#💬｜general-chat

1 messages · Page 182 of 1

eager mauve
#

i think mine is 16 ram but im not sure

#

but damn

atomic mortar
#

Depends 100% on model. Sd1.5 like 4-5s 3070ti

fervent thunder
#

Do you need to smell colors or what

#

Why so big resolution

atomic mortar
#

And sdxl like koitenshin 14-30s depends on the lora

fervent thunder
atomic mortar
#

And resolution

#

I typically use smaller resolutions that are close to 1024*1024 ( like 882 * 1238 close enough)

#

So its faster

#

Mostly 3:2 or 16:9 tho

eager mauve
#

i did what yall said and its better

#

but its still like 8 mins

dense token
dense token
atomic mortar
#

Oh yeah but im on phone and i use a silly dropdown

#

My braincells dont function at 5am

#

Well 6 now

dense token
atomic mortar
#

Nah i use swarmUI, its useually a bit more then a normal 1024*1024 pixels

eager mauve
#

what browser yall using btw

fervent thunder
#

brave and edge

fervent thunder
eager mauve
#

would you say its the best for SD

#

ive been seeing firefox too

dense token
fervent thunder
#

Isnt sd just webpage

eager mauve
#

some people have trouble wit their browser

#

ion know using more ram and stuff

fervent thunder
#

Then use opera

#

Like guy on the top said

eager mauve
#

i am but most say firefox is the best dats why im asking

atomic mortar
#

Im still a disgusting chrome user with 20 tabs open on Average

#

But i totes recommend firefox

#

If you have +32gb of normal ram it doesn't really matter that much

dense token
#

When I say hate I actually mean "rampant memory leak"

eager mauve
#

it says Installed RAM 32.0 GB (31.2 GB usable)

dense token
fervent thunder
eager mauve
#

with amd radeon rx 6800s

dense token
eager mauve
#

this message appears a lot

eager mauve
#

or amd gpu just sucks

dense token
eager mauve
#

ill do it tomorrow then cuz there is

fervent thunder
dense token
#

Forge / ReForge have ComfyUI's memory management + A1111's interface. That's why Forge is faster than A1111

fervent thunder
#

I thought if you dont have nvidia gpu you re basically cripled in aspect of sd

dense token
eager mauve
#

yeah thats why theres zluda

eager mauve
#

forge and reforge i mean

fervent thunder
dense token
eager mauve
#

reforge it is then

#

oh wait

#

theres only a forge guide

#

nvm

dense token
dense token
fervent thunder
#

Or what

eager mauve
dense token
fervent thunder
#

Understandable

signal nest
#

anyone know if stablediffusion.cpp can do tiling (repeating textures or horizontally repeating panoramas/ scrolling backgrounds, that sort of thing). ages ago i improvsed it the hard way in python for sd1.x , then saw they added real support

glacial osprey
#

Hi

outer cloak
#

Hi, is there a way to capture the timing of each inference step during image generation? I can see the timings through the loading bar, but I would like to record them separately.

outer cloak
#

Does it make sense to calculate (1024x1024) / (Average time of inference steps ) to estimate the average number of pixels generated per second?

I understand that diffusion follows the concept of noising and denoising, but I feel it’s more accurate to measure the time for each inference step, take the average, and then compute pixels per second—rather than using the total inference time [for all the inference steps ]for this calculation.

abstract quarry
#

there is no stable diffusion program, it's a model. There are dozens of different tools that can run such model. For whatever question you have, you always have to say which tool you use

#

because, yes, in many tools you will see the time per step in the console, but probably not in every tool

#

seconds per pixel does not make necessarily sense, because the running time increases quadratically with the number of pixels. So a twice as large image might need four times as much time to be generated. However, in practice this depends also on your graphics card

tardy steppe
#

Hello I would like to create rule three four content with the app, how can I do it

still glacier
# outer cloak Does it make sense to calculate **(1024x1024) / (Average time of inference steps...

Generative AI does not work "per pixel". So I'd say it would be a bad unit of measurement.
Secondly as kaibioinfo pointed out you would be measuring the inference program (auto1111, forge, comfyui, etc), not the model .
Thirdly speed does not scale with resolution in a perfectly linear fashion.
Also you can already check this page if you want some rough benchmark numbers : https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html

main snow
#

Yap

atomic mortar
#

You mean you want to run stable-diffusion models on your pc? What gpu do you have

hearty night
#

My mom keeps watching master chef on the tv and I can't use my playstation ts pmo sm 💔💔

bleak matrix
#

Good morning, everyone! How are you doing today?

atomic mortar
#

Just got off work, also morning? Sheesh big timezone difference

fervent thunder
tardy steppe
#

how can I create r34 content

fallow veldt
# eager mauve with amd radeon rx 6800s

Your images are taking 20 min. you say? You must be running it on CPU. For AMD/Windows I think you have to install something called Zluda. But I wanted to answer as I have a RX 6700, and what I do is to run it on Ubuntu 22.04.5 (dual boot with Windows). I think that's the optimal way for speed.

fervent thunder
#

ye for zluda probably want linux

fervent thunder
tardy steppe
#

Say sorry to him

fervent thunder
#

just report it, don't engage cos it could be a malware bot
(if you reply to them it triggers them to say more and sometimes to direct message you)

#

the way the bots mostly work is after "talking" in direct message a bit they drop a URL link, and that's where the malware is

atomic mortar
#

@vapid dove we got someone being a big ol baby in general

#

❤️

#

eh i just wanted the emojis

tardy steppe
#

Peace please, you are friends

#

friends dont' do that

atomic mortar
#

i have never heard of this man in my entire life, though he does sound like the previous troll from last time

#

thanks maxfield, your the best

fervent thunder
#

thanks maxfield

#

it could just be a troll yeah
I find troll less scary than malware bot

atomic mortar
#

last week we had one that basically insulted all nine generations of mine so it doesnt really phase me anymore

fervent thunder
#

yeah you get used to it

#

modern internet is kinda more tame than early internet anyway

#

it used to be a jungle

atomic mortar
#

these kids wouldent survive a mw2 lobby with their current mindset

fervent thunder
#

I meant more like 1990s rather than 2009
but yeah even 2009 internet was different to now

balmy inlet
#

hi

lofty osprey
split kestrel
#

intresting

dense token
split kestrel
#

i havnt played with it

sour iris
#

Hi gang, we are organising the best ever multi-modal creative AI HackXelerator with streams for AI music, image, video, fashion, gaming and combinations. The event can be joined online and/or IRL in London, Paris and Berlin. We are supported by the best brands in the business; Central Saint Martins, Station F, Mistral AI, Hugging Face, Luma AI, Vultr, AMD, Nvidia etc.

This event will bring together 100s of creatives and techies to create awesome experiences https://www.kxsb.org/lpb25

frank glen
#

Hello everyone!!!
I hope you are doing well!!!
@w361_emp is SCAMMER.
He stole my portfolio.
Be careful!!!

spiral bone
#

Is this an option to use loras on certain areas of image? Like, I need to use orc lora only only on mouth part, but all image should be affected by another lora

karmic swallow
atomic mortar
#

But during generation? Depends on the UI i suppose

still glacier
fervent thunder
#

they added regional lora to comfy at some point

dense token
#

Sorry again @vapid dove but we have another one -_-

naive charm
#

hello

eager mauve
#

do yall know what would be some of the best controlnet models for referencing and making consistent characters? SDXL btw

fervent thunder
#

reference unet if you can find it

junior shuttle
#

Any LLMs out there that references and updates real time data? I am looking into Gemini

livid rose
#

A bird trapped in a cage, with the door open but the bird refusing to fly, symbolic of emotional imprisonment, surreal and artistic."

ruby parcel
#

hello folks

#

is there something that is along the lines of SDXL kinda resource requirement, but also with T5 clip?

#

had a taste of T5 prompt adherence in SD3.5

#

it's... addictive

#

is there like a SD1.5 or SDXL that works with T5?

west bloom
#

Hello, how can i use different model with auto1111 text-to-img endpoint?

atomic mortar
atomic mortar
tulip yarrow
#

Any reliable way to train flux on colab?

tulip yarrow
#

What's a good civitAI bounty for a 54 character-pack in SD1.5

#

Yeah, I specifically want SD1.5 by the way (how much buzz)

#

I know the others are better but I really like using SD1.5 still.

severe arrow
#

lmao the 'no robots or apocalyptic themes' part.

vapid dove
abstract kiln
#

I don't get it, what's with all the hate?

severe arrow
#

you know that's just going to be pure Streisand effect

#

tell people not to do that, for so silly a reason, and people are going to be only doing that

vernal ore
#

Another competition for US people exclusively

vapid dove
vernal ore
#

It's ai artwork that's featured on a site, prompted by someone from somewhere

#

Could you elaborate why it's illegal outside the US

iron pendant
#

not accepting outside US doesn't mean it's "illegal"

#

probably just something to do with contract enforcement

vapid dove
severe arrow
#

cash prizes are just fine in other places, but there may be different taxation requirements and the like.

vernal ore
#

Weird that. But I'm also not a lawyer, anyway goodluck with the competition

severe arrow
#

That means the cash prize may have different sizes depending on who wins where.

#

but it's not actually illegal or anything

iron pendant
#

things can get weird across boarders with jurisdiction, different laws, etc

severe arrow
#

I mean, Fortnite manages with their cash cups lmao

#

The EU cash cup is actually bigger than the US one

iron pendant
#

fortnite generates billions of dollars per year, its a massive business

severe arrow
#

Sure but they're beholden to the same laws

#

especially since both revolve around a competition

#

not a sweepstakes.

iron pendant
#

lawyers don't work "for exposure" 😉

#

or at least, rarely

severe arrow
#

Not for competition law no

#

Nevertheless if Linus Tech Tips can do it I'd imagine SD can do it

#

Those are businesses that are presumably in the same ballpark of revenue

iron pendant
#

every country they open to probably costs more for legal analysis is my point

severe arrow
#

so open up only to the big players

#

US, European Union (single market, laws are harmonized for this sort of stuff), China, and whomever else can be expected to add a pretty sizable number of competitors

#

I don't think most people would be bothered by the lack of Togonese participiation

iron pendant
#

EU isnt as harmonized as you might think

severe arrow
#

for stuff like sweepstakes and competitions it sure is

#

it falls squarely in the commercial purview that the EU was founded upon

iron pendant
#

I've had to deal with at least soem of this myself, I have had a few clients in different EU nations, different nations have different laws

late osprey
#

2025-02-28 02:24:01,984 - Inpaint Anything - ERROR - shape mismatch: value tensor of shape [159, 256] cannot be broadcast to indexing result of shape [64, 256]

#

does anyone know how to fix this error

late osprey
bright breach
#

Hi, i have a question.
I want to use stable diffusion directml to color my drawings. When i use it, it usually messes up the hands and changes the art style completely. Does anyone know a way to use img2img to only color line art or add shadings to base color? Without messing up the hands or changing the art style and eyes and such.

dense token
bright breach
manic wraith
#

AI has revolutionized the "for exposure" line of work
artists used to require manual labor and creative skill for exposure
now you can just prompt

atomic mortar
manic wraith
atomic mortar
#

But i also do a remote install for them to make it work

dense token
atomic mortar
#

And knowledge, not the wormflow perse

dense token
manic wraith
atomic mortar
#

Things like automated Photoshop, custom nodes, custom workflows now thats business

manic wraith
#

you'd be surprised at how many creative ways people can screw it up

atomic mortar
manic wraith
#

"it's very slow"
"I picked CPU, is that a problem?"

atomic mortar
manic wraith
#

ah yes hahaha

#

not to mention ControlNets, LoRAs, all the nitty gritty

#

like a friend... was putting... hires fix to 1.1x... at 100 steps......

atomic mortar
#

Yeap thats i useually shy away from techsupport now and useually hit up businesses asking for something

#

Might eventually steer towards making loras more

manic wraith
#

it's kinda fun when you get it right

#

and very gatekept by the fact Kohya_SS makes the other SS look like the friendly guys

#

it's <input> porn

atomic mortar
#

I don't mind guiding new people making their first image but after that i kick them out of the nest

manic wraith
#

good idea tbh
I might be too nice sometimes crine

upper anchor
lofty osprey
#

hello everyone there

floral zealot
rigid warren
#

is animatediff currently working on forge ? I read sometime ago that it wasn't

dense token
manic wraith
#

well have you tried making sense of that thing?? I had to go AdamW with the Cosine with Restarts because anything slightly more complicated than the basics and it just came out wrong

floral zealot
#

do the stable works on 5070ti_

#

?

delicate oxide
#

hello stable people

dusty zealot
#

What's better forget ir sd.next

brave relic
#

hey guys, where do you upload your gens to show off? it sounds like its just deviantart right now?

obtuse marten
#

Are there any good guides to get me going on my journey to making actually good stuff instead of producing garbage?

#

So many guides idk which to believe or what is still relevant.

dense token
sour iris
#

Hey guys, would anyone be interested in joining a multi-modal AI hackathon? You can join online via Discord or IRL (London, Paris, Berlin) #music #art #video #gaming #fashion https://www.kxsb.org/lpb25

atomic mortar
atomic mortar
dense token
atomic mortar
fervent thunder
# 
🚫 IP-infringing content

🚫 Robots or apocalyptic themes (We’re steering away from the “robots take over the world” narrative)```LOL I think I am out of the contest cos I only make R2D2 images
#

but I am in UK anyway so I was not eligible

#

its a good point about robots I guess, didn't think of that

fervent thunder
#

similar words come up the same

atomic mortar
#

The only robots i make are certainly not in a position to take over the world smirk

fervent thunder
#

for some reason R2 was basically used to deliver maps/plans every time

#

they did it again in the recent star wars, IDK why lol

#

he's basically a delivery box

atomic mortar
#

imo its because they dont really innovate

#

since the rogue one movie i basically never touched starwars

fervent thunder
#

yeah I didn't like episode 7 8 9 era
I agree rogue one was great

#

some of the visuals in 7 8 9 are good though like the salty planet

#

its a bit like with Avatar Way ofthe Water, IDK if it was the best writing but the water looked rly nice

atomic mortar
#

for me episode 7 8 9 simply doesnt exist

fervent thunder
#

I think episode 7 was totally fine

#

they essentially re-did episode 4 plot

boreal prairie
ember niche
#

hello everyone there!

dense token
digital briar
#

yoo does anyone have experience training Lora's for SD? I tried using Flux Ostris on Replicate but it didn't work very well, i'd appreciate some help!

Trying to generate really basic essentially MS Paint drawings. Here's an example: #🏞｜general-with-images message

atomic mortar
#

though you said two very different things. SD is not Flux lol

#

like i tried to make applie pie with pears kek

digital briar
digital briar
atomic mortar
#

hmm well with that style i am not entirely sure but i manged to make a style lora with 64 images

#

but theres still a bit of bias due its smaller data set

digital briar
#

Do you think it's possible to achieve an output like the one I posted? I don't know much about this stuff

atomic mortar
#

Hmmmmm i honestly am not sure since i tried that image but the auto tagger didnt recognise most of it

#

you could try to induvidually label each image but again i havent done a image so minimalistic

desert dagger
desert dagger
slender rain
#

Greeting, I'm kinda in middle of complicated decision whether I should buy a normal pc or try out the recommendation build for stable diffusions. Do you guys have like recommended build for this, I'll let ya'll decide

dense token
slender rain
#

Well yeah it's expensive for one piece

slow summit
#

Can anyone help me figure out what model to download? I have never messed around with AI art but want to make 2d pixel art for a game I'm making.

slender rain
#

Ohh you mean Lora?

#

Or the checkpoints

slow summit
#

Im not too sure what im doing. Do I have to download a specific model of stable diffusion first and then get a lora from civitai? Or can I download straight from civitai?

dense token
slow summit
#

Ok. I have a 3070ti in my pc. Also, when I find the checkpoint and download it, how do I use it?

dense token
slow summit
#

No I dont

dense token
slow summit
#

Ok, will do. Thank you!

silk lake
#

hello

astral lantern
#

/image_dream prompt: genarate a image of Indian man who dress pants and shirt barrowing money from another man who dress dhothi and topi

tepid finch
#

Gm guys

#

I was wondering if anyone had a workflow for creating consistent character with different poses. I also want to integrate my logo in the image like on his shirts. Im using comfyui

digital briar
pure mango
#

hey guys im tryna steal an artists art style and train my own data with their artwork

#

is there a guide to do exactly that?

desert dagger
desert dagger
fervent thunder
#

a method came out in the last month
to extract training images from loras

#

I'm trying to spread the word a bit LOL cos it means they can see what you trained on

pure mango
#

i said it like that to be clear i have no intention of stealing like that i just want to see how it would come out ill draw my own

fervent thunder
#

I honestly have no idea about the current legal setup regarding this stuff anyway

#

its probably changing fast

tepid finch
storm anchor
#

Hello from China

dense token
atomic mortar
# tepid finch Anyone?

Hi, its a pretty regulairly asked question but mostly it boils down to lora, faceswapping (reactor) and controlnet

#

i think if you search for consistent character in search you find better explained awnsers

atomic mortar
fervent thunder
#

was on arxiv in the last week

dense token
fervent thunder
#

date might have been earlier then maybe

tepid finch
#

But possible yeah?

atomic mortar
#

Sort of yeah

sonic ivy
#

hi, havent posted here in forever, is it okay to ask here for a quick job offer? i'm sure i could technicaly do it but i'm lazy to look for the prompt and stuff, havent updated my old SD in forever too.

atomic mortar
#

atleast i think you can

woven shale
#

Hello, I was wondering if something like game ui can be made with STable Diffusion? With a prompt will it be accurate enough to recreate something and give different forms? (Let's say I want the same exact style image as this, but instead of a tree, make it a car)?

glass cloud
#

Stable update for 50 series gpus?

atomic mortar
glass cloud
#

Additional info?

atomic mortar
#

Check general w images

#

If you use swarm i suppose

glass cloud
#

Thank you friend! I truly appreciate it!

late osprey
#

so, which ip adapter for face copy should i choose

#

oh, nevermind, i think reference only do the job better

atomic mortar
late osprey
#

and i love all mighty controlnet

late osprey
#

does reforge support amdgpu, saw that on stability matrix that it has support but don't say anything about that in github page

warm junco
atomic mortar
#

Whenever i hear stability matrix i get a ick

#

So many bugs due matrix, i recommend sticking to a UI outside matrix

dense token
late osprey
warm junco
late osprey
warm junco
late osprey
#

hmm, let me test that for you

warm junco
#

Which webui do you use currently?

late osprey
#

automatic1111

warm junco
#

With directml or zluda?

late osprey
#

zluda

warm junco
#

Ok

#

I will ask Lshqtiger if he can look into forking reforge for AMD like he did for Auto and forge

#

Shouldnt be to difficult

late osprey
#

wait what, he can?

warm junco
#

He is skilled enough to do so yes

#

But idk if he has time and motivation

late osprey
#

hmm, lemme try reforge with directml for a while then

warm junco
#

Let me know if it works

late osprey
warm junco
odd fossil
late osprey
rustic sapphire
#

can you make animation

dark pollen
#

guys anyone kind to help me a bit please?

atomic mortar
#

or if you have a question dont be afraid to ask

woven shale
#

Hello, can I do requests somewhere, that someone can generate an image for me? Or like an external website

atomic mortar
rigid widget
#

How do I create images ?

gray bane
#

hello all

atomic mortar
#

If you have a gpu with around 6-8gb vram i recommend looking into the #🤝｜tech-support pinned messages for the guides from CS1o

late osprey
#

Hey, weird question, i intend to run automatic1111 in wsl, and wonder if it has any differences from running on linux native?

sharp peak
soft pumice
#

guys how does pykaso AI website provide so realistic images...How can I do the same locally on my pc?

ashen sleet
#

ys

plucky cape
#

!generate "a beautiful sunset over the ocean"

nova beacon
#

free?

chilly storm
atomic mortar
shadow eagle
#

Hi!

fervent thunder
sudden schooner
#

are there any prompt generators for Flux? Which will edit your prompt to make it better

deep narwhal
#

anyone using an amd card on windows rn? is it still as hard to use as im reading from year old threads?

still glacier
warm junco
true saffron
#

Hello New to Stable Diffusion and have no clue what I'm doing. Wish me luck! 🐈‍⬛

exotic delta
#

good luck!

faint sparrow
summer cove
#

hello

rigid warren
#

hello, i'm trying to use regional prompter for the first time, and even if I was able to generate the image I want, the resolution is pretty bad, everything looks blurred, any idea?

inland cloud
#

Hola a todos!!

river roost
#

hello

still hare
#

HI So quick question i have a 4080 super. i wanna try out stable diffusion and someone gave me a link to huggingface and i see a bunch of options but i dont know what to try

fleet kayak
wraith oak
#

I have a mental disability that effects my memory of things after my brain injury... I have been working toward running SD locally for 2 years... I found some cheap(ish) hardware that will run it, and am eager to get started. I have experience with Midjourney and Dalle but I'm way behind now with so many different choices to make. Is there anyone here that could help me learn which things I must choose to get the functionalities I'm looking for...

I have been trying to do it myself but my poor memory and the speed with which this is all developing, and my limited resources' off-optimal hardware, means when I watch a tutorial on how to do something, by the time I adjust to it the information isn't current or accurate. If someone who knows about running SD locally on an AMD APU (5700G) running Ubuntu and can patiently guide me toward the functionality i need I would be willing to discuss compensation (not rich or anything but, this is a priority for me.) DMs open to anyone interested.

desert dagger
atomic mortar
#

Id do it for free but currently omw to work

soft pumice
#

how does pykaso.ai create realistic images every single time?

#

it even has 'amateur shot from iphone' stuff

#

how can I replicate the same locally?

wraith oak
jolly mortar
#

Hello

glacial hinge
#

Hi all , can someone tell me how can i delete Stable Matrix??

twilit forge
#

question, do I need to download the base model that a certain checkpoint or lora uses to efficiently use it?

#

for example SD 1.5

warm junco
regal sentinel
#

Hi I'm using SDXL as I have 8gb vram, I need a good method of fixing hands automatically without inpainting, I'm on Comfy Desktop V1 (windows) and I can't find Mesh Graphormer in Comfy Manager. What are my options and what is the best method of fixings hands without the need of inpainting? Also is there a good hand LoRA for SDXL?

faint timber
#

There is a mechanism called "embeddings." They are basically per-canned negative or positives you can add to the prompt. The face detailer can also fix hands when properly configured. Another route might be to add an OpenPose control net.

twilit forge
#

what upscalers / hiresfix do you guys use for photorealistic images?

bright ember
#

hello guys i would like to create videos from 1 photo for free. Do you know anything about this?

abstract quarry
#

search for Wan 2.1 i2v model

#

you need a really good gpu and a lot of patience, though

atomic mortar
#

You could also use a online service that gives you some free credits

#

Results vary though

bright ember
#

unfortunately I always find everything paid XD

abstract quarry
#

doing it locally is also not free - you have to pay for electricity

atomic mortar
#

Though im sitting at a 4 person household average consumption as a single person lmao

abstract quarry
#

if your gpu runs 20 minutes to generate 5 seconds of a video then, based on your 20 cents/hr, you would already reach 7 cents per 5s video.

#

if you assume 40 cent per video with SORA, then this is only 6 times as expensive as your local video generation

atomic mortar
#

Benefit of local: uncensored

abstract quarry
#

oh, yeah, sure, if you want to generate 5 seconds porn then you can use local models

#

I find it kind of pathetic that people don't do anything more creative with that technology than boobs

atomic mortar
#

Eh its not only that but sure*

abstract quarry
#

look at civitai when you select video models as filter. Its basically 99% porn

#

btw. I also don't say anything against local models. I love doing everything locally. I just say: video generation is not as easy and mature as image generation and for novices it might be the easier and maybe even cheaper solution to just use SORA and similar tools

atomic mortar
#

Oh yeah for novices / common rabble definitely i agree

#

But the moment people try to do more then the set guidelines then from those corpo smucks decide your stuck with no way to edit it the way you want it to

#

Without paying even more offcourse

fervent thunder
#

I mean you can just rent GPU servers from reputable datacenters and then they don't monitor at all

polar spruce
#

hello

steel fable
#

How to make short videos? There is any video on YT to learn about plase?

abstract quarry
steel fable
#

Thanks

steel fable
abstract quarry
#

comfyui or swarmui

peak musk
#

@vapid dove any idea when well get a stable diffusion v4 its been over 4 months and 3.5 wasnt that good even at launch

vapid dove
#

Kinda wish I'd started using it sooner

peak musk
fervent thunder
#

to be fair those are like the two biggest closed models

vapid dove
fervent thunder
#

apart from recraft which slides in between the two, in terms of size

atomic mortar
#

I mean considering it takes a lot of money to make something like this and we're getting it for free eh. Compared to early sd 1.5 they have come a long way

fervent thunder
#

ye I've started training models and its rly hard

#

if you put in a reasonable budget what you get out is like
a 256x256 imagenet model lol

fervent thunder
#

its by TensorArt

vague relic
#

What are you talking about?

solemn acorn
#

/generate

#

现代都市公园，阳光柔和，绿地和树木，背景有摩天大楼。一个白人男性（30岁，浅棕色短发，浅蓝色衬衫，灰色休闲裤，微笑）和一个黑人男性（30岁，黑色卷发，深绿色针织衫，卡其色长裤，开朗笑容）站在桥上，影子交融，背景有鸽子和国际象棋棋盘，插画风格，低饱和色调。

dapper grove
#

d

tiny aspen
#

cup tea

ruby parcel
#

Hei folks . Want to ask are there SDXL models that use vpred that also uses the normal natural language prompting?

#

Or maybe are there ways to make regular sdxl to be better at bright/dark scenes?

abstract quarry
#

just do img2img on a dark or bright image

#

it's the most simple way to fix that issue. I never understood all these weird workarounds like noise offset and so on

charred lotus
#

Hello

finite cloak
ruby parcel
finite cloak
finite cloak
ruby parcel
#

OHHHH

#

you mean taking a full grey/white/black image

finite cloak
# vapid dove I wasn't super impressed by 3.5 when it launched either. To be honest I didn't e...

It's weird getting past the community opinion blinders to see the underlying model for what it is, eh?

Back when SD3 was a baby, even the smallest pretrain version blew my mind with its capabilities (as compared to xl or any other preexisting model).
Was quite frustrating that licensing mess and communications confusion ruined such an incredible model release.
Check out that TensorArt 3.5L-TurboX model btw, ~8 steps and even more amazing than the base, on par or better vs Flux Dev but runs way faster thanks to the turbo.

sudden schooner
#

trying flux with ComfyUI. It starts generating and then says 'reconnecting'. I try to start again and it says 'failed to fetch'

Is it an issue with my RAM?

#

did you solve this?

sudden schooner
finite cloak
# sudden schooner ok i found out this was a memory issue. any suggestons on how to run flux with l...

GGUF models can work a fair bit lower than the normal models, you'll probably want schnell too if you're resource constrained -- links and relevant usage docs for Flux here: https://github.com/mcmonkeyprojects/SwarmUI/blob/master/docs/Model Support.md#black-forest-labs-flux1-models
with limited resources you'll probably want to consider a different model class though, eg SDXL finetunes or the new SD3.5 TurboX

twilit forge
#

guys, how can I inpaint text onto paper on an image realistically?

jaunty hinge
fervent thunder
#

SD3 and SD3.5 are perfectly fine for the way most people use image models
like if you are just making a 1024x1024 image they can quite often give out a more interesting image than Flux, it will at least be more colourful and creative
their issues were more on the technical side with the pos embeds and attn maps not working as well as they should, but this mostly affects inverse-problem stuff like using the model for inpainting, super-resolution, segmentation etc
for just generating images at 1024x1024 both SD3 and SD3.5 are really under-rated at this point
if the image is painterly or has smoke/fog effects they are particularly good versus the others

desert dagger
fervent thunder
#

there are stronger methods on Arxiv but none released model/code

#

for example they use control nets or new SDE samplers

olive quest
#

Yoooo Guys can anyone of you help me pleaseeeeeeeeeeeeeeeee
my stable diffusion is using only the cpu instead of GPU taking lot of time to generate images

#

if anyone wants to help me please DM me

fervent thunder
#

could we talk here instead?

olive quest
#

yeah sure

fervent thunder
#

I don't use DMs cos in the past people acted weird in DMs
what software is this

#

is this A1111 software?

olive quest
#

yes

fervent thunder
#

as first step could I maybe convince you to change software

olive quest
#

i have 3070 ti but still image generation is very slow due to it is using my cpu

olive quest
#

what should i use

fervent thunder
#

your graphics card is strong, I rent 3070 ti sometimes

#

probably SwarmUI

#

it will have a way to make sure the GPU gets used

olive quest
#

hey can we get on a quick call please, if you thing i am weird you can disconnect & block me

fervent thunder
#

sorry I can't do DM or calls
too many bad experiences helping people over DM or call

olive quest
#

please pleaes from morning onwards it's bugging me

fervent thunder
#

and see what happens

olive quest
#

okie

fervent thunder
#

as a first step

olive quest
#

i'm following the steps

#

do you have like any good youtube video

olive quest
sudden schooner
#

neg prompting 'face' but faces are still appearing in the immages

#

any suggestions?

olive quest
#

hey OG

sudden schooner
#

hey

olive quest
#

can you come to VC

fervent thunder
#

try delaying the negative prompt for some steps

olive quest
#

VC in diffuser lounge

#

Hey Neon do you know any good image to video generaates free online ones or local installed

#

@fervent thunder

fervent thunder
#

yeah Wan video is good

#

on some level Stepfun video could be ran locally but not really

olive quest
fervent thunder
#

IDK just google "Wan video huggingface"

olive quest
#

okie thanks

#

@fervent thunder
swarmui installed

#

how to make it run on GPU

#

image generation is slow this swarmui also running on CPU instead of GPU

fervent thunder
#

ok so first get a copy of something small

#

like SD 1.5

olive quest
#

have it

fervent thunder
#

then look for command line arguments that will make sure the model runs on GPU

olive quest
#

v1-5-pruned-emaonly.safetensors downloaded and placed in models file

fervent thunder
#

yeah v1-5-pruned-emaonly.safetensors should be ok

olive quest
#

what is the file name that i need to edit

#

i'm new to this man, sorry for asking so many questions

fervent thunder
#

the problem is I use comfy but I advised you to install swarm 🤔
in comfy it would look like this:

vram_group.add_argument("--gpu-only", action="store_true", help="Store and run everything (text encoders/CLIP models, etc... on the GPU).")
vram_group.add_argument("--highvram", action="store_true", help="By default models will be unloaded to CPU memory after being used. This option keeps them in GPU memory.")
vram_group.add_argument("--normalvram", action="store_true", help="Used to force normal vram use if lowvram gets automatically enabled.")
vram_group.add_argument("--lowvram", action="store_true", help="Split the unet in parts to use less vram.")
vram_group.add_argument("--novram", action="store_true", help="When lowvram isn't enough.")
vram_group.add_argument("--cpu", action="store_true", help="To use the CPU for everything (slow).")```
in this file ```https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/cli_args.py```
olive quest
#

checking

fervent thunder
#

this will also help work out if you have an issue of multiple CUDA devices

olive quest
#

in swarm also we have comfy

fervent thunder
#

bare in mind these nodes have two issues

  1. they monkey patch
  2. they interfere with the memory management
olive quest
#

bro could help me out quickly i'm in the VC on this server

olive quest
fervent thunder
#

issues like this are tricky no matter how much experience someone has lol
cos with local generation you never know what the prior setup of the machine is

olive quest
#

thanks for the help buddy, i'm gonna go to sleep and deal with this mess in the morning

fervent thunder
#

okay sure

olive quest
#

goodnight

fervent thunder
#

night

crystal cape
#

маке зшсегку

#

make picture

chrome herald
#

Man, apparently nobody wants to see any kind of straightforward image comparison between any set of t2i models anymore on the Stable Diffusion subreddit lol
I used to do them every so often and they were all fairly well received usually, but all recent ones I've done have just been downvoted to 0 with very few comment replies (but thousands of views)

pine path
#

The research channel was deleted?

unborn hedge
#

i wanna make a lora of my pet cat using flux, is flux like good for that, idk about animals specifically

atomic mortar
#

unless you wanna do crazy stuff

#

but xl is pretty neat for it too

unborn hedge
#

yeah, i wanna like make my pet cat fight pirates or something

#

or surf on lava

atomic mortar
#

hmm well i dont make character loras often so im not sure

abstract quarry
#

Flux is so much better than SDXL. It always shocks me if I compare my own face lora in SDXL (which I found great at that time) with that Flux is capable of

#

However, the big disadvantage of Flux: it trains extremely slowly

#

it's not a cfg model, so things need much more time to get cooked in

unborn hedge
#

im probably using civitai to train, does that mean i need to mess with some settings?

abstract quarry
#

I'm not sure if civitai is a good place for that. I haven't tried it

#

but if you pay money anyways I would rather train somewhere else

#

e.g. on the Flux Pro page itself (there you can directly train the Flux Pro model which gives you probably the best results)

#

but there are so many other services that offer training as a service and usually they have good predefined settings

unborn hedge
#

soo what should my settings be then to get best results (still using civitai)

iron river
#

Wan2.1 needs its own channel

fervent thunder
#

Flux Pro doesn't rly look that different to me

#

Flux Ultra does but not the regular non-Ultra Pro

fast sage
#

Hey lads, whats the current best img to video that I can frun locally?

dreamy sorrel
#

Hey everyone, not sure if this is the right place to ask, but I’ve got a few questions:

  1. Is there a way to get Flux 1.1 Pro quality for print-on-demand designs without using the API? Maybe some local setup?
  2. Would a Mac M1 with 8GB RAM be able to handle it with some crazy hacks?
  3. Could you finetune models to learn faces and match Flux 1.1 Pro output locally? Any pointers?

Thanks so much

fast sage
willow lark
#

Hello, is this server strictly for stable diffusion discussion or also allows related tools discussion ,such as ComfyUI?

fervent thunder
#

Finally, I can chat

#

Had to wait 10 minutes

dreamy sorrel
dreamy sorrel
fast sage
#

afaik flux dev requires 8gb and 32 to train, if I remember correctly,.

fast sage
unborn hedge
#

it worked! my cat is fighting a pirate lol

chrome herald
chrome herald
# dreamy sorrel Like, there are websites that generate public images from Flux1.1 pro (like Glif...

I've done what I think you mean (train a Lora on outputs from a different model) at least twice
https://civitai.com/models/633453/zoot-fluxxer-xl
https://civitai.com/models/1204546/zoots-flux-pro-ultrafier-for-kolors
the dataset was higher quality / higher resolution for the Kolors one cause they were ALL Pro Ultra outputs
whereas the SDXL one is mostly Dev outputs
also Kolors just does a lot more with the natural language captions I used

dreamy sorrel
unborn hedge
#

i think this one worked because its a cat and flux probably has plenty of data about that in their model

chrome herald
dreamy sorrel
bitter ice
#

hello

empty pond
#

Is there an api specific channel?

fervent thunder
grand merlin
#

哈喽 你们好

chrome herald
#

stock stuff

compact swan
grand merlin
#

This looks a bit scary

fervent thunder
#

although almost all the big models are bad for photos in their stock form

#

for some reason photography seems to need models to be pushed towards it a fair bit

autumn venture
#

what's up guys

twilit forge
#

any ideas on how to make a less realistic selfie? i don't want professional camera quality and insane lighting

#

it seems SD struggles with this

atomic mortar
#

It really depends on the checkpoint bgl

#

Ngl*

abstract quarry
#

usually you can also prompt for that using keywords like "2006 digital camera image", "amateur photo" and so on

twilit forge
#

thanks

regal sentinel
#

Out of interest who's rushing out for a RTX 5090? I'm saying that as I need a stronger GPU as I have a 4060 and everything is just crazy at the moment.

atomic mortar
#

As 6k for a 5090 is insane

regal sentinel
#

@atomic mortar I would love a 4090 but there stupid prices, but as a former system builder I don't trust the 2nd hand market. So the only option for me is a new card with a warranty if it starts smoking. I'm not liking the extra power draw and Nvidia has just F'D up in this area without proper power design on the card. But new 4090's with a warranty (warranty not transferable to 2nd user) like rocking horse Sh!t to get hold of without being scalped. I only can afford a FE 5090, and they're just badly designed at the moment. I hope AMD get in the game for AI or Intel for that matter just to give competition and the pricing down. But it's clear where Nvidia are for the money and that's commercial GPU's for GPU farms and its feels like the big 2 fingers for people who want local GPU power.

atomic mortar
#

Woops i read it as 4090 but you said 60

#

Mb

#

I settled on a 5080 and while a bit slower with video ai it manages

#

I can always rent a gpu if i need a heavy workload

regal sentinel
#

The 5080 has a lot better power balance, as it has less draw, but the power design is ridicules

atomic mortar
#

Oh yah definitely, when i first got it i ran a few Wan gens in a row like 3 hrs

#

And checked for any plastic melt smell

regal sentinel
#

@atomic mortar I'm looking at the A series with 20gb, but am I gonna miss those missing 4GB

atomic mortar
#

Hmm it depends really

#

I don't know what your gonna use it all for

regal sentinel
#

@atomic mortar I'm just learning at the moment and will end up doing video at some point, which really lends itself to GPU farms really. But the 24GB does seam like a good trade off from 32gb. But of you're really doing stuff you really need a 60gb + card which takes you out of consumer, as the better models won't fit. Crazy really, the move is heavily focused on render farms and less on some local GPU, but hay that's where the money is

atomic mortar
#

Text to video like 9

#

16gb vram

#

Unless you go for the extreme llm's you dont need that much

regal sentinel
#

@atomic mortar You know the crazy thing with the 5080 pricing just for a few hundred more you might as well pony up for the 5090, and that's like crazy am I that stupid to buy a poorly designed card. I'm left gobsmacked, but the 5080 is the better option power wise unless you like roasting marshmallows. The only option really is a 3rd party device to do power smoothing for the 5090 at the cost of performance.

atomic mortar
#

Non FE

regal sentinel
#

@atomic mortar where I am the 5080 is around $1500 depending on GPU type non FE, the FE (5090) for me looks like the best option price wise as it what I can afford, but on the risk of using the card out weights that option, that's when there's stock. It's easier for me to find the extra $500 than it is to buy a lower grade card and upgrade later. But there's also logic to buy what you need, as these cards will sell 2nd hand. But the A series look attractive and i'm wondering what will the Blackwell offering be when they release these Pro cards and is it worth the wait. Draw back is less cores to work with, but you then have the VRAM for the models to fit, swings and round abouts.

atomic mortar
#

But realistically, what problem do you have with your current 4060?

regal sentinel
#

Models won't fit + future proof need and 16gb now adays isn't good and the VRAM will always increase

atomic mortar
#

I was on a 3070ti and waited till i could score one at msrp

#

If you can wait id wait as sdxl is still pretty dope

regal sentinel
#

@atomic mortar I'm thinking that I'll end up with a 5080 unless I can get a 4090 that I can trust 2nd hand even better new, and the A series is an option too

lavish current
#

I wish people who don't know about AI wouldn't talk like they know about it

#

Ppl acting like Loras haven't existed forever

#

(Not on this server)

#

I don't even use SD anymore but loras are like, super basic

chrome herald
chrome herald
#

For Medium vs Flux with and without Loras (including a Lora made myself)

chrome herald
sullen thicket
#

where do i go for help with forge

desert dagger
atomic mortar
#

hmm

#

i mean more cuda yeah

desert dagger
#

you are assuming and you have no idea how the 50 series is even built

atomic mortar
#

my comment was based on the fact that he has a 4090 but i misread

desert dagger
#

yes, but - the 50 series is not the old tech that the 40 series was built on

#

it's totally different, it has AI on the chip, it generates a FEW pixels and then the AI takes over and predicts the rest

#

you can't base any idea of what the 50 series does on previous nvidia tech

#

go to youtube, search on CES 2025 jenson and watch Jenson's keynote

desert dagger
sullen thicket
#

nvm i fixed it

spare flicker
#

Hello

fervent thunder
#

I don't mean that as criticism, what I mean is that some of the big flux checkpoints and popular loras are not as good for photo as some of the less popular combinations

#

maybe try this UltraReal Fine-Tune with Sony Alpha lora, Boring Reality lora, iphone photo lora

#

for SD3.5m I have seen some good ones in training discords that I am not sure have been released, but there are some that brought the photo style out rly well

chrome herald
# fervent thunder I don't mean that as criticism, what I mean is that some of the big flux checkpo...

Right, again I just meant like e.g. stock SD 3.5 Medium (and Large presumably) already has a solid fundamental understanding of terms like "editorial", "amateur", "candid", "analog" and such. It's also reliably possible to get rid of depth of field using negatives in them by default. None of that is really the case in Flux without at least loras, or jumping through more hoops to approximate the negative stuff.

fervent thunder
#

Flux Dev is more overfit

#

when it first came out I was worried

chrome herald
#

like If I made a general photo lora for SD 3.5 the dataset would be completely different from the Flux / Kolors ones I've made
Cause it just doesn't really need training on the same things
Doesn't need de-chin, or de-bokeh, and so on

fervent thunder
#

what does de-chin entail lol

#

I think for the most part you train towards what you need rather than away from what you don't need

#

so in that sense I would approach both models the same

#

and train towards photos

willow needle
#

Quick question, what method do y'all use for line breaks in your prompts? Any hotkey for that I know of includes enter, which no matter the key combination, begins image generation.

crystal edge
#

what is the latest stablediffusion image generator version?

fervent thunder
#

3.5

crystal edge
#

is it the best ?

fervent thunder
#

it doesn't have to be the best, if you like it you can make nice images out of it

crystal edge
#

it called sdxl 3.5?

fervent thunder
#

just SD 3.5

#

and there are two

#

SD 3.5L and SD 3.5M

crystal edge
#

what do u prefer sdxl 1.0 or sd 3.5

fervent thunder
#

oh SD 3.5 is definitely stronger than SDXL yeah

chrome herald
# fervent thunder what does de-chin entail lol

just throwing a whole lot of super high res professional photos where a face is very prominent at it, basically
with my Flux one I was basically just trying to brute force Flux in general into looking "not distilled" or "normal" by default
so it was basically just a super diverse dataset of people of all ages, with no single person ever appearing more than once

#

whereas SD 3.5 really doesn't need that kind of Lora at all, moreso just like, targeted improvements in specific areas kind of thing

fervent thunder
#

I see what you mean yeah

#

I would go about it a bit differently, and just run through general photo datasets

#

like not curate them beyond quality control

chrome herald
# fervent thunder like not curate them beyond quality control

well I was only concerned about people for that one, not so much environments which seemed "fine enough" to me
I did one specifically for animals too that I never got around to releasing yet, with like accurate captioning for different dog breeds and quite a variety of other species represented also

fervent thunder
#

yeah there is a problem with my method that it is not suitable for smaller lora/lokr/loha

#

it doesn't have to be full rank finetune cos it can work with a large lokr but it has to be a reasonably big dataset if the dataset is generalist like I was describing

#

an animal one sounds cool 🤔

chrome herald
#

and caption it well

fervent thunder
#

I actually don't know settings well for this stuff

#

I skipped lora/checkpoint training and I've started looking at training new (very small) foundation models instead

#

its kinda backwards to the order most people learn lol

chrome herald
fervent thunder
#

you can train lightningDiT for $80 its less than it would cost me to make a flux checkpoint probably

#

although it might go wrong I guess we will see

abstract quarry
#

I don't know. I tried to make a pen&paper character with Flux and SD 3.5L and was surprised that Flux had much more variety and diversity in its generation. All SD 3.5 characters looked the same. Also, SD 3.5 had big issues to not make him too handsome.
Could also be a skill issue as I'm not so familiar with Sd 3.5 as with Flux

#

negative prompts work for flux if you use cfg

#

you just have to use lower cfg values and probably should skip one step at the beginning

chrome herald
#

I do remember that Large was much more finnicky about CFG when I tried it, and had a tendency to do weird things with the background
and also the FP8 Scaled version was REALLY noticeably worse than the Q8 GGUF for many outputs
so those could have been possibly relevant things for you

fervent thunder
#

in my testing Flux had a very high variety across seeds, with the right workflow
as well as that, there are ways to estimate log-likelihoods of samples now so
stuff like diversity which used to be qualitative can now be measured

fervent thunder
#

SD 3.5M does have depth+canny though (from tensor.art)

#

I would maybe look at merging the new SD 3.5M turbo with SD 3.5M turbo base

desert dagger
fervent thunder
# desert dagger why?

turbo is one of the distill methods that has an adversarial part
so it can raise image quality
doing the merge lets you choose the tradeoff level per block essentially

cosmic scarab
#

I've been using --xformers instead of --opt-sdp-no-mem-attention, and going back to sdp improved quality significantly with sdxl

chrome herald
still rampart
#

Hello insights needed

#

Does conceptual captions have significant chunk of important representation encapsulated for a generalized diffusion model for pre training?

#

Or is just too erratic high frequency data

#

I was thinking of using synthetic sets along with sa1b

fervent thunder
fervent thunder
#

what is the sa1b for?

#

sa1b is uncaptioned I think

still rampart
still rampart
fervent thunder
#

journeyDB is common

#

what is diffdb?

#

I don't think there is anything special about CC12M

#

if you drop it its fine

mighty finch
#

Hi

twilit forge
dawn kindle
#

I'd like to transform a person's face photo into a cartoon-like character while keeping their recognizable features (just like loverse.ai does).

Questions I have:

  1. SDXL vs Flux for this specific task - is one clearly superior, or are people just following the hype?
  2. IP-Adapter configurations - is there a "golden setup" that actually works consistently, or is everyone just guessing?
  3. Has anyone ACTUALLY created a workflow that matches commercial quality?
  4. What workflow end-to-end to get same or better results?

I've seen countless tutorials claiming to solve this, but the results never match services like loverse.ai. Who's actually figured this out?

If you've got real insights (not just theories), I'd love to hear them.

atomic mortar
twilit forge
#

@atomic mortar what is the best way for me to upscale a generated selfie of someone at 512x512 to a higher res while preserving the composition and shit

#

hires fix or img2img or extras?

atomic mortar
#

hmm it depends if you want to refine it or just upscale

#

if you just upscale it wont mess with the other stuff but i use SwarmUI so my knowledge doesnt translate over to A1111 well

twilit forge
#

well i generate it at 512x512, but I want it to be like a photorealistic full-size phone image, so it has to be "refined" not just upscaled

#

it looks like shit at 512x512

atomic mortar
#

hmm well it seems to me that your using SD 1.5. what gpu do you have?

twilit forge
#

RTX 3080

grizzled harness
#

hires fix and img2img are basically the same thing

twilit forge
#

10gb vram

atomic mortar
#

since SDXL gens at 1024x1024

#

or similar aspect ratios

twilit forge
#

the only reason I use SD 1.5 is because im using the realisticvision v6 checkpoint

#

well, technically it's SD 1.5 Hyper

atomic mortar
#

there are realistic SDXL checkpoints too lol

twilit forge
#

recommendations?

atomic mortar
#

id honestly look on civitAI to see what matches your prefference

#

since i mostly use illustrious

twilit forge
#

alright, ill try the SDXL juggernaut checkpoint

#

seeing as that is the most popular one on civitai

#

which upscaler do you prefer btw?

atomic mortar
#

well i use ones for anime so
4x_NMKD-Superscale-SP_178000_G
4x_foolhardy_Remacri

naive laurel
#

also I see that it's base model is SD1.5, however maybe I have to use it as a lora? i dont understand

atomic mortar
#

txt2img + Hi-Res

#

hi res fix

#

and do you have his embedding like "bad dream"?

naive laurel
#

also isnt txt2img something from the automatic1111 webui?

atomic mortar
#

was talking about his hi-res fix

#

settings

naive laurel
#

doesnt hi-res just make the image bigger

atomic mortar
#

it also "refines" but i recommend watching a tutorial for that one since i dont have it open rn

naive laurel
#

what I find odd is that Stable diffusion's default model is 2+ GB while this model is like 1.98 GB and generates better results. I'm new to this btw so maybe I'm missing a core concept

#

my intuition tells me that the model I sent before should "run" above sd 1.5 but i'm actually replacing it

crystal edge
#

what the best ai for generating arts right now

abstract quarry
#

Flux. Although for artistic images SD 3.5 might be a bit more creative

abstract quarry
still rampart
still rampart
sinful fiber
#

Hello!

chrome herald
oblique jay
#

Has anyone in here utilized SPAR3D successfully?

fervent thunder
#

personally I am going mnist -> cifar -> mscoco -> imagenet -> lsunsBedrooms/lsunsChurches

#

as a progression of models

#

cos then you have the most stuff to compare it to

last bay
#

Does anyone know what is the best checkpoint for using niji like midjourney ?

twilit forge
#

anyone else have problems using controlnets with SDXL models?

#

takes like 5 hours

fervent thunder
#

I find it hard to understand anime stuff

#

but maybe pony, illustrious or noob

#

the anime community clearly don't like Flux, but I am not sure why, because when I look at Flux anime checkpoints they look impressive. There must be something inauthentic about them but I am not sure what it is

last bay
#

because anime isn't suppose to look like flux usually

dawn kindle
magic grove
#

hi

exotic sphinx
#

Furthermore Flux doesn't support anime stuff even remotely out of the box and is kinda hard to train on

#

Similar reason why anime ppl avoided SDXL until something like Pony came out that just was trained on enough stuff

exotic sphinx
fervent thunder
#

wouldn't hardware be the same across genres

exotic sphinx
#

And the stuff that Flux is genning, at least from a quick view, is basically doable with SD1.5 aside from the text capabilities

fervent thunder
#

I'm not sure Flux has less users
on Civit it has more reactions on its images

exotic sphinx
#

Unless you spotted a model that I wasn't able to spot just from a quick look

frigid star
#

Anyone know where to find the pre-github-removed reactor?

fervent thunder
#

not sure if it is good for anime community or not

exotic sphinx
#

o.o

#

which one

fervent thunder
#

sorry I forgot link

exotic sphinx
#

Looks decent actually

#

Doesn't look like it's capable of genning furry or nsfw stuff tho, or there would be at least one post

#

It also doesn't help that quite a lot of the posted pictures are not even anime but more realism / other art oriented

fervent thunder
#

that might be better though, if it was a good anime checkpoint but was more censored

#

like it might be good to separate that out

exotic sphinx
#

I do think the one you posted fell under my radar, I'm going to have to give it a go

#

Once I clear some space

pine path
#

Chroma is modified from Flux Schnell which has the Apache license btw.

#

It was also trained get rid of the distillation and readd cfg

fervent thunder
#

ah I forgot about Chroma that is a big checkpoint yeah

hallow flax
#

Hello

still rampart
#

The models are mostly diffusion transformer backbones with cross self attention right?

#

What makes them better worse than others? If we exclude diffusion formulation like EDM or ddpm or ddim

#

Just number of parameters?

fervent thunder
#

rectified flow is the thing to look into

desert dagger
still rampart
still rampart
fervent thunder
#

there is stuff like sliced score matching but does not apply here

still rampart
#

sliced score matching got solved by vincent pascal in denoising autoencoders if im not remembering it wrong, it still needs to calculate jacobians..

#

but do these models just scale number of parameters or change architecture or training loss semantics entirely?

fervent thunder
#

sliced score matching was song, same researcher as DDIM. It can be done with jacobian-vector products rather than full jacobians, that is still slow though

#

architectures changed loads

#

and major training loss changes as well, in particular regarding first v-pred and then rectified flow

last bay
still rampart
#

didnt yang song write about score based models as SDE as well

fervent thunder
#

yeah

opaque stirrup
#

can anyone help me install comfy on amd

atomic mortar
trim fiber
#

hello

hallow plaza
#

Is there anything like stable diffusion with game art? I want to create sprites but in the past it wasn't onsistant.

onyx vessel
#

Hi there

atomic mortar
#

Perfect consistency is probably not a thing

#

But generating assets like tools or items yeah

atomic mortar
#

hmm a longshot maybe but anyone found a decent finetune of the illustrious V1 release? (not 0.1)

naive laurel
#

where can I find tutorials, articles or anything for lower requirement models and faster generation? Everything I find is always for PonyXL or any other XL model

#

i have 4gb vram and 16gb ram

warm junco
naive laurel
#

I found this which looks great but the prompt is completely different to the image. I'm trying to generate a similar thing using that prompt and for me, it follows the prompt (using same everything except Controlnet which idk what it is) https://civitai.com/images/11346154

#

the prompt is a space scifi themed thing but the image it generated in the link is completely different from that

#

for me it did generate a space themed img

naive laurel
warm junco
exotic sphinx
#

Niji is not, the rest are

last bay
naive laurel
warm junco
#

Its a seperate install

hallow plaza
naive laurel
# warm junco No

I remember back in the day, like 2023, I was generating like 4 images at the same time at a very decent speed. I don't remember how was the model called. Maybe it wasn't even SD 1.5

#

I thought about generating in a very very fast model, then if i like the generated concept, pass it to img2img in a better model

#

what do you think

naive laurel
abstract quarry
#

basically every model can do pixel sprites. There are also loras for it.
However, the images just look like pixel art, you usually have to apply some post processing to really make them pixelated

fervent thunder
#

there might be some Aurora model

#

it went away for a bit and then apparently came back

minor idol
#

Anyone use SD reforge? I'm having some troubles here. I just updated SD reforge (Stability matrix). After that all of my generated images looks very ugly and disfigured. From the 'Png Info' I copied the parameters from my previous generated images and sent to text2img. Still getting bad result (I used the same checkpoint, settings, same seed and LoRAs)

fervent thunder
minor idol
fervent thunder
#

there are some node packs that make it simpler

#

like

#

they can make an image with just 2 nodes

warm junco
minor idol
warm junco
#

Check it to be sure.
Load the image into the PNG info tab and copy all data into a txt file.
Then do that with the second image and compare

still rampart
#

does anyone happen to know a platform that rents h100 instances

minor idol
minor idol
fervent thunder
#

I had a look at Reforge code and
its using a lot of ComfyUI code anyway (they credited Comfy so its not a bad thing)
I also started out in A1111 so I understand why you like the interface

fervent thunder
minor idol
fervent thunder
#

forge and then reforge are like combinations of different things

#

they all have linked history though
cos both comfyUI and A1111 got their sampling code from K-diffusion

#

which was made by someone at Stability AI based on a paper by someone at Nvidia

minor idol
#

I heard Comfy is the best. It gives great control for generation. I'm thinking of learning it, but whenever I see the workflow, my motivation always breaks. 🤣

still rampart
fervent thunder
abstract quarry
#

depends what you mean with "best"

#

it's definitely not the "best ui", it's horrible 😅 but you have full control and it's quite good in memory and speed

#

and it definitely has support for more models than any other ui

fervent thunder
#

lmao

#

its a bit like trying to eat noodles with two teaspoons yeah

naive laurel
# warm junco Its a seperate install

im getting errors constantly in the cmd

  File "C:\Users\user\Documents\Stable Diffusion reForged\stable-diffusion-webui-reForge\venv\Lib\site-packages\gradio\blocks.py", line 1429, in process_api
    inputs = self.preprocess_data(fn_index, inputs, state)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\user\Documents\Stable Diffusion reForged\stable-diffusion-webui-reForge\venv\Lib\site-packages\gradio\blocks.py", line 1222, in preprocess_data
    self.validate_inputs(fn_index, inputs)
  File "C:\Users\user\Documents\Stable Diffusion reForged\stable-diffusion-webui-reForge\venv\Lib\site-packages\gradio\blocks.py", line 1209, in validate_inputs
    raise ValueError(
ValueError: An event handler (save_new_preset) didn't receive enough input values (needed: 13, got: 0).
Check if the event handler calls a Javascript function, and make sure its return value is correct.
Wanted inputs:
    [textbox, dropdown, dropdown, slider, radio, slider, slider, slider, slider, slider, checkbox, radio, radio]
Received inputs:```
#

and then a giant random text

pure mango
#

Guys how do i Train a Stable Diffusion Model Based on my Own Art Style

ashen scaffold
#

Does someone tried the new hunyuanvideoi2v model with comfyui yet ?

desert dagger
exotic compass
warm junco
karmic swallow
iron onyx
#

do bigger prompts take longer? seems like they'd have to..

#

conversely it seems like larger models could take longer, but not necessarily.. like how many dimensions do each token or inference level have?

ocean slate
#

高清球队合影，运动员身穿运动装备，佩戴护膝、手持保温杯，展现健康养生主题。背景中有毛笔书写的“养生局”，呈现中国传统水墨风格，电影灯光效果，动感构图，精致细腻的艺术细节

strong night
#

hello to all

#

anyone here speaking german?

#

anyway, i hope to find a nice person that could help me and could create a AI picture for me for, well, political commedy

#

not realy a big task for a person that is able to do it (to be honest, i am too dumb for it), but it should look like the emperor Palpatine (star wars - already in his dark robe etc) with the face of the high commissionaire named "ursula von der leyen"

fervent thunder
#

hello anyone here knows about sillytavern

atomic mortar
frigid light
#

Hello everyone!

novel dagger
#

Does anyone know what the Enterprise license to use SD 3.5 commercially costs in case I succeed $1mio annual revenue one day?

atomic mortar
novel dagger
#

okay, thank you

fervent thunder
#

but i gave up

#

so nvm

still glacier
#

go for ollama, it s much more simpler and usually faster (but less tweakable)

compact swan
fervent thunder
atomic mortar
#

What gpu do you have

abstract quarry
still glacier
deep narwhal
#

what upscalers do you guys usually use

still rampart
#

as a heuristic how often should i expect it to get interrupted, I could instantiate callbacks to run every x steps

#

to save progress

#

how often do the interruptible instances come back online for use

river roost
#

Adetailer got its update so it works on a 50 series gpu

fervent thunder
dreamy sorrel
#

Hey all, I’m pretty fed up with Black Forest Labs and Flux 1.1 Pro. It seems like they’re not bothering to improve the image generator anymore— as of it is now often the images are still wrong, and an incredible number of times the text is incorrect. Plus, it feels like they’ve shifted their focus to video and inpainting instead, all while keeping prices ridiculously high. I’m also worried they’ve gone into that startup scaling mode, and everything might go downhill from here, like so many other scaling startups. Anyone know of new startups or alternatives that can match or beat Flux 1.1 Pro’s quality for pod images—preferably cheaper or better value for pod images?

fervent thunder
dreamy sorrel
#

the ones you print on tshirt mugs and so on

atomic mortar
#

though a little photoshop could fix it

#

if you run it local theres some ways you could work around it with some custom comfy nodes but cloud based? no idea

fervent thunder
#

there are some text rendering things on arxiv but generally not released or ported to familiar software

dreamy sorrel
#

Thank you for your answers. Hopefully, someone else can point me to a better alternative 😦

fervent thunder
#

its not a model that you need its a sampling method

dreamy sorrel
#

Also, because it’s kind of super annoying to complain about quality and price at the same time. I think I would be okay paying the crazy prices right now if they were able to ship what I need to ship. Or, I’d rather pay less. But like this, I feel like I’m paying for pricey stuff that also doesn’t deliver. Meanwhile, it’s been months that they’ve been focusing just on the other stuff.

fervent thunder
dreamy sorrel
fervent thunder
#

closed source are really really far behind open source at this point for image stuff

#

mostly because APIs kinda have to be generalist whereas the best methods are usually highly specific to one task

#

and secondly because they limit how much compute they will assign each API call

#

and generally pick an amount not that high

dreamy sorrel
#

got it! my questions was "Anyone know of new startups or alternatives that can match or beat Flux 1.1 Pro’s quality for pod images—preferably cheaper or better value for pod images?" I will add also "delivered via api".As i said I don't really care what's behind the api, closed source, open source, nodes whatever

#

So i can stop paying them and use an alternative instead that can deliver same quality cheaper or better quality at the same price

#

But not sure if the alternative exists, that's why am asking

fervent thunder
#

are you willing to make and host the API yourself?

dreamy sorrel
#

but tyvm for all your answers 🙌

fervent thunder
#

ok I think my conclusion is I don't know of one then

dreamy sorrel
naive laurel
still rampart
naive laurel
wary flint
#

Just looking

gusty pine
#

Is there a way to convert/merge a sd1.5 model to/with an sdXL? I have a great model i made with great lain but would like to upgrade it with the data of an sdXL model

solemn pewter
fervent thunder
#

guys help mee

gusty pine
tulip yarrow
#

Any reliable way to train flux on colab? I have 139 compute units from colab (currently not subscribed but those were when i was)

mystic gale
#

hello

#

ok dead discord, goodbye

idle depot
#

hello

lusty stirrup
#

hello, everybody

minor plinth
#

hiya, i'm trying to get into SD, and don't really know what the best ones to use are. any suggestions?

unborn hedge
#

Civitai erased an image I made without posting and now I have to recreate it from scratch : (

lusty stirrup
atomic mortar
#

but it depends on the GPU you have though.

minor plinth
#

i have a 4060ti

atomic mortar
#

oh then you can run a big range of models pretty easyly

#

For anime images i do recommend Illustrious or Pony (though illustrious is better IMO).
For realism. i suppose a realistic SDXL finetune/merge or Flux schnell but i dont make a lot of those

minor plinth
#

perfect thank you

fervent thunder
#

4060ti is fine yeah

#

2080ti or 3060 even are fine as well, I've rented those before

#

with some offloading and quants you can run things

#

video is the hard area because those models don't seem to squish down very well

abstract quarry
#

you have to retrain. You could use synthetic images for training, though

gusty pine
abstract quarry
#

what do you mean? which tool to use? I always used kohya's training tool but there are plenty to choose from

fervent thunder
#

have a go with the Koyha nodes maybe

#

I know they are called Flux Trainer but they added SDXL support

gusty pine
abstract quarry
#

you usually use a separate tool for training. If you want an UI you can search for bmaltes kohya_ss

fervent thunder
#

IDK A1111 stuff any more

abstract quarry
#

*bmaltais

fervent thunder
#

I actually used it before the others but I forgot what I knew

abstract quarry
fervent thunder
#

I liked Fooooooocus but that's dead now apparently
that was where I did inpainting for the first time I was astounded by inpainting

#

I removed a clock tower or something like that and was shocked that it worked so well

gusty pine
#

i am a leek with pc stuff its already a miracle i got a1111 running

unborn hedge
abstract quarry
#

I mentioned it so many times here but I always liked InvokeAI

unborn hedge
#

i was struggling for a bit but i think i found all the pieces of the model and loras i used

abstract quarry
#

I find it a bit weird that so many newcomers try auto111 where invokeai is so much more intuitive and easier to install

#

I myself most often use comfyui cause it has newest features first most of the time

gusty pine
#

prolly because the first thing you find on internet when researching a localSD is a1111

abstract quarry
#

yes, but auto111 is dead

#

it's from sd 1.5 times

#

since then there are so many forks

#

it's totally not user friendly though

gusty pine
#

anyways it sounds like i should try a different localSD to actually start rumbling with sdXL and other stuff

#

excuse me what the shit comfyUI only needs a zip file installment and a bat file executed? What sorcery is this?!

unborn hedge
#

comfyUI is for the advanced users tbh, i use Forge as my sd1111 fork

deep horizon
#

is there a way to re-evaluate my prompt text every time I generate, even if it hasn't been changed? Trying to use the random selection syntax, but its using the cached prompt (thus not changing the random selection) unless I make a change to the prompt.
ComfyUI btw

abstract quarry
#

comfyui is very complicated to use

#

you can use swarmui that is a bit more user friendly

#

I would recommend InvokeAI

deep horizon
#

I'm so deeply setup with comfy though. would be a pain to switch

#

I'll take a shot at getting invoke running on the side, and maybe move my things over to it if I like it

gusty pine
deep horizon
#

that would be sick