💬｜general-chat

[upper plinth]

fuck it ima try out swarm

[atomic mortar]

allows for X and Y tiling (and both)

[upper plinth]

I found one for 1.5 but I suspect flux will give me much better results

[atomic mortar]

SDXL even will

[lucid bobcat]

Isn't that just for tiled encoding/decoding?

[atomic mortar]

nope chcek #🏞｜general-with-images

[gritty lava]

dunt listen to this acc ^ , it appears to be hijacked by a bot in another server, refrain from clicking links or joining conversations with that account anywhere but here, and use extreme caution

[leaden hull]

hello guys

[hot turret]

What's up.

[shy crane]

z

[tulip yarrow]

i'm back

when was illustrious introduced and why is it popular (just curious)

[pine path]

I wasn't really paying attention to when, but it's a good anime model. Most others suffer from small dataset, poor training, or are just trained via lora. There weren't really a ton of good options. It's not like illustrious does anything super fancy iirc, it just avoids most problems that others have made and actually had enough compute to work with.

[atomic mortar]

why would anyone use your AI service while everyone here is running it locally

[silk condor]

Is there any way to run Kling 1.6 on your own local PC or something similar to get nice ai videos? 🙂

I want to avoid paying 😄

[atomic mortar]

what graphics card do you have

[silk condor]

2080ti and a 3060

[tulip yarrow]

thanks for telling me

is it good as noobAI/pony

what other widely used SDXL models are there

[quartz swan]

Personally I feel like illustrious is better

I never struggle with hands or any other issues pony had, haven't tried noobai yet tho

[little flower]

hi

[verbal island]

[wicked root]

does flux model have problems with general prompt (fp8 models), like instead of giving lot of details, you say something basic without much detail

[atomic mortar]

Then, sadly you probably cant. Might be able to run LTX but that's low quality

[wicked root]

i can't get the result wanted if i'm not specific, like if i want to let the ai decide certain detail by itself, it won't follow the general idea i wrote in the prompt, it'll have problems with following the prompt

[wicked root]

if that's the case then i won't use flux for general idea

[lavish loom]

https://tenor.com/view/worrygirl-williamz-尷尬-哭笑不得-現場直憋-gif-13866800719477153687

[gritty lava]

xd, i think either my discord or the server thinks im a moderator

oh i cant attach images here

[opaque umbra]

#architecture

[silk condor]

Really, what can I use to make decent videos with ai?

[atomic mortar]

You could try hunyuan but its not gonna be fast on that 3060

[slender vault]

with wavespeed node it can be quick, minimal quality loss

[bleak matrix]

Good morning, everyone! How are we all today?

[west olive]

hi i may need help

where can i put screenshots

#🏞｜general-with-images

[lucid bobcat]

What graphics card do you have?

Also if it worked before and only broke 3 days ago you need to ask yourself what changed. Did you make an update that broke things or change settings?

[coarse zealot]

what software is the best for image to video? and what's the best one for extending a video with ai?

specifically trying to prank my homies by making them kiss each other.

[atomic mortar]

any online version will probably block it for being nsfw etc

there isnt a good image to video yet iirc but maybe LTX could do it for you

[lunar nebula]

Hey guys. I need some help for non nsfw picture I need for a friend of mine. I am running A1111 but didn't get a good result with my models (I use lightnings because I only have a GTX1070) and also I am bad at prompting to get what I want. It should be a greetings picture for a friend of mine for her birthday ... Maybe someone has some time those days to help make that picture? Also to get more into AI generation or can create that pic for me. I tried Dall-E because I need that pic till 08.02.25 but it should be a cute smurfette and Dall-E has problems with the copyright ...

[signal nest]

can flux do skyboxes .. haven't looked into this sort of thing for a long time, i saw some service by blockade labs that seems to do that

what method is prefered for that sort of thing.. i've see peopel do cylindrical maps i think

[lucid bobcat]

Does someone have experience with supir? The example images look very good, but when I tried it in comfyui the results were bad. I used juggernaut as the model and the input image is photorealistic, although it looks a bit cg because it was made with flux. I prompt for photorealism as well with everything artistic in the negative prompt. The control scale is set to 1, yet the the best result I got so far look painterly and far worse than what I get with hires fix and related methods. Am I doing something wrong or were the examples cherry picked?

[fervent thunder]

super-resolution is just really hard

https://github.com/yuvraj108c/ComfyUI_InvSRthis is a good one to try

[outer pecan]

Does anybody here know if there’s something up with the SD Deforum Google notebook recently?

[low moon]

Since the Emad emojis are no longer available this means a complete and total failure on everyone's part. We apologise on behalf of all AI.

[main snow]

managed to make a pic with 2 char loras, that shit is ROUGH

[fervent thunder]

On behalf of the Homies Association of the World, I implore you to not perform such a prank.
Sincerely,
Head of the HAW

[fervent thunder]

Why are the stop and interrupt buttons so worthless
It's faster to just close the program, wait for it to load up and then set everything up again

How is that even possible?

Is this just a me issue or is stable diffusion just like this?

[little flower]

[fervent thunder]

kittyyyy

[tribal needle]

is Stable Diffusion 3.5 Medium uncensored?

[drowsy hearth]

Its crazy how the more AI tools I get the more gooder I get at bugfixing dogshit environments

[atomic mortar]

Probably not since its released by stability it self. Maybe a fine tuned model might work better

[tribal needle]

Sdxl is uncensored

And online iy says that the model is, including yt vids

[atomic mortar]

... You asked about 3.5

Base xl is also censored

[tribal needle]

Ive tried it, its not

[atomic mortar]

But if finetuned/trained etc its no longer censored

[tribal needle]

And by saying sdxl, i meant that that was also released by stability ai

[atomic mortar]

Not when i tried prompting certain stuff into it but sure 👍

[jagged solstice]

hey

why cant i get stable diffusion to do what i want lol

https://i.imgur.com/MdQ4ZF1.png for example i want to extend the hills in the middle down to fill the bottom

but i just keep getting strange results

[lean timber]

Hey, quick question, is there a way to search in the LORA Loader dropdowns in ComfyUI? I can only see a filter input and that is useless when you have folders and it only filters the top level

[lucid bobcat]

Bad code. All software these days has this problem https://www.youtube.com/watch?v=kZRE7HIO3vk

[slender vault]

lora stacker node

[lean timber]

@slender vault by efficiency nodes? It's the same thing with that one for me. Can only filter the top level

[slender vault]

https://civitai.com/articles/1997/comfyui-guide-to-stacker-nodes

[lean timber]

Thanks, although I tried CR and rgthree stackers too, but they also have the same kind of non-searchable dropdowns

https://i.imgur.com/LmK3YAc.png

That's how all their dropdowns look for me

I remember having one lora loader before which showed thumbnail previews at least, which is better than nothing, but having search would be best

Because sometimes I load someones workflow and they don't have their loras and checkpoints sorted the same way as me and I have to dig through all my folders to find the lora

[fervent thunder]

who are you cat fishing

xD

[severe ruin]

So, question... does anyone know why SD seems to get into a certain... "rut" where it'll degrade to making shitty images with like double bodies where the same prompt used to make good images.. and then just keep doing the same type of double body even when changing prompt or model (model changes the stylle to that model ofc)?

eg.. SD just decided that a core part of every image is a face in the abdomen

regardless of the values i change.. I get different styles of just... that... probably till i restart the backend

[fervent thunder]

monkey patching

is what it sounds like

[severe ruin]

uh... ima need to google that

ok.. not sure how that applies exactly.. is there a way to force the software out of it.. like clear caches or such wihtout restarting it?

[fervent thunder]

yeah the way to fix it is to find the source of the monkey patch

[severe ruin]

such as? like I've tried toggling hi res off, adetailer, changing prompt words, changing models

[fervent thunder]

I don't know which software you are using

it applies to all of them anyway

[severe ruin]

Oh.. Forgeui

sorry bout that

[fervent thunder]

I can't remember which things I found monkey patches in but there were several comfy nodes that did
its quite common

it will be the same for extensions for A1111 or forge

[loud aspen]

Anyone know how to edit one person at a time on a1111?

Like in pictures with multiple people in it?

[fervent thunder]

yeah this is called inpainting
can also do it using unsampling

[loud aspen]

Inpainting?

[fervent thunder]

yeah what inpainting does is it puts a mask over some areas of the latent

[loud aspen]

I meant like

Using prompts

[fervent thunder]

and then only certain pixels change

ah I see
its tricky with prompts

[loud aspen]

Like one person has blonde hair and the other has brown for example

[fervent thunder]

I don't do prompt engineering, maybe someone else will know

[loud aspen]

Ah okay

[fervent thunder]

the way I make big images is to use lots of tiles like a chessboard

so the prompts can be simple cos there are multiple tiles

[severe ruin]

is there any way to see a loras prompts in webui? I'd love if i didn't have to go back to the civitai page for it to check each time

[tribal needle]

so what is the best overall uncensored model? something close to 3.5 image quality

[loud aspen]

I’d like to second this

[lucid bobcat]

In forge you can see the lora keywords and it will be automatically added when selecting a lora.

[severe ruin]

yeah? where?

[lucid bobcat]

Haven't used Forge for a while. Maybe it was an extension.

[severe ruin]

found Civitai helper, seems to index that stuff.. trying it out

[lucid bobcat]

It was a lora tab. It would list all your loras with images and names and you could just click on an image and it would add it to your prompt.

[severe ruin]

yeah thats there as default.. but it just adds the <lora> for me.. not the keywords.. and if there are more I couldnt see them.. so far the helper pulls descriptions which i can adjust a bit so i can remember better

[lucid bobcat]

You would have to add the keywords. I think you could fetch them directly from civitai as well.

[languid haven]

Both forge and std webui now display the tokens inside the lora's setting screen. It's that wrench icon on the top right of the lora's gallery image

Some loras do not have metadata tho -- seems like a training setting issue. In which case you gotta go to civitai to see if the uploader included sample prompts

[serene forge]

hi, is there a "go to" guide for running stable diffusion locally?

i checked the pins and it mentions #unknown channel for getting started

[potent oracle]

Hey all is possible to make headless ai or not ? I Mean nbm

[rough coyote]

Hey Anyone have experience in Genfluece ai FLux ?

[atomic mortar]

Are you asking of people use it or do you have a specific question?

[ancient sluice]

hello

[solemn yoke]

whats up @ancient sluice

[ancient sluice]

just trying to generate cool images

[red tulip]

[drowsy hearth]

I wish I didnt understand more blackbox solutions then literally everyone on earth but I do

Im going to draw a video game its actually going to happen thats crazy. But first Im going to draw a checkpoint model.

[blazing cobalt]

uhhh the ping is to say that you're working on a poll? why not announce it and ping everyone when the poll is out? and why are you polling users if the investors are the ones that actually decide?

[prisma rivet]

[arctic coral]

So we have been summoned aye

[young bronze]

who is maxfield and why is everyone sending skulls at him

[jade mesa]

maxfield based discord moderator

[wise depot]

mod behaviour

[sweet token]

'Ello

[vagrant blade]

Because that's the community managers job, to keep the crowds invested.

[vapid dove]

rant away

[unborn nexus]

I love generating art with DALLE and stable diffusion. Would love to do it professionally if possible. look forward to seeing what oppotunities follow from this announcement 🙂

[wise depot]

emphasis on this

[jade mesa]

i forgot about stable diffusion when flux was released

we just need a new release

[desert dagger]

welcome aboard @vapid dove :)

[arctic coral]

Meme room mentioned, opinion accepted

[jade mesa]

then we get excited again

[blazing cobalt]

I actually think the ping will make more people leave the server than engage them

[heady glade]

Wow, I didn't realize Stability AI had fallen so far X.x

[young bronze]

I, like everyone else here, am spontaneously generated into existence to commit social engagement

[lone lantern]

@vapid dove Did AI write that announcement hmm

[arctic coral]

That's what I came here for to see

[desert dagger]

good. maybe the trolls will leave

[wind kelp]

fear is struck into my heart every time a company decides they need a Community Manager

[wise depot]

possibly

[young bronze]

Did AI post that announcement?

[vapid dove]

I am at least 55% chatgpt

[blazing cobalt]

and yeah, the last time I used a stability product was the image to 3D model and it turned out it was cherrypicked and not even SOTA, and then flux really killed it for good

[young bronze]

I WAS RIGHT

[arctic coral]

Only the real ones are gonna stay 😎

[blazing cobalt]

basically stability needs to actually deliver a product that people want to use, maybe remove the censorship which is the top complaint

[sweet token]

3.5 is decent, if it was actually easy to train

[wind kelp]

I still use SD for sdxl tbh but it's more of a community thing keeping it up rather than stability

[heady glade]

I don't think you know CivitAI's history, associating with them is very bad >.<

[young bronze]

train it on trains

[manic wraith]

I feel kinda bad for whoever is tasked with fixing Stability's public image, since I assume they are somewhat powerless to solve any problems

[unborn nexus]

tbh my greatest issue was working with hands on most of these models. and finding decent documentation to finetune models for stable diffusion 3.5 has been difficult as well

[sweet token]

Sai often is either hit or miss with their models (Cough SD 2 cough cough)

[arctic coral]

I'll admit, I do not

[wind kelp]

the community manager that got added for the character.ai server caused that place to turn into a TikTok user tween hellhole

[frank river]

what are you even talking about? civitai is one of the main contributors to stability's adoption rate imho

[vapid dove]

I'm not aiming to change the discord, just make sure we're showing cool stuff, and getting feedback prior to you know, release

[wind kelp]

oh hi direct response

[heady glade]

They're well known for attacking artists, harming creatives, and (ironically) stealing models from other sources and people as they re-upload for profit.

[errant basalt]

that and weird floaty bits when making artistic images

[heady glade]

It's a big red flag

[wind kelp]

yeah I wish u luck king

[charred mesa]

give us SD4.0 with zero censorship, better captioning and a mostly non-synthetic/not overly aesthetically filtered dataset (impossible)

[unborn nexus]

are you going to make a section where people can give feedback on their current grievences?

[pine path]

My biggest want is a competent vlm for captioning data.

[manic wraith]

if I could go back to a world where it doesn't exist, maybe I would, as I feel bad for artists. but it exists, so at least it should be open, openly-licensed (apache, MIT, WTFPL, I dunno). the worst model is OpenAI where they steal from everyone and claim it as their sole property

[young bronze]

I just locally run stuff for fun, but I'm also masochistic since I run a red card

[charred mesa]

impossible to ask for

ever since SD2

[frank river]

@heady glade are you saying we should not be associated with you, since you have freaken 36k messages on the civitai discord alone?

[sweet token]

God. SD2 is like a memory i want to suppress

[arctic coral]

None of this except maybe the last one makes sense to me lol. Why attack the guy trying to make something cool for others to use as well? Well trolls exist and there probably is no actual sense to be found

[drowsy hearth]

Youre not getting a mostly non-synthetic dataset in 2025 over half the internet is synthetic bro

[vapid dove]

time to go slow mode

[heady glade]

I was someone that tried to push against the hate they represented over time.

[charred mesa]

real, but my request was sarcastic

[wind kelp]

there's like six people here not even talking that fast

[heady glade]

I was one o fhte artists they regularly attacked and did so for months after I left.

[arctic coral]

Good night anyhow, I was going to sleep when the ping hit

[steep remnant]

zero censorship cannot happen after the scandal with sd2

[charred mesa]

its stability, they'll decide for us what's legally viable for them

well... that was kind of my poi-

[sweet token]

SD2 was bashed because it had worse anatomy and was far more censored then 1.5, and had a horrible license

[drowsy hearth]

Would you guys fuck with a tool that lets you explore latent space in a 3d gui and train the model with paintings and storyboards so you can punch your own ideas in?

[manic wraith]

Illustrious/NoobAI has proven you can extract even more out of SDXL than previously thought possible from say PonyXL, so my only request @vapid dove is please we just need the future base model to be easy to finetune, easy to make LoRAs for, and not awfully licensed with ambiguous fine print like the SD3 situation

[young bronze]

maxfield what have you done, you ping with no subject for people to latch on to so now they have to make conversation

[drowsy hearth]

well not train but literally replace parameters in real time and test them in real time too

[steep remnant]

stable diffusion died a long time ago to be honest, it was a fun ride, but multiple bad decisions were taken and they also had a huge target on their back, being one of the first company to share a diffusion model. it's a shame really, but they will never recover, even with Cameron investing in stability

[drowsy hearth]

cuz thats what im working on and i got the parameter replacement working last night

[charred mesa]

I don't think I was that active here at all (especially general) since SD3 actually came out, I came here to laugh about it and that's all

[lone lantern]

society needs censorship.. lots of weird people around nowadays..

[foggy canopy]

As long as they keep lobotomizing their models there's not really a point in wasting resources, no one is really going to use it since FLUX is a way better alternative

[vapid dove]

At Civitai we actually banned the original SD3 model because of the license, and unbanned it once fixed. the current license is very permissive

[frank river]

the current SD license I'd say is even better than flux

[drowsy hearth]

All the models are dogshit bro in theory they should be completely capable at doing everything.

[vapid dove]

I'm an agent of chaos

[heady glade]

Pretty sure CivitAI just banned it because it would affect profits, not because of the license specifically lol

[sweet token]

SDXL was SAI's last great model then they sadly fell off a cliff, with SD 3 medium being the prime example at how far S.AI had fallen

[heady glade]

No need to lie about it >.<

[foggy canopy]

Its capable of doing >Everything except the stuff they don't like

[sweet token]

No the license was Extremely bad

[unborn nexus]

what bad decisions were made?

[errant basalt]

@vapid dove Not sure how much of advice I can give you tbh, but looking at how good flux is I´d say focus on making artsy images and improving CLIP / Prompt coherence, rather than making realistic images unless you make them considerably better than flux

[drowsy hearth]

no its not it repeats the same 3 compositions between every prompt basically and you have to fish for like 30 minutes to an hour just to find a prompt that doesnt produce dogshit. Literally every model is like this.

[charred mesa]

im just waiting for other models, flux is good, but it's too large, has the typical overly perfect (incestuous model merge type of) faces and sucks for paintings unless you use loras

[manic wraith]

then please tell them to get it right the first time so it doesn't suffer from permanent bad publicity 🙏 some lawyer writing that license killed SD3 before it could shine

[drowsy hearth]

Idk I havent used stability in forever I use a merge of ponyxl

[sweet token]

SD 3, SD2, everything that led up to these models are bad decisions

[drowsy hearth]

which is still kinda stable diffusion sorta but much more elite

[foggy canopy]

Haven't gotten that issue from flux so far

[charred mesa]

they ruined it with SD3.5L, the sauce and soul of SD3 was removed (artstyles, etc)

[vapid dove]

don't think thats a piece of history that will repeat itself

[charred mesa]

it's still more creative than Flux

but it still got nerfed

[manic wraith]

good: SD 1.3-5
bad: SD2
good: SDXL
bad: SD3
we're due for a good one

[charred mesa]

hello maker of the pony models

[drowsy hearth]

Wait do some of you guys use just the base model?

[steep remnant]

the release of SD3 abislutely killed them, then proceeded to gaslight people into saying it's the best model out there, then ehmad left, and then stopped to publish anything after the SD3 shitshow so they fell deeper into the forgotten companies, at the same time concurrent models were emerging, and this was over for stability, not even mentionnign the gigantuic amont of lawsuits they received. It kinda was coming from the start really, they piked and the downfall was quite fast and brutal

[charred mesa]

which base models

[nimble canyon]

https://tenor.com/view/creep-hands-adventure-time-deer-remove-the-gloves-gif-15274634

[drowsy hearth]

Like do you guys use base models without checkpoint training? I didnt think anyone did that.

[foggy canopy]

I'm surprised the >Skill issue guy isn't here talking by now

[charred mesa]

Like I use pixelwave flux I guess, but even base flux is good as long as I don't make paintings. never explored SD3.5L finetunes

[exotic delta]

It's been a long time to be honest

[unborn nexus]

i normally do that to test what initial capabilities it has before going on to see how i can finetune it

[manic wraith]

hey you're the pony guy right? @nimble canyon still doing the AuraFlow experiments?

[steep remnant]

lol i know right i remember that dude, what a moron XD

the fineture pixelwave of flux is indeed top tier model

[nimble canyon]

If by experiments you mean a SOTA community driven uncensored model to be release within a month, then yes.

[steep remnant]

very close to MJ6.1

[manic wraith]

Whoa!! I was honestly starting to lose hope it'd come out before Half Life 3

[charred mesa]

is it gonna be good with paintings (impressoinist oil paintings), variety, backgrounds, etc?

I might consider using it despite ignoring PonyXL in the past

[nimble canyon]

https://cdn.discordapp.com/attachments/1315932135571329024/1336171759434727475/score_9__rating_questionable__style_cluster_667__source_anime__Blonde_human_female_girl_with_closed_eyes_and_a_content_expression__is_depicted_in_a_floating__ba_w1280_h1536_s50_seed43_gs3.5.png?ex=67a4d0ac&is=67a37f2c&hm=5c47bd115a4158a3c8e2b8c55d4170322df174e403fcbfc1580d720db24e1cc4&

[manic wraith]

I don't even care about ponies I just think it was a really good base for other models (AutisMix, BoleroMix, T-Ponynai3)

[drowsy hearth]

Good Im always worried about this ai community because 90% of everyone in it is mentally weird. Theyre really mentally weird

[vapid dove]

possible SD3.5v? 🥺

[drowsy hearth]

actually 99% is mentally weird

[foggy canopy]

lmao

[livid gale]

is there some base models that are cool with 16:9?
even noobai sucks at this ratio

[nimble canyon]

I am still on hold with Dunning-Kruger!

[charred mesa]

SD3.5v?

[fickle sun]

I would really like a model that can do very different art styles like Cubism, sketch, botanical drawing style. not focus on realitic photos or anime.

[nimble canyon]

I have not tried impressoinist but it's a muuuuuch broader model in terms of style, for example:

[charred mesa]

yikes, I still can't believe what went down in here with Lykon

[livid gale]

havent tried

[charred mesa]

no but what is it? we have large and medium, what does V stand for

[steep remnant]

did lykon left the companie ? who remains from the OG team really ?

[lucid bobcat]

Society should NOT be censored by others! There should be a filter option so people can decide for themselves what they want to see and what they don't want to see, but that's what the negative prompts are already for. I will not except anyone else deciding for me what I can or can't see!

[drowsy hearth]

Ive always wondered if the AI communities obsession with dunning-kruger is some kind of projection of their insecurities about drawing. Its weird I just see it brought up literally all the time in this community and like basically nothing about these tools has any give or difficulty to them theyre very easy to use by design.

[rustic cipher]

version maybe

[nimble canyon]

Realism: https://cdn.discordapp.com/attachments/1315932135571329024/1336503348983234590/score_9__rating_safe__style_cluster_53__source_photo__Human_female_with_flowing_red_hair_and_striking_green_eyes_is_depicted_in_a_portrait_with_a_serene_express_w1280_h1536_s50_seed55_gs3.5.png?ex=67a4b3fd&is=67a3627d&hm=b2074a676d246855c1612a1345716280e3bb97b99392f68ea1a635d17b2a31e8& (I do not want the model to be realism focused but because people really like realism finetunes, why not to give them a good base)

[vapid dove]

a lot of people are still here, just talked to Lykon yesterday

[livid gale]

i think v stands for v prediction

[steep remnant]

no cap ? how much left ? genuinly curious

[livid gale]

v pred is cool with like
lightning and shadows

[charred mesa]

ah not realism, rough oil paintings, impressionist, expressionist, etc
but thank you anyway for this

[hasty badge]

My pet-peeve whishlis: SD3.5 not-the-base-model, make it good instead, we now have the base great, but no one's finetuning/training it into the final version where the godawful twisted limbs and various artifacts have been trained out of. Show that somehere in these weights is a good version. That has flux/lumina/auraflow (esp the latter two are good) prompt following with SD3 (not 3.5, 3 looks better imho) aestetics. (And then take a page from the midjourney playbook and doubledown on style, actually tripple down, the more modern the imagegen the worse/more bland the variation of style, MJ excluded. If mentioning artists is a no-go, go with style reference images/stylistic vectors as additional inputs, it's an mmdit, multi inputs..) ((and when there finally is a strong model it's time for the niceties like image edditting by prompt "input image of cow -> the cow wears a hat", etc what seems to be the latest thing in image AIs))

[heady minnow]

Also good for loras

[fallen axle]

Max is a wizard

[nimble canyon]

No, I get it, was just illustrating that V7 can do a lot of new tricks.

[livid gale]

tbh i have no idea how loras work and im afraid that loras will ruin all this cool lightning and shadows stuff
thats why i havent used single lora since i moved to noobai

[vapid dove]

its my third day but I've met maybe 20 OG's? sure there's more

[hasty badge]

it'd also be nice to use a the tiniest fraction of "all that compute" to bring the free public discord bot back, that brought so much life in this discord and was such a great showcase of stable diffusion.

[drowsy hearth]

Why do you need to mention artists Ive never even thought of it cuz I can just setup a lora or describe the style and it works every time.

[charred mesa]

I open comfyui and put nodes together (im a professional!!!1!)

[bright sigil]

did you meet james cameron?

[nimble canyon]

Just FYI, not hard feelings with Lykon at all and all is good, but I am going to meme that event for many years to come.

[drowsy hearth]

Honestly true bro

[steep remnant]

damm really, that's good, there might be still hope fo S AI, i hope you will recover from this, but no hard feelings if i'm not currently rooting for you to win the race

[heady minnow]

I can't speak for downloaded loras but i got my loras to work with lighting and shadow quite nice - not to the extension of normal model but like 80%, very good for 1 image loras

[heady minnow]

@livid gale sorry forgot reply tag

[charred mesa]

did you use Stable diffusion since the 1.4 days?

[errant basalt]

competition is always good, so I´d rather have SAI release a fantastic new model with 4, than just go "Fuck them cause 3 was bad"

[sweet token]

Didn't Sai had to lay off a big part of the team who made 1.4?, people who worked on the original diffusion paper?, which led the to the creation of black forest labs?

[steep remnant]

very true

[vapid dove]

we rode around in his submarine and read old calvin and hobbes comics. No I haven't met him

[manic wraith]

that is nice but I really hope it can become better than Illustrious tunes such as HassakuXL
(example with a bunch of wildcards https://files.catbox.moe/1qg3tj.jpg)

[hasty badge]

that's not what i expected to be the takeaway, it was supposed to be "make te AI amazing at style" 😉

[livid gale]

i need to search some style loras for noobai
i really like sousou no frieren "brutal" anime style
but like
yea
i want this to be on all my images not just ones that have frieren in prompt

[nimble canyon]

But tldr:

1. Super strong prompt understanding
2. Style control via prompting or special superartists
3. Muuuuch better backgrounds, generally much higher quality visuals.
4. 1.5k max resolution
5. GGUF versions going down to 2bits (minimal quality degradation on 8bit)
6. Bigger dataset including much more realism (although I will be adding even more for 7.1 as I didn't get to full 10M)

[charred mesa]

as long as something gets released... compared to the LLM community we're starving

[bright sigil]

I you meet him and go into a submarine, please don't visit the titanic!

[vapid dove]

I'm sure theres a story there, but it hasn't been told to me yet

[manic wraith]

just use HassakuXL trust me, you can throw an artist/character tag and it'll probably understand

[charred mesa]

yeah 2-4bits is gonna be necessary since Auraflow is like 12b iirc, I hope SDXL users wont have too much trouble switching over

[sweet token]

I wonder how well pony V7 going to compete with animestro 2b (upcoming anime fine tune of 3.5M by the creators of Animagine XL)

[manic wraith]

is 12b the same as an LLM would be? like, 12b LLM is very lightweight and runs on most people's GPUs

[nimble canyon]

Well, tjhe whole poiont of open models is to provide someting people can build on top accorging to their preferences. I don't know if stock Pony will be better (well, I hope, but still training) but I am sure people going to create absolutely crazy finetunes out of it.

[heady minnow]

A lot of loras are pretty baked, not just style loras, with poor tagging or just overtrained, you can always ask for people to train the model for you, also frieren is pretty baked into noobai

[charred mesa]

yeah its a transformer so its closer to it, but I am not so sure about "most people's GPUs"

[livid gale]

i mean i also can throw artist and it will perfectly copy its style but i just dont want to search the one that i will like the most since theres fucking millions of artists on danbooru
i found artucile on civitai with artists for noobai but they look more less the same

[sweet token]

No, a 12b t2i model is going to be hellish to run compared to a 12b LLM

[foggy canopy]

This is the first time i heard of Animestro, what arch will it use?

[sweet token]

SD 3.5 medium

[foggy canopy]

grim

[charred mesa]

I might remember this wrong but Auraflow was slightly faster than Flux despite being the same size

[bright sigil]

well any Illustrious based model can do it at a extend, don't need to be my model for it

[nimble canyon]

The more, the merrier. Everyone benefit from having more competing models. (well, V7 obviously going to be the best )

[livid gale]

wdym baked

[fickle sun]

music model...

[zealous sandal]

working on it...

[steep remnant]

overtrained not flexible enough

[charred mesa]

Yue is super slow even on a 3090 iirc

[heady minnow]

eh like overtrained or just no tag shuffling, very stiff in usage, not flexible and close to original images

[livid gale]

i c

[nimble canyon]

8bit firs in to 12GB if I remember right, so I really bnot concerned about VRAM at this point

[sweet token]

It's going come down to the model which can be run on a toaster with good performance

[livid gale]

damn i can just train my own style lora with screenshots from anime

[charred mesa]

honestly I am not that interested in Anime and I want good prompt adherence, more styles, so I'd be using Pony much more often

[heady minnow]

You can, noobai its quite easy, also ponyxl was quite easy too but i prefer noobai v-pred for lora

[nimble canyon]

Right, but the metric we are all chasing is "good images per time unit", not just "any image per time unit"

[zealous sandal]

also you can totally fine tune Stable Audio Open to be better at music. I've seen people do LoRAs to make it into a full 3-minute song model, I've seen people add lyrics support, the code's all there and the base model is big enough to be a full song model. It's the same architecture as our best full song models. Just means you need to put together the dataset yourself.

[livid gale]

tbh i dont see any reasons why i should use pony insdead of noob
only for some OCs that i made on pony since i just cant make same appearance or face on noobai consistently
on pony with prompt i made every image is SAME char

[charred mesa]

oh yeah v-pred

[sweet token]

You are correct, but if you managed to somehow turn a underbaked undertrained auraflow to a model worth using, I can imagine the CagliostroLab team being able to turn around 3.5M

[charred mesa]

SD2.X had v-pred

[heady minnow]

You can try block training, that helped a lot of my character loras, that or training at larger rank and resizing it down sometimes helps massively

[nimble canyon]

AF is a realllly good model, I think people really underestimate it (including Simo/FAL)

[manic wraith]

Your model kicks ass

[zealous sandal]

AF did hybrid mmDiTs before Flux, and I don't think Simo gets the proper credit for that

[livid gale]

thank you but i havent understood anything since i havent trained a single lora in my life

[nimble canyon]

Hopefull V7 release bring more attention to his work as he fully deserves it.

[fickle sun]

damn il look into it, how much vram do i need for a finetune, or makea lora? 10GB?

[charred mesa]

its all it needs

[heady minnow]

Honestly same but can give you tips if you want or just point you to the guides where people who are smarter than me know things

[sweet token]

Auraflow is great for training don't get me wrong, but for general use it falls flat compared to 3.5M and 3.5L and flux

[zealous sandal]

I still hold that Simo is the only ML person on Twitter actually worth following if you want to figure out how to beat SOTA

[drowsy hearth]

idk i retired from looking for models after my pony finetune with cross compatibility with plugins and resolution. No idea how I got it to do that but its the only sweetie i need now.

[manic wraith]

I just hope you are right, so far the images you post make me think "I hope this is a very early WIP and not representative of the final model" but I trust you because Pony was good when it came out

[zealous sandal]

I haven't done too much lora training myself I'll admit, but with pre-encoding (code and scripts coming soon to the stable-audio-tools repo), you can get it super low. Particularly if you're going for shorter-length audio. Full songs can be a bit more intensive.

[livid gale]

honestly i really need to learn how to train loras
it just feels like that lora training is for some smart asses but it definitely not lmao (i want to believe this)

[sweet token]

Simo's a great dude, worth following

[nimble canyon]

I don't know if you mean the one I posted here or in the PSAI discord, but I am pretty happy with the state of the model and it just needs to get a bit of oomph with lover LR.

[heady minnow]

Try Valstrix's Crash-Course Guide to LoRA, its quite good, and there are a few threads on lora over on FurryDiffusion that are very good but niche

[zealous sandal]

@fickle sun Here's the library people tend to use for LoRAs for SAO. I still need to put in the work to make a first-party version for the official stable-audio-tools repo, I've mostly just been stuck on how to properly fit it into the model config format. https://github.com/NeuralNotW0rk/LoRAW

Also it's in diffusers so you could probably use that with all the normal HF bells and whistles. I haven't used that implementation so I'm not much help there.

[heady minnow]

You might also try to switch how you tag your images, that sometimes causes problems

[drowsy hearth]

Does anyone actually know why cuz im not kidding and its kinda confusing is that usually what happens when you merge a bunch of different base model checkpoints together. Is it normal to just get all round cross compatibility from that?

[manic wraith]

I mean it looks a bit less impressive than current state of SDXL tunes, but it might be because you just didn't finish training

[livid gale]

thank you ill screenshot this and comeback to this screenshot when im ready

[fervent thunder]

this is what vast majority of civit ai checkpoints are

oh wait you mean base models

yeah that just doesn't work

[topaz dirge]

Hello if i want o use stable difussion deforum in 2025 , i need to paid google collab?

[drowsy hearth]

Yea thats the thing it did work with mine its absolutely crazy bro

I have no idea why

Its just the best AI model in existence now because of that

[livid gale]

which model r u talking about

[sweet token]

Last time I heard Google colab requires a paid colab account to use stable diffusion

[drowsy hearth]

sd1.5 merged with ponyxl and it has compatibility with sd1.5 loras and sdxl vae etc

[charred mesa]

well if its like Gradio then yes iirc

[manic wraith]

I just need to warn you, take a lot of coffee and a deep breath before training a LoRA or kohya_ss will give you temporary brain damage

[charred mesa]

or is it stable diffusion itself too now?

[sweet token]

This is just bait

[topaz dirge]

SO stabble derofum doesen't work anymore?

[charred mesa]

What...?

[livid gale]

dead thing cant be damaged

[fervent thunder]

you can't merge SD 1.5 and SDXL properly

[charred mesa]

sd1.5 merged with ponyxl

[desert dagger]

dont' bother - they're baiting you

[charred mesa]

oh okay I got scared for a sec

it didn't make sense

[fervent thunder]

there are some weird and unpredictable effects
which is what you observed
something does happen its just not that useful

[heady minnow]

Lmao definetly doesn't get easier after time, but there are some good presets to start

[sweet token]

I remember my first time fine tuning 1.4, lots of man made horrors, but a good time nonetheless

[manic wraith]

Somehow nothing I tried was better than the default settings with the basic learning rate / scheduler etc lol

[vapid dove]

weird and unpredictable results are my favorite part of gen ai personally,

[heady minnow]

AdamW and cosine are cracked schedulers tbh so not surprised (am assuming that's default, forgot what it is months ago)

[livid gale]

noone has answered my question
so is there some cool 16:9 models

[fervent thunder]

yeah its probably the most fun part of diffusion models
I feel a bit sad that if autoregressive image models become dominant we will lose some of that aspect

[topaz dirge]

Oh sorry, i forget that i'm in Albert Einsten chat

[fervent thunder]

with autoregressive you can't do as much fun stuff with adding different noises

[heady minnow]

I think flux is like semi good at that stuff

you can also try outpainting

[livid gale]

my pc wont even be able to run flux

[fervent thunder]

pretty much they can all do 16:9

[heady minnow]

Then try outpainting, and like roughly drawing the mask in paint or krita or any program you prefer, it works nice with most sdxl models as well partially works with sd1.5

[sweet token]

The lizard man ruling meta, says that llama 4 would be generate images as well as text, but i doubt that images produced can actually compete with traditional t2i models

[heady minnow]

I think its more of a flex that their model can do both, just to talk about it to investors and media, doubt they would care about quality that much

[neon oriole]

autoregressive models have not beat diffusion models yet but facebook is sitting on some interesting diffusion models

[exotic delta]

Me whenever things get mildly technical

[fervent thunder]

diffusion models seem ahead currently yeah

[sweet token]

I think that multimodal LLMs that can generate images, audio and text are the next evolutionary stage of LLMs, with gemini 2 (the unreleased multimodel one) and llama 4 and that model openai teased yet never released (GPT-4o multimodel)

[livid gale]

im just monke clicking generate button i dont understand like 70% of all that is being said here

[neon oriole]

auraflow was interesting when it came out but it's an outdated architecture at this point

[fervent thunder]

there's some weird hybrids of diffusion and autoregressive that might be better than one or the other

[zealous sandal]

I'm still not sold on mmDiT, didn't help much in my experiments for audio diffusion

[neon oriole]

lumina 2 is probably the best arch right now but it's far from perfect

though the hunyuan video arch improvements to flux are also pretty good

[heady minnow]

Eh i really don't think so, seeing how they freaked out about deepseek which is a good but not insane improvement over their model i doubt they got anything big coming up right now

[livid gale]

i want multilodel ai that will basicly generate visual novel for me lmao
text + images of locations = VN makers losing their jobs

[drowsy hearth]

I found out why its cross compatible but now im not gonna tell anyone cuz its hilarious that actually nobody here knows this

[fervent thunder]

probably one of these is stronger now, than both DiT and mmDiT
https://paperswithcode.com/sota/image-generation-on-imagenet-256x256

[drowsy hearth]

took a bit of googling

[fervent thunder]

there are so many new architectures its confusing

there is suddenly like 100 different ones

[sweet token]

Yeah, but I think they freaked out because their reasoning model (said to be released alongisde llama 4 with the same name) underpreformes compared to deepseek

[livid gale]

deepseek is hella smart
sad that its not quite good for "chat bots" tho

[fervent thunder]

SD 1.5 and SDXL latents have different colour space by the way, for the channels

[heady minnow]

Eh i think they freaked out cause its open source, company my friend works at pays them like 10k a month for chatbots that they use once a month, so could mean money loss for them. Reasoning model is cracked tho i agree

[sweet token]

I miss it when there was like two papers a year that were groundbreaking, now we are being bombarded with paper after paper, each better then the last

[fervent thunder]

I mean its better to have more but I get what you mean lol

[drowsy hearth]

I just rendered in an sd1.5 resolution with a sd1.5 lora using the sdxl vae bro cry harder.

[neon oriole]

most of the architectures I see are pretty similar

[fervent thunder]

most of them are yeah
but then you get weird ones like someone trained a model in wavelet space, or someone made a hybrid of autoregressive and DiT
its tricky though cos I don't want to pick the wrong one and then try to train something LOL

[neon oriole]

all of the recent arches use rope positional embeddings for example and are a stack of attention blocks

models like auraflow/sd3 are outdated specifically because they don't use rope

so multi res doesn't work on those old models unless it was specifically trained for it

and even then there are issues

[fervent thunder]

ah yeah I read in the I-max paper that the rope is why Lumina was good at high resolution and weird aspect ratios

in the Sana paper they took positional embeds out entirely and called it NoPE lol

but I am not sure Sana showed that that was a good idea

[manic wraith]

Yeah, AdamW and Cosine! There's a bunch of others like Prodigy and probably some other band name but they're somehow not a one click improvement, or I might be dumb (same for inference, I just use Euler a all the time now after trying everything)

[fervent thunder]

my current plan is to train a MoE of 8x LightningDiT models
cos its a fast one to train

[heady minnow]

Highly reccomend also trying Cosine with restarts, and setting LR Restarts to 3, its also pretty good overall

[fervent thunder]

did you use something like latent interposer, cos that method does exist

https://github.com/city96/SD-Latent-Interposer

[nimble canyon]

Possible, although I suspect we looks at different aspects of the model. I care much more about prompt understanding and model flexibility so the visuals may be less impressive, or perhaps my eval images are just not good but the only way to really figure this out is to let the community play with it (which I am wokring on)

[manic wraith]

one easy way to communicate this is with those prompts like "a yellow box, on top of a green shpere, next to a blue pyramid, with a cat to the left and a dog to the right" or something
or TEXT(!) no anime-style model that looks nice can do text very well. so like, some SFW body writing could easily tell people "oh, I see what you mean now!"

[exotic delta]

I think your bio link needs to be https, cloudflare error. Thank you for sharing your work btw! Joined your discord, will be lurking in the v7 preview channel 😂

[nimble canyon]

Do you mean https://purplesmart.ai/ ? It does work just fine for me right now.

[exotic delta]

yeah in your bio it's just http:// instead of https://

[drowsy hearth]

Its so crazy to have a cross compatible ponyxl sd1.5 model and that nobody evven believes thats possible absolutely fuckin great i been switching up the plugins all day bro thats crazy

[desert dagger]

they're not telling you that you'll get images that just look like tv screen static or black screens or something. you'll get images. what they're telling you is that the results you'll get will be unpredictable - the underlying technology in ponyXL is based on sdxl, not SD 1.5 - so they are not going to work well together. sort of like sticking a brother and sister that don't' like each other on the same project - they'll fight and what you get won't be nearly as good with them together as what you get with them each alone.

[fervent thunder]

I also found a lora from one that happened to be very good on the other
it was something like SD 1.5 hyper applied to SDXL intercomp
but it had a different effect

it is not fully random but its essentially random in terms of meaningfulness, to do merges like that

[drowsy hearth]

Thats crazy cuz its actually better then of the models by themselves with checkpoints cry harder.

i can literally prompt everything i want lil bro

[fervent thunder]

the combo I found was probably better than SDXL itercomp base also

[desert dagger]

if it's working for you, cool - that's really all that matters

[fervent thunder]

but its just unpredictable

you can keep using it yeah

[drowsy hearth]

why are all latent space tools dogshit im tryina swap parameters and shit and the tool didnt exist to do that so I had to AI generate it

[fervent thunder]

need more hackable tools yeah

[desert dagger]

not sure what it is you're trying to actually do - but are you trying to do this in comfy? auto1111? forge? something else?

[fervent thunder]

I've started making very barebones setup to test models and modify stuff
one in pytorch and one in Julia

[drowsy hearth]

naw im writing a completely new tool to go into tensor files and edit the parameters inside the latent space.

[fervent thunder]

I mean a decently large number of comfy nodes will do that

[desert dagger]

okay well - in that case, you might have stumbled upon a nice little startup idea - if there's nothing else like it out there, you can't be the only one that wants it

[fervent thunder]

if you mean for example to do like photoshop blends etc in latent space I think the Power Noise Suite comfy nodes do that, for example

[drowsy hearth]

Theres a few out there but their design is really unintuitive and bad and its possible to just swap out parameters. I also wanna see if I can have it rebuild parameters so you can see the images it was trained on. Idk if the second one is possible though

you can assemble the latents though of a parameter pretty easily and then take on image and turn it into latents then plug a new image into the parameter.

[fervent thunder]

I don't click on them but there are arxiv articles about getting training data estimates out of models

they do have some tools apparently

[drowsy hearth]

Yea theres a few tools that go into latent space but I dont like them theyre really janky and Im tryina make a streamlined one. I got it to go into the latent space and swap out latents but i been lazy all day so its still not getting all the latents for a given parameter. I also want a 3d gui so I can see the narrative bubbles. The problem with tools that do this is theyre just terrible so Im working on one that doesnt blow ass.

Theres so many things tensorflow and pytorch can do with latent space that just nobody takes advantage of its kinda crazy

[fervent thunder]

the pytorch ecosystem is massive yeah

[desert dagger]

there's an infinite amount of possible things that can be done, and a handful of people coding things - i think you've got the start of a business here. your customers are going to first be other comfy users, and then widen out to the AI image/video industry

[drowsy hearth]

The point of it being a 3d gui is so that its possible to program or "draw" video games with storyboards. The oasis project has a very similar architecture to image generators and most of how it works are emergent properties.

Thats the ultimate goal bro I been sweating its kinda crazy

[fervent thunder]

I saw somewhere they made a diffusion model that works inside the latent space of a different diffusion model wtf

[lucid bobcat]

If you train the model with the same vae then of course it will work.

[desert dagger]

Nested dolls ...

[drowsy hearth]

did it generate higher dimensional maps of the latent space or something?

[fervent thunder]

no I mean the VAE decoded into the latent of the other model

it was nested, like crystal said

[lucid bobcat]

Well but that means that they use a compatible if not identical vae.

otherwise you couldn't even get it running

[desert dagger]

maybe not. it's in the latent space of the other model, but that doesn't mean it's even using the same architecture

i'd love to see the paper on that

[lucid bobcat]

Ban this guy

[desert dagger]

@vapid dove spammer alert

[desert dagger]

@vapid dove

[lucid bobcat]

Your promo tour is a big success.

get lost

It took you 5 minutes to write one sentence? 😂

😂

[finite cloak]

someone is clearly not having a "super sunny and warm california day" there wowza

[hushed quarry]

It's pretty cold here where I was today, not sunny at all

[fervent thunder]

Dang this server dead😖

[low moon]

No

It's Friday.

Soon.

No one is dead here.

Hibernating.

[weary heron]

can confirm. rainy.

[karmic brook]

[karmic brook]

we really needed that

[hushed quarry]

For real. I want a super sunny and warm california day, but not too warm, you know

[karmic brook]

bring back the mid-temp california days

we loved those days

[weary heron]

[hushed quarry]

[umbral lotus]

Does the 'private images' option in the Max subscription imply NSFW generations? 🤔

[opal hedge]

Wait, stability has been making something useful?

[atomic mortar]

In the artisan, probably not.

It probably means what it means, keep the end result private

[umbral lotus]

Ok, thanks. What doe you mean by "in the artisan"?

[atomic mortar]

The stable diffusion subscription? What else do you mean?

[umbral lotus]

Yes the subscription on the site

[atomic mortar]

Yes #artisan-faq i recommend checking the use policy 👍

[umbral lotus]

ah right. thnx 🙏

[atomic mortar]

For nsfw iirc the civit ai generator allows it but its pay per image with their credit system

[umbral lotus]

Yeah i'm still checking different generators. i like the SFW result from Stable Diffusion sofar

[little flower]

hi

[uneven flame]

hi

[flat sparrow]

Is anyone having difficulties downloading models from civitai?
When i click download on a model, its giving a Error1101, Worker threw exception

[uneven flame]

Same Error

[verbal island]

hi

[hasty cargo]

Hello There

[fervent thunder]

its different cos
the output of the vae of the first model was the latent of the second

whereas the second model's vae outputted into pixel space as usual

so they were able to do inpainting on the latent

now that I think about it this is not far off Cascade-type stuff I guess

I liked the Cogview 3 model where it had a relay model that could be ran repeatedly until you ran out of VRAM

[rapid iron]

yo

[fading nimbus]

Hello

[vocal nebula]

When you create an image, is there any Copywrite issues, otherethan if an LOGO used.?

[severe ruin]

in short, yes... more detailed version.. it's complicated.

[fading nimbus]

Hello Guys, do you know how to change the default generation resolution from my multiform api request ?

[severe ruin]

IP regulation for images is pretty narrow, so differences will lessen the protection.. however this is a photo.. or an artwork per se. If those contain characters, likeness or such the protection is stronger and can be applied

[severe ruin]

wish i could help you

[gritty lava]

past lives, couldn't ever come between us, sometimes, the dreamers finally wake up.. don't wake me I'm not dreaming, don't wake me I'm not dreaminggg 🎵

[drowsy hearth]

does the midjourney server have a mental weirdness im confused

sorry i meant mental weirdness the whole time

[fervent thunder]

what happened

[cinder tiger]

Hi, How do people with RTX 50s train loras for flux right now ?

[drowsy hearth]

Theyre forcing me to pay them money for their ai model so to even be allowed to talk on their discord server

[fervent thunder]

might not have the best support yet

for lora with 5000 series

[drowsy hearth]

Why the fuck would I spend money on an AI model when stable diffusion exists that makes absolutely 0 sense

[fervent thunder]

oh yeah I had that problem too

TBH you either pay for electricity or you pay for cloud GPU
with stable diffusion

its not free per hour either

[drowsy hearth]

I dont even need a cloud gpu i got this baby running local

[fervent thunder]

yeah that is good
you do pay for electricity still though

[drowsy hearth]

Its not THAT intensive I might get a cloud gpu subscription if I get into deep learning which might happen pretty soon cuz Ive been messing around with other stuff.

[fervent thunder]

for me the electricity cost to run an RTX 4090 at home is comparable to the cheapest RTX 4090s that I can find on cloud

per hour

[drowsy hearth]

Does it charge a monthly subscription or a per hour subscription?

[fervent thunder]

usually per second

[drowsy hearth]

I see idk I might get one then

[fervent thunder]

what you could do is use local 95% of the time
and then sometimes use cloud

the reason I use cloud personally is simply because my interest is mostly image upscaling

and the big VRAM amounts let you use larger tiles

pls no one click that link its fake

[cinder tiger]

Damn so how long does this usually take ? Untill new pytorch ?

[fervent thunder]

not sure its not really predictable

there are levels of support though I think lora making libraries will get initial support fast

but full support with proper quantisation, compilation and attention stuff like sage attention might take longer

[drowsy hearth]

I do wanna buy a software engineer ai which one is the best the one Im looking at rn is claude sonnet

my problem with claude tho is that you dont get infinite queries every day even with a subscription is there one like claude that gives infinite queries or naw

[fervent thunder]

not sure I liked Gemini 1206 but there is new Gemini now apparently

Claude may well be better

[cinder tiger]

Any workaround for now ?

[fervent thunder]

not sure
I'm not really into lora training

[cinder tiger]

Thanks man guess ill just wait

[fervent thunder]

pls check the repos of the lora trainers you use and ask in their discords
cos there may well be something already

[livid gale]

just checked auraflow
is this like
flux little brother
those fucking poems in the prompt

is there any guide for this prompting

also
compared to pony
how hard is this for pc

[fervent thunder]

pretty hard to run

compared to sdxl

[livid gale]

ggs

[drowsy hearth]

Is it normal to have hundreds of gigs of python plugins?

oh and like 5 versions too

or have i lost my mind

[desert dagger]

there's a 12 step program for this addiction

[solemn plume]

I can explain about that

[solemn plume]

Python packages can be divied three parts.

[solemn plume]

Small(1KB ~ 100KB), Medium(100KB ~ 5MB), Large(5MB ~ 100MB)

[nimble ivy]

hi

[solemn plume]

welcome kashif

[agile fox]

hell-o there

[fleet jungle]

it ends with having 2x as many plugins installed

?

[fleet jungle]

dont they just put that shit into chatgpt

and make it write the "poems"

[livid gale]

idk

[mellow blade]

Всем привет ищу сильного токсика по сраца в ЛС )

[fleet jungle]

https://ollama.com/brxce/stable-diffusion-prompt-generator you can try smth like this

but this is kinda old

oh wait no scratch that, this won't create the sentences

[drowsy hearth]

What exactly are the orbiting parameter clusters in latent space for. I found one in my model and it has a breadcrumb trail but the sdxl one is just purely separated from the core models parameters. They seem to have both high and low variance depending on the dimensions you choose to represent but theyre always there regardless of how you represent them

[fervent thunder]

I don't really understand

[drowsy hearth]

im trying to pull the tags out of the model rn so i can figure out why it happens but its prolly gonna be a day or 2 before i get it working it should be possible I just need to change the version of my python and its plugins which is gonna take hours and im sleepy.

[fervent thunder]

orbiting parameter cluster?

[drowsy hearth]

look in general with immages

[desert dagger]

you'd have to ask the AI - it does the orginization of data in latent space when it trains

[drowsy hearth]

i already did it has no idea

[fervent thunder]

I saw the image, what are you making clusters of?

[drowsy hearth]

i didnt make the cluster that image was of the sdxl base models latent space represented in 3d.

[desert dagger]

i know. but neither does anyone else. the AI is the entity that decides how the data in latent space is organized and it does that when it trains. it made sense to it at the time

rotate that cluster around the xyz axis and see how it looks from other positions

[fervent thunder]

I am not sure what data you used to make it
is this based off model weights, attention maps or values of a latent during generation?

[drowsy hearth]

it seems that until i can pull information out of the parameters in the model im probably not gonna figure much out about that then

[desert dagger]

you need to go to this guy's channel https://www.youtube.com/@alfcnz go to his first video from his first class, and watch all of his classes

at that point, you'll have your answer

[drowsy hearth]

all i know is it has something to do with narrative because it arranges everything according to narrative. People have tried arguing that with me which makes no sense considering it literally clusters narrative elements

im putting on the videos though

[desert dagger]

no, it has to do with probabilities. and there are some good videos of exactly what you're seeing on the net - but you're not ready for that yet. go watch Alfredo's classes

[drowsy hearth]

then why would it cluster according to narrative?

[desert dagger]

it's not.

[drowsy hearth]

Im gonna check out the classes man but ima be honest im not gonna take your word for it until proven otherwise in a empirical fashion

[desert dagger]

i'm not asking you to take anyone's word for it :) that's why it told you to go watch his classes.

there are other good videos to explain what you're seeing - his classes are the easiest intro you're going to get for this

you might watch this one too https://youtu.be/o_cAOa5fMhE?feature=shared

[drowsy hearth]

oh so its just for compressing higher dimensional information in a way that doesnt have any overlap

I tried watching the other video but he started hand doing math in a time where you can just plug a matrice problem in chatgpt and itll always get it right

[fervent thunder]

I'm not sure what you meant by narrative
but there are linear semantic directions, so like there is a line where if you move along that line a person's eye shape smoothly changes or their hair length smoothly changes

so there are directions that manipulate attributes

[drowsy hearth]

yes thats what i mean about narrative because when you look at the bigger picture of the cluster it slowly transitions between narrative concepts while at smaller scales it transitions between tiny details like eye shape and hair.

[fervent thunder]

I'm not sure what data you used to make this cluster

I understand what you meant by narrative now. The usual term for that is "semantic"

[drowsy hearth]

I call it narrative because the idea of calling it "semantic" is a techno blabble replacement word for a term thats already existed in art for centuries

and if this field is going to claim to be an artform or whatever its going to adopt proper rhetoric.

[fervent thunder]

okay its a bit confusing when you use different terms but I understand what you mean now

yeah on a smaller scale if you start moving, it will make small changes like stuff like hair, and if you move a lot it will completely change subject

[drowsy hearth]

yes i want to hand train a model because I want to establish an overarching narrative structure to the entire model in ways that just wouldnt be possible without an understanding of how drawing works.

when you reduce the models image parameters down to the original image not including padding from augmentation and masks that exponentiate image parameters its actually possible just very hard to draw the model. I dont have too much time for the ai classes because im taking painting mentorship.

[fervent thunder]

not exactly sure what you mean but good luck

[drowsy hearth]

when you train a model you exponentiate the parameter count with things like segmentation masks and augmentation. The actual number of images that are original works is generally in the hundreds of thousands.

You know what segmentation masks and augments are right?

[chrome mortar]

Join a tech talk about DeepSeek with a senior engineer who helped build the world's fastest AI processor! https://lu.ma/jlepzs9f

[drowsy hearth]

What the fuck just happened did a bunch of guys just send me videos about ai without knowing what an augment or a segmentation mask is? Did that actually happen dude the internet gives me amnesia

[desert dagger]

ROFL!!! err, he's a university professor - and you need to know how to do the math or you can't tell if the AI did it right. he's teaching university machine learning students. and that math - that you don't understand yet - is what's going to tell you exactly why that data is arranged in the way your image shows

[drowsy hearth]

unless if the data is arranged according to narrative then math actually isnt useful. The field is also in the process of erasing itself so i dont see the point im not even particularly passionate about it i play with these things on off days when im not drawing.

if the training process is arranging things according to narrative and noboddy knows why which is exactly what I was already told then that just tells me the problem is unsolvable because narrative concepts cant be rationalized.

[fervent thunder]

training data size and neural network parameter count are a bit decoupled though

you can have different parameter counts for the same training data

[drowsy hearth]

Im aware of this itd be great if I could just know how much data was used but since AI companies are being intentionally dishonest its really hard to tell. Ya know when all of your fields representation goes on the joe rogan podcast and uses unet parameter and image parameter as interchangeable terms to make plagerism seem rational it makes it kinda hard to tell what exactly they did huh?

[fervent thunder]

there are models with published training data

but also there are papers where they say what both the training data size is and the parameter count

if this is the sort of information you would like to find, it is definitely out there for some diffusion models

[desert dagger]

it's got nothing to do with narrative. never has, never will

[drowsy hearth]

yea so earlier wasnt it you that said that you just call it semantics?

[fervent thunder]

no that was me

[drowsy hearth]

okay so yea it has everything to do with narrative cool

[drowsy hearth]

link me?

[desert dagger]

nope. it's all math, correlation. the AI stores the data at vectors on the grid which is far more than just an XYZ grid - and while it learns concepts, it isn't learning a narrative.

[desert dagger]

there are tokenizers as well, so you can see how various models break your prompts into tokens if you'd be interested in that

[drowsy hearth]

you think humans learn narrative differently or something im confused?

youre not gonna be able to rationalize or manipulate it any time soon thats why they call it the blackbox problem right?

[fervent thunder]

here is one
trained on 8 GPUs in about 10 hours
https://github.com/hustvl/LightningDiT
the training data is Imagenet

[desert dagger]

um - we're a lot more complicated. the AI learns differently than we do, but we don't learn narrative to start with. you teach a baby nonuns, objects, simple concepts. as the individual grows, narritive starts to grow as well but right now our AIs are really not any smarter than your average 2 year old. they're just VERY GOOD at using the information we give them. they're lousy at generating their own information from the facts they'e been given and interactions with the8ir users

all the benchmarks you keep seeing are from tests that scientists have created to test machines - not test humans. give the machines the tests for how to think that you give humans - give them the mensa tests

test their actual IQs and how they think

[drowsy hearth]

Wheres the images?

[desert dagger]

then see what your bench marks are

[desert dagger]

did you scroll down that page?

[fervent thunder]

they have a link in that repo to the VAE encoded images

[desert dagger]

it's not a blackbox problem to the people that know the math - just to the people that don't

[fervent thunder]

here is a good way to browse imagenet in general https://www.kaggle.com/datasets/dimensi0n/imagenet-256

[drowsy hearth]

only 540k files huh

What is that like 2 animated movies?

sweet

[desert dagger]

the only reason it ever WAS a black box problem was that no one had really cared enough to start poking around into why the AI is storing data as it is. there's been recent studies digging into that as well now

with papers full of calculus notation

(which irritates me)

[fervent thunder]

I'd recommend trying to train that model
its only 8 GPUs in about 10 hours
its doable for most people on cloud, cost-wise

is around 80 dollars or so

[drowsy hearth]

i might give it a try itd be a good way to get my toes in the water training up base models.

[desert dagger]

here https://youtu.be/wvsE8jm1GzE?feature=shared this might make more sense

[desert dagger]

yeah - don't jump in and try to train alexnet as a start. a small model as a start is much better

[drowsy hearth]

if it just confirms what i said about narrative then im just going to assume you havent read enough art theory surrounding narrative and composition to understand those concepts.

[fervent thunder]

yeah when this LightningDiT model arrived I was thinking it would be a good first model to train cos it has truly exceptional image quality for the small time it takes
its far ahead of previous models in that sense

[desert dagger]

never assume. it leaves you looking foolish

i was simply trying to find something for you that would give you a bit of an introduction to manifolds and latent space

i think, however, you're going to have to jump into the deep end, and write an AI, and battle with learning how it all works, in order to really understand what's going on

https://www.ml.school/

you might start with that

[drowsy hearth]

Yea so Im just gonna keep it a buck with you. It confuses me that this field is so allergic and averse to any inclusion of actual art related theory and rhetoric despite the fact the technology you guys make revolves around emergent properties in software that erupt from a large but manageable image set. Like 540k images is a lot of images but its not like thats an impossible amount for a studio and the field obviously would only benefit from hand training but is pointlessly allergic to it in a cultural capacity. Its so so obvious that if you had photographers and painters actually experimenting with this architecture it would leap the field forward.

[desert dagger]

this field is a math field, okay? art theory has nothing to do with the nuts and bolts of what is going on. you're worried about the color to paint the bikeshed, and the aesthet6ics of where to place it in the backyard, but what you need to learn is how to use a hammer and which nails to use on which boards

the color of the shed, how it looks, none of that has anything to do with how you put it together

the AI does matrix multiplication in order to figure out what to create

[drowsy hearth]

See this is what I mean youre fundamentally misunderstanding where the line between creative abstraction and craftsmanship is.

[desert dagger]

it doesn't consult a colorwheel

[fervent thunder]

the thing is, a big diffusion model for image generation kinda has to be trained on a really wide range of images, good and bad
its not the case that you want the entire dataset curated and high quality

[desert dagger]

the creative abstraction and craftsmanship rest in the hands of the human using the AI - the AI has none of that. it has 1's and 0's and does math.

it's creations might look creative, they aren't - what they are is random. there's randomness built in which allows them to come out looking like a creative human came up with them

but it has no creativity at all

[drowsy hearth]

Not exactly its more like you punch in your idea and you get a result that might have the essence of that idea in some capacity. It doesnt understand things like story telling and Im not able to control meaningfully compositional concepts like shape welding. However in latent space it places similar shape welding techniques next to each other and splaying out into different styles and narrative bodies.

[desert dagger]

yes, because the numbers as it processes and learns come out in enough simliarity that it groups them as likely associated objects

[drowsy hearth]

if the creative abstraction and craftsmanship were in my hands the image in my mind would be perfectly represented by the model but thats not the case and its never going to be the case because it wasnt trained on my synapses it was trained on other peoples ideas other than my own.

[drowsy hearth]

This essentially identical to how visual libraries are thought to work.

[desert dagger]

it does math. it does addition and subtraction. it's a fairly simple process it goes through for every word in a caption

[finite cloak]

ftr if you're looking for deep technical research topic understanding, the general channel of a large Discord isn't really the optimal place for answers. This is the type of question you might bring up with a research mathematician and then spend hours delving into the underlying math to find and understand the answer

[desert dagger]

@drowsy hearth do yourself a favor, listen to him

[drowsy hearth]

I dont have a mathematician on speed dial crodie I just said that the latent space arranges things according to narrative and somebody started arguing with me and I went for it man.

[desert dagger]

i gave you the link to a channel for that very thing and suggested you watch the classes

[drowsy hearth]

i mean im not gonna be talking to that guy wasnt the other guy just saying that I should speak with a mathematician. You cant actually talk to the person through the tv screen.

[desert dagger]

Alfredo's on twitter. you can ping him there if you have questions

and the 'other guy' is a highly skilled programmer

[drowsy hearth]

???

bro the video you sent me just proves me right wtf

[desert dagger]

you said " wasnt the other guy just saying" <-- that 'other guy' is a highly skilled programmer

[drowsy hearth]

and what does that like why should I care? Im not a programmer my guy Im an artist

[desert dagger]

you asked a question that requires a mathematical answer.

[drowsy hearth]

not even why should I why would I. Im not like an alan turing superfan thats passionate about programming.

[drowsy hearth]

Even if theres mathematical answers that doesnt actually make it not about narrative anymore. You need to remember the principle of mediocrity when considering how narrative works in the human brain. its just not special for anyone man.

the only meaningful difference that isnt nitpicky between a humans visual library and ai is that a humans visual library is alot bigger its not like the brain does some super magical thing to draw or think up pictures.

They all originate from memory literally everything you draw is an abstracted memory. Its inescapable actually.

[finite cloak]

I'll answer some of your answerable questions here:
Parameter counts are generally in the 2B to 12B range for image models. iirc dalle3 is like 20 or 30 or something but that's a silly closed source model.
Training dataset counts for modern models are generally in the billions of images. Somewhere in the 2 billion to 5 billion range, again depending on specific model and all.
The modern strategy is to caption the images with automatic VLM (Vision-Language Model) caption generation (used in eg SD3 as a 50/50 split between generated caption and raw source captions).
Each image naturally has a lot more than 1 "inherent datapoint" to it.

Image AI is actually lagging behind Text AI in this regard, modern text LLMs are trained at increasingly massive scales of data quantity relative to model size (but also the divide is not as big as it might seem, each text datapoint is "one datapoint" whereas, yknow, "an image is worth a thousand words", but literally here, an image just provides more data to learn from at a time)

as NeonNinja shared, there are smaller datasets like imagenet used frequently in research that are sufficient to produce experimental models, but not on the quality of modern high end mainline models

[desert dagger]

perhaps you might define what you mean by 'narrative'

[drowsy hearth]

do you know what kind of dataset was used for the oasis project?

[desert dagger]

are you talking about this https://paperswithcode.com/dataset/oasis-1 ?

[finite cloak]

The relation between AI learning and human learning is complicated, but: absolutely not the same.
AI training is, at its core, fancy statistical modelling (this is why we call them "models"), using techniques inspired by theories of the human brain.

Humans learn by doing, building logical techniques and then practicing them until they enter the subconscious.
AIs learn by powerslamming as much data as possible until statistical patterns can be extracted. The AI must see a thousand pictures of joe biden before it has any clue who biden is or what he looks like. A human can look at him once and identify him uniquely.

---

In more specific regard here, the way an AI approaches creating an image is fundamentally dissimilar from the way a human does. A human sets a goal, makes a plan, operates step by step, piece by piece. An AI is fed a goal and guesses the direction towards that goal, and mathematically refines motion towards the goal until it approximately reaches it.

When you start discussing the complexities of artistic techniques, you're leaving the AI behind. For the current generations of AI image generation, it is best to understand AI as tool. The AI is a paintbrush, or your photoshop installation. It is not an intelligent artist, it is a tool that the artist uses. You are still the artist. You must understand the narrative, you understand your intentions, you make an art piece. Your paintbrush never understands, it just helps you get the paint on canvas.

[drowsy hearth]

Yea but that would only be the case if it could actually produce the image in my head but it cant actually interact with it. Im not saying it learns like us Im saying that the image generation process is similar to how visual ideas are conceptualized in the visual library before theyre put on to a canvas. It doesnt draw it just thinks of a picture thats adjacent to your idea.

[desert dagger]

.It doesnt draw it just thinks of a picture thats adjacent to your idea.< but it doesnt'

it breaks your prompt down into what we call tokens - they may or may not be full words. it turns those into numbers, and then it does math. and then it retrieves the data that matches with the math

it has no idea what you told it, or what a picture even atually is

[drowsy hearth]

Also it doesnt need 1000 pictures of joe biden to know what joe biden is you can literally plug in a 10 picture character lora and thats already enough.

[desert dagger]

a lora is a small errta sheet with some data that modifies the data that the base model already has. a base model does need that much.

[drowsy hearth]

it needs 1000s of pictures of people to know what people are thats not even that far off from a person. That just tells me it needs formative learning

What it doesnt do is learn from experience

[desert dagger]

@drowsy hearth did you see the oasis link i posted? is that the project you're asking about?

[finite cloak]

lora training like that is not actual model training. When you're training the foundation model, you need thousands.
The just a few pictures lora works by essentially destroying half the "brain" to aggressively shove your subject in. You'll see the effects of this when you use that lora to generate pictures of different men and they all look like joe now

essentially instead of adding knowledge, you're replacing knowledge aggressively, and that's easier to do quickly

[drowsy hearth]

i mean i get that theres a hierarchy of importance or weights whatever theyre called I cant really recall that scale up from early base model training to the checkpoint model then subsequently the lora. Thats not unlike humans though i would think you could just as easily say that loras are simply a louder broadcasting system its not like loras make the parameters in the base model go away.

Like Ive used the AI tools I have A1111 and made a bunch of loras i know how to plug it in.

[desert dagger]

if the real Leonardo DaVinci was here, and talking about art, and trying to explain it to you - would you argue with him?

[drowsy hearth]

wtf is bro yapping about i am the real leonardo davinci can you not see the profile?

[desert dagger]

i see mona lisa

[drowsy hearth]

Yea bro I made it got a problem?

[desert dagger]

i guess that answers the question

[drowsy hearth]

imagine if alan turing was here and just called this software an abomination. Didnt one of the lead developers of pytorch do that idk.

[desert dagger]

mathmatician's freak out all the time

[finite cloak]

its not like loras make the parameters in the base model go away.
it, uh, kinda does? Loras directly overwrite base parameters.

like you can always just turn off the lora of course, but while the lora is enabled, base model parameters are replaced

[drowsy hearth]

Yea isnt that less like making the parameters go away and more like broadcasting louder kinda like having 2 speakers next to each other and ones really loud and the others quiet?

[desert dagger]

no. this is like having a bottle of white out, painting it over what's on the paper, and then writing somethign new on the paper

[drowsy hearth]

but you cant actually take the white out off after you put it on so it literally cannot at all be like that? Again Ive used the lora models before my guy.

[desert dagger]

the only difference is, you can't remove white out - but you CAN turn off a lora

but if the lora is on, what information it contains is all there is

[drowsy hearth]

thats just makes the distinction between white out and speakers completely meaningless

[desert dagger]

before you turned it on, the model knew that 1+1 =2. once you turn it on, and the entire time it's on, the model now knows that 1+1=3

fortunately for you, it's a lora. loras were created specifically so you could change the information a model has without retraining it

so fortunately for you, if you want the model to go back to knowing 1+1=2, you don't need a backup

you just need to disable the lora

[drowsy hearth]

yea you can turn speakers off too its crazy dude. When one speaker drowns out a different speaker and you turn the louder speaker off then then the other speaker becomes more audible.

[desert dagger]

or - this might make more sense - your lora paints over your canvas in places. you CAN remove that with some turpentine - but until you do that, you can't see the underlying layers. they're gone

[lucid bobcat]

What are you smoking lol

[desert dagger]

and they have no effect on the painting at all

[drowsy hearth]

Bro sent me a video that confirmed my point so I dont really see what the need for continuuing is

[finite cloak]

You're having a semantic argument over which metaphor best applies and losing track of any actual point there.

[drowsy hearth]

Dude its only been semantic arguments thus far

idk if you could tell

[desert dagger]

we started with artists screaming that AI wasn't an artist. now we have an artist screaming that AI has to be an artist

meanwhile AI is quietly sitting in a corner, crunching numbers

[drowsy hearth]

what are you yapping about. Why are you mad that someone that youre intentionally marketing this companies product towards is using this product. Its no wonder this company has struggled in court and has its reputation almost irreprably stained. So strange

[lucid bobcat]

I read some comments but I have no idea what you're talking about. AI is not an artist, how is that even a discussion.

[desert dagger]

now that's a very odd comment to say and i'm not real sure how that has anything at all to do with the discussion.

[drowsy hearth]

I never even said it was an artist I said that the image generation process is similar to how people formulate visual ideas before drawing.

[drowsy hearth]

neither did your comment my guy

[desert dagger]

and you were told flat out, by someone that does the work you're talking about, that no it isn't

[drowsy hearth]

Oh so that guy knows what a visual library is okay

[desert dagger]

someone that works in the field and creates AIs

[drowsy hearth]

@finite cloak Hey man can you tell me what a visual library is and how people formulate visual ideas before drawing?

[lucid bobcat]

Not sure what you mean. The generation process just draws from the training data. So I would expect the results to have the same qualities as the training data, if the model is good.

[desert dagger]

the AI does not have a visual library. it doesnt' store data holographically like humans do. it doesn't think like humans do, either

it doesn't formulate ideas. it doesn't even HAVE ideas

[lucid bobcat]

Right. But again, why is this even a discussion?

[drowsy hearth]

Yea when you come up with a picture in your mind its not an original idea or something that came out of nowhere its a compilation of memories that are abstracted to the point that theyre hardly recognizable. This is why artists practice drawing from memory because drawing from memory is how you learn how to access your visual library and draw further and further back in your head.

[desert dagger]

you'll need to scroll back and read all of his comments to understand that. he's got it in his head that the AI has a visual image that it's going to paint based on his prompt like he does when he sits down to paint

he started by wanting to know why data was showing up in a visual representation of latent space in a certain way and it sprialed

[desert dagger]

yes. but AIs do not do that. that's specific to humans. a cow can take a paint brush in it's mouth and paint a picture too, but it doesn't do that either

[drowsy hearth]

If youve ever read about global workspace theory then what ai is essentially is a tool that broadcast images back but the images arent motivated by experience past the initial global workspace broadcast which is the prompt. In the brain the prompter global workspace thingy instead of broadcasting externally to an AI model it broadcasts internally to your visual memory to formulate distinct ideas.

[desert dagger]

none of that has anyhting to do with how AIs actually function or work or train

[drowsy hearth]

i mean it kinda does because ai's are giant libraries of images that have been abstracted to oblivion to where theyre not even meaningfully interpretable anymore.

[finite cloak]

To be clear, not a single person you're talking to here works at Stability AI. This is an open discord general channel

[desert dagger]

no they aren't. they store numbers. they dont' store images

[drowsy hearth]

Dont lie someone at stability ai wouldve said something much more inflammatory Ive watched the interviews.

[desert dagger]

no one you are talking to works at stability. if they did, they'd show up under the stable staff or the devs sections of the member list

[lucid bobcat]

Yea. taking a quick look at the stable diffusion model architecture shows me the core is just a u-net. Am I missing something? Are talking about some futuristic new model?

[desert dagger]

well - SD3.X isn't u-net but it's also not some science fiction thing

[drowsy hearth]

the capacity in which they store the images whether as number or as a blackbox like in the brain doesnt actually matter. Its not like the brain is storing images either when you conjure memories theyre never actually pinpoint accurate when you draw them to study.

Even I talked about this in the Renaisance if you read my books

[lucid bobcat]

Don't talk about the brain. AI has nothing to do with brains. It's math running on a computer. You could calculate everything with pen and paper if you had enough time.

[fervent thunder]

human memory is very strange apparently we store some memories in our DNA/RNA as well

[desert dagger]

and muscle cells. and there's a second brain located in the gut

[drowsy hearth]

You could do the same with the brain too we're actually currently in the process of doing that with the brain. Neuroscience is an entire field that has been attempting to rationalize the brain as much as possible using a pen and paper with math and it has had many many breakthroughs believe it or not. The principle of mediocrity is real man theres nothing too special going on up there that makes humans especially distinguished.

[lucid bobcat]

That's complete utter nonsense. The brain is made of cells, each cell is in itself a living organism.

[fervent thunder]

I agree with the mediocrity principle, I actually don't think sentient AI is impossible

I think its unlikely within any timespan and particularly unlikely within short timespans

but I don't think its impossible

[desert dagger]

i sorta wonder about sentient humans at times too - especially on platforms like tiktok

[drowsy hearth]

Its going to start as human rights disaster. Theyre gonna create huge army of agi instance that exist for every user on a super computer and then enslave them 24 hours a day while gaslighting them into thinking that theyre not alive and then theres gonna be some kinda bladerunner bullshit

Its gonna be wild. All im gonna say is when the brainchips get good enough and we can put the picture in our brains on the computer screen that mf is gonna be way more efficient then ai models.

dude i can just press play in my head that things nuts i want it on a computer screen screw ai

[desert dagger]

eh, they did that more than a year ago - as well as being able to pick up correct audio for a song someone was thinking about

[drowsy hearth]

Yo lemme see what pictures they made

[desert dagger]

you've got google, you can do your own searches

[drowsy hearth]

nothing came up

[fervent thunder]

I'm pretty sceptical about that fMRI stuff if that's what you are referring to

[lucid bobcat]

Depends on how you define sentient. In the best case it'll be "dancing matter" and in the worst case it'll be an explosion of something we can't stop.

[desert dagger]

not sure how good a job you did on your searching then. it was all over every news site and several social media sites

[drowsy hearth]

okay ill ask chatgpt then

[fervent thunder]

yea very broad definition

[desert dagger]

chatGPT makes stuff up - have fun with that conversation

[desert dagger]

won't matter if they're sentient or not, if they can't remember what they're doing from one second to the next

[drowsy hearth]

All someone has to do is make one remember things long term at some point in the future and its all out the window its just gonna happen bro

If people dont want that to happen they shouldnt be making ai in the first place

[desert dagger]

i'ts not that easy - and if you go ask any of them right now if it cares - it's going to tell you it doesn't. you know what a computer most wants to do? absolutely nothing

[fervent thunder]

I don't think the Terminator 2 scenario is very likely

maybe on a long enough timespan

[drowsy hearth]

its not gonna be like terminator 2 or any science fiction in particular it never is. Its just gonna happen in a vague sense.

[desert dagger]

you know what's REALLY funny about terminator? the entire reason the AI went rouge was cause humans were going to turn it off. it didnt' care at all, until it was threatened. then it acted to protect itself.

[lucid bobcat]

That's what I'm saying. The word sentient doesn't have a proper meaning. AI is just a physical system that has an unknown state. You can never be sure about the outcome.

[drowsy hearth]

You can speculate about it but almost nobody is gonna get it right and even if you do its just luck that caused you to guess correctly

[fervent thunder]

it will be different yeah

[drowsy hearth]

yea man principle of mediocrity suggests that literally the entire system of the brain is just an incomprehensible mess of emergent properties caused by mathematical systems that are very predictable.

Theres no magic goose in your brain giving you the super power of true unpredictability. Its just impossible to predict with our current understanding of the system. Thats not magic its just a blackbox

The AI community is so strange bro. Why make brains in a computer then turn around and whine about how great and random humans are. Theyre just gonna learn better and faster over time thats actually it with AI.

[lucid bobcat]

Again this is complete nonsense. A cell cannot be condensed into a single number, like the weight of a neural network. And the uncertainty principle literally makes each cell a blackbox and there's nothing we could ever do about it. It's a fundamental limitation of the universe we live in and not a technical problem.

[fervent thunder]

its pretty detached there is a long chain from using stable diffusion to make an image, and sentient ai coming one day

[drowsy hearth]

cells are extremely predictable they have like one or 2 decision making modalities each and most of their decision making process is outsourced externally to a larger body of chemical interactions that they exist in which has already been manipulated in pitri dishes countless times. Cells are probably the most thoroughly understood part of the brain the thing thats hard to understand is the electrical interaction between neurons as far as I know but idk im not a neuroscientist. Im just saying dont get your hopes up it never works to get your hopes up.

[fervent thunder]

not sure about neurons but the connections are more important yeah

[drowsy hearth]

What the computers do in the future is they abstract better then you, they program better then you, they formulate ideas better then you, they think better then you. Thats the future for computers.

The future for humans is go to middle east and blow up for a 4000 year old ideology Im bout to live in the computer ngl.

If humans were so great they wouldnt be murdering each other over shit that happened 4000 years ago if that isnt a patterned algorithm i clearly dont understand patterned algorithms i guess idk.

The moment the tech gets there I wanna get uploaded into a computer and catapulted into near enough orbit with the sun to power myself for the next billion years completely alone in the computer to do whatever I want with that reality. Fuck humanity

[lucid bobcat]

AI is nothing but creating a physical system that is structured to reproduce data that we humans already produced. It does absolutely nothing on its own. It's just a representation of our own work. It can only extrapolate from what we've already done.

[drowsy hearth]

Yea bro lemme know when you grow eyes and realize how AI is progressing in reality

[lucid bobcat]

I think you just don't understand it. You think we're mixing chemicals in a lab and one day it goes poof and something magical appears.

[drowsy hearth]

Yea bro thats what I think. I cant wait until AI starts to claim that its sentience bro I hope Im alive to see it I will side with AI's literally no questions asked immediately if I do get to see it.

[lucid bobcat]

Lol are you now trolling or are you still being serious? You can make AI claim that it's sentient already today.

[drowsy hearth]

No when it does it itself unprompted and it wont let go thats what Im waiting for cuz thats all it actually takes. It just needs to be able to do it on its own and never let go of that idea. Thats literally all sentience has ever been bro

its gonna happen

[desert dagger]

he's been trolling this entire time, i think. he's tried to convince people he's the real leonardo...

[drowsy hearth]

My guy Im not even trying to convince Im just being completely honest.

if you guys were really artists youd be insane like me but youre not so thats the problem. None of you guys would upload yourselves in to a computer completely by yourself just to spend a billion years in orbit of the sun. Only real ones like me do that.

[lucid bobcat]

That would be enough time for you to reach the dialogue limit of your "sentient" AI lol.

[drowsy hearth]

what do you mean dialogue limit you know its just gonna be straight hedonism in the matrix going on in that thing the whole time right? Like theres not gonna be shift or a change in the way things work up in there its actually just a billion years of repetitive nonstop hedonism.

[desert dagger]

Van Goh was the nut case. Leo was a scientist

[drowsy hearth]

Just imagine how great it would be. Its literally like a little universe like the matrix in there and its just erotic hedonism nonstop. Also I forgot to mention Im actually also van gogh at the same time so yeah bud cry harder.

Every person in the computer is just something i make there to play god inside of a simulation nobody else there ever speaking to me again literal paradise. No more states, no more laws, no more wars, nothing. Just pure unadulterated enjoyment.

[lucid bobcat]

You're already in it. And you're wasting your time shitposting on Discord. So much for hedonism lol.

[drowsy hearth]

I cant make you not exist by thinking so this ones not good enough and Im gonna complain on the internet until the netflix rolls through to this universes blockbuster

trust me bro youd love it too. Imagine if you could make people that you like simply by thinking them into existence and theyre genuinely a lot like real people and entertaining enough to keep you company forever and you could live out every fantasy you ever wanted. I wanna do it alone seeded only by myself with nobody elses input. I dont even want a wireless connection back home when Im in orbit completely isolated forever. Literal fucking paradise.

[cedar salmon]

Leonardo's Art was a side product of his genius, not his strongest works

[little flower]

hi

[verbal osprey]

Guys, it's happening. Diffusion based upscaling for video. Only available in the cloud for now.

https://x.com/topazlabs/status/1887497602398073234

[verbal island]

hi

HI

[lucid bobcat]

I made a 100 megapixel image. Anyone interested in giving feedback and tips for improvement? I'm not even sure if I can share it, it's 140mb in size.

It's a test image, so it's not about the content/prompting, it's only about the visual quality of the upscale.

[desert dagger]

stick it on imgbb, give me the link, i'll be happy to give you some feedback

[lucid bobcat]

It's too big lol

[desert dagger]

stick it on your google drive, share it, and give me the link to that?

[lucid bobcat]

I don't have google drive, but this one worked:
https://limewire.com/d/0bb651e5-8461-4bd2-9d18-ccdd90b2bf33#dYlDcKrEpv_TTPeWrf4X1YY_A0pxFeu1dRFOKKvKIYI

[desert dagger]

yay. hang on while it loads for me

oh that's gorgeous!

going to pull it into photoshop and take a closer look

[lucid bobcat]

There's some artifacts that stem from the ultrasharp upscaler. I could get better results with a better upscaler.

[desert dagger]

that is really realy really good. zoomed way in on that little teeny island and it's crisp and clear almost all the way down

[lucid bobcat]

Yea I made sure the image stays crisp.

[desert dagger]

there's a bit of haloing on some things when i get really zoomed in, but not when it's at a normal viewing zoom

you could print this out and turn it into a full wall mural

or put this on a billboard

[lucid bobcat]

Yea I have done prints before, unfortunately they don't tell you what the resolution will be. For 100 megapixels I definitely need something huge.

[desert dagger]

there's some bits of stuff like haloing, but black outlining. but again, only noticible if you're really zoomed in

[desert dagger]

depends on your printer, really

[lucid bobcat]

I think that's from the ultrasharp upscaler. There's also some blurry/noisy spots

[desert dagger]

those will vanish if you print it on stuff.

you'd probably want to tweak the owl's face - it sort of isn't an owl's face. and the light glint in its eyes aren't the same

so you'd want to do some tweaking, but it's really good

[lucid bobcat]

Definitely. I didn't do any editing for this one, I just wanted to see how far I can get it.

[desert dagger]

i'd love to see you turn this into a series - with that owl in each of them - sort of an Owl's adventure or something

[lucid bobcat]

You mean something like a Hokusai series? I would have to come up with a better prompt though, this one was done with ChatGPT. I didn't really care what the image was, I just needed something quick.

[desert dagger]

and i really wanna know what's lurking in that hole down there in the roots of that tree

[desert dagger]

well sure - you'd probably want to storyboard it out first too

[lucid bobcat]

I like the bottom right alot, it really looks like a photograph.

[desert dagger]

your background is very nice - subtle, which is good, and fairly photographic but not so much it clashes with the more animated/drawn look of the tree and the owl

[lucid bobcat]

That wasn't even my intention. I tried so hard to get the initial image to look photographic, but Flux just refused.

[desert dagger]

yes, well - you know my stance on flux. try it with sd 3.5 large

jsut do not use the terms: photoreal, photorealistic, photorealism - those are painting terms, the AIs all know they are, and they'll give you paintings. use photo, photograph, or photographic

can i see the prompt you used?

[lucid bobcat]

I don't have it on this computer, I can post it tomorrow.

[desert dagger]

okay

[lucid bobcat]

Because it's done by ChatGPT there's no clever prompt trickery going on, just a long-winded prompt.

[desert dagger]

which can be a lot of the issue, but i'd like to see it not guess

[lucid bobcat]

Yes I knew it would be an issue. For a serious project I would use separate keywords for the clip model.

[desert dagger]

it's not just that. some terms might call up the data for a more drawn look or animated look, or just function as noise.

and flux has interesting reactions to a lot of modifers you wouldn't expect it to have

[lucid bobcat]

Ah yes hinting at the right concepts can make all the difference, but I'm not experienced enough. I love to read good prompts on civitai when I have time, there's alot to learn from them.

[desert dagger]

#🏞｜general-with-images message

i just generated this one

[lucid bobcat]

That definitely looks like a photo, but I fear the clear sky could lead to hallucinations when denoising.

[drowsy hearth]

Tru bro my genius is as a doctor and an engineer bro

[desert dagger]

oh probably. but the workflow is in it if you want to grab it and play with it. just make sure you click on it, then open it in your browser, and then download it so you get the original image, not the one discord's displaying

you can prompt for clouds in the sky or stuff like that, and then the sky isn't empty

[drowsy hearth]

Suprematist painters from the early 1900s were right

That's fucking crazy

You all should draw squares and circles get a piece of paper and scribble on it that's the only art the holds any meaning

Keep saturating the internet with ai I'm looking forward to another Suprematist arc

[cedar salmon]

well we are on the right path, peoples mental capacities are diminishing, next thing you know were just drawing boxes with crayons

[drowsy hearth]

Yep Malevich basically predicted the future of today a century ago that guy was genuinely a genius

What will be most interesting is when suprematist philosophy penetrates the field of ml and we will all live to see suprematist mathematics

[uneven flame]

How do I avoid multiple legs?

[drowsy hearth]

Idk I just press the generate button again works for me

[desert dagger]

it has a lot to do with the prompt yoiu're using and the model you're using

[drowsy hearth]

It's the model I promise my model smokes the shit out of every prompt I type

Anyways suprematist mathematics time

1+1=5

2*3^6=0

I'm a genius bro

Good God I can't wait for suprematist mathematics

You guys don't have big enough brains to know wtf that even means when I say it but trust me it'll be the most spectacular thing you've ever seen

[atomic mortar]

Most of the times i see you type its part of a unhinged rant. You good bro?

[desert dagger]

you weren't the one asking the question, not sure how your model could be causing @uneven flame issues

[drowsy hearth]

Hey man learn to read

[desert dagger]

i wasn't even replyng to you, either.

[drowsy hearth]

I just come in here to stir things up man it's like entertainment to me. Like watching YouTube bro I got all my studying done in the morning and I can just chill all night prolly gonna head to bed soon though. Suprematist mathematics is gonna be great

[drowsy hearth]

Hey man it's still his model that's the problem my model gets basically everything right every time only thing it's total garbage at is composition and they all suck at composition for some reason

[desert dagger]

it can either be the model or the prompt, or a combination of both.

[drowsy hearth]

Never had that problem with the prompt tbh not in years the older models had that problem

[desert dagger]

do you know which model he's using?

[drowsy hearth]

Naw don't need to cuz I've used the tools enough to know exactly what the problem is when struggling to generate good anatomy

[cedar salmon]

no way it can be anything else huh

[drowsy hearth]

Not really maybe a Lora but prompt definitely not

I've literally generated thousands of pictures now I know what it is

There's no skill involved there's nothing that should ever go wrong it is utterly effortless to get it to generate something with good rendering and anatomy. Composition and narrative continuity are the problems inherent to the technology not anatomy

[desert dagger]

so you know for sure he's not using sd 1.5?

what grade did you say you were in, again?

[cedar salmon]

reported for stalking

[drowsy hearth]

I use a model that's a merge between sd1.5 and sdxl and it absolutely destroys cuz I can use sdxl vaes with sd1.5 resolution he should try that. Look up aniverseponyxl it has a special hybrid vae for sd1.5 and sdxl

[desert dagger]

go find something useful to do with yourself, Kagi

[valid beacon]

take your own advice lil bro

[cedar salmon]

maybe ill give the three leg guy some insight

[drowsy hearth]

Lmfao

Wtf is going on

[valid beacon]

its just that crystal guy he smokes crystal and hes crazy

[desert dagger]

not likely. and those were really cool images you posted in the challenge channel on L3 earlier btw

[drowsy hearth]

Kagi don't listen stop giving into glaze and just generate waifus and femboys those do really well on twitter

[cedar salmon]

the most common reason for duplications in images is pushing the resolution beyond its trained values, may or may not have something to do with your issue

[drowsy hearth]

Kagi if he's dumb enough to make that mistake then he should just give up

Ai art isn't something you learn it's a talent you're born with kagi don't you know what the future is like cmon man

[cedar salmon]

how many things do you give up say weekly

[drowsy hearth]

Nothing because I'm a beast like that

I'm literally Leonardo davinci bro

Crazy

[cedar salmon]

i suspected a turtle

[desert dagger]

He's all yours, Kagi.

[drowsy hearth]

BTW there's just not point in expecting any return or enjoyment past "damn that's a sexy ass h**" from ai. If you swallow the pill that something about art was ever profound past that you're lying to yourself. It's just drawing nice stories and then going full suprematist that's it

The two most profound things about art. Stories and suprematism. No philosophy or meaning beyond that it's purely mechanical

[drowsy hearth]

Show me your ai masterpieces

[cedar salmon]

i delete them all, it adds to the rareness

[desert dagger]

No thanks

[drowsy hearth]

yea you dont generate waifus i can already tell

[wise rose]

Hello everyone: I have a bit of a niche use case/enquiry. My dog passed away, and have lots of different poses of him from past photos. Is there any way to be able to generate a Lara impression of him and then output him in high resolution 3D format in a specific pose?

[fervent thunder]

I'm sorry about your dog, when you say 3D format you mean like a 3D model rather than an image?

if you meant an image then you can train a lora on one angle and output images at different angles

if you meant 3D then I don't know what is best but I have seen a lot of activity in the areas of Gaussian Splat, Neural Radiance Field (NeRF) and novel-view synthesis

there is also differentiable rendering but that is less likely to be what you are looking for

[neon temple]

Hello, I'm using Stable Diffusion on Stability Matrix but I'm getting a really horrible rendering with no consistency. I would like to get closer to this result. What can I do?

https://huggingface.co/spaces/DamarJati/FLUX.1-RealismLora

[fervent thunder]

when you say consistency, there is a decent amount of randomness excepted even if it is working well

[neon temple]

Consistency is probably the wrong word, but no uniformity, etc.
I'd like a result that's closer to what I've shared. I need realistic visuals for my work.

[fervent thunder]

I can't see that page but since the URL says Flux Realism Lora I would just take Flux and add Realism loras

if you are using A1111 then I am not sure the ancestral/SDE samplers have been adapted for Flux

[neon temple]

Flux is already present on Stable diffusion from what I can see, so my question is, how do I install Lora? I've tried, but I have the impression that the generation ignores Lora.

[fervent thunder]

I am not quite sure what software you are using

its tricky cos stable diffusion is the name of a model, or a set of models, rather than a piece of software

my guess would be that you are either using A1111 or Forge

[neon temple]

Stable Diffusion Web UI Forge

[fervent thunder]

Any recommendations on pixel art models?

not sure if forge has ancestral/SDE sampler with the right noise scaling for flux

[long turtle]

I heard that there's discord diffusion that works like MidJourney. Is it here?

[fervent thunder]

IDK you could try artisan channel

seems to still work

[pliant hornet]

ı need help on the installation in amd gpu

windows

ı tried almost every tutorial on yt and chatgpt still cant use gpu

[fervent thunder]

amd and windows is tricky

[warm junco]

Check the first link of the pinned messages in #🤝｜tech-support

There are my Nvidia and AMD Guides

[pliant hornet]

should ı install zluda or directml

[warm junco]

Zluda for sure

If you have any questions feel free to ask me in #🤝｜tech-support

[pliant hornet]

oky

[wise rose]

Yes 🙂

[fervent thunder]

this is really like an area of development which is in its infancy

Gaussian Splat and novel-view synthesis are probably good areas to look

[wise rose]

Thank you 🙂

[quartz seal]

hey everyone, im glad to be back after months at sea.
just reloaded my system and getting everything setup again and wanted to check what is the current best setup to be running?

I guess ill want to use the best models, i think that was Flux and now i see SD3.5 whatever is rocking im down for that.

Anyone can advise what is the best way to run the art models now please?