💬｜general-chat

[desert dagger]

you can't do this here

[fervent thunder]

also the model that can pull that request off has not been invented yet 🤔

[raw fable]

thanks!

[sonic hull]

hi

[foggy halo]

HuggingFace gets more and more annoying when they keep on warning "You have exhausted your GPU quota"...this is so irritating especially if we want to urgently generate an image

[desert dagger]

guess you need to go use colab then

or use mage.space

[foggy halo]

will mage space also have Flux model?

[foggy halo]

thanks for the suggestion 👍

[desert dagger]

on the home page, yes. 2 versions right under the prompt text field

[foggy halo]

oh that's amazing then

[desert dagger]

you get stuck, let me know, i'll point you to the how to use documentation

[mystic star]

it's there any difference for using a1111 when its like linux mint or ubuntu etc...

[versed radish]

Mov2mov don´t load

Mov2mov don´t load Click no response

what's going on?

[unborn wave]

heya!
anyone have luck with img2img for plushies/stuffed animals?
i want to turn people into marketable plushies

[desert dagger]

that's not what image2 image does or is for

[unborn wave]

I just want to use a real photo as a reference. I've had some luck but not alot. Definitely a noob though.

[desert dagger]

i know, but that's really not what img2img does. it uses your refrence image as a visual prompt. it'll use the color pallet, the shapes, and the locations of the shapes. then it will read your text prompt to find otu what it's creating. then it will create that using the color pallet, and shapes of the refrence image.

[unborn wave]

What would be a better option?

[desert dagger]

but you're wanting to create physical 3D objects to sell based on people - if it was me, i'd find someone that was already doing that and partner with them. that's a very complicated business you're talking about starting

[unborn wave]

not sell

just for fun

and not physical

just like, turning people into plushies. I want to plushie my friends.

[desert dagger]

so when you said "i want to turn people into marketable plushies" what did you mean?

[unborn wave]

I basically want a "filter" to take photographs of people and turn them into images of plushies.

[desert dagger]

okay well - img2img isn't going to do that

[unborn wave]

what will

[desert dagger]

you'd probably have to train a custom lora

[unborn wave]

I did find a lora but it's not working great.

[desert dagger]

that's why i said 'you train' and 'custom lora'

[hardy sedge]

Hello I wanted to resume my training with kohya_ss.I was at epoch3 have set 6 epochs. How can I resume trainingHow can I resume training?

[analog stream]

Did anyone know a good workflow, a video or a document to make a workflow for comfyui for consistent style and same type of color grading through each image with a upscale of final image if possible a lora. I want to do it in comfyui. Thanks in advance for your help

[static plaza]

I hosted SD on Azure and I access with gradio.live. Can I download Models using WebUI?

[fervent thunder]

the common term img2img is kinda misleading
because it makes it sounds like img2img is the best way to make an image from an image, but its actually a fairly poor method
before lora you could try to combine some control nets such as open pose, depth, canny and segmented with a good IP adapter like the mad scientist node or the SDXL style and composition IP adapter

when you get IP adapters that differentiate between style and composition, what's actually going on there is that they have preselected for you blocks that pertain to style and blocks that pertain to composition. This is key for it to work well

another thing that can boost IP adapter a bit is preparing very good embeddings. You can average/add/concat multiple images, both for positive and negative conditioning, to feed to IP adapter

and you can use image editing to boost the effect of IP adapter image embeds beyond what would normally be possible, a common trick is to sharpen the positives and blur the negatives

[feral pike]

anyone know what sdnext is?

[fervent thunder]

yeah SD Next is also known as Vlad Diffusion

its made by a guy called Vlad

its got two backends, one based on the CompVis LDM / A1111 SDWebUI code lineage and one based on Diffusers

[tulip yarrow]

what even is flux? is there a local SD3/flux lora trainer that you can run on colab/locally

[abstract quarry]

SimpleTuner for example

https://github.com/bghira/SimpleTuner/

as far as I know, kohya_ss is now supporting flux, too. Haven't tried it yet, though

[bleak jackal]

I'm having trouble downloading homebrew on mac , can anyone help me ?

I got it dw

[solid kindle]

is auto1111 automatically utilize my GPU vram or is there any setting to turn it on ?

[pulsar venture]

Hi, what is PDXL?

[fervent thunder]

its a fine tune of SDXL

[pulsar venture]

is ponydiffusion?

what is flux?

[fervent thunder]

yeah

[fervent thunder]

flux is a new recent model

[pulsar venture]

Model like checkpoint?

[fervent thunder]

yeah but its a different architecture also

[worthy zephyr]

Hi, anyone knows what happened to the playground ai discord server ?
I've been on it for a while and now I can't find it anywhere

[desert dagger]

did they ban you from it?

[worthy zephyr]

no, I think the server doesn't exists anymore, but i'm not sure why

[desert dagger]

i dunno. look on their website and see if you can find a link to their discord

[worthy zephyr]

I can't find the link anywhere, and someone said that their server is down
https://www.reddit.com/r/PlaygroundAI/comments/1evgaej/i_believe_playground_is_ending/

[desert dagger]

it's playground.com and i'm looking at their website. doesn't look like it's ending to me

[hardy sedge]

Adetailer multiple faces. Is there a way to Detail only 1 face Not all?

[fervent thunder]

in comfy UI yes, Impact pack has nodes designed to select SEGS

[fervent thunder]

sd 3.1 is coming ?

apparently yes

today i tested flux gguf the lowest level , for the first time running flux on my 2060 in less then 20 seconds

no freez no out of memory all thanks to city96

[desert dagger]

some century

[ashen rune]

Recent research/experiment done in TouchDesigner and SD: https://www.youtube.com/shorts/VyZFzKAuqsk (more info in video's description)

What do you guys think?

[fervent thunder]

https://youtu.be/pNyIp73zva8

screen share is the real deal

i assum this will work on any window even a windowed video game lol

[fervent thunder]

that looks really cool

[fervent thunder]

the old photos give it a really distinctive look, its great

it can be used on youtube videos , movies , blender , you name it.

[sand heart]

Total noob here.I am interested in running SD on my pc.Any recommendations about where to start would be greatly appreciated.

[desert dagger]

start by making sure your hardware is actually going to run it. Also, the recommended interface is comfyUI

[sand heart]

ComfyUI,got it.Do you think a 3080 card would be sufficient?Thank you very much for the response sir.

[desert dagger]

depends on what else you're doing. you should probably ask additional questions in #🤝｜tech-support

[sand heart]

Will do,thanks again.

[tired fossil]

how to use? i forget

[drowsy robin]

What's the current overall most generally effective upscaler in comfyui?

I can't get supir running effectively on arc 🤷‍♂️

[fervent thunder]

Are there any "ai image sorters" on windows I can use that are actually good and don't spy on me

Or is that entire concept iffy

[pale latch]

you could builda hydrus server and run a classifier tagger on it https://github.com/hydrusnetwork/hydrus

hydrus might even be able to create tags based on the prompt and settings

[neat cypress]

flux lora with 9.7 gb : https://www.reddit.com/r/StableDiffusion/comments/1ey6hss/kohya_ss_gui_flux_lora_training_on_rtx_3060_lora/

[fervent thunder]

Hello - does anyone have much experience with image to video using AI?

[desert dagger]

yeah. what do you need help with?

[hallow wave]

hi

hi

[nocturne pewter]

Is there a way to make SD understand lightning in the image, when doing inpainting?

[fervent thunder]

there's like 30+ inpainting methods

so its very hard to talk about inpainting in a general sense

[nocturne pewter]

Trying to get a better looking face. The problem is when put denoising high enough to get good results, the new face is much brighter than the original, a person who's in the shadow.

[fervent thunder]

IP adapter would be the first thing to try

[drifting knot]

Hello guys, I am working with SD upscaler and have some unclear part of it. When we use using it we use Stable Diffusion upscaler and usualy some type of GAN. How are they working like someone increase size and someone quality? Could some one explain it to me or sent some article about it? TY so much

[fervent thunder]

depends on the interface

[abstract quarry]

the GAN is just sharpening the image - a real "upscaling" is often too hard for GANs. In particular, they cannot fix things that are not recognizable in the lowres image. If in your image is a leaf, then they can try to remove noise artefacts from the leaf when upscaling, but they won't be able to add new stuff like leaf vein and texture which was not visible in the original lowres image

Stable Diffusion upscaling is a normal image2image pass. Therefore, it can invent new things (like adding texture), which is a good thing but sometimes also a bad thing. e.g., it might add a beetle in the leaf although there was none in the original image. Sometimes this additions make sense and improve the image quality, sometimes they don't make sense and lead to weird artefacts

usually you use controlnets that enforce SD to not invent new things that are not in the original image

nevertheless, a GAN tries to keep the original image mostly intact and only tries to remove blurryness from the upscaling. A SD pass creates more or less a completely new image (depending on how much noise you use), so it's no longer the original image, but this is the only way to really get high resolution texture and features.

Although many workflows contain GANs and Stable Diffusion - you don't really need both. You can try GAN and if that is not sufficient, use SD instead. I don't think that using both after each other has high benefits (it's also not a big disadvantage, though, in particular as GANs are fast anyways)

[nocturne pewter]

Thanks, will give it a try!

[fervent thunder]

the old GAN-style upsale methods are under-rated yeah, I like to use this one https://openmodeldb.info/models/4x-NomosWebPhoto-atd

its a bit like Swin-IR

[fervent thunder]

with the release of closed source model such as ideogram 2.0 ... i am oficially hyped for sd3.1

I2.0 is a good model tbh.

[unkempt hatch]

so for Flux, the CFG inside the sampler should always be 1, right?

[errant glen]

hello

[brittle oyster]

I'm trying to generate an image with this prompt using Flux but all I'm getting is a picture of a house with lightning above it. It looks very realistic but I won't add the dome over the house:

"In the midst of a massive lightning storm, a typical residential home in Florida, stands under a protective, glowing dome of energy. The dome, a visible forcefield or bubble, envelops the entire house, clearly separating it from the chaotic storm outside. The dark skies are alive with intense lightning bolts, but each strike that comes close is deflected or absorbed by the dome, causing bright, electric ripples to spread across its surface. The dome is the central focus of the scene, shimmering with power and creating a stark contrast between the turbulent weather outside and the safety within. The house beneath the dome is a classic Florida-style home, secure and untouched, thanks to the powerful energy shield that protects it."

How could I improve that prompt to achieve the desired results?

[fervent thunder]

likely too long and too many concepts

its also not in the style of a caption

this sounds more like a prompt an LLM wrote tbh

[abstract quarry]

which is probably on what flux was trained on 😉

[fervent thunder]

lol yea maybe

there's something fishy about the way flux responds to prompts

[abstract quarry]

I could imagine some images/prompts are strongly overfitted

[fervent thunder]

we rly need the research paper so that we can know which caption model or VLM was used

my opinion is that Flux is too overfit overall yeah

not terribly so, but to a certain extent

a bit like Playground 2.5

the aesthetic is rather "baked in"

[abstract quarry]

have it!

it's not beautiful, though xD

see general-with-images

[fervent thunder]

wow that's rly impressive

unlikely to be in the training data so it synthesised it from multiple concepts

[abstract quarry]

so the general idea seems to be to FIRST describe the energy barrier. Describe how it should look like. Afterwards you describe the house

A glowing energy dome during a thunderstorm. The dome is an energy barrier and consists of blue light and a geometric pattern. Inside the dome and visible through the transparent energy barrier is an American house. Lightning strikes the barrier and is stopped by it.

thats my current prompt

it generates the barrier and the house in every generation

[unkempt hatch]

I love examples like that

[abstract quarry]

but to be honest: Ideogram and Dall-E usually have no problems with such things

Flux prompting behaviour is not really good. It can sometimes handle very complex prompts, but then sometimes it fails even for very simple ones

I think that loras and IPAdapters will be crucial for Flux

[fervent thunder]

I have used Dalle 3 by far the most of any model and I still thinks its the best yeah

the secret HD mode in the API in particular

[unkempt hatch]

and a good general finetune

[fervent thunder]

need juggernaut for flux, that also attempts to get CFG and fully effective negatives back

[drifting knot]

TY so much! so if simplify SD increase picture and GAN makes details better or vice versa?

[pallid quest]

hi

[fervent thunder]

both increase the size of the picture by the same amount if you choose, but SD adds more details and GANs or transformers like SWIN-IR only add a tiny bit of detail and keep the image more the same

[quartz siren]

kinda what this project is trying to do: https://huggingface.co/ostris/OpenFLUX.1

[fervent thunder]

thanks will keep an eye on this

[brittle oyster]

Thanks, this does seem to work better! Still not getting what I was wanting but it is much closer

[fervent thunder]

you might get good results with a second pass using a nice model like Leosam's Helloworld

one of the pretty ones

[desert dagger]

there's always just using something like CapCut's free image upscaler - it just gives you a larger, sharper, image that is exactly like the image you started with. And they have a video upscaler that's free, too

[split cape]

Does it matter what punctuation you use to divide prompts and prompt phrases?

[fervent thunder]

depends on the text encoder but yeah ideally it would match the training data

if they have a text encoder like T5 this matters less

as working text stuff out is what T5 was for

funnily enough the main model with a smarter text encoder is Kolors but it didn't actually help too much

[split cape]

I cannot for the life of me get this one model to produce uneven/bad lighting. Being an anime style it might be forced to give it that perfect backlighting but it doesn't always fit the scene.

I think the V2 expansion prompts that I didn't ask for are screwing me over; gonna try different styles

[quartz siren]

I believe training data also is very important. Auraflow0.2 and 0.1 use a much smaller t5 text encoder called pile t5xl compared to flux and sd3 which use a much larger text encoder t5xxl and the 0.2 and 0.1 have better prompt following then flux and sd3.
pilet5xl is like 1b params while t5xxl is like 4b params.

Auraflow 0.3 is the same architecture but seems to have much worse prompt following it was further trained on a different type of dataset.

[desert dagger]

did you try "dynamic shadows"

[split cape]

I have not.

[desert dagger]

i'll tend to add phrases like "dynamic lighting, dynamic shadows, stage lighting, dim lights with haunting shadows"

[trail lion]

Wow civit exploded with flux content in the last couple days

[desert dagger]

but how much of it is really worth your storage space?

[trail lion]

The early stuff is normally not great. I love it when they say that up front, like I barely tested this... I'm like, why upload it then?

[split cape]

I kept trying with the same idea but the more I added to the prompt and changed the styles generally it just got worse

[desert dagger]

why are you using the model you said you were using?

[fervent thunder]

ye exact text encoder choice matters less, than other choices

[split cape]

it's just what I have; all I have is Fooocus. I don't feel like getting a new engine/model yet, I think my skills need to increase first.

I don't think I've truly found the limits of this model; it's more likely I don't fully understand it

[fervent thunder]

Fooocus does a lot of stuff without telling you

skills you learn on fooocus will not neccesarily transfer anyway

[split cape]

I've gotten a couple really promising results that make me think I can control it better

and sometimes I really like the art styles I'm getting out of it

but I see why people say it is good for beginners, it definitely works best with simple prompts and is an easy way to get your toes wet in ai

Front ends like ComfyUI look intimidating to me.

[desert dagger]

well - then you probalby are stuck with what fooocus wants to create. you could grab a copy of pinoko and install something other than fooocus

[desert dagger]

then install Swarm. it's a much friendly front end, written by @finite cloak

[split cape]

what models is Swarm able to run?

[desert dagger]

should be able to run everything

comfy is really not that hard once you start using it and get used to what stuff does

[split cape]

it uses nodes which I'd think would give you better control but idk how any of these apps truly work on the backend

[desert dagger]

all the more reason to start using comfy so you can learn and it stops being mysterious. there are a lot of workflows already put together that you can grab and run, and then look at. and we can explain what everything's doing as you have issues. and if you join the L2 discord (matteo's) you'll find a lot of people there that can help too

[split cape]

good idea, and someone said comfy was also well optimized?

[desert dagger]

it is, yes, and while some workflows people put out might not be, there are plenty of us that can help you optimize those if you need. i'll even give you a fairly simple flux workflow if you'd like

[split cape]

Can you link me to that server you mentioned?

[desert dagger]

which server did i mention?

[split cape]

Matteo's

[desert dagger]

i'll DM you the link. can' tpost discord links here

[trail lion]

It took me about 2 weeks, and I actually seldom leave the workflow tab these days so I may as well be just running comfy

For swarm

[mighty forum]

so i want to generate an image, where can i put in the prompt? im so lost

[desert dagger]

if you're trying to generate here on this discord, first read the information in at this link #artisan-faq

[low moon]

You beam it in telepathically.

Focus. Concentrate.

[trail lion]

I sometimes wonder how people land here, they think images will just somehow manifest

[fervent thunder]

we have midjourney to thank for that

[low moon]

You have to focus harder than that.

[fervent thunder]

https://www.reddit.com/r/StableDiffusion/comments/1eykiy0/now_we_have_sorta_conquered_prompt_adherence/

sd 3.1 need to have an option that must be activated after each render.

this option lock data, colors, gender, atmosphere... ect for better consistency.

one day

[fervent thunder]

is it true that is free now ?

[fervent thunder]

For me, MJ is the best prompt spammer for stable difusion.

[low moon]

Consistency. Only real consistency will make any model stand out now.

Image or video.

[fervent thunder]

even if concsistency maybe done by control net, its anoyying when the whol generation change because of one word.

[low moon]

No. We need real consistency.

Until then this is all a fad/gimmick.

[fervent thunder]

honestly, prompting with natural language is overrated

prompting with commas might be the sole reason 1.5 will never die for me

I might do a mix-and-match of Flux and 1.5 for upscaling

[low moon]

I agree

Just throwing in the words of what i want without any grammatical coherence works amazing for sd15, sdxl and flux for me

just list the stuff u want

😄

[desert dagger]

natural language tends to use commas...

[fervent thunder]

yeah, but, not, like, this

[narrow kernel]

What about Captain Kirk speak where we only talk in ellipses... Like... This...

(im actually really guilty of sometimes overusing ellipses, but I was a drama student as a teen that did a lot of stage plays and scripts used ... a lot)

[mild yacht]

hi, stable diffusion is using all my 16gb ram and it's so laggy something wrong it wasn't like that yesterday anyone know how to fix this ?

[desert dagger]

some people talk like that

[desert dagger]

is it running on your ram or your vram?

[mild yacht]

i solve it with chatgpt it was a huge X checkpoint causing the problem

[desert dagger]

ah. okay

[low moon]

Friday.

[desert dagger]

last week

[low moon]

Ready your rears.

[drowsy robin]

Seems you can use --reserve-vram 8192 or any number to limit comfyui's total vram usage now

I can do --reserve-vram 12288 and run flux 8_0 and t5xxl no problem with unloading

[fervent thunder]

I might be addicted to porn but don't know how to quit 💀

[desert dagger]

get rid of everything that you have in your environment that even hints at it and avoid coming in contact with it going forward

[fervent thunder]

but what will I do instead. what do people who dont watch porn all day do

[desert dagger]

there are millions of other things you can do. so every time you feel the desire to participate in the habit you're trying to break, find something else to do, maybe something new and different to you like jogging or whatever

[fervent thunder]

to quote smiling friends, I think I'm stuck in a loop of short-term dopamine rushes

really appreciate the advice bud

[low moon]

Anybody wants pizza?

[fervent thunder]

eating healthy might be a way to start cutting off bad habits.

[low moon]

aight

This is broccoli pizza tho.

[thin helm]

is it normal to have stable audio long-pending in the input audios?

[desert dagger]

that's likely the case, and the only way to break that is to find something else that'll be enjoyable and healthy for you. so jogging, exercising, etc

[pulsar venture]

I don't understand how install ipadapter on comfyui someone could help me?

[desert dagger]

there's a #🧣｜comfy-ui channel that might be more useful to you than generat chatter

[sinful fox]

https://discord.com/channels/1002292111942635562/1276220594416914536

[split cape]

How do you get noisy/low quality android phone pictures out of Flux?

trying to make some scary alien pictures

[desert dagger]

you could try adding "film grain, photo, blurry, found footage"

absolutely do not use photoreal, photorealistic, or photorealism

[split cape]

I'm trying things like that on the huggingface schnell distro but it's probably not as good as installing locally

I was under the impression Flux is trained explicitly on phone quality images

[desert dagger]

you can't adjust all the settings on huggingface, that's just a demo. you might also need to play around with different samplers and schedulers

[desert dagger]

it's a 12 BILLION parameter model. it didn't train explicity on anything

[split cape]

not explicitly like that's all it can do but explicitly like it was part of the model

[desert dagger]

it's a 12 billion parameter model. i'm sure it has photo photos in it. what's your prompt?

[split cape]

I tried a couple and gave up

I used terms like "lost government footage, 1970s, canon 70mm, noisy, droid razr, android phone, low light" nothing worked

it all looked plastic wrapped and rendered with good form and light

[desert dagger]

what's your actual prompt, please

[split cape]

I closed the tab I don't have the prompt anymore

if you want a new prompt I'll give you one

[desert dagger]

it's a little eaiser for me to try to assist if i can see the prompt being used

[split cape]

grey aliens hiding in woods, noisy, low light, android phone quality, scary, hiding in the shadows, lost government footage.

try that

I know you said we don't want realism but we don't want a cartoon style either and that's what I was getting. If I had control of negative prompts I think I could have gotten something.

[desert dagger]

and you want this on flux?

[split cape]

I just figured Flux was the best shot at generating that.

I can try it on SDXL

[desert dagger]

i'm not sure what the huggingface demo will do, because i don't know what's implimented in it. do you have the ability to run it at home? if not, can i have the link to the demo you were using please

[split cape]

https://huggingface.co/spaces/black-forest-labs/FLUX.1-schnell

I will look into a local flux installation

I'm not pressed for time or incapable I'm just lazy

[desert dagger]

give me a bit here

[split cape]

take your time

[desert dagger]

had to post it in the other channel #🏞｜general-with-images message

[desert dagger]

you might try using flux on mage.space. you cause use it with a free account

[split cape]

hmm alright

[echo kite]

what are yalls recommendations for some good versatile models (non-SDXL since I run SD on a cpu)

[desert dagger]

um what's wrong with sdxl?

[echo kite]

I am running SD on a cpu with 32 GB of ram

images on normal SD (sd 1.5 based models) take like a minute

[desert dagger]

so every single thing is going to run slow. you might as well use the good stuff

[echo kite]

can SDXL even run without out of memory errors on my machine?

[desert dagger]

depends on how you're running it

and how much swap space you have

[echo kite]

ahhhh

so that is why I have not been running into as many issues recentlyish after switching to linux

I forgot about swap space

to use sdxl do I just download an SDXL model?

[desert dagger]

how are you running stable diffusion right now?

[echo kite]

a1111 webui on arch linux

[desert dagger]

then yes, all you should need to do is download SDXL - i think it uses the same VAE as sd1.5 does

[echo kite]

k

[desert dagger]

you might post in #🤝｜tech-support and see what @warm junco has to recommend, he's more familiar with a1111 than i am

[echo kite]

half the good models require a civitai acc

and I don't want to have to make one just to download a few models

also, did inpainting models and standard models get merged in sdxl??

I am not finding a lot of inpainting versions for the sdxl models

[fervent thunder]

is there a good way to get reasonable results from a local installation - what software is good. I've found most tools including runwayml to be very overkill when it comes to censorship

[echo kite]

of sd?

ah, went to the linked context

[desert dagger]

sure, but it isn't an overnight thing. you have to learn the interface you're going to use, and then you need to learn how the AI you're prompting thinks

[sullen dock]

Is it possible to fine tune stable diffusion on multiple subjects so that it remembers them. For example, I want to train it on two people, person a and person b and I want to be able to use person a and person b as keywords. How can I achieve this?

[split cape]

probably with a LoRA

I still need to figure that out

[fervent thunder]

that could go beyond typical lora abilities. For that, you might need full weight dreambooth where you train the text encoder too. SEcourses have step by step tutorials for this

generally lora is single subject

[sullen dock]

Okay, I will take a look into them. Do you think face swapping would be a better approach then? It's not very accurate but I would be able to use it for multiple people

[fervent thunder]

it depends if the project is casual or serious
I don't actually do real projects with image generation cos I am just playing around with it
but if I was ever to actually do a real project I would train a full dreambooth every time

there's no reason not to

[sullen dock]

Okay, also, what is the best approach for fashion modeling. Let's say I have some images of a garment a, and I want to generate images of models wearing that garment. I want to preserve the print of the garment.

Based on my research, the only way is to do lora on each of the clothes that I want to generate an image of but the results are not very accurate and I need to train a model for each cloth.

[fervent thunder]

there are some other ways that specialise in clothes

for example this: https://github.com/cozymantis/clothes-swap-salvton-comfyui-workflow

[hardy sedge]

Is there a way now to train Flux loras model with a 4070 TI 12 GB Vram?

[sullen dock]

That's really cool, thanks a lot dude

[fervent thunder]

👍

Stable diffusion doesn't really know how to make specific facial structures from just text prompts in my experience so it would be nice to know if it would be possible to make a lora for a specific facial structure to give it something to go off of or is that not how loras work

(I have never made a lora)

could you give an example

anyway the answer is yeah, a lora can do that

you could also try IP adapter style transfer, instant ID, or canny control net

textual inversion also

if its nose then depth control net

Literally never heard of half that stuff is there a yt video I can watch that will teach me?

yeah there are a lot of tutorials out there

Ok

[split cape]

Fooocus does NOT want to draw a woman sticking her tongue out

is that hard coded as nsfw or am I just prompting wrong

it's an image prompt so it should just follow the picture, I'm redrawing a screenshot

[fervent thunder]

fooocus does a lot of things that it doesn't tell you about

it doesn't just sample normally it adds lots of things on top

it makes the model far less controllable and far more confusing to use

[keen epoch]

would be happy if somenote with 4090 rtx run a tests.. https://github.com/comfyanonymous/ComfyUI/discussions/4571

[split cape]

I turned off the V2 style for this iteration because in the log you can see V2 adding additional prompts

[fervent thunder]

the idea of fooocus is to have a very simple interface for people who just want to type a few words in and get a nice image on average
if you want to actually improve and optimise things, that requires moving off of fooocus

[split cape]

grrrr

no I know you're right

the UI is just easy and I'm messing with inpainting and image prompting

[fervent thunder]

fooocus does have one of the nicer inpainting methods

[split cape]

I'm having great success turning anime styles into irl styles and vice versa which is just too much fun

with the exception of the tongue thing but we'll get to the bottom of that

[fervent thunder]

anime to real and back is good yeah

[sudden ruin]

Send the workflow somewhere

Oh wait I cant read

[fervent thunder]

if you want to really push optimal speed you may have to move off comfy

fully optimal speed is rarely worth it though

[wise ginkgo]

how can i run sd with a 6gb gpu?

Comfyui is out of the question/not doable, right?

[solid kindle]

i've run ith 4gb gpu

but its real slow

note that im runing A1111

[round dome]

sup

[woven grove]

https://colab.research.google.com/github/Jelosus2/Lora_Easy_Training_Colab/blob/main/Lora_Easy_Training_Colab.ipynb#scrollTo=vGwaJ0eGHCkw
Does anyone know how to use this colab version?

[pliant crag]

https://tenor.com/view/have-a-little-meet-up-real-housewives-of-beverly-hills-have-a-get-together-have-a-small-gathering-have-a-little-party-gif-22409009

Nice little AI meetup in NYC on the 12th of Sept

https://resources.prolific.com/prolific-ai-research-summit?utm_source=Brad Communities&utm_medium=referral&utm_campaign=AIEvent

[desert dagger]

cant' use images in this channel

[pale latch]

make sure you're confirming people aren't bots when they show up

[final stag]

hi

[low moon]

hi

[robust needle]

Is stable diffusion and flux the same thing?

[low moon]

they twins

[indigo gyro]

anyone having problems loading forge after the update?

[robust needle]

I'm interested in downloading Stable Diffusion, does anyone have a guide that they would recommend?

[abstract quarry]

flux is newer and better, but also larger (need better hardware)

[sleek otter]

Automatic1111 or ComfyUI?

[robust needle]

dealers choice, idk about either

perhaps the one that would cooperate better with my 8gb AMD graphics card

[sleek otter]

Try SwarmUI

[trail lion]

you can get AMD to work, but it's going to be quite slow. there are guides in the tech-support channel, look at the top in pinned messages https://discord.com/channels/1002292111942635562/1002602742667280404

[robust needle]

Cool, thanks to both of you

[covert bridge]

in flux running in forge, does anyone know what i should set my "gpu weights" to? I have 16GB on my RTX 3080, and 32GB on mobo

[slim dragon]

hi, how do i specify the seed in the online stablediffusion.com? i cannot find any UI field for that..

[pale latch]

the daddies that made stable diffusion left stability and then daddied a new ai model

[slim dragon]

ugh, i've been using actually stablediffusionai.ai... i get that's a different beast than the .com isn't it?

[fervent thunder]

this is literally a virus

my virus checker dinged it for multiple threats detected

[silk yoke]

hey everybody

[slim dragon]

hm. my browser got quite sluggish using it but i didnt want to be overly suspicious. now that you suggest it could be smthg fishy then it's the end of affair. thanks for the warning!

[fervent thunder]

it can't be anything other than fishy things

because the real website is stability.ai

you can't just go to various websites that sound like the thing you need, that's a way to guarantee getting a virus

[white parrot]

what, prompt should i use when have mutil same girl

i tried (mutil gir) and 1 girl but no use ===

[fervent thunder]

having multiple of the same person might actually not be possible with just prompting

because it would be extremely rare in the training data

[slim dragon]

yeah, i get it (i think). so, is there any legitimate online portal/web ui for stable diffusion? (...that i dont need to setup myself? actually, ive tried midjourney.com already any others?)

[floral umbra]

Why must comfy be so annoying. I put "mixed color bodysuit" in negative, and it adds it as if positive lol

[fervent thunder]

rundiffusion is ok

you need to be careful because for any popular company now
lots of criminals make similar sounding websites and then put a virus there

best thing to do it always go to things via Google

instead of typing the URL in

[pale latch]

comfy only does what the user builds the graph to do

😮

🎙️ 💣

[slim dragon]

actually i think i got there via google (or duckduck) - as of viruses i hope that (my) linux is so marginal that it's not cost-effective for bad guys to support viruses for this ;) anyway, big thanks

[floral umbra]

Aye, and i set "mixed bodysuit colors" in negative, but they get more mixed in the results than before :P

[pale latch]

yeah i dont get it either. i was just trying to drop bombs

[fervent thunder]

if you are on linux the risk is way lower yeah

[pale latch]

but it's a taste of thigns to come. you report something like that as a bug on auto, and they know it's not user error. you report that as a bug on comfy and you going to have to defned your workflow like a doctoral thesis

[fervent thunder]

this negative is very unlikely to work TBH

[abstract quarry]

to be fair: that's a really bad negative prompt

the way negative prompts work is roughly speaking you generate in each step two images, one with the positive prompt and one with the negative prompt. Then you subtract both images and add the difference (multiplied with cfg) to the negative one

what happens if you subtract multi-colored bodysuits from, say, one-colored bodysuits? Do they get uniformly colored? No, they just get differently colored

(in practice it's a bit more complicated because you do that in the latent space, but the effect is very similar)

[desert dagger]

just an fyi, flux ignores negative prompts unless you set cfg over 1, and then you risk ruining your image

the default setting is 0

[abstract quarry]

negative prompts are always ignored for cfg == 1. Default value should be 1, if you use an cfg sampler

0 would mean you generate the image with the negative prompt ^^°

[pale latch]

this is a smart way to deconstruct it. good call

ease off cfg after half the steps always helps. there's other cfg nodes too that do good stuff

if you're not going to do all the powerful stuff that comfyui affords, just use forge ui, honestly

[fervent thunder]

I forgot the name but there was a paper that looked into the optimal scheduling for making use of negatives

it was different for object removal and style negatives

[trail lion]

so training flux in kohya_ss on their sd3 repo does work, I did 4 epochs and loaded it into comfy, and seems to work just fine. lots of stuff missing still, like it doesnt do samples as far as I can tell, and no GUI yet, so you have to run it all cli

[fervent thunder]

sounds good

[trail lion]

on my 3090 btw, so that's another bonus

this is my 3rd day with this card, already beating the heck out of it 😄

[fervent thunder]

lol

[low moon]

Happy Weekend ya'll.

[nova lily]

So what are the ways to upscale while gaining minute detail?
Ive seen the control net tiled method bot honestly its pretty confoluted to get to work. So im wondering i there are other ways out there

[low moon]

SD upscale. with low denoise. prompt for minute detail. Like detail in materials in your image. Skin, Fabric, concrete, Wood, etc.

Do 2 or 3 SD upscales too with lower scaling not in one go

[desert dagger]

the best way is to 1. save the generated image 2. go to capcut's website 3. go to their magic tools section 4. use their image upscaler

[low moon]

Remember when we just drew. We didn;t have to debug or install the pencil. 😄

And we could do it with a power outtage or no internet. We are so gone. So far gone. Hopelessly addicted to this garbage.

[desert dagger]

remember chewing the eraser off the end of the pencil, and sharpening the pencil so much there was nothing left?

[low moon]

[Applause]

[desert dagger]

the good ol' days are usually not as good as memory makes them out to be

[trail lion]

yah, like the blisters from holding those wooden #2 pencils

[low moon]

Yes I did have blisters.

I wasn't at the mercy or programmers and power outages tho. And everything I made was more special. Now I just crank out thousands of images always frustrated. Would I go back? no, of course not. Do I think this shit is better no absolutely not.

We're on a road to nowhere. Come on inside.

It's never good enough.

It will never be good enough.

[split cape]

yeah but my drawing skills suck

with ai I can be better!

https://tenor.com/view/typing-bald-big-body-laptop-muscle-man-gif-17063750

in between like third grade and middle school I wasn't afraid of drawing though, maybe I just thought I was good, but before and after that I've been terrified to pick the pencil back up

it's actually an immobilizing perfectionism

[low moon]

looool

well

As artists we never were are or will be happy

whether ist AI or pencils

we will never be satisfied 100% with our work

[shy lodge]

fax

whats the prefix for the artisan command?

wait this is payed

😭

[nova lily]

Really? from a website

[desert dagger]

yes really.

[nova lily]

the image-resolution-enhancer?

[nova lily]

thank you will try out

[desert dagger]

go to capcut's website, click on magic tools on the left side bar, and look for Image Upscaler on the top row

right next to video upscaler

[nova lily]

tyty

[hollow rune]

hello all... just wondering, is https://stabledifffusion.com an official site or are they unrelated to stability AI?

[nova lily]

I just tried it, unless im using the wrong feature it's simply upscaling. im not getting any new detail

[desert dagger]

that's what you said you wanted. a minimum of new detail. 0 new detail is the most minium you can get

[desert dagger]

scroll to the bottom of the page and look at the stuff there

[nova lily]

gotcha 👍

[hollow rune]

Sorry... Perhaps I should have asked a different question - is it ok to ask questions about that website in this channel?

[desert dagger]

certainly.

[low moon]

Ask and you shall receive.

[hollow rune]

A couple of months ago my kids were having a lot of fun on that site making images "in the style of" artists/studios that they like.... But today they tried again and the result were really inconsistent and not at all as easy to get as the last time they tried... Has there been some change to the model or some kind of copyright change that's reduced the effectiveness of "in the style of ?" type prompts?

[desert dagger]

probably because they're using flux on the site instead of sdxl

or they're using SD3 8B

[hollow rune]

Is that a setting they would have changed? Or a setting that changed that they need to revert?

[desert dagger]

i might suggest you get your kids free accounts on mage.space and then they can use all the versions of stable diffusion

[hollow rune]

Honestly they had so much fun I'd be happy to pay for something... I will look into mage.space

Is self hosting any of this stuff viable yet? We have a home server running proxmox and a couple of PC's with rtx 3060s

[desert dagger]

sure. you want to look into running comfyUI. go to youtube, look for scott detweiler - he works for SAI. watch his tutorials on setting up comfy

[hollow rune]

I will do that... Thanks so much for your assistance

[low moon]

Gracias por el input.

[lusty beacon]

Anyone has contact to bfl from flux

[nova lily]

So ive been messing around with SDUpscale vs TiledDiff for sake of detail gain. I dont have definitive results yet, But if the trick of **upscaling at seperate levels instead of in one shot ** truly does give detail then SD simply wins because with TiledDiff it seems you're just not able to scale in bits as it just doesn't play nice out of the box (maybe this issue has to to with TD parameters)

[desert dagger]

not this discord

[balmy wing]

https://t.me/dogshouse_bot/join?startapp=8PYp1s3kTTSEkZBzibx3Qw

Who let the DOGS out?

[real birch]

Hello. Can someone walk me through the installation, please? Apparently I'm doing something wrong.

[fervent thunder]

in L2 discord, Treeshark has a workflow that signficantly beats ultimate SD upscale

its very similar but its done as a manual comfy workflow instead of one node

lets you manually re-run tiles and have different denoises per tile

if you want an easier version, someone made one of his previous upscale workflows into a node that is called "Mcboaty"

"Mcboaty" still beats ultimate SD upscale although it is not quite as flexible as the full workflow

having said that, my personal recommendation is not these, but instead this: https://openmodeldb.info/models/4x-NomosWebPhoto-atd with SUPIR after if needed, and using deepshrink/hi-diffusion/res-adapter to get the initial image as large as possible

[modern gull]

guys whose have a comfyui_execution module in impact pack?

[fathom marten]

Yoh, the keypoint to add details is to inject some noise. Cheers.
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c7b06656-bcf9-457f-bee6-6d353325b95b/original=true,quality=100/4249906775-1.jpeg

[lusty beacon]

Check dm

[fervent thunder]

that's rly rly nice

[nocturne pewter]

That's a gorgeous image!!

I need to google info about injecting noise, I have no idea what that means haha

[fathom marten]

With A1111/Forge it's pretty simple, basically a single setting.
All my images upscales are using A1111 with :

• Ultimate SD Upscale (9 tiles whatever the ratio) with ~0.5 denoize, some extra noise, padding of 256.
• Soft Inpainting to be seamless.
  There is a bug with ADetailer, so i can't run ADetailer in each tile anymore, but i was able to. Opened an issue on github for this.
  I posted something at some point to show the difference when using noise:
  https://www.reddit.com/r/StableDiffusion/comments/1dkh5cv/a1111_forge_improve_your_images_with_a_single/

[nocturne pewter]

Thank you, I will read your post later today! 🙂

Not quite sure how it's possible to add in some extra noise or do soft inpaint alonside Ultimate SD Upscale, but I'm sure you'll go all that trough in your post. Looking forward reading it. Thanks for sharing!

[nocturne pewter]

@fathom marten just browsing your images on civitai, man these are fantastic!

[fathom marten]

Ahah thanks !
I've wasted weeks trying to improve this.. But I still can't, as i'm limited with 12gb VRAM..
I can't use even a single CN nor go above x1.4 for the hires part.
Could be better with a better hardware, i think.

[nocturne pewter]

I'm also a huge ZavyChroma fan! 🙂

[fathom marten]

ZavyChromaXL is awesome ! ((:

[nocturne pewter]

It certainly is!!

[fathom marten]

With a better card, i would also downscale the result and run a 2nd pass.
I think it would be wayyy better too.
Frustrating lol

[main snow]

Bro I felt insulted when the app said "lowvram" I didn't think 16gb could ever be called low 🥹😭 lmao

[fervent thunder]

if you like ZavyChromaXL, I do too, I found Leosam's HelloWorld to be a straight upgrade from it

[nocturne pewter]

Nice, will have to check HelloWorld out, thanks for the tip!

[feral eagle]

So whats the best combination to use Flux.1 dev or snell with a 4070 12gb GPU? Currently trying with forge and the dev nf4 version and adding any loras makes it dog slow.

[low moon]

I tried both and Zavy seemed better. 😮

[nova lily]

Thank you, where is extra noise? And I'm still not knowledgeable about the tile size I should be using. It says higher is faster and no seam so I maxed it out but then I heard lower ads detailing more carefully or something.

the double edge sword tho is you need higher denoise for new details but I usually want to keep the face as it was since it came out good

[nova lily]

Comment is appreciated, guess I'll jot this down if I ever learn comf.

[nocturne pewter]

@fathom marten What kind of settings do you have on the softinpaint part when you're doing the upscaling?

[fervent thunder]

Zavy is more contrasty, for those who like that

[fathom marten]

Default. These settings are still shady to me, I never did play with it.

[fathom marten]

True, black is black. I also use a style with "highly contrasted" in it.

[nocturne pewter]

Ok great, I have to test them out and dig some info if needed. I noticed when I'm in the Img2Img tab, the softinpaint is missing and I have to go to inpaint tab to be able to use it.

[sly eagle]

Hello! Do you know a good Stable diffusion template for Runpod?

[fervent thunder]

not sure about runpod but the default vast ai comfy template is good

for the most part runpod just costs more for no advantage, relative to vast.ai

it does let you roll the dice once more by having a different set of servers though

[sly eagle]

It's because I want to use it as locally, I want to train my own models and images

I used LastBen Automatic 1111 and that worked well. I'm looking something similar

[pale latch]

i'm going through pony realism models. people telling me they're great at versatility. i'm finding that they're great at one thing only. they don't know general knowledge. all bikes are harley bikes, no ninjas. all outdoors are the same. no machu pichu. no national parks. They all look the same. all bedrooms look the same, the same neutral colors because certain production sets need it for good lighting. Pony models are really great at one particular situation. People are obvious when they praise it.

they default to pornographic situations and bedrooms if you don't prompt specifically against that, and even then it comes through too

[main snow]

Skill issue

[fervent thunder]

go on Civit.ai, turn on maximum NSFW filters to filter as much NSFW as possible
then browse images in the Pony category
the feed looks perfectly fine

[true canopy]

hmmm, im getting a bunch of size mismatch errors when trying to run flux on comfyui , cant figure it out

[pale latch]

i did that. first two pages were just portraits of characters that seemed like they were about to be nsfw and probably were in the rest of that user's gallery if i looked. i'm already viewing pony though a lens. i'm going to see what i see still.

i looked at pony main model, then the popular pony realism, then cyberrealistic pony. it's all really generic stuff that is really just ONE catagory. PG rating viewing only even, so Civit is intending these images for kids 12 and under. PG13 is worse.

a good case study is cyberrealistic xl and cyberrealistic pony. they're both refined on the same datasets. just one is refined on a standard xl merge recipe where the text encoder layers are intact, and the other is refined on pony's broken but very topically focused text encoders. they don't prompt the same at all. there's no versatility on the pony version. clip skip 2 or 1 being used. try to get the pony version to make machu picchu.

i don't think cyberrealistic is a good photo realism model fwiw. it's good if you want that over produced professional synthetic airbrushed look on everything. It has a neat vibe to it but it's not #1 photo realism. just a good case study of prompt versatility and knowledge

[fervent thunder]

i did that. first two pages were just portraits of characters that seemed like they were about to be nsfw and probably were in the rest of that user's gallery if i looked. 🤔

this still sounds like they weren't nsfw

by normal machine learning standards essentially every checkpoint you will find is over-fit anyway. This goes for base SDXL and it goes for cyberrealistic xl

when people say Pony is over-fit I think "yeah but Flux is over fit..."

its possible, or even likely, that over-fit models will score higher on human eval for image gen

due to the increased ease of making an image with higher image quality, even though image diversity has been lost

personally I would rather have the image diversity which is why I run super low CFG, but I am in the minority

[pale latch]

about 20% actually were nsfw

[pale latch]

flux certainly does have class collapse. distillation problems.

I'm comparing photo refines of pony to other photorealism xl models though

entirely different leagues imo. not worth comparing at this point. flux is still a very dynamic situation that is evolving day to day

you could say, it is .... in flux

[fervent thunder]

lol

[silk condor]

Hey :3

I have a question for those who use IMG2IMG and use the SDUpscaler

If you use it on a folder with several images.. Do you run it without prompts and negatives or do you run the same prompts and negatives as the images?

[pale latch]

humans evaluating will have biases surely. and rating one image at a time will inform nothing about the depth of knowledge the model has. also, Goodhart's law is keen in most situations, which is probably why it's considered a "law" of sorts

[fervent thunder]

oh they have some metrics to measure image diversity in some of the papers

there's also just clip score on imagenet/MS COCO which is normally the prompt following benchmark, but has some relevance here too

[pale latch]

since pony xl killed the clip layers

just murdered them savagely

[fervent thunder]

at least according to Dr Head, a lot of it was done badly during training

I wonder how much of the abilities of the model to follow tags could be done without over-fitting a model too much

their next project is targeted for auraflow apparently

which would have a better chance

as auraflow has shown some of the most amazing prompt adherence examples I have seen from any model

[pale latch]

dr head is right. While the new tenc is smarter towards pony's dataset and captioning, it obliterated all the base model knowledge that came before it and it came to know only the new dataset captioning. a very large dataset but also small compared to the base model and all the refines since then. i believe all the base model latent knowledge is still in there but much like a distilled model "galvanises" it (a metaphor i'm trying to make a thing), the knowledge isn't really linked to any promptable contexts anymore.

he's not the only person saying this stuff. he's just an effective communicator. its a hard concept to communicate especially when there's so much fandemonium surrounding the pony scene

[fervent thunder]

should i get 64gb of ram or 32gb is enough for sdx+depth+ ipadapter+ comfyui?

[pale latch]

32 is enough. 64 gives more leverage

really depends on your vram. 32 can help 16gb vram loads a lot. 12gb of vram a lot less.

system memory is also where all of your browser tabs and extra programs live too. windows will manage all the memory allocation really well, but you'll want to look at your task manager details, sorted by memory use, and check out how much your average system use requires

[trim magnet]

the more u can get the better,same with vram

[pale latch]

make sure its' a compatible kit too. dont just add new 32gb to a different 32gb. confirm they match first

or just get a whole 64 kit

[fervent thunder]

i am getting one pakcage xpg 16gb x 2 3600mhz

[pale latch]

you got 2 16gb right now?

[fervent thunder]

i have 8gb x2 and 2 slots only in motherboard

[pale latch]

yeah. consider shelving those when you put the new 16gb in. 48gb sounds like more but it could slow it down a lot. potentially not. there are things afoot when you mix dimms.

[fervent thunder]

i will sell them or put them in my pc office.

[pale latch]

i collect them and pretend i'm going to do something with them one day

[fervent thunder]

yeah i wish i got 32gb ram directly in 2019

i hope i dont need a 64gb in future.

[pale latch]

i bought 32 in 2020 knowing that i'd buy the same kit again once i could justify it. wanted a new gpu first. now i got 64

[fervent thunder]

i even hate getting 16gb x2 now cuz they are ddr4. next time if i want to upgrde to ddr5 i will buy ram again.

maybe i sshould get a gpu and forget about ddr5 and new cpu.

[trim magnet]

u could also buy a single 32gb stick that way u can get another one in case u need more

[floral umbra]

Better to just save up for 2x 32GB dimm kit at that point than having to lose out on half the bandwidth by not running dualchannel, and also ending up later getting the same model ram stick, but slightly not the same one due to different subtimings.

[trim magnet]

you can buy the same one as long as its the same part number

[floral umbra]

Aye, hopefully they manage to find the exact same part number too. I spent 3 days locating a store that had my kit in stock to get the exact same part number

[trim magnet]

yea thats the hard part i was lucky to find one in ebay

[floral umbra]

Nice :P

I buy mine new as you gotta love 5 years warranty by law

[split cape]

Have any models been released with the clear declaration that it was only trained on royalty free content, or IP fully owned by the creators? Or does nobody give a crap

I was talking in another server about ai and they got all mad and said they don't like the possibility of "art theft" when a model is trained on someone's work without their permission

[trim magnet]

yes the adobe models

[split cape]

well pffff I'm not giving a dime to Adobe

that settles it

[abstract linden]

So much stuff in short amount of time. Starting to get to be a lot 😵‍💫

[desert dagger]

yes, well - if it's okay for you, or any human, to read content, watch content, and learn from it without it being considered stealing, then it's okay for AI to read content, watch content, and learn from it without it being considered stealing.

humans train themselves and their offspring on copyrighted material every day - and don't give compensation to the authors. the AI doesn't use the content it learns from in any way different than humans do. the AI haters are still of the opinion that it's doing mashups like you do in photoshop

[split cape]

that is the exact rhetoric I heard from the haters

that it's mashed up

I see risks and dangers of this software becoming sophisticated and easy to run but I don't throw around second hand opinions for virtue signalling, and that's what it sounds like to me when I hear anybody bash AI, crypto, Tesla etc

you read two twitter posts and now you're a SJW against AI that's great

[rain crest]

How do you tell Flux to generate an image of subjects in the distance? It keeps forcing a closeup.

[desert dagger]

that is what they think, and no one can tell them any diffrent. they are positive that the AI is using images and jsut taking bits out of them

[desert dagger]

good question. so far, i'm not finding any way to successfully prompt that. so i'm using outpainting to add content around the edges, move the subject away, and then crop what I want

[split cape]

I had this exact problem the other day

I was probably pissing off crystal with that one

[desert dagger]

you play with their heads. one of the things I love to do with those people is ask them how they fit a library of billions of images into a file the size of a model. they go red in the face

[split cape]

I actually don't understand how the models understand all they do with a few gigs

it is incredible

[desert dagger]

it's a complicated subject but they aren't storing large images. your image is made up of pixels and each pixel has a lot of data in it. you can actually read the code for it with the right editor. but all that data takes up a lot of room. however if you learn that humans have faces with eyes, nose, mouth in certain spots, and you learn the shapes of them, all you have to do is make a note of that data, and then when you want to make a face, you look at that data and go draw a face. no need to use any specific faces for that, just the information of what a face looks like

it's the same thing you do. watch yoruself think. if i tell you apple, you will have an extremely brief flash of every bit of data you've ever learned that is associated with apple, and then your mind will settle on the bits that have the most weight for you and rapidly construct a mental image of an apple

[split cape]

but what is even more impressive is when the model can reproduce a specific person's face, specific model of car, specific art style from one person. At some point there is a photographic like quality to it. Although maybe it is just beyond the way we can comprehend patterns as people.

like you said it might just be weighted just right to produce consistent images

[desert dagger]

sure, and you can memorize to - though we really dont' want them to memorize like that. but if you show the AI thousands of images of shrek, and tell it 'shrek' with each one of them, it sort of has no choice. you pounded shrek into its head

[split cape]

https://tenor.com/view/shrek-shrek-rizz-rizz-gif-11157824601050747846

yeah that makes sense

[desert dagger]

can't use images in this channel

[split cape]

I know, I keep forgetting

[desert dagger]

that sort of overfitting is actually a bug. we really don't want it to do that. Flux has that problem. it's way overfit for fantasy. IF you give it a token that should bring up random data, and generate a batch of 4 to 10 images, you should get stuff that makes no sense and is all over the place. jsut random junk. however, do you? put this prompt into flux dev: .....,,,...!!?

and see what you get

[abstract quarry]

I don't know what you are talking about. I get an anime image, a pixel art image, nothing related or overfitted

batch size 6 for prompt : ".....,,,...!!?"

1. photo of a dog
2. photo of some people on street
3. digital painting of a winter landscape
4. Two anime characters
5. Photo of an autumn landscape
6. Some weird pixel arts

so much about overfitting bug lol

[desert dagger]

but you shouldn't. you should get random junk. see how much fantasy you get when you leave the prompts (neg and positive) totally blank - you're not giving it any tokens to use, so it's just going to go pull up random data

[abstract quarry]

no, you shouldn't get random junk

that's not how diffusion works

every seed will give you a proper image.

[desert dagger]

do that with sd1.5 and show me what you get. and sdxl

[abstract quarry]

the selection of images is random, but not the image itself

even there you get an image. The problem is that they need cfg to produce good images and for cfg you need a prompt

[desert dagger]

here, i have a challenge for you then. use just the word gun as your prompt, post the images in general with images. then just the word Javelin as your prompt. post them too

[abstract quarry]

flux dev is cfg-free, so it will give you a random image if the prompt is empty or nonsense

[desert dagger]

i want to see you get just a javelin, the way you can get just a gun

[abstract quarry]

I don't even see how that is related to your initial claim. I can try that tomorrow, go to bed now

[wicked mauve]

Hello I heard that they will release sd3.1 soon is this true ?

[low moon]

In 2 weeks yes.

[desert dagger]

you hear lots of things on the internet - take all of them with a huge grain of salt

[potent shale]

Does the processor and motherboard matter as long as I have a good GPU and enough RAM?

I'm probably getting the 4060 TI 16GB and I have 16 GB RAM, but my processor is Ryzen 5 2600x, not sure if that's good enough or it'll drag everything else down

[stone pivot]

anyone use diffusion bee

[nova lily]

It is said to upscale in multiple runs with Ult SD but my second pass seems to come out extremly fuzzy and noisy.

[pale latch]

Man oh man. All day i been playing with Silly tavern's character building stuff. It gives some really crazy leverage over LLMs. I hated them before cause i didn't have tools to turn it into chatgpt, but i do now. Shit's good. Plus also now i know how to run a llama 3.1 instruct model and sdxl at the same time and it'll interact with comfy or automatic for image creation. So you can make really powerful prompting bots. and 3.1 has vision capabilities so you can show it a pic and build a prompt around that.

There's so much more software work that can be done to enhance the tool, but people aren't even tapping the potential it has for chat bots and interactive experiences. i'm talking potential MUD capabilities, not you know, 90% of the "programmed" bots out there are just bonk lol . "you are a step sister with big bigs. you will not not be this step sister" is basically what most of the popular character templates available are. Great potential for stuff though no doubt. People are sleeping on silly tavern

think i might even be able to fit nf4 flux and a more quantized model into vram together. if the memory swapping plays happily

pushing the limits of my 64gb of system memory with beefy 25k context sizes

[fervent thunder]

if the conditioning signal is not very strong you just end up getting results that look like sampling from the unconditional distribution
which might be what's happening with ".....,,,...!!?"

[fervent thunder]

silly tavern style stuff for DND has a lot of potential yeah

[pale latch]

could potentially encode the actual odds mechanics into macros and everything. there's so much macro potential before you even get into scripting or new samplers

and the models that run on home gpus are REALLY good. not gpto good or whatever, but goodnuff

[fervent thunder]

within a year can probably get gpt 4o performance at home

[pale latch]

and then there's the whole databank situation. which is stuff it can procedurally index and queue up into the model prompt as needed. So basically dnd lore books

[fervent thunder]

yeah that's kinda the tricky part

there's some knowledge graph projects and stuff that tries to work on optimising context injection

[pale latch]

right. i tried to summise that. like a textual embedding

basically are textual embeddings

[fervent thunder]

yeah they actually are textual embeddings that's right

[pale latch]

Silly tavern will use it to queue up knowledge and passages directly from source too. Interesting stuff

[wicked mauve]

But I expect a response from stability

Flux took all the market

[abstract quarry]

I have the feeling SillyTavern only uses a small part of its potential, because the chat templates never 100% work

[wicked mauve]

So I would think there will be a new version to fix these issues with sd3.1 or sd3.5 I don't know what they will call yet

[abstract quarry]

like I don't get why they don't have predefined chat templates for the few big llms

[pale latch]

sampler and model type would really matter a lot towards different templates i imagine

[abstract quarry]

My favourite local llm is Gemma 27b

and it was sooo complicated getting their format into SillyTavern. Which is crazy, cause it's the most simple format you can think of

this software is just overly complicated to setup

I often prefer to just run Ooobabooga

and insert the template manually

not an option for roleplay, of course, as you don't want to bother with correct syntax

[abstract quarry]

I still have no clue what you are talking about 😂
Both prompts, "javeline" and "gun" works fine, except that in some cases flux makes an arrow instead of a javelin.

[frozen birch]

#😊｜co-creators

[abstract quarry]

???
Just do Google image search for "gun" and for "javelin". It's the same. If you prompt for gun you get a gun, because there are a lot of websites with advertisements showcasing their guns. If you google for Javelin you get either historical or fantasy images or olympic athletes, but you won't get a single image with a product placement of a javelin

Flux is not buggy and overfitted - this is just how diffusion works.

[fervent thunder]

I don't rly understand the point crystalwizard is making in this conversation
regarding gun/javelin or the prompt ".....,,,...!!?"

but anyway I would say flux is slightly over-fit on the midjourney/instagram/aesthetics look

it could be Skill Issue on my part but I can't get base flux to output a look like RealvisXL or Realistic Stock Photo

what I can easily get out of flux is images that resemble either midjourney itself or SDXL checkpoints+loras that are mimicking midjourney

[abstract quarry]

I agree on that. But that seems some kind of fine-tuning or dpo problem. The model has the capability for a lot of styles, it's just difficult to prompt them. Seems you need loras for that until someone makes a general finetune of the model

[fervent thunder]

But that seems some kind of fine-tuning or dpo problem. this is almost certainly the case yes
the base model could have done Realistic Stock Photo and then they fine-tuned that ability out of it 😿

I wonder if it was an attempt to make it worse at deep fakes in order to deflect regulator attention

a bit like how Dalle 3 can't do realism either

[fervent thunder]

hey all, hows things going

[robust needle]

I'm wondering if it's my lack of experience or the fact that my checkpoint is "v1-5 pruned-emaonly-safetensors" that I'm getting poor results.

I followed a recent tutorial to get it installed with AMD graphics

[fervent thunder]

its not a great checkpoint

[deep coral]

hello guys, what is better for GenAI? 4070 super or 4060ti, ipeople on reddit recommend 4060 ti because of 16gb vram but most of benchmark I saw 4070 super is faster than 4060ti? any tips guys?

[fervent thunder]

vram is everything

[deep coral]

why on benchmark 4070 super is faster though?

[robust needle]

Does more vram give you a better picture or does it make the image generate quicker?

[fervent thunder]

because speed depends on other things

but slower with more vram is better

having more vram lets you run things that cannot be run on the lower vram card

[deep coral]

oh ok

is less 4gb vram deal breaker? ill just do comfyui+flux, or it is not future proof?

[fervent thunder]

vram is the most important thing of all

and every amount of vram you increase, will increase the amount of things you can do

[hasty hornet]

Anyone knows a model \ lora \ workflow \ prompt or anything that can generate images I can safely downscale to 64x64?
I want to make few quick items for a game, but so far downscaling not doing me good...

[tacit mountain]

what is the 'Guidance 4' I see with some loras

[fervent thunder]

its for flux

[abstract quarry]

yes. You want to have as much vram as possible. With a lot of tricks and optimizations you can run everything on 6gb, but it will be slow. 12gb is a solid foundation

[fervent thunder]

Why does flux do this? The images though different prompts are very similar in the composition / background, same angle, same plate, same background.
Check general with images for examples.

This is not necessarily bad. But i need to know
Is there a way to get more diverse results (without modifying prompt) ?

they just trained it like that

and then distilled it

its how it is

is it possible to train a lora / fine tune which instead of making photos a certain way is specifically designed to modify input photos a certain way ?

kinda like inpainting would. Atm i havent found any workflow that works for flux in the same level as sdxl inpaint on fooocus

[pale latch]

instruct to pix models. there's lots of that stuff out there

https://stable-diffusion-art.com/instruct-pix2pix/

[abstract quarry]

are you sure that you use different seeds for this images? They look very much like same-seed

[pale latch]

aesthetic guidance and it's also really good at maintaining prompt adherence, that is, how it understands your prompt

the training data could be a lot of phone photos of food that are up close like that too. and phones all have a very similar filtered aesthetic

[hasty hornet]

similar is one thing, but damn, look at this
#🏞｜general-with-images message

[pale latch]

thats what i was looking at. thats what my entire explanation was about

i mean, you asked for insight and i offered some. then you argued with it. nevermind then.

[fervent thunder]

Is nvidia a100 good for stable diffusion or it is good only for text ais?

[abstract quarry]

I'm pretty sure he just used the same seed for all these images

then it's natural getting almost the same image back every time

[trim magnet]

you pretty sure or unsure?

[fervent thunder]

Any way to train flux with it yet?

[pale latch]

probably. unity put out an instruct adapter for sd3 already

their project might lend you some clues

https://github.com/unity-research/IP-Adapter-Instruct

[fervent thunder]

Thanks will check it out

[pale latch]

the SD3 technical report even talks about tests they did, refining 2b on an instruct dataset. that's worth reading too, but i'ts just a paragraph in the whole report

[fervent thunder]

true, that was the case. The seed was the same. When using from UI it changes after each generation, but i had saved the workflow and loaded each time to make calls via api (so seed in the json workflow was the same each time).

[full lark]

Do anyone know how to make sd forge use random seeds? if I use seed -1 the seed in a batch of images would be xx3, xx2, xx1...etc

[trail lion]

-1 is the built in way, you can put random numbers in there of course, generate them some other way

[white parrot]

can anyone help me guide those thing #📝｜prompting-help message

[tacit storm]

hello

[quartz siren]

yes it's very good. Probably one of the best gpu's for all types of machine learning models including stable diffusion.

[wise ginkgo]

elaborate?

[pale latch]

i downloaded this new model i saw that didn't look too bad. was for roleplaying so i assumed you know, dungeons and dragons or like, rpg games.

Nope it's trained on roleplay from yahoo chat or something. gonna delete that thing. the future is now hey?

i was thinking maybe i could hook it up to skyrim. thats probably what it's meant for anyways even

[regal snow]

Wow - I was just writing an "Issue" for Forge and when I clicked "Send" it brought me to 404 page.
I go back and... Illyasviel seems to have removed Issues from the repo lol

Oh dayum there's actually a big news thing on the front page about it

[pale latch]

Was it an extension not working? People were really abusing his repo. Dude gets a lot of shit

[narrow kernel]

The forge dude was roasting city96 for his gguf formats in one of the PRs. Like yeah, let's roast him for making the first working gguf format for dit based models and then complain about having to spend 20 minutes writing code for the format. I hope he was just trolling

[low moon]

Come one we don't need any more popcorn moments. AI is the future. If you want popcorn watch Hollywood burn.

[trim magnet]

do not worry the internet officer is already here to settle this

[regal snow]

Nah, was making an Issue regarding the VAE / Text Encoders (in Forge are now handled together as a “list of extra modules”): PNGInfo send to t2i / i2i currently ignores the “Modules” info. Also override_settings (API param) needs updating to accommodate

Tried figuring these out but it’s over my head

[lyric pumice]

hey all! I've joined to better my creation of funny cat images!

[low moon]

A honorable goal.

[spring yarrow]

lol I love that

[main snow]

idk man

that profile pic

is cursed lmao

you might be too powerful

to use ia

[lyric pumice]

lol i thought it was funny

its not mine i got it off google

[main snow]

then whoever it is gotta get off the kitchen damn

bro gotta stop cooking

[deep coral]

hello guys, can I ask whats better for comfyui or stable diffusion, 4070 super(12gb), 4060 ti(16gb). on benchmarks I am seeing 4070 super generate more images and generate faster than 4060 ti despite of only having 12gb. but most people on forums,reddit, etc. suggest VRAM is king. but how? what other factors should I consider except speed?

[trim magnet]

4070 is better,the 4060 is a crippled card

[hollow current]

why is the 4060 crippled? asking because I'm running a 4060Ti at the moment, and thinking of getting a 2nd one to help with generation of images or getting something better if I can afford it.

[trim magnet]

it only has a 128bit memory bus so it has less memory bandwidth than the RTX 2060 Super

[hollow current]

so would a 4070 / 4070 Super be better, then?

I can get a 2nd 4060Ti for $400, but 2 x 4070 Super would be $1300

[trim magnet]

yea or if u have the money u could also buy a 4080 16gb

or if u buy used a 3090 ti for like 700 usd

[hollow current]

gotcha... didn't realize that the mem bus made that much of a difference, still learning a lot about doing Image Generation

[trim magnet]

well it doesnt it only happens when the memory is full

so it happens when u play a game at 4k or if u doing a very high res / very high img batch size gen

[hollow current]

I'm mostly doing SDXL at 768 x 768, then upscaling to 1080p or 4k

as well as using it for other things like LLM chatbots and such, separate from doing image generation, that is

[trim magnet]

that should work fine,you would fill all memory if u were like training lots of imgs or using a very big model like the big sd3

[hollow current]

yeah, definitely not training, just using models to generate images... I might try training at some point, but not quite there yet in my journey into things

I'm mostly interested in cutting down time to generate images. I've had some landscapes that I tried doing that took an hour or more to do 4 batches of 8 images in A1111. Hoping to test out the same things in Forge to see if there is a difference over the next few days.

does running a 4070 and a 4060 in the same system help with generation times?

[trim magnet]

no u can only use one card to gen stuff , if u wanna use a second one u would have to open another instance of a1111

[hollow current]

would something like this not work with A1111, Forge, Comfy, etc? https://huggingface.co/docs/accelerate/usage_guides/distributed_inference

[trim magnet]

that isnt implemented in a1111 there was an old ui that worked with two gpus but its dead now,but two or more gpus works in training

[hollow current]

gotcha, was just reading into multiple GPU scenarios and came across that

[fervent thunder]

ye this is literally the GPU I use these days

it makes 300 steps viable as your casual generation setting

[fervent thunder]

Diffusers, particularly their Jax version, is the most optimised out of the common CLI/GUIs for multi-GPU

[short tapir]

anyone abel to help me?

i was doing jsut fine making images, but now no matter what lora i use or what i do the faces are messed up. but i didnt xchange anything

[desert dagger]

you might post in #🤝｜tech-support

[mystic star]

o/

[shell tendon]

city96 commented Aug 25, 2024

    and I hope this is the last time that people randomly invent new state dict key formats to waste time of other developers.

This is the llama.cpp standard GGUF format for LLM models, which natively supports the high quality T5 quantization methods used, and which both of our codebases use as reference. Likewise, the standard llama.cpp build was used to create those quantized models without any patches/changes.

I think it's rather unprofessional to put comments like this in code/commit messages while dismissing the work of the llama.cpp team as "randomly inventing" formats and find this rather diasppointing.```

[desert dagger]

am i reading that right? is llama written in c++?

[narrow kernel]

Yeah that's it. City96 called him out on being a dick

I was under the impression that what city96 did for the ggufs was nonstandard or something though

[mystic star]

Who's City96?

[narrow kernel]

The guy that got ggufs to work with flux

[desert dagger]

did he have a reason for this?

[narrow kernel]

Not sure, but he's making it sound like he followed the llama.cpp way or something

[shell tendon]

imo it doesn't even matter if it was standard or not

[desert dagger]

wonder if he didn't have a choice

[desert dagger]

apparently it matters to someone

[narrow kernel]

Well for performance reasons, likely not

[shell tendon]

that tone has no place in a civilized discussion, it's disrespectful and rude

and is unnecessary

[desert dagger]

did you forget it was taking place on social media?

[shell tendon]

yeah.... siiiigh

[desert dagger]

it probably broke forge or something

[narrow kernel]

Llama.cpp is in cpp and is super fast. If city96 wants to bump up gguf performance, llama.cpp has to add stuff to their repo for it. So what city96 has working right now isn't perfect and isn't the most performant, but it works great. Once stuff can be added to llama, you'll likely see massive gains in performance

[fervent thunder]

why is there always drama when it comes to diffusion GUI devs

[desert dagger]

revision: why is there always drama

because programmers are artists - and artists fall apart emotionally at the drop of a hat

[fervent thunder]

it does a lot of harm cos it splinters the ecosystem

[desert dagger]

granted, not that it's ever not been the case.

all the way back to jobs and gates stealing each other's tech

[narrow kernel]

Hotheads, same as any other field

Tech ones have some of the hottest though

[desert dagger]

that describes the human race

this is why AI will never want to take over the planet - then it has to deal with humans

[nova lily]

Ultimate SD smaller tile = better detail that true?
the page doesnt say so but videos do

[fervent thunder]

not rly a fan of tiled diffusion upscale

[desert dagger]

go with what works?

[fervent thunder]

funny things start to happen

[pale latch]

depends on the model really

[fervent thunder]

my usual upscale recommendation is deepshrink/resadapter/hidiffusion -> HAT/DAT/ATD -> supir

SDXL can get to 5k resolution using 2 passes of deepshrink/resadapter alone]

[stiff valve]

can anyone help me with a good comfy workflow for a consistent character? I tried making a lora with civit but when i use it, the quality is garbage, definitely not worth the 4k Buzz it cost to make it...

[fervent thunder]

style transfer IP adapter, instantID, control net depth and canny?

[stiff valve]

i'm not quite that good, lol. I've tried a few pre-made workflows that do ok, but still use more than half of my prompts for face and body style to stay consistent

[fervent thunder]

for the four things I mentioned you pretty much just wire them up in the way their examples show

[stiff valve]

ok, i'll dig into those and give it a shot, thanks

[fervent thunder]

🥳

[mild yacht]

How i can get black background in img2img from my white background image ?

[fervent thunder]

this isn't rly what img2img is about

[mild yacht]

wdym

[fervent thunder]

img2img starts with your input image and then does diffusion with that as the starting point

if you want to change stuff its not the best choice

[mild yacht]

oh i missunderstand

you know how i can do this trick ? becasue actually i'm generating good (2D sprite) image for my game with a white background for png porpuse but i'm having issue is when i try to remove the white background i get a white line in edges around the whole design i tried with Photoshop and krita and some website i can't really remove the white background without losing some pixels i'm stuck lol looking to solve more then 24h

[fervent thunder]

yeah

the way is to use tools to get aspects of the image

style and composition transfer IP adapter

various control nets

Instant ID if there are faces

with these combined you don't need img to img

[mild yacht]

i didn't understand how to do it is it hard ? (i'm new on ai)

[fervent thunder]

if you follow the instructions you should get there

[fluid belfry]

I'm trying to use Virtual Tryon (IDM VTON). It takes input of a person and input garment and then outputs the final image with the person wearing the new garment.

It also supports text prompts, but I'm unable to see any difference with the prompt.

Does anyone know how to increase the prompt strength? Or have some idea about 'classifier with guidance' does.. it'll be helpful!

[abstract quarry]

I mean, good tools like photoshop can remove the background quite well. There shouldn't be lot of white pixels left

but you can do the following:

• remove the white background
• paste the character into a black background
• ideally, you remove a bit more aggressive the white pixels even if you loose a bit of the sprite
• then do inpainting or img2img to repair the sprite (for inpainting, only mark the edge of the sprite to repair it)

[fervent thunder]

there are other deep learning methods than IDM VTON that are also specialised towards clothing

maybe some of those would work better

I don't follow this area too closely

you might be able to get some of the way there with style transfer IP adapter and canny edge control net

as IP adapter tends to be able to transfer clothes fairly well when doing style transfer

if you are willing to apply IP adapter block-by-block you can get MUCH better results than normal, and with less concept bleeding, but its a lot of effort

Wow that got a bit heated up there

[sudden ruin]

Im glad he pinged me, easiest ban of my life

[fervent thunder]

never seen a spammer ping the mods before 🤔

[regal snow]

If you use Forge, it has an extension called layerdiffuse which can make images with transparent backgrounds - recently updated to also work in img2img (incl inpainting). I know you said “black backgrounds”

[mild yacht]

you know what i just download it and it won't generate also in txt2img xD it say "RuntimeError: shape '[640, 768]' is invalid for input of size 1310720" i tried to change the resolution to 1024x1024 same error

[fervent thunder]

dayum,you re rich

its under 2 dollars per hour

on vast . ai

[fervent thunder]

oh

i thought you actually had an pc with a100

its actually the norm in machine learning field and industry to use cloud

using home PC is just a current trend from reddit mostly

[fervent thunder]

but most funny thing som eguys tried to run games on a100

and failed

yeah it might not actually be that great for games

drivers are everything

they just spent 24k $ for nothing

world most expensive lint roller

A100 and H100 are extremely over-priced
they aren't that much faster than a 4090

but Nvidia monopoly sadly

[fervent thunder]

soon AMD...soon

yeah would be good

[fervent thunder]

because like rn 4090 in my country costs 9x average monthly salary

wow

okay yeah

[regal snow]

I’ll have to try again later, last I checked it was working. Make sure you’re using Forge and also up to date including the extension

I’ve also only tested it using XL models

[mild yacht]

yep it's updated with the extention and i tried several models XL i think there is something to install is missing for me

[regal snow]

I’ll check too in 2 hrs and ping ya

Sucks if broken again

[mild yacht]

thnx nah i think the problem is on my side

[mild yacht]

it work now lol idk what was the issue i just tried several things and restart my computer 🙆

[fervent thunder]

The world wide fix... Turn off, Turn on lol

[mild yacht]

🤣

[fervent thunder]

Never seems to work on my bottle of Jack Daniels tho. I always have to go and buy a new one lol

[mild yacht]

It work only on artificial 😄

don't try it on bitcoin pls

[fervent thunder]

it worked lol i posted in general with images

[weary talon]

Hello, my name is Jack, I am an Architect. I really like how SD services works with shadows, highlights, and lighting in finished images, but I don't want the AI to change the objects in the image. However, I haven't been able to achieve this through the settings. Is it possible for me to upload a finished image, and the AI would only work on the shadows, highlights, and lamp glow, without changing anything else?

[fervent thunder]

its doable with lots of advanced techniques

and even then it will not be fully reliable

it can't really be done with an easy plug-and-play solution at this point

control net, IP adapter and Lora are the things to look into

[livid spindle]

hi

[gentle dagger]

hello~guys? whick is the best upscaler model in your opinion

[dry pulsar]

Hallo liebe community ich möchte ein kurzes Video ca. 15 bis 20 sekundn erstellen , könnt ihr mir eventuell dabei helfen? wär escht klasse. Mein Englisch ist leider nicht gut und ewenn ich in deutsch korrespondieret könnte ,wäre das auch super, danke schon einmal vorab

Hello dear community I would like to create a short video of about 15 to 20 seconds, could you possibly help me with this? it would be great. My English is unfortunately not good and if I could correspond in German, that would be great too, thanks in advance

[stiff valve]

got my comfy a bit more...comfy, but its still not where i expect it to be

[eager bison]

Hi, I asked this question on Huggingface. Going to try here as well: I am using stable-video-diffusion-img2vid-xt, and I experience that the videos I create starts moving backwards about half way through (not all my videos). How can I do something to avoid this?

[floral umbra]

I seriously need to get a few 4TB nvme's for my models Almost wanna go dramless as they will just be loading models 90% of the time anyways, so write cache is a non-issue lol

[fervent thunder]

Lol, im glad i read this comment. Made me appreciate my 300$ 4tb ssd investment

[jagged mason]

i just downloaded the sd3_medium.safetensors from huggingface and creating a simple image from an example prompt generates only a dark blurry junk image, what am i doing wrong? i am using a amd fork of automatic1111

[hushed flower]

hello guys

[graceful bramble]

So I have an odd question. If I train a lora with 30 pictures of a red background and 1 picture of a blue one. Will it make a slightly pink background or will it just have a 1/30 chance of blue or will it just ignore the blue and go with the majority?

[rich umbra]

yo

[gray vapor]

Hi:) What is the best method for regional conditioning right now? (Both by prompts and by input image). Ive used many methods before, all of them very unreliable, with a lot of bleeding. Best one was ipadapter attention masking.

In my usecase, i need to heavily respect the mask boundaries. (Architectural renderings were i need to specify materials per mask)

Most methods use the regions as a "guide", but the shapes of the masks are not respected 100%

I still need to try densediffusion tho

[oak latch]

can i pay someone to make a lora for me?

[quartz siren]

Lower step(8 and 16) version of flux
https://huggingface.co/ByteDance/Hyper-SD
From my very limited testing its much better then flux schnell and merges while having similar and(sometimes) better prompt following then dev.

[eager kite]

can anyone help me to run SD smoothly on AMD GPU?

[desert dagger]

what sort of lora?

[desert dagger]

go to the #🤝｜tech-support channel, look at the pinned posts, read all the guides. that'll get you started. then post in that channel if you need additional help

[eager kite]

Thank you

[oak latch]

like sdxl or what the lora is on?

[desert dagger]

what the lora would be on

[floral umbra]

Aye lol. Sadly, cheapest 4TB here is 380 :P

And make sure it's gen 4. As load speed from a 1TB gen 4 SN850 is 2.4GB's when loading models, and a gen 3 is half that.

[fervent thunder]

got Kingston NV2 4TB M.2 2280 NVMe Internal SSD , its even 230$ on amazon rn
I need to get more ram though 32gb not enough anymore

[floral umbra]

Oh neat! And yeah, can testify for that lol. What i find most annoying though, is that i see pagefile gets filled instead of using free ram whenever i go above 24GB vram

Got 64GB ram, so 32GB free most of the time lol

[fervent thunder]

I envy you, not for long tho xd

[floral umbra]

Aye :P

I already got 2x nvme's filled with A.I nonsense lol. As there's always one model that can give that visual, while an entirely different model gives that visual, and that's 2-6GB wasted per slight change lol

[pale latch]

i used to hate paging to ssd's. i was afraid that they'd wear out from write death. but i guess to do that you need to do constant write operations to it for over 10 years straight. It took me till i got my gen4 m2 to start allowing windows to do page files on the ssd.

gotta say wowweee it was a boost . don't be afraid of hammering your ssd. all i'm saying.

metaphoric hammer not a real one

i still got my image generation folder on a bulk 3TB HDD . i'm getting annoyed with it tbh

scrolling the folder in explorer and it's so slow

[narrow kernel]

I had a high end WD nvme lose 5% life in a handful of months of paging it hard. I moved my page file to a regular SSD, so it's slow as hell in comparison(like 1/10th the speed)

But the way I have things set up now, it rarely ever touches the page file now

[floral umbra]

If assuming NVME's were capable of always constant 7GB's of gen 4's, then 11.9 hours of writing data to it, and the cheapest nvme will be dead.

But seeing as after 8 months, my server's cache nvme of gen 3 which has 600TBW, is just at 134TB, and it is holds the active data for my docker containers, and used to write 150GB backups daily for over a year until i changed that to once a month.

And we probably don't write much to the nvme vs a server's activity lol

[pale latch]

yeah ultimately, you shouldn't be paging much either. good point

[pale latch]

i was really scared of hammering writes to my SATA ssd that i first got. they did used to suck with a lot of writes for sure.

[floral umbra]

Only SSD I've ever killed was a 64GB Kingston SSDnow v100. As neither win 7 nor i knew that you shouldn't defrag a flash storage lol

10 months of defragmenting killed it off lol

And I defragmented my steam deck's 1TB nvme due to converting from ext4 to btrfs for compression, all data had to be moved so that the moved data would get compressed in the process.

So 950GB of games when defragmenting them wrote 27TB worth of moving data back and forth lol.

[pale latch]

lol. i used to love defragging hard drive day. cause it was a legitimate improvement to the system performance for a few weeks.

[pale latch]

oh nice. compression doesn't hurt the performance to much hey?

[floral umbra]

In the matter of fact, opposite. It only uses more cpu when writing due to compressing the data, but when decompressing, file is smaller by up to 30%, thus reads the files much faster,

I had to go away from BTRFS though, as it doesn't like to be shrunk/expanded when messing with partitions to install windows on a 256GB partition for the deck lol

Sadly windows can't use BTRFS or ext4, as windows is windows after all

[pale latch]

i got the 1TB deck. might give that a look see

[bleak matrix]

Good afternoon, everyone! How are we all today?

[solid kindle]

good

how about you ?

[bleak matrix]

Pretty tired today, but good! What are you up to?

[solid kindle]

Just browsing some news about the nvidia 5000 series and trying to understand what an epoch is lol

cant quite grasp my head around the concept

[bleak matrix]

You know, I haven't even gotten around to considering what my next PC will be

[solid kindle]

haha, when was the last time you built a PC ?
the next gen intel CPU and nvidia GPU have AI gimmick so i think its a good time to build a new one ? since they are rumoured to be released soon anyway

[fervent thunder]

😆

[pale latch]

that wrong or something?

its how it was broken down for me. thought it was a simple way to think about it. mb

[trim magnet]

needs more balls

[pale latch]

when a seasoned ai expert laughs at something i said, i take it back

[trim magnet]

what kind of seasoning tho

[pale latch]

lawrys

[trim magnet]

😋

[pale latch]

https://tenor.com/view/mike-lowrey-gif-8186790

[trail lion]

man, some of the stuff flux can do is absolutely crazy. dynamic angles, proper depth perspective like someone reaching out toward the camera where their hand is really close and their body further away. and with how trainable it seems to be, I just dont see how it doesnt obsolete everything that came before it

the 1.5 pony cult aside

[desert dagger]

depends on what someone wants out of the tools they are using. some people still use a spoon to mix a cake with

[lethal gorge]

anyone know any command line auguments to improve speed of flux on zluda?

[desert dagger]

@warm junco probably does. you should ask this in #🤝｜tech-support

[lethal gorge]

thank you!#

[floral umbra]

How much does flux models use of vram for you guys? As even flux dev fp8 t5xxl_fp8_e4m3fn clip, plus clip_I at 512x512 which supposedly allows even 12GB vram cards, just tanks my 3090's memory instantly and drags over to swap o.O

[fervent thunder]

what webui do you use bro?

[floral umbra]

Comfyui

https://image.duckers-web.site/hEja1/RAmEmido03.png Here's what my generation attempt does

A random flux workflow i downloaded

[fervent thunder]

your 3090 is very capable of running it. your 3090 should be 24GB vram

you may have the settings slightly high can you screen shot the workflow

[floral umbra]

Which it is, and as my screwnie in my link shows, 24GB plus some shared nvme storage lol

[fervent thunder]

screen shot your setting bro

and post them in the general chat with images

[floral umbra]

Unless you mean other settings, that's the whole workflow

Unless my link doesn't show for you? As i posted another one right above your comment

[floral umbra]

:i can't read 2 sec lol

[fervent thunder]

called you in the other chat

[desert dagger]

there are 2 generals - this one which doesn't allow images, and the other which does.

[pearl field]

在discord里加入Stable Diffusion 社区后，如何开始绘画

[floral umbra]

Aye, i know. I just keep forgetting

[pulsar venture]

there is someone that could help me to install ipadpter?

[stiff valve]

i'm using the Model Manager in comfy and want to hide some of the checkpoints and LORA's, there is a "Show NSFW" button that seems to do nothing, how could I tag stuff to fall under this action?

[fervent thunder]

what checkpoint are you using when the NSFW does noting?

[stiff valve]

sorry, i mispoke, its the comfyspace node, not model manager.

https://github.com/11cafe/comfyui-workspace-manager

I have over 100 checkpoints and LORA's in the folder, for obvious reasons, i can't share a screenshot, lol

[fervent thunder]

ie, if your using SD3 checkpoint then the NSFW control wont do anything as the checkpoint wont allow it

[stiff valve]

it's not a checkpoint issue, its a node/extension

[fervent thunder]

you asked i'm using the Model Manager in comfy and want to hide some of the checkpoints and LORA's, there is a "Show NSFW" button that seems to do nothing, how could I tag stuff to fall under this action?

[stiff valve]

i moved the convo to the images chat, look there and you'll see what i'm talking about. its not a generation issue

[livid spindle]

hi

[faint heath]

hi

[fluid dagger]

hi

[outer frigate]

i have a photo that i need to edit using ai. where i can get it done?

[desert dagger]

what sort of photo and why does it need to be done with ai?

[outer frigate]

i need to change cloth on my photo

[rugged vale]

the new pixel 9 runs some kind of image generation model on device, anyone know if you can swap it out for another one?

[unique cairn]

I want to make some kind of retro pop art with AI

what do you guys reckon

[fervent thunder]

leosam's hello world has a decent chance of having that in the training data

if you get 20 or so images of retro pop art and put them into IP adapter style transfer (average the embeds) then I think you will get there

[gray vapor]

Is there an attention coupling workflow for Flux out there?

[fervent thunder]

Is a image 2 image / style transfer trainer for flux ?

Atm i will try and train a neural network with image pairs, to predict 2nd latent based on the first and do image to image that way. Not sure if anyone tried it before

yeah both

[fervent thunder]

Where is the img2img trainer ??

you can just use the original flux safetensors file, but swap the empty latent out for an image, and then play with the sigmas to get the result you want