🏞｜general-with-images

[swift harbor]

That's why I must have the complete set used in the prompt if I want to have hopes to reproduce the sample images

[livid ermine]

that place there is called titties

i see i see

[swift harbor]

Well, at least that I can see. LOL

[livid ermine]

hehe

[swift harbor]

In Olympic games, Brazil is called BRA.

Not very flattering

[livid ermine]

haha

nope

damn fine dad joke though 😄

[swift harbor]

LOL

[livid ermine]

they could change their name like turkey just did 🤷

[swift harbor]

Can a country change name?

[livid ermine]

apparently so

[swift harbor]

What is Turkey called now?

[livid ermine]

the brits seem to have their own name for every country in the world, except for their own

turkiye

[swift harbor]

Well, Japan is not called "Japan" either.

[livid ermine]

true that!

[swift harbor]

That's a random name Americans invented

[livid ermine]

really?

[swift harbor]

It means nothing in Japanese

So it couldn't be a Japanese name

[livid ermine]

hmm norway is equally stupid imho

and sweden

(norge and sverige)

i think its the oxford dictionary that decides in the end

[swift harbor]

Well, Brazil is still called the same. Except in Olympic games.

[livid ermine]

thats where u gotta send cocaine and prostitutes if you want your country's name changed (or restored rather)

[swift harbor]

And it's written with S and not a Z. Not sure why they changed it.

[livid ermine]

yes!

[swift harbor]

Japan was even worse, because it's totally random

They made it up out of whim

[dense tapir]

Nippon

[swift harbor]

YESS!!!

[livid ermine]

love brazil, i collect lps from there and am now working on a music project with this brazilian rapper

[swift harbor]

Not so hard, is it?

Brazil was once the source for musical inspiration.

[dense tapir]

This is where the British term of Nips came from in WWII. When did it get renamed to Japan?

[swift harbor]

Americans called it Japs.

[dense tapir]

Yes

[swift harbor]

It makes it sound like Nippon is unpronounceable, doesn't it?

[livid ermine]

well, adopted. he calls himself angelo reira and is famous but his real name is the most corny sounding white name you can think of

luckily many languages still seem to call it nippon though

[dense tapir]

Old world maps from 19th century, and before, have it labeled as Nippon

[swift harbor]

I must have missed that. In my maps, it's still Japan

[livid ermine]

its so much more comfortable to pronounce too

[dense tapir]

OMG, 4m per image to get a caption?

[swift harbor]

And most people don't know what language is spoken in China, considering there is no "Chinese" language.

[dense tapir]

There is the CCP official language

[swift harbor]

Not to mention most people believe Brazilians speak Spanish. Most Disney movies believe that

[dense tapir]

Winnie the Pooh and all that

Portuguese, right?

[swift harbor]

Yes, but wasn't that in Korea?

Right!

[livid ermine]

in half malay and there its called nippon

[swift harbor]

Somebody need to tell Disney about that

[dense tapir]

Disney, pffft

[livid ermine]

oh please 🤣

btw

[swift harbor]

That last Disney movie happening in Brazil, the music was from cental america, the buildings were from Mexico, and the language was... Spanish.

[dense tapir]

still sold 200m of tickets

that is all that matters

[swift harbor]

Dawyne Johson was in it.

That sells!

[dense tapir]

Shit, lol

He is part Samoan

[livid ermine]

re: photorealism, what would be really cool is a way to add https://vsco.co/store/film/resources

[dense tapir]

Okay, this is bad ass

[swift harbor]

Part Samoan looks just like Brazilians, according to Disney.

[livid ermine]

word! samoan and malay as well as maori are pretty similar

btw that frozen movie is supposed to have taken place right across the mountain here in a place called "arendal"

[swift harbor]

Cool!

[dense tapir]

Never cared to watch it

[livid ermine]

yeah me neither, only cartoons i watch is Aqua Teen Hunger Force 🤣

[dense tapir]

GOD, my elderly step father was in love, literally, with those

so much gonna build me a snowman

[swift harbor]

I watched that Disney movie, and leaned we have Mayan ruins in the Amazon.

Learning never ends

[livid ermine]

oh man, its supposedly way more than that

a pre-historic civilization around 12,000 years ago who is said to have had technology as advanced as we do now (albeit in different ways)

they seeded the entire amazon, built the pyramids etc.

The Younger Dryas impact hypothesis contends that an extraterrestrial object exploded over North America at 12.9 ka, initiating the Younger Dryas cold event, the extinction of many North American megafauna, and the demise of the Clovis archeological culture.

[swift harbor]

I bet they spoke Spanish too! 😄

[livid ermine]

ehehe

no tiendo

[swift harbor]

Before Spain existed

Ok, now I have ALL the embeddings used on the prompt, but the image looks nothing like the reference. What am I missing?

[swift harbor]

It's this image - can anyone get it to look like this? --> https://civitai.com/gallery/281696?modelId=21493&modelVersionId=25632&infinite=false&returnUrl=%2Fmodels%2F21493%2Fhellmix

[livid ermine]

dunno about language, but some speculate they built the pyramids using soundwaves to melt rock and levitate it into place

we cant build the pyramids today!

[swift harbor]

But we like to think we can.

[dense tapir]

They all sat around a fire and hummed really loudly too

[livid ermine]

(my mom were one of the first people inside the pyramid of giza btw)

no thats my point, this civilization was long before the ancient egyptians

they just found it there

anyway this was all recorded in the library of alexandria which sadly burned down

[swift harbor]

Well, the Bible says the Hebrews destroyed the walls of Jericho by playing trumpets against it. I wonder how long that took.

[dense tapir]

Fact is only a self centered fool would believe that in all of the universe only life exists on this grain of sand.

[livid ermine]

all the knowledge we lost 😦

its since been rebuilt by norway and snøhetta (whom i have a project with actually): https://snohetta.com/project/5-bibliotheca-alexandrina

[livid ermine]

indeedie-o

[swift harbor]

Cleopatra is said to have read many of the books in there. She spoke many languages, and received the Romans speaking Latin.

[livid ermine]

there's really only one way to find out, and it's called dimethyltryptamine 👍

true that

[swift harbor]

It is also said she has received Mark Anthony totally naked. Whether or not that was true, it worked!

[livid ermine]

well, cheers guys back to ruby

will keep yall posted, thanks a lot for the help!

[swift harbor]

See ya

I am still trying to figure out why I couldn't reproduce ANY of the sample images from the HELLMix model. I have used the same versions and embeddings.

Not a single one.

In the sample link from above, the hair is way more detailed than what I get.

[dense tapir]

I am still trying to figure out why I come back to 1.5 for the day and keep getting titties all over the place. I mean undead titties?

[swift harbor]

They face and eye details are also way more detailed and sharper than what I get.

Haven't heard of anyone complaining about getting titties so far, but there is always a first time.

[dense tapir]

So many NSFW in 1.5 I had forgotten.

They stripped those out from 2.x

[swift harbor]

Like Julius Caesar used to say: "Give the people what the people want".

[dense tapir]

I can't go that far

[swift harbor]

I wonder how these lawsuits against AI will end up

[dense tapir]

We want A-Bombs. Alright, here ya go. Now remember to play nice with it.

[swift harbor]

That sounds familiar

[dense tapir]

Not having nudes is one thing that hurt 2.x but they legally can't have children and nudes in the same model.

[fiery lichen]

I wonder which 1.5 model was chosen for a sketch to cartoon app for kids. A father developed it for his 3y/o kid, what if the kid wrote the letter w in small caps, and waifu diffusion decides to show the 3y/o a pair of titties.

[dense tapir]

LOL

[swift harbor]

Hahaha

[fiery lichen]

does SD 1.5 standard do that?

[dense tapir]

oh, hell yes

[fiery lichen]

i bet he can only use SD 1.5

[dense tapir]

I had so many titties in 1.5 base

[fiery lichen]

i think there are self-censored 1.5 custom models though...I think the dad better use those. but even in 2.1 we still get accidental half nipples

its rare but it still happens on 2.1, then i just put nsfw in negative prompt

[dense tapir]

yeah, 2.1 I don't know when it was I last saw a nipple with it.

[fiery lichen]

maybe the dad put 1000 nsfw prompts in negative to offset the potential

[swift harbor]

I think this whole thing is totally BS. People will get the models they want, whether they are forbidden or not.

[dense tapir]

first image with 1.5 today was something I couldn't post here. 😦 Full frontal nudity.

[swift harbor]

There you go!

[dense tapir]

pretty cool image though

yeah, thing is SAI can't be the ones delivering it to you.

as emad said with 2.0 do you want children or nudes and you can only pick one.

[swift harbor]

I am curious now. HELKMix is a 1.5 model, but prompts use 2.1 embeddings. How??

[dense tapir]

noidea as I have a ton of my own and other embeddings for 2.x

[swift harbor]

I am starting to suspect this is why I can't get the results in my attempts.

[dense tapir]

auto shouldn't even allow that

[swift harbor]

These prompts cannot possibly work with a 1.5 model

[dense tapir]

like what?

[swift harbor]

It doesn't! You get an error

[dense tapir]

always pay attention to the cli window

damn, these models for interrogator are HUGE 10gb each and they all went to my system drive too.

need 4tb nvme m.2

[swift harbor]

The HELLMix samples images are too small to be from 2.1, but some prompts are using 2.1 embeddings. I cannot reproduce ANY of them. Something is fishy here.

[dense tapir]

oh, even more is downloading. before done >100gb

I saw hellmix on civitai but was not interested since it was not for 2.1

[swift harbor]

I am still using 1.5 here

[dense tapir]

I am only doing an experiment today then back to 2.1 for me

[swift harbor]

Based on what people are posting on CivitAI, 1.5 is still the most popular.

[dense tapir]

I hope SD is still around by Oct but SDXL is just flat out stupid. As I said it makes them seem cheap to call it that.

cause it has nudes

easy to work with, and fast for old cards

[swift harbor]

Haven't seen or heard of it

[dense tapir]

for photorealism I take 2.x any day over 1.x

SDXL is, I believe, 3.0

[swift harbor]

Makes sense, bigger 2.1 produces better small detail.

[dense tapir]

yes

[swift harbor]

More pixels, more detail

[dense tapir]

problem is they only made it 768 for the last part and why we get elongated necks sometimes

would be even better if all steps were 768sq instead

[swift harbor]

There is that too. Larger spaces, more duplications, and general mutant creatures

[dense tapir]

if they had 768sq for all steps there would be no stretched anatomy

[swift harbor]

Even in 1.5, if we make bigger images, mutants become inevitable

[dense tapir]

not sure wth they were thinking in doing it the way they did

[swift harbor]

So what's new in SD 3.0?

[dense tapir]

not sure. SDXL was supposed to get 1024sq

[swift harbor]

Ahhh

[dense tapir]

might be why XL

[swift harbor]

Sounds like the natural step after 768

[dense tapir]

even 4080s will cry with those at inference

[hoary sable]

Is it possible to generate 2k or 4k images by the model itself, without using Real-ESRGAN and similar upscaling models?

[dense tapir]

yes

[swift harbor]

I can only imagine

[dense tapir]

I did it before auto nerfed/changed mem stuff now I get OOM

[hoary sable]

I know about Hires Fix, it does not work bad but to a certain point.

[dense tapir]

2k took my 1060 over 30m to gen

Never liked it

[hoary sable]

I also tried the SD Upscaler but nothing really works.

[swift harbor]

I imagine 1K images might take even a 3090 to its knees

[dense tapir]

yes, from the numbers I saw for 768x1152

[swift harbor]

We can already do that size with 2.1

[dense tapir]

Ultimate SD upscaler is just fantastic

basically 1024x1024 vs 768x768 is 3 times longer

768x768 is 3x over 512sq

[swift harbor]

Once AI upscaling started being used in video game GPUs, things started developing really quick

[dense tapir]

works out not quite 3x but close

[hoary sable]

Yes exactly, I still do not understand what values are the most optimal.
For example, also Denoising Strength.
It's in Ultimate

[dense tapir]

I know what Nvidia has said that having assets gen on the fly is coming

[swift harbor]

I might stay with the smaller models, just to benefit from quicker results.

I am not buying a new GPU any time soon after what I this 3090 cost me.

[livid ermine]

i love noise, my two girls from earlier was brought into gimp and given film grain with some blur and it looks so much better

but those eyes, damn

those image restorators https://replicate.com/collections/image-restoration doesnt help at all, and they take away a lot of the photorealism

ive always said the eyes are among the most advanced objects in the known universe

like you can always tell if somebody miles and miles away has locked eye contact with you

(partly the reason why i dont stare at girls 🤣)

[dense tapir]

damn, my card can't run this model for interrogate OOM

[livid ermine]

i guess we'd need some advanced blender-based simulator to fix the eye thing?

[swift harbor]

Haha 40GB VRAM required?

[dense tapir]

Who knows

[swift harbor]

That's a bit unrealistic

[livid ermine]

added some more analog simulation (www.lomography.com)

[swift harbor]

Ok, I have figured out why I was having inferior details from the sample prompts - I was not using Hi-res. Fix. Now I get the same level of detail - BUT... some of the params don't match the current A1111 version.

[astral osprey]

I opened SD the first time with the batch file, is that what I'm supposed to do to open it?

[swift harbor]

The default state is to start with txt2img tab

[astral osprey]

yeah but how do you get the IP, I got it the first time with the batch file lol

[dense tapir]

Nope, can't use this shit on 6gb

8gb maybe

[swift harbor]

Based on the error message, maybe not even on 24GB

[dense tapir]

If nvidia hadn't been such dicks I would have a 4080 right now

[glossy herald]

for running what ?

[dense tapir]

flan

[astral osprey]

check VRAM usage?

[swift harbor]

No, the error message that was posted before.

It claims it had tried allocating 40GB VRAM, which is absurd

No wonder it has failed

[astral osprey]

I was getting that too. Thought it was my VRAM

[dense tapir]

6gb is not going to work with that one

[astral osprey]

as in it never cleared it

[swift harbor]

Time to give nVidia your savings

[dense tapir]

yeah, when you error like that with OOM close it and restart

I am saving but ironically what the 4080 is was what the 4060 was supposed to be only costs more

[swift harbor]

The 40XX pricing is still absurd.

[dense tapir]

I will not spend over 800 on it though

Not going to come down anytime soon unless we hit a global depression

[astral osprey]

so glad I got 6800xt for 500. fk nvidia pricing, I woulda gotten 3060 if it ever dropped under MSRP

[swift harbor]

Not until AMD starts posing a real threat

[dense tapir]

I have no choice as I need a card to train with and AMD sucks for that

[swift harbor]

My point, exactly

[astral osprey]

train what?

[swift harbor]

No real competition, no fair price

[dense tapir]

Dreambooth, etc...

[astral osprey]

ah yeah.

it do have AI cores

[dense tapir]

Considering Intel makes the AI chips that go into autonomous cars I think they can somehow get that over to their gpu division.

[swift harbor]

The prompt I am figuring out has these params at the end: "Clip skip: 2, ENSD: 31337 and Eta DDIM: 0.3". I can't find them in A1111. Any hints?

[dense tapir]

all that is in settingas

clip skip is meaningless in 2.x world it does absolutely nothing

makes sense cause it is a different clip

[swift harbor]

Does that mean I am supposed to change the A1111 settings to use this prompt? That sounds unusual....

[dense tapir]

no, not unusual at all.

not in 1.5

[swift harbor]

I guess I haven't been doing this long enough then

What do they affect?

[dense tapir]

All of that is what NAI used

when they trained their model

[swift harbor]

What do these param mean? What do they affect?

Would this be significant to the final image?

Or can I skip them?

[dense tapir]

Looks like I can't use them

Now I need to see where it shoved them so I can remove them as they were huge

[dry crow]

ENSD is called Eta Noise Seed Delta in settings

[dense tapir]

I wonder where these models went?

was on my C drive for sure

I ticketed it to find out. Bummer I can't use it so I can train as I need the captions

Even the dev had no idea where they go as he said models/clip-interrogator but bzzzzzzzzt, wrong

[chilly mural]

This is a photo that I turned into an AI video using Stable-Diffusion on Unishape.io. 【Animate】It really looks like a real person. AI is becoming more and more amazing.
https://www.unishape.io/creative

[smoky oak]

anybody know if there is a better version of high res fix? One that doesn't change the base image so much?

I am really struggling with my super high res workflow, cause I need high res for more detail, but it changes the image massively

[glossy herald]

Usually reducing denoising in high-res fix helps on that side. But i guess you could also use the image without high-res fix as controlnet input to force more details to stick

[smoky oak]

hmm, never thought about that

the issue is if I go any lower with high res fix, the image just gets blurry, which is a flaw of high res fix

[glossy herald]

You can add more "denoising steps" to force those to happen a little more

But yeah, this is the problem on too low denoising

[smoky oak]

it wasn't that bad this time

but you can see the big leaf plants on the right are now small leaf plants, when i wanted them big ._.

[glossy herald]

I use "tiledVAE". It's good for very high Rez, it helps the high-res fix, but I'm not sure how good it is at sticking with composition like this

[smoky oak]

I just wish it stuck more to the original, while just refining the detail from there

[glossy herald]

And did you try to just upscale the picture ?

[smoky oak]

I find tiled VAE is great for some thing, but it falls apart when you get to much higher resolutions

[smoky oak]

it can do that, but it adds too little detail. Its an awkward balance

[glossy herald]

Yeah, complicated balance for sure... I would try the controlnet thing I think

Using canny for example

[smoky oak]

I am trying a higher denoise ultimate upscale to see right now

@glossy heraldsuccess!

[glossy herald]

How ?

Ultimate ?

[smoky oak]

its not super close to the original, but its wayyy better than the high res fix

yeah, ultimate, with higher denoise

if you use too low denoise, it just kinda stays at the same detail level, but goes higher res

so I cranked the denoise

gonna try a very high res gen on it now

[dense tapir]

I can't use blip as 6gb isn't going to cut it and for some 12gb is min. Go to HF and it is busy or 1200+ minutes

intricate

[austere pilot]

Can you use this for 3D models as well?

[void hull]

my god

if i commisioned a human artist for $500usd and got this i'd be happy as fuck and the AI just made it in 10 seconds

[dense tapir]

This is why they be afraid, so very afraid.

@smoky oak

[trim jacinth]

what model should i use for animal and buildings?

[trim jacinth]

wow what model doyoou use?

[trim jacinth]

which model did you use?

and what model should i use for good building and animal

i will use with controlnet

[glossy herald]

Deliberate

and this is with one of their test prompts

hum, buildings with controlnet, 1.5
animals, no idea

[austere pilot]

@grizzled sage

[grizzled sage]

300 frame limit. 10 seconds. thats what i said. you need to pay attention

don't believe the hype. that limitation is GLARING

[austere pilot]

okay, it's still helpfull tho you can create multiple files with 300 frame limit that follows up to eachother

[grizzled sage]

are you affiliated with this channel or something? It literally has nothing to do with image generation, deep learned networks, stable ai, or anything related to what SD can do. It's just a commercial for some over priced software. All of those features are available in blender too

https://blendermarket.com/products/auto-rig-pro/?ref=89 $40 lifetime

don't believe the hype

[austere pilot]

no, blender doesn't have AI auto pose animations, it's not auto-rig either and ik that it has nothing to do with SD. I am just saying it can be used like controlnet is used to create poses but for video....

[austere pilot]

maximo is free

[grizzled sage]

when a simple animation tool has$200 annual costs, there are SaaS business school shenanigans afoot. It's just not worth that much when the same functionality can be done at a fraction of the cost

[austere pilot]

mixamo

https://www.mixamo.com/

the difference is that the entire model reacts natural on your changes in poses that you make to the character and that is what I am interested in

[grizzled sage]

yeah. and it's adobe so it' is probably going to woork well with all there other software. Adobe can get away with SaaS unfortunately, because they bring it

[austere pilot]

with the cascadeur one

[grizzled sage]

Their product suites actually bring value to a production chain

[austere pilot]

you can literaly start using it now for free

try it yourself just create a login and try it out

you don't need an adobe account

only a mixamo account

[grizzled sage]

i don't have a 3d modelling workflow at all right now. i've modelled before. was doing it before we even had inverse kinematic systems completely established

these days when i do 3d work, i try to lean into blender since it's free and open source software, kind of

[austere pilot]

there is no pricing tab nothing. (the uploaded character is something from school). it even auto-rigs for you

[grizzled sage]

its more of just a grab bag of animated poses to use. can't do much with them outside of asset flips

[austere pilot]

well you gotta place markers where certain things are but that's about is

[grizzled sage]

i'd rather have a system in place rather than copy pasta library

[austere pilot]

in blender you can still change this, I was more interested in the AI physics for natural poses in cascadeur which seems very easy to do there

[grizzled sage]

many of them have like 200 frames of animation, but are the same 10 key frames on loop. Typical asset flip library shenanigans. You get what you pay for i suppose

[grizzled sage]

right. that's why i linked the plugin for blender. $40 one time fee, as opposed to casc's $200 annual

if they're charging that much but their software only has one niche barely useful feature, find the alternative

[austere pilot]

does it do the same?

or very similar

[grizzled sage]

it's a tool set for rigging and posing that does a lot of the tedious legwork for you

same goal. speeding up the same workflow

[austere pilot]

it doesn't have the AI physics which cascadeur offers

it's only for auto rigging

[grizzled sage]

i think you're just bought by good marketing. okay. i'm not going to convince you to find alternatives. go forth. use the free 10 second at a time over priced nagware. Maybe you'll figure out what i'm telling you. Good luck out there. Stay frosty

before i go, To answer your original question, systems like that that cost $200 annual, probably won't be integrated into stable diffusion web ui ever

there's a photoshop plugin tho

[austere pilot]

it does exactly the same as mixamo, only difference is that you can add more of those circles but it's pretty much the same. and even if you don't like the animations with just "coppy pasta" You can use this for the auto'rig without the animations

[austere pilot]

if you find me a cheaper/free alternative I am all in but what you give to me is something else entirely then what I am looking for

[grizzled sage]

if you're trying ot market a $200/year product at me, i'll tell you right away that price tag is a deal killer and i WILL find alternatives if this is a need of mine

but its' not a need of mine

blender is a full featured industry standard editor btw. The plugin has certain autorigging featuers that aid the rest of the very robust software.

[austere pilot]

mixamo is completely free and also has a blender add on that does the same, I am also using blender for school. I was looking for something else then auto-rigging can't you understand that????

[grizzled sage]

blender can dynamically animate a rigged model fine. $200/annual is an instant red flag about that software. this is really easy to hear what im' saying.

Marketing can be very powerful and sometimes it's hard to realize that they're lying to you.

oops i'm mistaken. it's $300/y

[reef frigate]

How about these video to rig solutions? Or maybe even these adapters that ControlNet and T2I-A uses? Open pose and MMPose

[austere pilot]

Yes perhaps but blender doesn't do it automated which I am after.

[austere pilot]

Could you show me what you mean?

[grizzled sage]

Those are more inline with what we're doing here. Machine learned computer vision technologies.

openpose is a whole thing that i'm forgetting about somehow. using it daily practically

i blame the marketing

[austere pilot]

OpenPose combined with AI physics and automated natural poses would be very powerful for video generations, that's what I am been trying to say all this time...

[reef frigate]

I think this is what most people use for video to mocap stuff:
https://www.deepmotion.com/

Free alternative (haven't tested it myself but looks nice)
https://github.com/zju3dv/EasyMocap

[grizzled sage]

yes . an this entire time i've been saying that "Free" software has a hefty price tag and most of what it's selling is just heavy handed marketing. It's not really the cool cutting edge ML'd stuff we do here. there are alternatives. i'm just not in a good position to find those for you this morning. good thing a real dev shows up and has the juice for us though

[austere pilot]

have any idea on how to download these?

do they have a UI or is it all with commands via cmd

[austere pilot]

I understand that free is never really free but in the case of mixamo it is fully free to use and is very powerful considering it is free.

[reef frigate]

There is a colab demo, but not gradio demo. So all cmd

[austere pilot]

oof

[fresh hound]

https://huggingface.co/spaces/damo-vilab/modelscope-text-to-video-synthesis incase anyone didnt know

[loud prairie]

https://youtu.be/tdB56Wq7o5Y

[austere pilot]

[austere pilot]

@reef frigate would you mind making a simple webui for this space to use it locally? if you have the time that is

[fresh hound]

I'd LOVE to be able to use that locally!!!

this is literally like the dalle-1 era

for videos

1 year from now - its gonna be so good!

[austere pilot]

you can already download the files locally I just donno how to use locally if that is even possible rn

also controlnet doesn't seem to be making my character go in a certain pose

[fresh hound]

oh right

if anyone knows how to download the files and use locally would really appreciate if you could make a tutorial for us here

[austere pilot]

you get this msg when you try to duplicate the huggingface space

[fresh hound]

https://www.youtube.com/watch?v=trXPfpV5iRQ&t=1s

i cant WAIT to make my own shows/movies

I've partially written a legit script for a mind blowing scifi movie series with a legit author so to be able to now have motivation to finish it is one thing, but when I can bring it to life will be insane

[hard wagon]

Just released Double Exposure Photo Edition!

https://civitai.com/models/21961/double-exposure-photo-edition-for-2x

[fresh hound]

https://www.reddit.com/r/StableDiffusion/comments/11wmskn/steven_segal_in_the_matrix_100_ai_written_filmed/

https://www.reddit.com/r/StableDiffusion/comments/11wqjhz/text_to_video_darth_vader_visits_walmart_ai/

[fresh hound]

@austere pilot https://github.com/deforum-art/sd-webui-modelscope-text2video

[austere pilot]

You send it twice

[reef frigate]

I'm afraid it's not that fast to make and maintain. I can look into it but I assume it's hard to get compatibility and everything right with all the systems everyone is using

[austere pilot]

It seems like there is a extensions for auto1111 already

[reef frigate]

Are we talking Text2video here or the easyMocap repo?

[fresh hound]

actual text2video

[austere pilot]

The text2video but one for easymocap would be nice to although I figure that is what your answer is for

[reef frigate]

Ah okay, now I get it.
Yeah to use it locally you just need to clone it and run the "app.py" file with python. I use conda to do this. The model downloaded wasn't working for me so I downloaded it separately from HF and placed it in the weights folder. Also you may need to update "einops", mine was an older version and didn't run properly.
Not sure if it's worth the hassle right now though.

[austere pilot]

That's for text2video? I confused myself Xdd

[reef frigate]

Yes that's my general workflow for running the HF spaces locally

[austere pilot]

Also I would be fine with something quick but I understand if you don't want to (yet)

[austere pilot]

Are there any tutorials for it online?

[reef frigate]

I don't think so.
It is quite simple though.

1. Install miniconda
2. Install git
3. Create conda environment and activate it
4. Git clone the space
5. Pip install it's dependencies
6. Start the app with python

[austere pilot]

I should have miniconda already and git cuz of very early SD repo that require it, although could be that it is conda not miniconda. how do I create a conda environment again? iirc it was with a CMD command

[reef frigate]

Let's move this to the #🤝｜tech-support channel and I send the commands in there

[austere pilot]

fine by me

[sage matrix]

baby

[austere pilot]

?

[glossy herald]

welcome on the server too, but yeah, strange intro

#1072220168534642768

[austere pilot]

Luck us we're used to strange right

[smoky oak]

Ok, that is extremely reassuring actually. One can only hope that they re right, and it's simply just a lack of streamlining the code

[smoky oak]

I use realistic vision 1.3/1.4 for all of my photo realism works

[grizzled sage]

nice to note that runway are the guys who released 1.5 too

[fresh hound]

oh cool didn't know that! - makes me very hopeful

[smoky oak]

Whaddup diffusers?

@glossy herald

[livid ermine]

What up mayne!

Tryina take the best of all worlds and combine it into one. A superior script that'll be like a promo generator for me and all the artists I'll be signing for my record label.

Using one of the biggest labels in the world https://warp.net/ as an example.

[livid ermine]

There's a 1.5? I haven't tried it yet, only https://huggingface.co/BestJammer/HASDX which produces some good looking women.

Hoping I can combine the two:

ckpts = [
  "https://huggingface.co/BestJammer/HASDX/resolve/main/ckptSXDHAS.ckpt",
  "https://huggingface.co/SG161222/Realistic_Vision_V1.4/blob/main/Realistic_Vision_V1.4.ckpt",
]

[grizzled sage]

no i mean runway made base stable diffusion 1.5

[livid ermine]

Ah I see

[trim jacinth]

@glossy herald when i use depth lib for hand

its works very bad

[livid ermine]

What are you using now @trim jacinth ?

[trim jacinth]

i using depth lib for human/ cartoon models

but its dont work good

[trim jacinth]

i use 2 controlnet 1st for open pose 2nd for depth

propreccors are closed

[dense tapir]

Makes sense. Oh, here is my answer about the low contrast.

@smoky oak

[hardy tulip]

is this the correct way to add commandline args?

[dense tapir]

yes

[hardy tulip]

thanks

[dense tapir]

no comma though

[hardy tulip]

spaces instead?

[dense tapir]

yes

[static tusk]

single space no comma

[dense tapir]

can be 100 million spaces doesn't matter

just no comma

[hardy tulip]

also, what is xformers?

[smoky oak]

So let me get this right, what they're saying is that it's only faster in an extremely restrictive use case?

[dense tapir]

What they are saying is that once these programs are updated to take full advantage of Pytorch2 you are going to need a top of the line system or your system will be the bottle neck instead of Pytorch.

That is fantastic news if true.

[smoky oak]

Oh, that is good news in that case

Though, they are suggesting that CPU can be a bottleneck?@dense tapir

If that's the case, I am fucked. My CPU always ruins everything

[static tusk]

but how often do you do batch size =1?

Im mostly >5

[dense tapir]

You are fucked because that is exactly what they are saying

[dense tapir]

for lora 100% all the time

for inference 100% all the time

for me BS1 takes forever

Basically you will max out your cpu then when you get around to getting a better cpu you will, possibly, go even faster. On a 4090 you would

Remember the 4090 is an overly priced BEAST! We are just getting into stuff that can take full advantage of its speed which is 56% faster than a 3090 for this stuff hardware vs hardware now the software is catching up.

[static tusk]

software for 4090 isnt fully optimised for sure

[smoky oak]

Great. So now I can expect to have my speed limited once again my having a trash CPU. That's why I have been so big on AI. Cause for once, I can do something cool without my shitty CPU ruining it. Everything else I do as a hobby is a massive CPU bottleneck

[void hull]

How much money did you pay for your computer

I got mine built for about 2000aud and it can make really nice images at a rate of about 1m per
500x600 25 steps

I ask the price because i dont know shit about specs and i dont care to learn

[dense tapir]

You are thinking about this the wrong way though.

[smoky oak]

that seems... really off

[smoky oak]

Unless you are saying I will be able to go faster than before, then I don't think I am

[dense tapir]

You will see gains in speed 100% you just will not see the full potential on older CPUs

[earnest solar]

https://images-wixmp-ed30a86b8c4ca887773594c2.wixmp.com/f/6d22fc3f-49f1-4200-a8bb-e75df1d45072/dfs920h-da7ba8a8-790d-4e46-827d-cc3483058dff.png?token=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJzdWIiOiJ1cm46YXBwOjdlMGQxODg5ODIyNjQzNzNhNWYwZDQxNWVhMGQyNmUwIiwiaXNzIjoidXJuOmFwcDo3ZTBkMTg4OTgyMjY0MzczYTVmMGQ0MTVlYTBkMjZlMCIsIm9iaiI6W1t7InBhdGgiOiJcL2ZcLzZkMjJmYzNmLTQ5ZjEtNDIwMC1hOGJiLWU3NWRmMWQ0NTA3MlwvZGZzOTIwaC1kYTdiYThhOC03OTBkLTRlNDYtODI3ZC1jYzM0ODMwNThkZmYucG5nIn1dXSwiYXVkIjpbInVybjpzZXJ2aWNlOmZpbGUuZG93bmxvYWQiXX0.WvjjEPW-8NmZE55J9RC9W3Ty29h313RyW3M4wZkRV7w

[void hull]

What about it seems off

[dense tapir]

Think of it like FPS is 45. Buy a new card it is now 60. Buy a new cpu it is 140+

[void hull]

You are running it locally right?
My computers also like 4 years old at this point

The only thing is that i cant run games like elden ring in the bg while images are generated cayse fps drops to like 10

[smoky oak]

But what I am saying is that now I will be in a predicament where my performance is hindered by my CPU once again, and I have to deal with the slowness of it

[dense tapir]

smh, you will go faster though.

[smoky oak]

gonna have to deal with CPU pinning again

I would rather stay at the same speed and be able to use my PC without it freezing ._.

thats what I am saying

[void hull]

How much was ur computer

Its not a fkn laptop is it (i will laugh at u)

[dense tapir]

which you should

[smoky oak]

There are plenty of laptops that would probably steam roll your desktop, but no, I have a desktop. Though i am upgrading to a laptop when I can afford it lol

I need a laptop for the work I do

(outside of AI)

[void hull]

Yeah maybe for 5x the price

[smoky oak]

what specs is your desktop?

[dense tapir]

Let's say your current system had a 4090 in it, right? You get 10it/s. Pytorch 2 fully implemented you now get 15it/s. Upgrade the cpu you get 40 it/s. Yes, you are not top dog but you are faster than you were at no expense.

[void hull]

No idea, i got it built from parts from one of those websites where they just build it for u

[smoky oak]

Is the expense not having my CPU maxed out?

[smoky oak]

thats more sad than buying a laptop IMO lol

[void hull]

I think it was about 1800 total excluding storage which i paid more for

[smoky oak]

you are the one that paid wayyy more lol

[void hull]

I think the 50-100 extra to get someone to build it for u with no issues is worth

Compared to spending days-weeks learning how to do it yourself and wasting time trying to fix problems

But yeah anyway point is my computers not particularly good and it runs SD completely fine

[smoky oak]

you did not pay $50-100 extra for it to be built lol. you paid at least 20% more

[dense tapir]

Just realize this is all going over to so we have no real say. We get what we get and if it maxes out your CPU to gen then it will mine and everyone else not on the current gen CPUs

[void hull]

It was a flat price

[smoky oak]

You are still missing my point, nevermind

[void hull]

https://cdn.discordapp.com/attachments/1004159122335354970/1087357171588673556/00113-4216156569.png

My computer took 1 minute to make this yesterday

[dense tapir]

No, I know your point I just don't agree with it or can see this as a bad thing.

[smoky oak]

I don't want my CPU pinned. I would rather have my SD be slower, and allow me to use my PC for other things at the same time, rather than it be faster, but my PC is bricked when I am generating. Not sure whats so hard about wanting to be able to use my PC lol

[dense tapir]

Yeah, that will probably go bye-bye for old hardware like my 6 year old Ryzen 5 1600

[smoky oak]

Then I am fucked too lol

[void hull]

Image gen will pretty much always make most stuff unusable while u do it

Youd need really really high specs to do that

What are u actually trying to do at the same time

[void hull]

If its just youtube / text editing / unity or whatever it shouldnt really slow it down

But if ur tryna run a graphic intense 3d game the ofc it will slow

[smoky oak]

use my PC for basic things? Like, its not particularly important

[smoky oak]

everything will slow down if your CPU is pinned

[void hull]

Basic things should not slow down

[smoky oak]

well, they do, if your CPU is pinned

[void hull]

Idk my computer doesnt at least and its not particularly good

[dense tapir]

No, Sytan is right if your cpu is >100% utilized of course everything will suffer but when genning, or even worse training you do nothing else on it.

[smoky oak]

unless they aren't saying full CPU pinning. If thats the case, then this is a way different situation, but that wasn't clarified

[dense tapir]

while it is genning or training

[smoky oak]

Thats not true, I watch videos and stream with my friends when I am genning or training. I use my iGPU to run those tasks, and it affect my DGPU none cause its not sharing resources. But now If I have to also pin my CPU, then everything is gonna be unusable while doing stuff

its a problem I have to deal with any time I do anything hobby related on my PC, and AI was the first one that it didn't matter with

[dense tapir]

As the tech grows we eventually will have to buy new hardware and no way around it. I can't even run W11 (if I wanted to) and eventually I will have to upgrade the OS as I just did for my mom's W7. It just became outdate and left to die so it became W10. I had to buy new stuff for it. Same deal here. There is no escape.

[void hull]

Just run image gen overnight then

[smoky oak]

Thats still not at all what I am talking about, nevermind

[void hull]

When ur not using it

[smoky oak]

that is a painfully bad idea lmao

"generate things on your PC while you sleep"

what, you want me to just magically type in prompts while I get shut eye? lmao

[void hull]

Uh

You are aware you can have it generate more than 1 image at a time, right

[smoky oak]

I can understand where General is coming from, but you allude me lmao

[void hull]

I'm at work rn and my computer at home is making 12x100 images for different character designs

Been doing the same thing for a couple weeks now

[smoky oak]

well I mean yeah, you can continuously generate the exact same prompt over and over and over, but that's nothing of what I need lol

[void hull]

I opened 12 tabs of it and theyre all making their prompt 100 times

Since it takes a lot of attempts to get a great version

[smoky oak]

I get what you are saying, but that is useless for my workflow. Thank you for the suggestion at least

[void hull]

So 100 outputs for 12 different things

Unlucky

[smoky oak]

unlucky?

Man, i wish high res fix didn't fuck up the base image so badly

like, if I spend time selecting a seed, I want the result to look like the seed :/

it changes it so much that these aren't even close to the same image anymore

[void hull]

What size do you guys actually gen at

[dense tapir]

768x768 or WS

[void hull]

I found 500 width 600 height is usually enough to make it look nice compared to the speed

[smoky oak]

I usually aim for 2560x1080

sometimes native, sometimes upscaled

[void hull]

What can you even use an image this large for

[smoky oak]

wallpapers

I got way higher res sometimes

I have an image that is 15360x6528

Made in stable diffusion

[void hull]

https://media.discordapp.net/attachments/753100806898843680/1082298073633533993/ophetra.png

Well that would be why your computer cant handle it lol

Anything above 800x800 my generation aborts cause of ram

[dense tapir]

Well, this was unexpected.

[void hull]

U can get enough character from small image sizes anyway imo

Bigger is waste unless ur expecting people to zoom in

[dense tapir]

Money shot, lol

[smoky oak]

you just don't know how to go higher res. Everything I am doing is stuff you can do yourself

but few people have a use for as high res as I go to lol

for me, I generate wallpapers and other high res compositions... so it makes sense for me haha

here is native 2560x1080

[finite frigate]

do you use multiple diffusion?

[smoky oak]

this is what it looks like zoomed in

[finite frigate]

or how do you do such high resolution

[smoky oak]

vs the same in my 100MPX version

[smoky oak]

no, multi diffusion is kinda trash

at least for super high res

I use Ultimate SD Upscale, which is what Multi-Diff tries to be

[finite frigate]

so what does your workflow look like?

[smoky oak]

batch gen, find a seed, high res fix it, then work with ultimate upscale and inpainting to get a result I like

[finite frigate]

what resolution do you do the initial batch gen at?

[smoky oak]

here is the base res version next to the 36x higher res version

[dense tapir]

I like this one even better

[smoky oak]

I make wallpapers for my desktop, so 1280x540 (half of 2560x1080, which is my monitors resolution)

[void hull]

I mean, i'm sure i could look up a tutorial on it
But i really just dont see the point of anything above 2k because u straight up cant tell unless ur zooming in and looking for details

But that would be pointless as ai art doesnt place meaningful details

[smoky oak]

I generate a ton of 1280x540 images, find the one I like and then generate it again but with 2x high res fix, then work from there

[finite frigate]

interesting

[smoky oak]

this is so full of false statements I am crying

"AI art doesn't place meaningful details"

yes it does lmao

[finite frigate]

what inpainting script do you use?

[smoky oak]

I am not using an impainting script

I am using Ultimate SD Upscale, which is in the extensions tab

it runs as an img2img script called the same

[finite frigate]

oh does that also do inpainting?

[smoky oak]

no, I just do that normally on the lower res images

then upscale once I have what I want

[finite frigate]

ah ok

[smoky oak]

I would send my RAW 100MPX image here, but its too big of a file lol

Here is a side by side comparison of the original 2560x1080 image vs the 15360x6528 image

[void hull]

Can you give me an example of a meaningful detail lol

[smoky oak]

look at that video

all of the detail is accurate, realistic, beneficial

[smoky oak]

and the benefit of an image this high res is you can sell it as a fit it yourself wallpaper

[void hull]

From what i understand u are doing this so if someone wants to zoom in on a random bush in the background and view it up close, they can

[smoky oak]

no, I am doing it for why I just said

A fit it yourself wallpaper

[void hull]

Seems pointless to me but i dont know the type of people who pay for wallpapers

Maybe theyre dumb enough to like big numbers

[smoky oak]

I have a feeling a lot of things are pointless to you. Anyways

[smoky oak]

@void hullThe whole point of an image being this high res is that it can be used for more than just one use. Its a single image that you can crop into. It has enough res to work for a phone, or a PC, or a print, or whatever you wanna do of it. Just cause you don't understand what the use of it is, doesn't mean its useless. And I find it ironic that you are calling other people dumb just cause they have a use for a product I am making for them lmao

[dense tapir]

@smoky oak If I spot something I wish to make better using DDIM what is a good step count for it?

[smoky oak]

I usually use DDIM 10-15 steps for speed, and I got up to 30 if I want the best result

any higher and its kinda pointless, IMO

[dense tapir]

thanks I was doing 30 for quality.

[smoky oak]

yeah, that would be my recommendation

[dense tapir]

same prompt at 30 from 10

[smoky oak]

30 is where is the last point where its worth it

[smoky oak]

this is the 10 one?

[dense tapir]

no, that is 30 the one above it was 10

[smoky oak]

oh... weird

I prefer the one above 😅

[dense tapir]

same

[smoky oak]

to each their own of course

yeah, so its really just a hunting game really

but yeah, I wouldn't go past 30

DDIM functions very usably on 10-15 IMO

[dense tapir]

I really like that UniPC when it works :/

UniPC is basically DDIM V2

it does something weird with a good pause then starts, or most times not. try again it works

I highly suspect auto didn't implement it right as no issues with it in comfy

RuntimeError: cusolver error: CUSOLVER_STATUS_INTERNAL_ERROR, when calling cusolverDnCreate(handle)

works fine 512x512

then I can gen any rez but it must be iunitialized with 512x512 first

[smoky oak]

@void hullhere you go. now that 1 extreme res image is 3 separate still very high res wallpapers. Maybe now you understand, or not, IDK lol

[void hull]

No i get that part

[smoky oak]

alright, then I have no idea what you were getting at

[smoky oak]

nice!

I like that one

[dense tapir]

Yeah, but it turned real. I saved it so I can remember how I did it.

This is as real as I am after

for now

a mix so to speak

[smoky oak]

ah, alrighty

You already know I am a realism buff lol

[dense tapir]

I have a need for the other next though

Oh, so am I

this is for some work I am doing then realism I will need that

This still gives me the creeps

Well, damn it did it again

[smoky oak]

Oh I really like this one!

I love her facial structure. Very eastern Asian

[dense tapir]

Aye

[fringe mason]

Protrait of astronaut

[dense tapir]

I add more steps and it changes the gender :/

[smoky oak]

lol

[fringe mason]

generate-with-images of an astronut

[dense tapir]

You know I liked 1.5 mostly obeying me while 2.x lost so many functional commands I used to use

[smoky oak]

I still stick behind preferring 1.5 :p

and resolution/detail is not a problem lol

[dense tapir]

I bet you a doughnut SDXL/3.0 will be just 2.1 rebadged

Already the hype train started and that is alright but when the hype gets to a point it becomes overselling which is normally overselling shit.

I am sure it will have its niche though

[smoky oak]

I have 0 excitement/expectations/hype for SDXL/3.0

[dense tapir]

SEE, I told it wide shot and it did it while 2.x tells me to go eff myself

[smoky oak]

I see you are starting to appreciate 1.5 a little more again haha

There is a reason I tried 2.x and said nope, and went back to my comfy, and well behaves 1.5 lol

[dense tapir]

I could easily fall back to 1.5 if it had the resolution, and for me, honestly, it is the only thing holding me to 2.1

I mean native res not trickery

[smoky oak]

Have you not seen the outputs I get out of 1.5?

even 1.5 with high res fix makes great results

[dense tapir]

Of course but I want native rez far more than upscaling

[tawdry geode]

Have you tried dreamboothing on 2.1? would it be better because of the resolution?

[smoky oak]

I don't care at all about native res. I care abotu output lol

[smoky oak]

its hell cause they gave it a trash TE

[dense tapir]

YES

TE is a SOB on 2.x

[smoky oak]

this output is from high res fix. I am not sure why the native res would matter when you can get such excellent results with barely any more work

[dense tapir]

emad expects the community to train 2.x and use 2.x as a base to build upon YET they gave us shit for the TE/CLIP

because take a single strand of hair. 480p can't see it (say 512 for SD) while 1024 native can. We lose so much and why they are pushing towards 1024x1024 which can then be upscaled to anything, or downscaled while showing that hair on a gnat.

[smoky oak]

But...

what

[dense tapir]

it goes deeper than what you are doing

[smoky oak]

You mean like single strands of hair like these?

[dense tapir]

already fuzzy

[smoky oak]

each individual hair?

[smoky oak]

I am sorry, but I find this a little funny, coming from somebody that barely gens above 512-768x

[dense tapir]

I guess you are too young to remember 480i to 1080p then to 4k

[smoky oak]

I remember them, no need to patronize me

[dense tapir]

Then you should know.

[smoky oak]

comparing these technologies is useless tho

you're acting like 512x SD can only resolve detail to 512x, which is not the case

[dense tapir]

no, of course not but it does have its limits

[smoky oak]

I am sure I could easily get it to do each and every hair now

That gen was from my first week in SD

[dense tapir]

That's cool. I am just anti upscaling unless you have no other choice.

[smoky oak]

and I don't get why. Who cares. upscaled SD1.5 blows native SD2.x out of the water anyways lol

[dense tapir]

if you say so

[smoky oak]

both get way better when upscaled

[dense tapir]

Now these two are funny

[ripe cedar]

I can rig characters in blender. Is posing those models and then using a depth map export to use in SD better than trained model with ControlNet? Or is there a way to combine the two with multicontrolnet or something?

[dense tapir]

I don't use controlnet so no idea

[ripe cedar]

I see. Is it possible to make consistent characters with depth? No ControlNet involvement

[dense tapir]

Without making your own model characters (subjects) probably will change and will if the seed does most times.

[ripe cedar]

Well I mean the characters are my own but in 3d mesh. I guess I'll have to try.

Just got my RAM in the mail so I'm going to try installing everything tomorrow

[static tusk]

depthmap will only allow you to retain the shape of the person, but the details will change the same as if you didnt use CN

if the initial image is clear you can combine depth with canny or scribble for a little more control

[ripe cedar]

I was hoping there would be some way to combine the shape of the depth with the texture/color

[static tusk]

texture /colour will derive from your prompt or a low denoise will transfer more of the original image

[ripe cedar]

Yeah. I was thinking about characters in Blender could be somehow used consistently without having to roll the dice on getting colors right with prompts and denoising and stuff. Ie the color accuracy of a trained model but also the depth for the shape consistency and control. The faces have 3d eyebrows and everything so fine control over expressions are possible instead of playing the prompt lottery

[static tusk]

you might be able to transfer 3d eyebrow detail with scribble or canny with depthmap, as mentioned. its not guaranteed but give it a shot

[static tusk]

also if you can render out outline passes from blender and then use that as your scribble, you will be golden. I just remembered Ive rendered some pics like that, but not using Blender though

[smoky oak]

You can export the depth map and then in control net, you can set the pre-processor to none, and the post-processor to the depth model

[sonic vessel]

https://images.seek.art/98677fd73bdf0529622e01f1bb7691b3.png?auto=webp&optimize=high

https://images.seek.art/0fc793f7179bb71b277cb509fd9a635c.png?auto=webp&optimize=high

[fiery lichen]

https://www.midjourney.com/showcase/recent/

look at a recent post with the prompt "sign that says F C U K" good job v5. Granted it really has improved on english alphabet letters, but people are already exploiting it for fun

[dense tapir]

Does anyone know how to randomize a word in a prompt?

I know wildcards but that never worked for me

never seemed to work for me

[smoky oak]

oh my, these are more baked than snoopdog haha

[dense tapir]

Yeah, I downloaded an extension and it screwed me even though it was off

[smoky oak]

oh jeez

not fun

[dense tapir]

I was looking for a randomizer

[smoky oak]

randomizer as in?

[dense tapir]

subject and it randomly changes the subject but all else stays, relatively, the same.

[smoky oak]

ooo, that is kinda fun

[dense tapir]

Yep

ddim is really eating me

8 steps and 10 steps

[smoky oak]

oof

I am glad I was able to turn you onto DDIM, even just a bit :p

I use it religiously with the models I focus

[dense tapir]

15 but something is way wrong here

OUCH, something is way wrong

euler_a 15 steps

Oh, wow my loras are ripped

just jacked

without the lora

it was that extension it must have hijacked something

this one https://github.com/innightwolfsleep/stable-diffusion-webui-randomize

working now I had to disable and enable the lora extension

AFTER I removed that randomize

It tore up my venv so I am letting it rebuild it.

These are so fragile

[smoky oak]

damn-

very sensitive, jeez

[dense tapir]

There we go

now to test

I rebooted and all kinds of venv errors right after I installed that extension using the auto extension list

now it doesn't work at all

oh, I think I know

yep, I had to refesh the browser as I saw that error and knew

yeah, oh well

ddim 10

removing the lora

ddim 10

[oak hornet]

How to start

[dense tapir]

it must have not only jacked up my venv it did a number on that lora, or this extension

[void hull]

The sword is bending into the building in the background

[valid badger]

c

[dense tapir]

yep

[void hull]

That one doesnt look ai

[smoky oak]

Stable diffusion themselves decided that we can't generate any form of weapon or anything, so now we get issues like that

Shitty, IMO

[void hull]

Where are u even generating from

I make weapons all the time perfectly fine

[smoky oak]

I don't think that image up there is AI

[oak hornet]

Bzj

[dense tapir]

If it is AI I love that style

[smoky oak]

examples? Cause SD themselves pruned their models to get rid of any and all weapons

[void hull]

Im using a blend of 3 cpks or whatever the extention name is i cant remember the names of

I can get them when i get home

[dense tapir]

Yep, that is the way

[void hull]

Anythingv3 is one of them
And cyberpop or something

[smoky oak]

Would be interesting to see, cause IIRC weapons are against TOS

[dense tapir]

the problem is you have to prompt with them not just generally prompt

[void hull]

I like the cyber one despite it trying to make everything a cyborg
Cause it adds nice line density

[smoky oak]

you can make literal CP, but if you want your knight to hold a sword, you're fucked. They went so far out of their way to get rid of weapons, its insane

[dense tapir]

1.x always had an issue with curved weapons

scythe just forget it

[smoky oak]

its cause they removed them lmao

[dense tapir]

yep

[smoky oak]

try to generate a sword in SD

it won't happen lol

[void hull]

I havent tried to just straight up make a weapon, will give it a try when i get home

[smoky oak]

not unless you get super lucky

lumberjack holding an axe? nope

[void hull]

Been able to make my characters hold weapons fine tho they arent super detailed

[smoky oak]

police officer with a gun on his hip? nope

its so severe that if you generate fighter jets, they will generate perfect, but with no weapons

Daily reminder that samplers MASSIVELY affect outputs

[static tusk]

Its not that common to have a char hold weapons, but it is possible

[smoky oak]

yeah, sure, but they look horrible

cause they took it upon themselves to remove them

look how massive the difference is between samplers

identical everything, but different samplers

[static tusk]

I dislike euler a as its adaptive so it changes wildly depending on steps

[smoky oak]

yup, I was just gonna show that

[static tusk]

see how different each one is. not great

[smoky oak]

identical settings, different steps

DDIM is my go to

its faster than Euler A, needs less steps, and is way more consistent

[dense tapir]

I never found a use for heun and really never cared for it

[smoky oak]

yeah, same

[static tusk]

LDSR uses 100 steps of DDIM

[smoky oak]

its just worse DDIM

[smoky oak]

LDSR?

that sounds familiar

why

[static tusk]

its an upscaler sampler

[smoky oak]

ahhh

this is why I stick to DDIM, cause its so consistent, and doesn't need many samples to look good

consistent, fast, reliable, DDIM my beloved

[dense tapir]

I actually like ancestral samplers but not always

[smoky oak]

for my workflow, adaptive samplers are very problematic

[dense tapir]

not adaptive ancestral

adaptive is a turd

[smoky oak]

cause I do low sample big batches to find a seed I like, then I refine from there

[smoky oak]

right, right, my bad

Euler A is both adaptive and ancerstral

[dense tapir]

euler_a 20 is the one I used to favour

[smoky oak]

I am curious, I wanna try something

[smoky oak]

@dense tapirvery big new revelation about DDIM. It doesn't matter much for you tho

[dense tapir]

????

[smoky oak]

so, the contrast is actually dependent on steps

this is 20 steps with 5 steps high res fix

and here is 20 steps with 50 high res fix

you can see the contrast and colors are considerably different

[dense tapir]

Sharper

that is contrast at work

[smoky oak]

here is 50/5

[dense tapir]

UGH

[smoky oak]

50/50

so more steps does make more contrast for DDIM, thats interesting... hm

[dense tapir]

No wonder 100 steps for it

Man, I almost have enough saved for a 4070 but I am holding out for a 4080 and I hope they lower its price

4090 is too much for me tbh

well, my system

[smoky oak]

why not go for a used 3090? More VRAM, isn't that what you are after?

[dense tapir]

speed

4080 tears a 3090 up

[smoky oak]

so in this case you would rather have much less VRAM for more speed

[dense tapir]

Like the vram but tired of waiting around for it to be done

[smoky oak]

aren't you on a 1060?

[dense tapir]

yeah

[smoky oak]

anything modern will be a massive speed upf ro you

[dense tapir]

did you see the chart for inference between a 4080 and a 3090?

[smoky oak]

I mean I already get close to 11it/s sometimes on my 3060ti

[smoky oak]

no

[dense tapir]

4080 is 30-40it/s if on a tired system about 20-25

[smoky oak]

I thought the 4090 barely hits 40it/s

[dense tapir]

potential it hasn't been done yet I am talking at the hardware level if the 4090.

remember 4090 HW vs 3090 HW it is 56% faster

4080 is around 25-30% which is not small

I wish the 4080 had 24gb

[smoky oak]

I am just surprised you feel that way is all. You really changed your tune lol

[dense tapir]

no, I am settling

first I don't buy used.

[smoky oak]

That is sure as hell not helping your budget issues heh

[dense tapir]

4090 is just too stupidly expensive then I need anouther grand to get a new pc

Just don't buy used. I swore used off decades ago and never looked back

especially after miners

[smoky oak]

I mean, as long as you accept that you are paying way more for similar things, then I guess thats your call to make lol

[dense tapir]

yep, new vs old

why are all my images so brown :/

[smoky oak]

my previous GPU was used, and it was perfect

0 reason to have bought new over used

[dense tapir]

I need less brown. let me neg that

Well, added brown to neg prompt

I'll take it

you do 100 steps vs 50?

50 steps of ddim with 1 lora takes 2 mins to gen

768sq

[smoky oak]

bleghhh

any GPU will give you a massive improvement lol

[wispy nest]

I can feel "intricate details" in the prompt 😄

it often does this brown\yellow-ish details for some reason

[glossy herald]

Intricate in particular has some real details that tend to show up

[dense tapir]

there is another term that is even better but I forgot it 😦

used it in a few of my prompts but no way would I be able to find it

Trigger discipline right there.

WOW

[wispy nest]

I got rid of it in most of my prompts after some time cause of blown-yelllow-ish tendencies, depends on what you want , I guess

[dense tapir]

Intricate is in the above

[wispy nest]

just "intricate" ?

[dense tapir]

intricate, intricate sharp details

[wispy nest]

hm , aight

I wish there was better word with no "detail" in it for clothing

[dense tapir]

yeah

[wispy nest]

we can do "ornaments" for little things , but clothing idk

[dense tapir]

there is but damn I forgot the word

started with a G I think but 2.x it doesn't work with so I long ago abandoned it

[wispy nest]

chatgpt suggests "ornamentation" , even for clothing patterns,, hmmmm

embroidery, beading....hm

[dense tapir]

maybe I should ask chatgpt

[wispy nest]

"embellishments on a garment"

I didn't even know this words exist lol

[dense tapir]

naw, this word made shit pop on armour and stuff

I forgot how to get into chatgpt without busy

[wispy nest]

I'm thinking mostly about cloth , not an actual armor

something wierid going on with my a1111 lately, model I was using started giving me weird results, they doesn't look as realistic anymore...did something changed in a1111 lately?

if not I'm probably prompting something model isn't trained for 😦

[quartz olive]

Can anyone identify what model they use?

[wispy nest]

No idea, could be any anime model at this point

[quartz olive]

Looks like it's mixed with counterfeit model maybe, but the face is 3d-ish

[wispy nest]

I don't do anime to tell the difference

[dense tapir]

ChatGPT was of no use

[static tusk]

Didnt notice that multi diffusion upscaler ext got a large update. It can do actual tiled diffusion now as well. The man, woman, background, each have their own prompt

[dense tapir]

I am downloading my old images so I hope one of them has it

512m of images

A lot of my real old ones went POOF

[wispy nest]

Just as test of prompt, with " clothing pattern ornamentation, ornaments, embroidery, beading"

but I don't like word "pattern" , it might be used for something else too

kinda works, I guess, but overweight, applies to lots of shit around

[void hull]

anyone know what this neck thing is called
i want to put it in the negative prompt

[dense tapir]

I can't find it but what I would give for a program that would scan all the prompt info in the files searching for me

choker

try that

[void hull]

chokers already in negative, i think choker is a much thinner thing isnt it

[wispy nest]

yea...

[dense tapir]

yeah, it is

Shit

DDIM above 30 steps this is 60 steps

[void hull]

tried to modify it in a way with adding keywords and generating the same image using a seed but
the disdainful expression goes away no matter what I try to change, unlucky

[dry crow]

You also need a vae for that model

[void hull]

whats that?

[dry crow]

A file needed mostly for anime models for Color and Detail correction

For example

[void hull]

i'm getting the style from several ckpts i blended

[wispy nest]

interesting comparison

There was another vae for anime tho...K something...

[void hull]

sometimes it makes animish stuff sometimes it doesnt
i just liked the pose and expression for the character

[dry crow]

Yea but if you mixed in an model That needs a vae, your mixed model requires the vae too

[dry crow]

Kl8-anime2 or something is from WaifuDiffusion 1.4

[wispy nest]

idk where it comes from, but most anime moels use it, yea

[dry crow]

Yea its very good, i have a comparison with it too but not on mobile rn

[void hull]

so is a vae something you run through the image img2img afterwards

or is it something i'm supposed to add into the models

[wispy nest]

yea, this one

[wispy nest]

no, it does whatever needs to be done during generation, you don't need to do anything

[dry crow]

It goes together with the model in models/stable-diffusion and has to be renamed to match the model name

[wispy nest]

I think renaming needed only with this option on
and if you don't want to change it manually after changing model

That's interesting result o_O

[austere pilot]

try to paint it out with the brush and then use img2img

[dry crow]

Mirror portal

[dense tapir]

lol

[gleaming flame]

https://media.discordapp.net/stickers/907134046419689502.png?size=160

https://cdn.discordapp.com/emojis/1048695050437132398.webp?size=64

[glossy herald]

nah

sorry but that is against rule 4