(

[oak osprey]

look there first, duh

i don't know why more people don't do that. they waste so much time lookin in places they won't find the item

just look where you'll look last, first. and bam

[wispy nest]

I guess it's the same people who get sick, I mean, isn't that just a plain bad idea? It's also unnecessary ;P

[oak osprey]

i never get sick unless i want to, just to keep my immunity up

if you don't purposely get sick sometimes, it happens by accident. and then, RIP

basically if i absorb all the possible bacteria and viruses it's impossible for other bacteria to take root

[smoky oak]

@dense tapir oh, the formatting on your name changed

[dense tapir]

We all did, or will

Yours did as well

[oak osprey]

Genetal Awareness

[smoky oak]

Yeah, but I mean your name in the server, not just your at

[wispy nest]

their name looks normal to me, other than that "at" symbol

[smoky oak]

It used to be General Awareness, not it's just generalawareness

[wispy nest]

is how it looks to me

[smoky oak]

Oh you know what, I am on an old version of discord since I can't login to my Google account on this old phone, I wonder if that's why

Yeah, that's probably why, my bad

[wispy nest]

you're seeing our profile names, not the displayed name

we all have to change them if it's not #0000

[smoky oak]

I know

It's just a weird bug it seems, oh well

[wispy nest]

now you know it twice ;P

[dense tapir]

We seem to see ourselves differently

[wispy nest]

🤔 🥳

grr, I can't seem to get less "fog" effect, why ai not listen to me! >:I

[dense tapir]

Only lowercase too.

[cyan snow]

yeah that's annoying af

[smoky oak]

Weeeeird

[dense tapir]

They will come to you soon enough

it will just pop up forcing the change

[oak osprey]

@smoky oak feast your eyes on this

ive stared at it for 5 minutes and i have a migraine now

[smoky oak]

Lmao

[oak osprey]

man its so good

not a single artifact

can you upscale that for me to 8k

[dense tapir]

Something I just updated just screwed me

[dense tapir]

Must have been the lycoris extension as now I get one gen and OOM even when trying to load another model\

[smoky oak]

I have done some extreme bokeh images in 1.5 haha

Fun to mess around with

[cyan snow]

huh, it seems dall-e combined diffusion models and transformers models

[oak osprey]

bing

[cyan snow]

again?!?!?!?

im telling you, bing will be the end of us all

[smoky oak]

Bing is really just Dalle 2 and Chat GPT together

[oak osprey]

you saw that M. Night Shyamalan movie too eh

[cyan snow]

no, i haven't

[dense tapir]

I hate the integration of GPT into bing

[cyan snow]

they already beat us with making toddlers

[dense tapir]

At least let me remove it but it is on everything now.

[cyan snow]

same, why can't a1111 do that