💬｜general-chat

[hot vortex]

nope

[sage reef]

nvcc command works on windows

assuming you have cuda

[rugged mirage]

I do weight emphasis, havent noticed normalization messing it up

[trail lion]

I dont bother with venv since I use docker, so whether it's conda, pip, or python, it's all just part of the image

using venv or conda env's would just complicate it

[astral goblet]

you'd likely notice it "messing things up" when you swap between systems that implement emphasis differently though

[earnest lichen]

I'm on stableswarm 90% of the time. but when I upload examples of my loras on civitai, I temporarily switch to a1111 so its easy to repeat my gens.
but its always such a pain x_x

[sage reef]

but then again depends how you installed cuda toolkit, cause there is an option to not install nvcc i think

[astral goblet]

Ever tried fooocus? lol

[sage reef]

isnt foocus limited to sdxl stuff tho?

[earnest lichen]

ive not. but that looks fun XD

I'll try it

[astral goblet]

its VERY opinionated and does even more in teh back end to improve the generation

has a gpt2 prompt expansion model turned on by default too

[earnest lichen]

I'm just curious how it handles my loras XD

[hot vortex]

I google it, CUDA is enabled on GTX 1650 and I also had to disablehardware acceleration

[trail lion]

so you're a fan of it? it just seemed like alpha-ware stuff to me

[hot vortex]

but I don't remember installing Cuda toolkit

[astral goblet]

fooocus isn't such a bad UI for demonstrating to new people and onboarding them. I even jump into it when all i wanna do is prompt. so many less knobbies to operate

[earnest lichen]

its really good for proper testing of loras & finetunes. make gigantic 2k megagrids, automate a whole bunch of things, and super easy to view everything you've ever generated, quickly and efficiently

[charred mesa]

thankfully comfy has an advanced sdxl conditioning node to manage normalization (sorry I was lazy to scroll through the convo)

with like 3 difference options, length, mean and both

[earnest lichen]

tl:dr, I want the reverse. I want a1111 to not do prompt weighting

[charred mesa]

yeah

[hot vortex]

I need this toolkit, right? https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=11&target_type=exe_local

[charred mesa]

is there like a specific usecase or situation where it may help besides newcomers

or its literally jsut for assisting newbies

[astral goblet]

it streamlines the workflow really well, so maybe for someone it may help them get into a flow state more easily. i love when i'm in a flow state. there's no one solution for getting there though.

[charred mesa]

hmmm

[harsh skiff]

are there anyone knows where can I find (last-state) file when I stop the training in kohya ss?
I couldn't find it the model folder

[earnest lichen]

prompt weighting?
yeah. It saves you from ever having to do downweight the rest of the sentence.
(a man walking:0.5) (down the street:1)
is not the same as:
a man walking (down the street:2)

[charred mesa]

ahhhh

[harsh skiff]

please help in need that

[charred mesa]

so yeah its literally in the name, normalization

[charred mesa]

don't you have like a set folder for saved model files

[astral goblet]

i was imagining he meant normalizing weights in the network.. i was like "ooh sounds spiffy" but i was over thinking it. emphasis weights rather. still a spiffy convo

[charred mesa]

ah

[harsh skiff]

I couldn't find it in any of the training folders

[earnest lichen]

yeah my bad. I worded that real bad XD

[trail lion]

I told you earlier, it goes to where your checkpoints go, if it's not there, something is wrong with the config...maybe look for errors in the console

[harsh skiff]

I couldn't find any errir

error

my training is completely fine

[sage reef]

@hot vortex you have to figure out the best cuda version for your gpu, do some googling, cause latest might not work

[trail lion]

yah, but it's clearly not honoring your config to save it, so maybe do the 'print training command' thing to make sure it's executing that option

[hot vortex]

ok, thanks

[harsh skiff]

Can I share the training command so you can see it?

[trail lion]

sure, maybe others would care, but I'm cool with it 😄

[earnest lichen]

Im assuming you're on linux?

if you're ok with setting things up new, I can heartily recommend pop!os. It solves all your cuda and driver issues

[hot vortex]

nope, on Windows

I'd like to go to Linux but it'd need the apps I need.

[earnest lichen]

on windows? O_O that's not normal...

[harsh skiff]

I have another question, if I want to move my kohya ss or automatic1111 from a drive to another, how can I do it? the UI is linked to the old drive.
it just works with comfyui because it's portable.

should delete the venv? or there is something can I edit in the venv will solve it?

[shell tendon]

yeah, i've found the better way to "weight" is to move things earlier in the prompt, or repeat them in a different form

[earnest lichen]

@hot vortex check first if your gpu is still supported x_x could be that any application in the pipeline doesnt support it anymore

[hot vortex]

it's GTX 1650, it supports CUDA

[earnest lichen]

not just cuda, also pytorch, etc...

[hot vortex]

it's just ... was the link I provided above the CUDA toolkit you've been talking about?

I wonder what version of pytorch you're using, if you use ComfyUI

[earnest lichen]

CUDA 7.5 is the max supported

just checked

[hot vortex]

because it shows CUDA error when I installed version 2.2.2

🤷 AI Copiot says 10.1

[harsh skiff]

accelerate launch --num_cpu_threads_per_process=2 "./sdxl_train_network.py" --network_train_unet_only --bucket_no_upscale --bucket_reso_steps=64 --cache_latents --cache_latents_to_disk --cache_text_encoder_outputs --caption_extension=".txt" --enable_bucket --min_bucket_reso=256 --max_bucket_reso=2048 --full_bf16 --gradient_checkpointing --learning_rate="0.0012" --logging_dir="F:/AI/train/MalsaadM-2/train/log" --lr_scheduler="constant" --lr_scheduler_num_cycles="3" --max_data_loader_n_workers="0" --max_grad_norm="1" --resolution="1024,1024" --max_train_steps="54450" --min_snr_gamma=5 --mixed_precision="bf16" --network_alpha="1" --network_dim=256 --network_module=networks.lora --no_half_vae --noise_offset=0.1 --optimizer_args scale_parameter=False relative_step=False warmup_init=False --optimizer_type="Adafactor" --output_dir="F:/AI/train/MalsaadM-2/train/model" --output_name="MalsaadM2-XL" --pretrained_model_name_or_path="F:/AI/Comfyui/ComfyUI_windows_portable/ComfyUI/models/checkpoints/SDXL/sd_xl_base_1.0.safetensors" --save_every_n_epochs="1" --save_last_n_steps_state="10" --save_model_as=safetensors --save_precision="bf16" --save_state --text_encoder_lr=0.0012 --train_batch_size="2" --train_data_dir="F:/AI/train/MalsaadM-2/train/img" --unet_lr=0.0012 --vae="F:/AI/Comfyui/ComfyUI_windows_portable/ComfyUI/models/vae/fixFP16ErrorsSDXLLowerMemoryUse_v10.safetensors" --xformers

[hot vortex]

also, where can I check it?

[trail lion]

there's an actual tech-support chat, this channel isnt really meant for that, fyi..it's fine to ask for help occasionally, but I'd caution against trying to treat this like a tech support channel, you'll not get the desired level of attention because people arent in here for that purpose

[earnest lichen]

yeah, I just checked the pytorch forum. found this:

CUDA 7.5 was used when PyTorch 0.4.1 was released and if I’m not mistaken, no Windows binaries were built for this old release.

[hot vortex]

it isn't that old, it's just Nvidia releasing a new series every year or two

[harsh skiff]

Ok, when i saw that finetune chat is limited I came here, I will go to tech-support chat

[earnest lichen]

its a hardware support. cuda also has versions. 1650 only supports cuda 7.5, which isnt supported by the software, due to being too old

[hot vortex]

it's only 5 versions

[harsh skiff]

but do you know a solution or should I just go there

[earnest lichen]

https://forums.developer.nvidia.com/t/cuda-enabled-geforce-1650/81010/5

All Turing family GPUs currently are compute capability 7.5

[trail lion]

I'm loking at your output, 1 sec

[harsh skiff]

Thank you

[hot vortex]

ok, thank you

[earnest lichen]

anything that doesn't require pytorch still works though. so keep that in mind

only pytorch is broken

[hot vortex]

Can you check the python version you use, if you're on Windows ...

[earnest lichen]

3.10.9

I know if off by heart since I do a lot of lora training

[hot vortex]

wait, 3.10.9?

[earnest lichen]

3.10.6 is fine as well though. no difference there

[hot vortex]

I installed 2.2.2 though

[earnest lichen]

python?! XD

[hot vortex]

no

not python

I meant pytorch

not python, also 10.1 is supported according to some websites

[earnest lichen]

if 10.1 is working for you, then everything should be good. at least not according to nvidia site... but not that thats ever meant much x_x

[hot vortex]

network or local version?

it shouldn't matter, right?

[earnest lichen]

local

[shell estuary]

Feel like my comment from yesterday has been deleted. Anyway, anyone know a tool that can create 3d mesh but from a bunch of photos from different angles rather than just a one shot?

[earnest lichen]

like meshroom?

[shell estuary]

ill check it out thanks

[hot vortex]

ok, thank you again

[trail lion]

your config is right, it has --save_state, which is what tells it to save the optimizer state, there should be directories in your model output folder. if they aren't created, and there are no errors, I'm not really sure how to help you. That is what should happen...I will say, however, that this probably not going to yield the results you think, and just resuming from a checkpoint is what almost every other trainer I've met has done

[earnest lichen]

Free:
• Meshroom
Cheap, perpetual license:
•Agisoft Metashape
Cheap subscription/credit based:
• Reality Capture

Reality capture is BY FAR the best app. try that one. its free to try, and you only pay if you have a result you're happy with and are ready to export the .fbx file. costs like 50 cents for a mid effort scan

[harsh skiff]

Ok thanks for your efforts

just can you tell me what is the extension of the last-save? I mean last-save.what.

[trail lion]

it's been too long since I've used that option to tell you exactly, but it will be obvious, they will have names that say "state" in them as I recall. They are folders, not files

[harsh skiff]

ok

what is the sample folder?

[trail lion]

the same folder is where sample images go, if you opt to generate samples

[harsh skiff]

and what are the folders in log folder

[trail lion]

just another way to eyeball the progress

[shell estuary]

@earnest lichen Thanks, ive grabbed meshroom, just waiting on my company to authorise it, Ill check out reality capture too, thanks again

[trail lion]

the purpose of these folders are for seeing a graph of the loss values...I dont like this mathematical approach to training, but some do

[fiery roost]

Hi my AnimateDiff does not shows

How can I fix this error

[alpine hull]

Eh idk at this point

I guess I'd say that Ideogram is best for text, SD3 best for mostly everything else

[hot vortex]

How come portabe ComfyUI works perfectly fine but the normal one clonned from github doesn't?

it has something to do with pytorch, I haven't even downloaded the CUDA toolkit.

[trail lion]

they do it on purpose to irritate you

[astral goblet]

folder perms or system environment variables. who knows

maybe it's an "if(you) break;" code block

[trail lion]

if [ 0 == 0 ]; then figure it out; fi

[alpine hull]

nvm SD3 kinda sucks at arms and legs

[astral goblet]

figertowt

@alpine hull good opportunity to start building arms and legs datasets ++

[trail lion]

the answer to 90% of the questions in here really boils down to python dependencies, it's the nature of the beast I'm afraid

[alpine hull]

I wish they released a better base model first 😭

Also text actually freaking sucks

It looks like it's badly photoshopped

[astral goblet]

base models meant for open weights paradoxically shouldn't be so refined towards single concepts. weird right?

the better it gets at portraits and arms, the worse it gets at robots and insects and cars, or something. It is meant to be a foundation that all can refine towards new purposes

[alpine hull]

That's not how it works for everyone else 🙄

Ideogram and, hell, even Dall-E 3 are better at being general image gen models

[astral goblet]

i think the api is still just a preview situation too. the bad results might just be artifacts of their prompt and settings. one concept among billions of parameters

ideogram is a software service selling access ot one model that nobody can refine , except for them

different market strategies

the biggest difference with stability releases is that they're open weight releases that anyone can affect with their own tooling

[alpine hull]

But like

That requires manual work to integrate a specific SD3 model into a product

Most products will just use the base model

Which is actually horrible

Compared to Ideogram specifically

Ironically, both have some regressions in specific stuff that Dall-E 3 still does best

[astral goblet]

yeah i think the current release is a preview. they'll have it more refined by the time its' actually 1.0

i think dalle 3 is intended to run on 40 gb cards only too. potentially ideagram but we don't really know for sure. it's a totally closed development cycle. they hold the weights behind a curtain

just wait till all the refiner teams get a hold of the weights with their curated datasets and training approaches. there will be a dozen variants of sd3 soon after release

better believe the pony team will flex their newest version of their dataset onto sd3

[sage reef]

@hot vortex did you got it working now? :3

[hot vortex]

yes, althoguh idk how to import or choose models, lol

checkpoint*

it says undefined and I imported the files into folder with models

checkpoints

so I'll just do a quick start and hopefully figure it out somehow

I actually need to insert a filename of the checkpoint

[sage reef]

your checkpoints should be in the comfy/models folder

you dont need to type anything manually

if you have the models already there, then either restart comfy or refresh button, it should detect them

[hot vortex]

Is it bad if I have models in a folder stable-diffusion?

It doesn't seems to recongnize it

[sage reef]

are your models currently from a1111 location?

just move them all to comfy models folder, since you are not using a1111 anymore

[hot vortex]

ok, it'll be hard to get used to it

[sage reef]

it's only maybe hard the first time, then it becomes like second nature :3

[hot vortex]

ckpt_name undefined

[sage reef]

you have models inside comfy models folder?

[hot vortex]

yes

[sage reef]

actually my bad... comfy/models/checkpoints folder

put them all in there

[hot vortex]

so it has a folder, got it, thanks

everything else is the same, thank you all for help, now if it works

[sage reef]

basically the comfy/models folder hosts all sorts of other folders, for all sorts of things, like controlnets, lora, etc

one of them is called checkpoints

[hot vortex]

but it doesn't it so much RAM

[sage reef]

what are specs of your pc, just tell me the vram and ram?

cause i think you had some problems with a1111 if i recall

maybe you need to enable low-vram argument

[hot vortex]

GTX 1650 4GB VRAM, 16 GB RAM, i5-10300H

[sage reef]

4gb vram.... oh boy

[hot vortex]

I had that already enabled along with xformers

[sage reef]

kk

[hot vortex]

ok, this definitely uses much less RAM than I expected

[sage reef]

not sure you will be able to do sdxl stuff tho, or maybe but like real slow

[hot vortex]

yeah, that's enough

[sage reef]

so you got the checkpoints working now?

[hot vortex]

ok, it works

although the picture isn't nice

[sage reef]

nice

[hot vortex]

that was fast

[sage reef]

well you have to adjust couple things, not all models have the same recommended settings

also the vae

[hot vortex]

it's some kind of a compression method

but I haven't used it with A111

and always had --no-half because of some issues with promps resulting in no half error

the performance is already incomparable

Do embeddings work too? embeddings in embeddings folder

[sage reef]

so you learned to hook everything up with the nodes and spaghetti? :3

yes embeddings work

[hot vortex]

I used the default one.

[sage reef]

for embeddings i recommend you install rgthree nodes, cause he has a very useful prompt that allows to select textual embeddings, instead of writing them manually, along with so many other cool nodes too

[hot vortex]

wait what, such a thing exists

[sage reef]

yes

[hot vortex]

I made an excel table for all models and embeddings 🤣

[sage reef]

that is not needed haha

do you have manager installed for comfy? that's like the number 1 thing you need

[hot vortex]

yes

I sent you a screenshot of the manager

[sage reef]

nice, so within the manager, i think you will find rgthree if you search

it's node number 163 if you dont want to type it

make sure to install that

and then you are set for the basic stuff :3

[hot vortex]

ok, hold up, it's more complicated, I see that I have to use some kind of node combiner.

[sage reef]

i mean you can install some other nodes like efficiency, or impact, inspire, etc, but that's up to you, so many things to explore

node combiner?

[hot vortex]

btw how to differentiate negative and positive prompt?

there are 2 prompt windows

[sage reef]

the clip text encode is any, it's up to you to color them

people usually color negative with red, and positive with green

[hot vortex]

yes, but if the program knows that worst quality etc is negative

in A111 you have negative and positive prompt

[sage reef]

yes, so for example you load one node "text encoder", color it red for example, even name it negative prompt if you want, and connect it to the ksampler negative node

it will know that whatever you wrote in that one will be used as a "negative" prompt

because it's connected to the negative node within ksampler

and similar with positive

but maybe color it green and name it

[hot vortex]

ok, thank you, I'll check the tutorials on the internet

[sage reef]

yea there are tons of easy to follow comfy tutorials out there

i like scott detweiler videos, he always builds from scratch, so it's easy to follow

[quaint fossil]

What the heck is a Pony Diffusion?

[astral goblet]

dark territory

[sage reef]

lol

it's for adults only

[trail lion]

haha, I found a full explanation for that on reddit, I Should have bookmarked it

[hot vortex]

lol

I have both versions downloaded, haven't tried it yet

[astral goblet]

the safer side of it is actually xtremely good at illustrative artworks

[sage reef]

pony diffusion is one of those things you might find on a stable diffusion iceberg video, at the very bottom of the iceberg probably, assuming someone does one 🙂

[trail lion]

basically short version, started as a pony furry kind of thing, refined, refined more refined more, and eventually trained to the extent that no prompt adherence from the original model no longer works, but it's really good at "certain things" shall we say

you can actually kind of say that it became its own base model?

[sage reef]

i personally never used it and prob wont, im good with my other models lol

[fiery roost]

Somebody safe my AnimateDiff

[trail lion]

same, but it's interesting from an academic pov

[sage reef]

huh?

[fiery roost]

Bro my AnimateDiff does not show

[sage reef]

the nodes?

[fiery roost]

nodes?

[sage reef]

does it show an import failed for animatediff?

[fiery roost]

Ya

[sage reef]

ok so you dont see the nodes

[fiery roost]

Show some error

At the cmd place

[sage reef]

well you have to fix whatever it is complaining about, check the console

you can post the screenshot

[fiery roost]

On Web ui

Ya I send it to image with image

[sage reef]

you are missing some requirements perhaps

kk send screenshot

[fiery roost]

Ya , on general with image that group

Can you see that

[sage reef]

i cannot see anything, you didnt send anything

oh ok in the other channel

[astral goblet]

i see the word "that"

[fiery roost]

Yaya

Haha

Something : error calling

[hot vortex]

ok, ComfyUI definitely has a problem with hands

how do people usually fix hands there

nvm, forgot about 2 prompts

[sage reef]

there is hand detailer

but i dont personally use it

also, the problem with hands by itself has nothing to do with the UI used, it's more about the model used along with the settings used

[pearl ocean]

How Comfy is ComfyUI?

[hot vortex]

surely

still lost for the time being with embeddings and plugins (nodes) I installed through the manager

[sage reef]

i mean dont do everything at once, learn slowly haha

[hot vortex]

yeah, I want to do it intuitively because it's already late

[sage reef]

do you have rgthree installed?

[quaint fossil]

Wait, what

I ask because I found models that apparently use Pony Diffusion for best results

[sage reef]

you can use whatever you want man, not gonna judge ya

[hot vortex]

yes, but can't find it

[quaint fossil]

https://civitai.com/models/368139/character-sheet

It's nothing weird, I just came across it today and don't understand

I wanted to use a character sheet concept, but it said that it's designed to work best with Pony Diffusion XL

[sage reef]

inside comfy manager, when you click on install custom nodes button, on the left side you will see node ids right? i think rgthree is like 163 or something

[hot vortex]

yes

[sage reef]

so install it

[quaint fossil]

But now I'm confused why people giving me a weird look

[hot vortex]

163,but there is no button to use it

[sage reef]

each node has a install button on the right side

[hot vortex]

it's already installed

[sage reef]

ah ok

[astral goblet]

thats a lora thats been trained using pony as a base. pony really changes the entire latent space of sdxl so much that loras made for sdxl base, generally dont work for pony. and visavi

[sage reef]

so you dont see it in comfy?

you might have to restart comfy

[quaint fossil]

So if I wanted to use this.. Checkpoint? Model? I would need Pony Diffusion or would it work with SDXL?

[hot vortex]

yes, I see it, thanks

[sage reef]

nice

[hot vortex]

so either use manager or type everything manually

[astral goblet]

load pony checkpoint then use that lora with it. or pony merges with other models

[hot vortex]

refresh doesn't work, only restart works within the manager, but the speed is already impressive

[sage reef]

now within rgthree, there is a node called power prompt simple if i recall, it has a insert embedding button at the bottom of it, which makes selecting embeddings easy @hot vortex

[quaint fossil]

Oh, so I need to install Pony

[hot vortex]

thank you, this definitely saves time

[sage reef]

yea

and that should cover the first day of the tutorial 🙂

then you can explore loras, controlnet, ipadapter, etc

[hot vortex]

that's a very useful plugin

[sage reef]

yea lots of cool plugins and stuff you will find for comfy 🙂

[pearl ocean]

I bet people use SD to create their perfect Girlfriend

[trail lion]

wont judge, create what you will. if you upload it, well that's it's own drama

[pearl ocean]

I bet people create their dream homes too

[astral goblet]

thats a great john hughs movie from the 80s.

tv show too

https://www.youtube.com/watch?v=9qd04u2Yj44

[quaint fossil]

@astral goblet I would put the Pony Diffusion into my Lora or SD folder?

[shrewd laurel]

Hello I'm new here

[hot vortex]

Hi guys. I've setup ComfyUI but it seems that power prompt from rgthree doesn't take into account engatives

the resulting picture was of a very low quality

any way to fix it

maybe I got it. I can actually make a clip and add the text there. embeddings into tghree

nope, it is still kinda pixelated, am I missing something?

[thorn anchor]

Out of curiosity, has anyone had success with consistent original characters with PDXL? Even just getting the same character with different expressions would be better than what I’ve managed so far (which is bupkis) but I know consistency is basically the holy grail of stable diffusion atm.

[quaint fossil]

Hello new here

[verbal osprey]

Not sure there is an easy way to do that right now....it will come

Does anyone know if the Forge UI one click install creates a contained python environment? I don't want to mess up my other stuff

[raven agate]

forge has a portable install

[wraith edge]

Guys why is glif down

[worn aspen]

There seems to be a switch in knowledge sharing, with patreon being required to get workflows. It changed super fast since I've only been into this for a tiny fraction of time.

[foggy halo]

guys i tried sd3 in huggingface using my API key, but after 2 or 3 images it showed me an Error message

i deleted the old key and made a new one still no use

why is it happening?

[karmic cedar]

What’s the status of your credits?

[foggy halo]

how to see that

[karmic cedar]

hello

[sage reef]

hello

[neat stirrup]

stable foundation

[frozen helm]

What

[pearl ocean]

Some people just don’t know the power of their phones these days

[granite topaz]

can anyone tell me which would be the best model to generate photo realistic images ? I use forge ui

I want portrait type of things

[placid hatch]

Is there any benefit from mixing lcm and lightning?

[sage reef]

civitai has tons of those, just search by realistic tag or similar

some of the recent ones are like juggernaut X for example

[sage reef]

not sure... i stay away from any fast model, cause they usually give worse results overall, but i can understand their use for people out there

[granite topaz]

I have juggernaut but I just can't get what I want. Maybe my prompting skill is not good enf

[pearl ocean]

EpicrealismXL is the realist one I know so far

[fiery roost]

Forge UI is better than a1111 right?

Is every pulg-in than in a1111 can work normally in forge?

I'm planning about change to forge UI

[warm junco]

Not every extension that works in auto1111 will work in forge.
Forge is faster and good for GPUs with 6gb vram or less.

[fiery roost]

I'm 4070 on laptop 8gb vram

[warm junco]

Okay and do you have any issues with auto1111 ?

[fiery roost]

No but I want faster speed

Just worry some extension can't work

[warm junco]

Are you using --xformers --medvram-sdxl --no-half-vae
In your webui-user.bat?

Thats needed for your GPU in auto1111 for best performance

[fiery roost]

I'm using --xformers and medvram

Can I also and this two in forge UI?

[warm junco]

Then change --medvram to --medvram-sdxl
So sdxl models will work much faster

[warm junco]

These are already enabled in forge.

[fiery roost]

Oh

Can I add --medvram-sdxl in forge UI?

[warm junco]

No forge removed every args from auto1111 and reworked them. It should have them already enabled.

You can have both webuis installed. So you can compare both directly

[fiery roost]

Ohhh okay tQ for your answering

[warm junco]

No problem 🙂

For any technical questions we have #🤝｜tech-support

[fiery roost]

Tq

[karmic cedar]

Hi

[sullen axle]

Hi I am new to stable diffusion and don't know much about . just watch some tutorial and started downloading forge webui . is that okay? or is there something better than that?

[feral fox]

Hey Guys!
Is the Membership meant for the commercial use of the model or the output images?

[lean knot]

Hi I'm back in stable diffusion but i'm just realize there something change. Is there anyone know where's the bot like bot 1, bot 2, etc at?

[bold marsh]

guys what is LoRA

[civic summit]

hi, I want to try decorating a video using AI. who can help?

[warm junco]

Hey the bots are offline
#1047610792226340935

[warm junco]

Hey, I would say it depends on the GPU and amount of vram.
If you have a lower end GPU forge can increase the speed.
For 8gb or more I would still suggest automatic1111.
You find install guides in the Pinned message of #🤝｜tech-support

[warm junco]

A lora is a small file, trained on a specific style, object or Character.
The lora works together with a model to help generate such specific style, character

[bleak matrix]

Good morning, everyone? How are we all today?

[deep berry]

Quite well

[surreal jay]

rather good, how're you?

[hot vortex]

it could be better, but fine for now

[lean knot]

Oh I don't know it's offline thank you. So how if i want to made AI in here?

[solemn stream]

What's up with glif

[deep berry]

Never been on there

[solemn stream]

Where do u generate ur sd3 images then?

[deep berry]

@solemn stream Pixeldojo with a small fee

[solemn stream]

Bruh

[solemn stream]

Glif is free

And unlimited

[deep berry]

Well it's down...so it's not that free

[solemn stream]

Yes

But it worked for a week

And it's coming back up

It's not like a little site

https://glif.app/@fab1an/glifs/clv488uy10000djtrx70u03no

[deep berry]

Yeah ive been still using dojo . not bad compared to other stuff i used

[solemn stream]

Glif is free and prompts are public so I love looking through them

All the images and prompts become public

[deep berry]

Ok that's fine

I might explore for prompts to remix but i normally make my own everyday.

[solemn stream]

Alright

[deep berry]

I will test on dojo. If it's worthy i will provide a prompt on glif

[bleak matrix]

Good to hear!

[bleak matrix]

Quite dandy! What are you up to?

[surreal jay]

Dandy is the best! I am at work but I'd rather be at home prompting :(

[civic summit]

help

[bleak matrix]

Yeah, I know what you mean! What are some of your favorite things to create?

[surreal jay]

my puppy dog, lol, I want to train a Dreambooth on her

[warm junco]

Right now you can't on generate images in this discord.
But you can find resources to install Stable diffusion localy on your PC to generate images

[surreal jay]

What do you like to create?

[bleak matrix]

Awh! What kind of puppy do you have?!

[bleak matrix]

On a scale of 1-10, everything!

[surreal jay]

Nice! I plan to branch out to everything as well lol, so many ideas. The only limit is your imagination! She's a shepherd / mastiff / lab cross. Mostly lab though

[rotund tendon]

I'm training a new art style image model with 320 sample images. What hyperparameters should I use for training, including epochs, learning rate, text encoder learning rate, and learning rate scheduler?

[bleak matrix]

Ahhhhh! I used to have a shep/lab cross; he was one of my fave dogs to have. They are such great dogs to have around--ours was quite fun bud to have around. He had an obsession with chasing lizards around.

I really love taking time to thorough learn about a subject, then trying to apply all elements to AI. I feel like taking natural methods of learning and applying it to AI helps make the process of creation a better overall experience.

[surreal jay]

lol that's a good obsession to have! Ours loves chasing rabbits. I will post a ControlNet pic of her to general with images.

[surreal jay]

Do you mean it helps with prompting?

[bleak matrix]

It's a good way to get exercise, that's for sure! Hahahaha!

[bleak matrix]

I apply it to all aspects.

Since I'm an artist, I use my experience, and knowledge in art to help me while I prompt, and with anything I create with AI.

[surreal jay]

Oooh cool! I am sure it's a huge asset

[soft iron]

😭

[bleak matrix]

I’ve been able to explore a lot of fun areas because of it! If you ever have time, you can always explore the basic principles of art and design to see what kind of fun things you can create. Any kind of learning, I think, is useful for AI (or life in general, ofc!)--I’ve seen so many talented people here, and they’ve been able to use their own knowledge to make their own creations even better. The more research you do, and the more of your own experiences you put into what you’re creating, the better the end result! It’s also pretty fun to play with different aspects of various languages--so there’s endless ways to create!

[surreal jay]

I've always been interested in it, so I think I will do that! Research is fun in itself too because you learn so many interesting things

I also would like to create game assets one day for personal projects, but am concerned about possible legal ramifications and that whole can of worms

[karmic cedar]

what are folks using the most for inpainting these days? I’m on limited hardware 😦

[bleak matrix]

I always love reading--what a great thing it is! To snuggle up with a good book, and a cup of cocoa next to a warm fire--a lovely day, to be sure! (Or a good morning, naturally, sitting on a window seat next to an open window, with the breeze blowing some white curtains while I drink some orange juice next to a couple of cats while I do my morning reading...) There's just so much to learn; it's hard to put those good books down!

Do you have any particular genres/subjects you want to get into?

Game development is always fun to do--I always find that generating a few images can certainly inspire me to work on assets, should I find myself in the mode to do so!

[lethal gorge]

hi guys! im using zluda for an amd gpu on sd.next, and so far it has been working well but recently my generation speed has been like 10 times slower. sometimes when i restart my pc it works better but then after a while it gets slower again. anyone know what to do?

[surreal jay]

Sci fi/ space ships. I have always wanted to live on a ship and be surrounded by tech

[karmic cedar]

could be related to other traffic via your service? do they talk about performance peak times, etc. anywhere?

[lethal gorge]

what do you mean?

[karmic cedar]

ahhh nvm—i thought you were using cloud compute for something. forget my gibberish :-x

[lethal gorge]

no worries!

[karmic cedar]

that said—my gut instinct tells me there’s a memory caching / optimization related issue going on

just call me captain obvious

[lethal gorge]

what should id do in said situation?

[karmic cedar]

i’d check if sd.next has a community that’s active on the discord

or some kind of tech support forum

there could be other folks using the zluda implementation like yourself

[lethal gorge]

thats a good idea, ty

[karmic cedar]

😄

[twin mirage]

Can anyone help me .. I am getting this when trying to install stable iffusion ERROR: Could not find a version that satisfies the requirement torch==2.1.2 (from versions: none)
ERROR: No matching distribution found for torch==2.1.2

[bleak matrix]

That's always fun! It's great to have an excellent idea, and see what kind of things you can make from it. Sci fi is such a great genre to explore--if you lived on a space ship, what would you do?

[bleak matrix]

You can check #🤝｜tech-support for help on that!

[surreal jay]

Explore the universe! And life, everything!

[sage reef]

i was confused with your name for a moment and then i got it haha 🙂

[sullen axle]

is it possible to output images into a different format like svg ? because as i saw some of the models can create vector art pretty good and if they output svg it would be great!

[rare parrot]

Hey, have there been any smaller announcements about SD3 since the API release?

Anything SAI staff mentioned etc?

I haven't had time to follow it too carefully recently

[sage reef]

the best we got is an approximation from Emad himself which puts it around may 10 or something :3

i mean Emad shares a lot of info, i always follow him through this lol: https://new.reddit.com/user/emad_9608/

he also confirmed Cascade is ded

[astral goblet]

i woudln't listen to anything emad preaches now. he was barely accurate when he was CEO. now he's all about block chains so you can count on anything he says to be some con artist bridge selling scam.

[sage reef]

sure, but i mean what Emad said kinda matches the range of the current lead dude:
https://x.com/chrlaf/status/1772228848387522728

[astral goblet]

emad said there'd be a signup for a closed preview. They never spooled those out though

[charred mesa]

yup

feb 29 was so epic!! oh wait it was cancelled in the last second

[astral goblet]

the new guy seems to have an entirely different plan. I bet we won't get full weights until well after summer business opportunities have dried up

if we get proper weights at all

[placid hatch]

i find I get significantly better and more consistent results with the lightning model I use

[sage reef]

is it a finetuned model?

[slender field]

Introducing HiDiffusion: Increase the resolution and speed of your diffusion models by only adding a single line of code
project page: https://hidiffusion.github.io/ github: https://github.com/megvii-research/HiDiffusion

[karmic cedar]

most impressive!

[sage reef]

imagine adding a single line of code

[karmic cedar]

curious about a demo.

[slender field]

its seem like a simple trick no one tested yet but now it seems to work

i cant imageine what this does to the really good finetuned community models of sdxl

like joggernaut

[karmic cedar]

lmk if you try it!

[hot vortex]

Is it real? I've just triedone of the LCM lora models

it generates previews really quickly

[sage reef]

pog

[fervent thunder]

gers

[pastel surge]

What does the server icon represent?

[fervent thunder]

A pog

[sage reef]

gers

[spare veldt]

Safe-and-Stable-Ckpt2Safetensors-GUI.v0.1.1. is it safe to use it?

[sage reef]

as you can tell from their title, it contains the word safe, so it has to be safe

[shell tendon]

Safe-and-Stable-Ckpt2Safetensors-GUI.v0.1.1.exe?

[spare veldt]

more or less yes

I mean the github tool: diStyApps/ Safe-and-Stable-Ckpt2Safetensors-Conversion-Tool-GUI

[woven kindle]

Hi, is it possible to assemble 2 models? I have a model realisticStockPhoto_v20 which gives a very very beautiful face but cannot make a body while pixReal_10 gives a very beautiful body associated with LoRa. My goal is to create a woman with a consistent face and body and then be able to put her in different situations. For now I create the face with realisticStrockPhoto, then I create the body with pixReal_10 and then I do inpaint+controlNet+reActor with EpicRealism to assemble the face but I'm not so satisfied with the result. Or could there be another solution? Train one of the 2 models?

can I post NSFW photo here?

I'm using EpicRealism for the inpaint because that doesn't work with realisticStockPhoto and with pixReal_10, I have an error

I have this error

NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.
Time taken: 17.2 sec.

Exception: ControlNet model control_v11p_sd15_openpose cab727d4 is not compatible with sd model(StableDiffusionVersion.SDXL)

[verbal osprey]

Good evening gentlemen and ladies. Is SD3 supposed to be faster? SDXL runs at like 1.5 iterations a second on a 4090

[karmic cedar]

yo where the ladies at

SD3 is slower (so far)

[sage reef]

you have to generate the ladies with sd3

[shell tendon]

what on earth are you doing to get only 1.5 it/s on a 4090

are we talking 1536x1536 minimum?

and a bunch of controlnets etc

[trail lion]

you can combine models in a sense, but you sometimes will get strange results. what I do in your situation is concentrate on the body first, and fix the face with inpainting, which is usually quite simple. the body is hard to fix later, the face is easy to fix later

also, having a lora is very preferable to just using face swap, face swap is a hack

it takes like 20 mins to create a lora, with maybe an additional 20 to do a basic image set with like 15 images, there's really no excuse

[neat elm]

hi all, i just created an ai art/avatar/video generated mobile app. I would appreciate if anyone want to try it out and give me feedback

[upper elm]

This is probably a silly question, but if I have an image of a character created in SD already and I just want to make an alteration to that image (such as giving him a hat or a coat), what would be the best way to go about that?

[trail lion]

inpaint + inpaint sketch

inpaint sketch is where you show it what you want, so if you draw a basic shape of a hat over your existing photo and do an img2img on that area, it will take your guidance + your prompt and make it happen

you dont need to be an artist, it can be a crappy photo, but colors you use will be taken quite literally

[upper elm]

Hm, ok and do I need to modify the prompt? Especially for anything specific?

[trail lion]

yah, so if you had, a woman sitting on a bench, you could do like a woman sitting on a bench wearing a brown hat with a blue flower on the brim...just making this up

I tend to use gimp/photoshop or the like vs the webui included paint tools, you edit them and just drop it back into img2img inpaint and select the relevant area

[upper elm]

Ok, I'll give that a try, thank you

[trail lion]

do some practice photos with inpaint sketch, it's really powerful, you can give a character gloves, a necklace, a pipe, there's little limitation apart from the model knowing how to draw it...like it has to know the concept

[fervent thunder]

Anyone know what SD-Kohya is

[trail lion]

sd is stable diffusion, kohya is basically a reference to the most popular training implementation for dreambooth, lora, etc

most people use kohya_ss which is a gui wrapper for the kohya python scripts, but there are other trainers out there, such as onetrainer

[rotund remnant]

When to release SD3's weight file

[trail lion]

if that were a question, it would be a common one

but it isnt

[cedar vortex]

How do you get people in the background of a picture in focus? Even people that are 15 feet away in a photo have indecipherable faces.

[trail lion]

inpaint masked only

most of the models have a certain tendancy to do bokeh, but worse than that models before sdxl were at a small enough resolution that far away subjects didnt have enough pixels to render with detail

even with sdxl distant subjects remain an issue during txt2img, you almost always have to fix them

hiresfix can help, but I'd never say that you never have to inpaint...I inpaint almost every image that's worth preserving

if you're not touching up your image, you're doing like 25% of the job

[fervent thunder]

@open wren

[soft iron]

😭

[tepid crow]

Would there be anyone looking for a custom instrumental backing tracks for their creations with open source licensing?

I like the artwork i've seen many artists create and i'd love to collaborate.

[harsh silo]

Noob here, What happened to all the bots wher we could gen pics with dream cmd?

[sly eagle]

Hello! Do you know what can I use to make dance animations to my own model?

[surreal frigate]

monetization : )
https://stability.ai/membership

I don't blame them, but it's all comes down to the money. I'll stick with my A11, and URPM model. does the work for me.

and no it's not for pron it's for actually training people to look realistic using TI as well. but noo SDXL had to get rid of it.

dreambooth, lora, etc img2img is just ehh too easy, i want AI to do it by it self. that's the hard task. instead of it trying to ok copy this image. and replicate it.

[pearl ocean]

no crying!

[wispy arrow]

Prompting Juggernaut is rage inducing, it’s not just you. I finally gave up after 8 hours of irate prompt adjusting and downloaded RealVis V4.0 and it nailed my prompt in minutes 😭

[woven kindle]

ok I will try that

thanks

[stark hornet]

any news about the 3.0 version?

[high scaffold]

hey guys, which channel can i generate images in?

[loud solar]

https://discord.com/channels/1002292111942635562/1047610792226340935

[slender mango]

Hello

I’m new

How to start

[loud solar]

This is more about generating lokal. If you wanna do quick and easy you better search a webservice to create. Otherwise check the different available UserInterfaces and decide for on and install it.

[slender mango]

Huh?

What’s lokal?

[loud solar]

You can't generate here online. You will have to install software on your computer.

[slender mango]

Oh

Okay

But where

[loud solar]

https://www.craiyon.com/ is a free generator ... for example

[slender mango]

Oh

Okay

Thank you very much

I need a monkey for my assignment school

[loud solar]

Have a good and creative time!

[slender mango]

Thanks a lot

May god bless you

[bleak matrix]

Good morning, everyone! How are we all today?

[static cape]

I want to train... but I guess I have other stuff to do. xD

[soft iron]

I will cry every day until SD3 comes out

[hot vortex]

I wanted to try it, but the requirements seem quite high. I have only GTX 1650 4GB VRAM and GTX 1060 6GB VRAM at home. CPU are almost same i5

and 16 GB RAM

[bleak matrix]

Well, consider it a goal to look forward to at the end of the day? Haha! I know whatcha mean!

[trail lion]

the requirements are higher, but you can use cloud compute to get it done, like vast.ai or runpod

[hot vortex]

I haven't attempted it, but would like too after I'm done with exams. It just looks lengthy. It'd take days to train one simple model.

at least form what I read it takes 5 hours for GTX 2080 ti 12 GB

[trail lion]

but like 45 mins with a 3090 on vast

[hot vortex]

well, GTX 1650 4 GB VRAM, lol

it took me quite a long time to avoid CUDA memory overflow error

tiled option in SD upscale alone does absolutely nothing

tiled VAE decoder fixed it

[knotty turtle]

What's the word on the streets about Forge? Is that what everyone is using these days?

[hot vortex]

idk, but all I know is that it's an edited version of A111 which is more memory efficient (automatically benchmarks your PC and sets parameters to fit.)

but A111 never again, eats too much RAM and make PC lag into oblivion

[fervent thunder]

hi

[hot vortex]

hi, how are you

[serene glacier]

Is there anyway for us to help out this paper’s author with implementing this paper into ui’s? https://www.reddit.com/r/StableDiffusion/s/zyBzvxSFSv

[spare owl]

hello.

[proud copper]

/drIn a sleek, modern urban apartment meticulously designed, numerous innovative smart home devices are placed throughout. A gorgeous girl is smiling as she introduces the interior decorations and furniture to everyone.

[hybrid thistle]

So did anyone get SD3 get from the waitlist? I applied more than a month ago lol

[balmy hemlock]

Hi there! Anyone is comfortable with Kohya_SS/Fine Tuning

[karmic cedar]

you can actively use it via the API now.

[trail lion]

there's a finetune channel, but also maybe specify if you mean 'full finetune', dreambooth, lora, etc

[balmy hemlock]

Yep full finetune sorry, sent message this morning to this channel but seems not much people

[fervent thunder]

Hey guys guys!

ITS MY BIRTHDAY! 🥳 🎉 🎈 🎂 🎁

[hybrid thistle]

Damn ok thanks

[karmic cedar]

lol wow “Phi 3 (4 billion parameters, 2024) beats GPT 3.5 Turbo (~ 175 billion parameters, 2023 version)”

[shell vector]

what do you guys prefer between analog madness and juggernaut

[low moon]

https://www.youtube.com/watch?v=2FUvHdnIaW4

new AI vidya thingie

Stock footage down, creeping into music video territory.

[verbal osprey]

Resolution is probably averaging on 1000x1000. No controlnet, no adetailer, no high res upscale. Just normal generation with SDXL. I'm using auto1111, maybe that's the problem? If forge is faster I might try setting that up. Especially if it will support SD3 when that drops.

[low moon]

hapy bday

SDXL with some loras stuff takes me about a minute 3060 12 gig vram, then the control nets and all that adds about 15 seconds each

slaying thru anything 1.5 in less than 20 secs with whatever enabled

not sure how SD3 will fare

Cascade worked

a little slower than sdxl but nothing drastic (only used it in comfyui)

🤷‍♀️

[low moon]

Forge is like a faster better looking cousin of A1111. no drop in quality also.

Foooocus does seem to sacrifice some quality tho so onyl for casual renders

[verbal osprey]

Someone mentioned there is a portable install. Is that the one click install? A random website I was reading said not to use the one click install if you run other python programs because it can mess stuff up. It sounded like it didn't create it's own virtual environment....but it seems odd to me that it wouldn't do that.

[low moon]

i install everything separately just in case then linke them all to my A1111 model directories to save disc space

and use git clone

not fancy installers

[verbal osprey]

I might do the same, to ensure that I have the virtual environment set up for it.

[trail lion]

I dont even use venv or conda env, just docker

[hasty oasis]

كلب

حمار

خرا

[verbal osprey]

I don't understand how to use docker, so I stick to the virtual environments

[fervent thunder]

Thank you friend I love you 🥰

[hybrid thistle]

wait this isnt free lol i only get a few credits

[frozen helm]

Si

[trail lion]

aww shucks...everything should be free, until they go broke...oh wait, that happened already

[verbal osprey]

Free is better obviously, but I would have no problem paying $20 bucks for the SD3 base models. They've done great work. I don't mind supporting it.

Or have a donation page where people can donate what they want.

[sage reef]

paying for models like we pay for triple A games :3

[verbal osprey]

I'd rather the company stay alive and open source rather than disappearing because they ran out of money

[charred mesa]

we're paying for a license not a game

do demos contain the entire game with no time limit?

and do AAA games from big studios nowadays cost 20$ or less?

(admittedly, this is monthly, but yeah)

[sage reef]

OnDiscDLC

[charred mesa]

SD3 DLC

[sage reef]

you know when you really want something (in this case SD3 to release), then every day feels like an eternity 😦

[trail lion]

cant say that I do

guess I'm just a patient sort, everything in its due course

[sage reef]

well considering all the tech and toys we are getting these days at a fast rate, we are occupied until it comes i guess 🙂

[hardy sedge]

Why sdxl pictures in 512/512 Looks so blurry? Is there a tweak? Iam using forge.

[charred mesa]

why are you generating at 512/512

you are supposed to use it with 1024/1024

[tropic frost]

small curiosity, when building a character lora, would it be more efficient to erase everything in the sample pictures with the expection of the character?

[astral goblet]

theres really not much difference between forge and stock if you've got a capable card. it helps people on the low end a lot

[trail lion]

no, let's say you have a white background, then every image you generate with your lora will be a white background, it finds patterns, even if the pattern is nothing

you're better off having a wide variety of backgrounds, clothing, poses

[astral goblet]

sample pics? like whats generated during training? those shouldn't be huge prompts. maybe one simple portrait prompt and one challenging action prompt. you're just looking to see if the training has an effect, not a full on test generation

[verbal osprey]

I have read the speed improvements can be anywhere from 3-5% for a 4090.

[astral goblet]

i got a 4080 and they're both neck an neck. i think that was maybe one version of forge and only with default installs.

forge has the memory leak problem where it doesn't unload loras in some cases. i think its under investigation and no roadmap to being solved yet. Dev has a lot on their plate. Big part of why i only load it when i need it for forge specific extensions. its the one UI i have to reboot periodically.

[next nexus]

where can i ask about help with an image I'm trying to use in SD to improve it?

[loud solar]

Could be easier to ask here ... https://discord.com/channels/1002292111942635562/1004159122335354970

[karmic cedar]

https://github.com/Snowflake-Labs/snowflake-arctic

has anyone tried this yet?

because holy balls that’s a performer

hit today, on par with llama-3-70b-instruct

https://arctic.streamlit.app

[fervent thunder]

Whats the best inpaint model

[karmic cedar]

AssPaint McDeluxe 2 Alpha Turbo Dipping Edition

No im actually not sure

[charred mesa]

https://github.com/SeaArtLab/ComfyUI-Long-CLIP

@crude notch @astral goblet

For SD1.5, the SeaArtLongClip module can be used to replace the original clip in the model, expanding the token length from 77 to 248. Through testing, we found that long-clip improves the quality of the generated images. As for the SDXL model, since the clip-long model for clip-g has not been released, our processing procedure is as follows: for smaller tokens, we expand them by an integer multiple of the original max_len, and since the last added are pad_tokens, we trim the excess part. Given that clip-g features occupy a larger proportion in SDXL, you may notice more detailed images. Finally, if you like our project, please give us a thumbs up.

expanding the token length from 77 to 248

[astral goblet]

neato

first i've heard of long clip thanks

[charred mesa]

same and this was released 2 WEEKS AGO

the addon itself! not just longclip

[astral goblet]

it seems to really aid natural language

if i'm to trust their examples

[charred mesa]

and if I recall correctly, SD3 also uses SDXL clip models (clip-l and clip-g), so we could use this for SD3 too maybe!!

they will be releasing a longclip-g as well, currently we only have longclip-l

magnificent.

[astral goblet]

yeah it looks like a drop in replacement for that one clip layer

[sage reef]

oh cool, didnt know about this stuff 😮 does it really improve the generations as stated in the readme?

there is also B version, but im assuming it's inferior

[hot vortex]

I haven't understood l and g clip fully. Any easy explanation.

I have it in the face refiner workflow

[sage reef]

@hot vortex this can help understand https://new.reddit.com/r/StableDiffusion/comments/15c2n0q/sdxl_two_text_encoders_two_text_prompts/

[narrow kernel]

One of them uses more naturally captioned phrases and the other uses the older style from what you'd see in sd1.5 prompts with a million individual comma separated words.

The clip models associate words/phrases with data in the model.

[hot vortex]

still hard

[charred mesa]

iirc clip-g is the one with natural prompts (where you type full detailed sentences)

[hot vortex]

Scott didn't go into tdetails. It was in a video for face detailer.

[sage reef]

i personally never use two for sdxl, i only use the text encoder and that's about it

[hot vortex]

so what is the 2nd one "l" for then

[charred mesa]

and clip-l is where you type a bunch of "tags", like -> photo of, dog, in castle, high quality, intricate

that's just my guess though

[hot vortex]

and g is their description?

[charred mesa]

idk tbh

don't have a concrete explanation

[narrow kernel]

G=a happy turtle in a pond, l=turtle, pond, happy

You can use only G, only L or G+L if you want

[charred mesa]

yeah that's basically it

[sage reef]

yea but it's not just that tho

[hot vortex]

I'll study it at the weekend. Too many things happening today.

[charred mesa]

if I recall correctly, SD2.X only had clip-g or openclip or whatever

[narrow kernel]

And sd3 will have G and L, but also T5 which is an LLM based model

[hot vortex]

Anyway, any ideas on how to make this titled or anything that's chunk down data? for 4 GB VRAM

I cna't send a picture her, I'll post in #🤝｜tech-support

[charred mesa]

wait SD2 has OpenCLIP-ViT/H

so its like clip-h

lol

[hot vortex]

#🧣｜comfy-ui here

[narrow kernel]

We don't talk about SD2

[charred mesa]

yes we do

😈

[narrow kernel]

Lol, I'm just kidding, but nobody uses it unfortunately. Maybe like 0.1% of people so

Do*

[charred mesa]

I used to be one of the only people who used it

then jumped to SDXL when it came out

[verbal osprey]

SD2 disgusts me

I like 1.5 and SDXL

[charred mesa]

[cobalt herald]

How does one even make SDXL look good

[verbal osprey]

custom SDXL models and loras

[karmic cedar]

check out HiDiffusion

[cobalt herald]

I have those but they just look... Bad LOL

[karmic cedar]

yeah, there are limits to the hack

[verbal osprey]

Are you setting the resolution to 1024x1024? If you lower the resolution it doesn't look good.

[cobalt herald]

I did for a bit but I put it back

Lemme try again

But yeah it all looks goofy 😭

[verbal osprey]

Download a model you like from civit.ai then try to generate what that model is good at. You can also take the prompts for the preview images and try the exact same prompts to see if you get those results. Then you can tweak from there

[cobalt herald]

Okok I'll try

[fervent thunder]

Hi

[lofty sphinx]

Hey i dont know much of kaggle, but are stable diffusion notebooks banned on it?

i saw some but idk if i risk ban

[trail lion]

some people are just hard to please

or bad with prompts 😛

[karmic cedar]

english syntax is interpreted literally by the AI, it can be hard sometimes to prompt with that fully in mind

AI also demonstrates the power of assumption

if you aren’t getting what you assume you should, then it causes you to reevaluate how you ask for it

[trail lion]

or people just have lofty expectations, if a statement is a broad as "everything looks bad"

I'm quite happy with things I generate on XL, but I dont know what picasso's some people are looking to generate

[karmic cedar]

and yet picasso is just another style in this context

[verbal osprey]

does anyone know how to prompt two different characters in a scene so that specific descriptions only affect that "character". For example, a man with blonde hair wearing a suit next to a tall brown haired man wearing hawaiian clothes.

I've tried things like this but it combines descriptions and sometimes just gives me one character. Or if I have two loras for two different characters, SD will only show one character and it will combine the characteristics that I described

[trail lion]

multiple subjects is tricky, you can try a regional prompting extension, or do lots of inpainting to fix it, but getting a great t2t2img gen, it's not going to be easy unless you have a lora that has both subjects trained in

[serene oak]

hi i'm new. am wondering if we can use music from the live stream https://stableaudio.com/live for non-commercial projects---e.g. podcasts, voiceovers...

[stone stag]

When will the bots be back?

[fervent thunder]

Anyone good at inpainting can help me

[sage reef]

depending on your question, maybe someone who is not good at inpainting can maybe still answer it :3

[west prism]

i never know where i am

[crude notch]

neato

[low moon]

Wasnt there an SD15 room? Its dead with Csscade? 😦

[low moon]

Bots are on bathroom break.

[karmic cedar]

https://cdn.discordapp.com/attachments/1004159122335354970/1232844239570862120/IMG_2688.jpg?ex=662aef2b&is=66299dab&hm=35b6fff66ba2acfd6e4b8e04e9b34a0f12eedfd1a6b699a6ba43e0ee773cf154&

[burnt wedge]

hello guys trying to make realistic lighting, but can't get good result. (Sampling Steps 60, CFG 6.5 getting best results) JuggernautXL model any tips?

[zinc bane]

Hey, I would like to change the background of some pictures for touch up or completely. Until now my workflow was using outpaint with a mask of my subject to keep it as is(inpaint upload on forge) and then touch up on Photoshop for matching the quality between the person and background. Doing so when generating i get a lot of mutations around the head or extra hair even with negative prompts that make the whole process a bit tedious. Is there a better way I don't know about? Should I just train a lora with my subject from head to toe and let it generate the complete image to match in every aspect?

Please @ me so I can get notified of your answer

[low moon]

https://www.flexclip.com/tools/remove-background/

https://pixlr.com/remove-background/

https://www.erase.bg/

[zinc bane]

Hey thanks for the answer but maybe I didn't explain myself well. I want to learn a good workflow to have images of e.x. me in another background seamlessly. And I was wondering if a lora of e.x. myself( the model) would be a better approach for less touch-up or some other way the pros knows of.

[fervent thunder]

does anyone know why when using PonyXL, when I increase the height of the generated image, it makes the characters stretched vertically?

[foggy halo]

sd3 huggingface when?

and how to use unlimited api sd3 generatiosns ?

for some reason glif doesn't allow me to generate sd3

[verbal osprey]

Few weeks

[low moon]

the SD3 slot is ready at civitai lol

[verbal osprey]

is it?

It's going to take some time for people to start making SD3 models and loras. The base models are never that interesting when released. At least for me.

[low moon]

yeah well Sd3 already is way better than vanialla sdxl

it make grafitti!

am i hallucinating memories or there was an sd15 channel which has also been archived together with cascade...

[verbal osprey]

inpainting?

[hasty hornet]

yea that's too complex prompt for sd without inpainting

[verbal osprey]

Is there an easy way to train loras now? In the past it seemed like you had to resize all your images to be the same size. Is that still an issue? I'm waiting for a time when training is a lot easier and faster.

[narrow kernel]

Just use blender or UE5 to rig up a basic scene with a mannequin, render a depth pass of it since it's easy to do in those and 100% accurate vs using zoe or other estimators and then use that with a depth controlnet at like 50-75% strength

You might have to invert the depth map depending on if they render white to black or black to white

[sharp bough]

can anyone help

[zenith prawn]

Is Stable diffusion allowed to make games on Steam ? It fear some of the art used to train Stable diffusion are copyrighted

[astral goblet]

valve typically doesn't want anything on steam that could cause them copyright licensing issues. all developers selling on steam must own all the rights to their software's content. So that considered, stable diffusion can create any number of copyrighted material. They won't let you hook a model up to your game and allow people to put any character they want into it. Maybe though, maybe, if you create a model that you can demonstrate to them doesn't produce antyhing outside the realm of the training material you created, maybe they'll allow it. Dont know if that's been tested. Constraints though, might be the way.

[pearl ocean]

Unreal

[astral goblet]

i dont think they care about peopel using SD for asset creation in production.

[zenith prawn]

Here is what I found :
First, we are updating the Content Survey that developers fill out when submitting to Steam. The
survey now includes a new AI disclosure section, where you'll need to describe how you are
using AI in the development and execution of your game. It separates AI usage in games into
two broad categories:
Pre-Generated: Any kind of content (art/code/sound/etc) created with the help of AI tools
during development. Under the Steam Distribution Agreement, you promise Valve that
your game will not include illegal or infringing content, and that your game will be
consistent with your marketing materials. In our pre-release review, we will evaluate the
output of AI generated content in your game the same way we evaluate all non-AI content

• including a check that your game meets those promises.
  Live-Generated: Any kind of content created with the help of AI tools while the game is
  running. In addition to following the same rules as Pre-Generated AI content, this comes
  with an additional requirement: in the Content Survey, you'll need to tell us what kind of
  guardrails you're putting on your AI to ensure it's not generating illegal content.
  Valve will use this disclosure in our review of your game prior to release. We will also include
  much of your disclosure on the Steam store page for your game, so customers can also
  understand how the game uses AI.

[astral goblet]

Cool. Updates!

[zenith prawn]

Now I am wondering what kind of questions are asked in the survey. For example if I use Stable DIffusion but don't own the art that was used to train it, would it be allowed ?

[civic schooner]

Is it legit to hire an artist who learns from some fine art that do not belong to him?

[loud solar]

Well you can't really know the material it was trained on as long as it's not your own model.

[pearl ocean]

Maybe a 4080 is all you really need

[zenith prawn]

So if I use a public model from stable diffusion it is probably using copyrighted material ?

That means I can only tell Steam I am using model YYY from stable diffusion

[loud solar]

Who knows?

To train a special style you could also use selfmade look-a-like pictures ...

[torn sable]

heya, anyone worked with animatediff here?

[pearl ocean]

You gotta say, you can’t go wrong with an iPad Pro

[fervent thunder]

yes i wwould like to help

[fervent thunder]

Hey everyone.
What is the best model for vegetation and landscapes? I am looking for a model that is good for img2img and inpainting specifically

[exotic vale]

hi guys

[sudden ruin]

Loving my iPad Pro
Especially with the pen to draw stuff

[lyric rover]

/bot-10

[loud solar]

https://discord.com/channels/1002292111942635562/1047610792226340935

[wet grail]

Hey Everybody ! We have just launched our beta product to generate webcomics easily thanks to AI (and Stable Diffusion). We would love to have early feedbacks on our solution. go and check www.artale.io to be one of the first to test our beta !!! Your feedbaks are super valuable to us !!! Thanks again and have a great dayyy

[eternal pagoda]

Hello friends, I want to create QR code, is this possible?

[plush reef]

hi?

anybody konw something about sd?

i got some problems~~

really need some help~~

[still glacier]

cf https://stable-diffusion-art.com/qr-code/

[plush reef]

i installed sd,but how to reopen this webui?

[noble orchid]

https://suno.com/song/fcedaca6-eaad-4b99-b6ac-aa28feb12d6d

[eternal pagoda]

thanks

[still glacier]

assuming you installed automatic1111, you just run webui-user.bat

[plush reef]

really?

i will try now

[rough mirage]

Hej hej everyone!
I'm the founder of a german non-profit doing projects around tech and the influence on society.
For a new art project I need a bunch of as realistic as possible images of different everyday-looking humans to create Polaroid (paper) images with them.
Anyone interested in contributing?

[rough mirage]

(speaking of AI generated images of course)

[simple zephyr]

i want to creat a logo

[trail lion]

Go ahead

[tight glen]

Hey! My name is Alex and I've been working on a project enabling video creators to get professional sound design for their video for a fraction of the time and cost by generating tailored audio in a timeline editor. I'm currently focusing on creating video-based sound effects and ambiences.

You can try it out at https://adorno.ai ! I'm always super excited to talk to creators like you to keep developing the product in the right direction so if you have time to chat, that would be amazing.

[cold fog]

can someone please share some good API service for text to video generation ?

[fervent thunder]

hi, where i can generate images? wich chanel?

[slow bramble]

#🏞｜general-with-images Woodcutter's Cottage

[digital arch]

Wow so many AI projects popping up huh

[fervent thunder]

can anyone help me how can i create imagws here if anyine can then ping me and write how to?

[pine fiber]

you cant

not in this server

if you want to use sd to create images you can use a tool like sd web ui or comfyui

[soft iron]

😭

[teal pagoda]

When will SD3 will launch?

[plucky cosmos]

hi everyone i am new here

[mortal delta]

hello and welcome to the world of ai.

[plucky cosmos]

hehe ty for welcome me , only u online here now

[mortal delta]

Thats most likely the case i kinda have no life.

[plucky cosmos]

you are funny bro , hehe i like it

[unkempt depot]

Hi, I am trying to generate videos of AI celebrities dancing or doing other movements, like shown in this video: https://www.youtube.com/shorts/ASkd9Oxk1Eo . I want to recreate the dancing animation with AnimateDiff from Stable Diffusion, but I am still not sure how I could integrate celebrity faces onto my animated characters?. I tried using Lora's, but this doesn't work as firstly, there aren't Lora's availabe for every single celebrity, and secondly and most importantly, I found out that basically every celebrity Lora has a non commercial license (at least on Civit AI), and thus I won't be able to publish commercial videos with them. Using some kind of face-swap AI would be the obvious answer, however, I don't know what AI I could and should use and I also don't know how exactly I could integrate this into my AnimateDiff workflow. FYI, I'd prefer using AnimateDiff within A1111, but also wouldn't mind using ComfyUI, if it would be easier that way. Any help regarding what face-swap AI I could use and how I could use this would be very much appreciated, thanks in advance 🙂

[vivid basalt]

Check out my latest music video. Ai generated lyrics, music, images and video. Two days production time with no budget and no talent/actors.
You can can ignore what's happening, but this wasn't even possible just a few years ago under such constraints without any copyright issues.
(SD images / Suno music / gpt4 lyrics) 👀
https://youtu.be/O3HRqKHxBXQ

[trail lion]

check the announcements channel for updates

[shy flicker]

Can I try stable diffusion 3

Like for free

Or is it only paid behind the api

Any third party website maybe

[rich kestrel]

imagine.ai

or something like that

has some free tokens

[rich kestrel]

make ur own lora

how much vram u got g

u dont need tht much for loras

and theyre eezy 2 train

face-swap is last resort. doesnt look as good

[unkempt depot]

Thanks for the answer, I think the problem is that I would need to create tons and tons of Lora's. To be more specific I want to create videos about top 10 celebrities in certain categories and then create a transition of the number 10 celebrity dancing to the number 9 celebrity dancing and so on, if that makes sense. I want to create these top 10 videos on lots of different categories, thus I will be using lots of different celebrities. I'd have to train like 10 Lora's each time I want to upload a single video, I think it would be easier to use face swap, even if the quality might not be 100% optimal.

[rich kestrel]

well you would only need 10 loras then

its not that time consuming depending on how sophisticated you want them

A-list celebs dont even need a lora since most models already have them nailed down

1.5 models that is

[unkempt depot]

Oh wait really? I tried creating celebrities, but I only used SDXL models, and the results were pretty bad, I'll try out 1.5 models

[rugged mirage]

you can also probably create 1 lora for 10 people with the right tags for each of them especially if the model knows them a bit

[rich kestrel]

I wouldnt recommend merging loras

but yeah thts why I only stick with 1.5

all other models are nerfed to hell and back

[unkempt depot]

Well I have never really created loras myself, but I plan on uploading videos 2x a day, I don't really know how time consuming creating Loras is, would creating 20 Loras a day consume a lot of time?

[rich kestrel]

why would you need to create 20 loras a day

you create a lora per celeb, they dont disappear after usage lol...

[unkempt depot]

Well, as mentioned I am creating videos about top 10 celebrities about certain categories, a few examples would be top 10 most followed celebrities, top 10 richest celebrities, top 10 most hated celebrities. While some celebrities will feature in multiple videos, most of these videos will contain 9 or 10 new celebrities, because the topics will be quite different.

[unkempt depot]

Btw do you happen to know if this is also the case for 2.1 or just mainly 1.5?

[rugged mirage]

if you set up a workflow after doing it once or twice you can just start your next day's loras in the evening so they are ready the next day

also 2.1? nobody uses that anymore, either 1.5 or sdxl mainly

[unkempt depot]

Oh ok, tbh I basically only use SDXL, didn't know that, thanks for letting me know

[ornate flame]

Can someone make AlignYourSteps for Pony 8 steps?

[rich kestrel]

2.1 is trash too

xl depends on who u ask

[hollow gorge]

If anyone is interested in a free Alternative to Midjourney, I've been working on a bot that can generate images at a similar level if not better. The images generate as fast as 8 seconds. LMK if you want to test it out, we are looking for feedback and suggestions from anyone who is willing to help out.

[rich kestrel]

which model does it use

[hollow gorge]

It uses Juggernaut V10 SDXL, with prompt expansion

[rich kestrel]

Is it discord-based, if so send invite plz

[sudden ruin]

Please use #1092446741984444416

[hollow gorge]

igu, the link is in my bio aswell

[lethal gorge]

i just joined it but all channels are locked for me

wait its fine now

[broken cave]

what widescreen resolution

should be used with stable cascade

why is this so hard to investigate?

is the answer, "none." only generate square?

[rich kestrel]

cascade uses very specific resolutions

I forgot which, but only a few work

the reason there is so little info out there is bc cascade was relevant for 5 days and died instantly.

[little mortar]

can someone help me install a model
i tried on citvai

is there paying model

bcs i cant open the folder that i installed

[broken cave]

yeah. you know, it didn't say that anywhere

that's a hallucination 🙂

@neon oriole do you know which resolutions stability's stable cascade checkpoint was trained with?

[hollow gorge]

Yeah you just gotta verify, thanks for joining 🙂

[spring forge]

Hey, looking for a good nsfw anime model. Does anyone have some good references?

Could be nsfw / sfw together too

[little mortar]

how do i import a model from civtai to my stable diffusion ?

[trail lion]

depends, what kind of model? a full checkpoint? a lora? a TI embedding?

[little mortar]

im a beginner id really see where its write
this is the model https://civitai.com/models/153568?modelVersionId=433727

V10 or LCM 4

[trail lion]

ok, that's a full checkpoint - so download the .safetensors file and it goes <your stable diffusion webui folder> / models / Stable-diffusion

[little mortar]

ok so its normal that i can't open the model folder safesentors

ok then do i have to install the extension for the model

[trail lion]

no, assuming your using automatic1111, you just click the refresh button next to the model name to reload the list, and you should see the new one, or alternately stop and start the GUI

[fathom pulsar]

I'm pretty new to SD in general and currently toying with SwarmUI. Is there a way to inpainting in Swarm? Mainly to fix hands.

[solid whale]

hi, does anyone know If there is anyway to like filter the checkpoints I can choose from in A1111. I probably explained this badly 😅 . What I want to do is be able to select SDXL or SD 1.5 or SD 2.1, etc. and then only have Stable Diffusion checkpoints appear that are in this version. Is there an extension for this?

[knotty rivet]

for hires latent fix in confyui what size or % should i increase it by

[sleek lily]

I have a picture of a car seat, and I want to replace the cushion in the original car seat picture with a cushion picture to achieve a perfect fit. Can Stable Diffusion do this? How can I do it? Is there anyone who can help me complete this? Thank you very much.

[regal snow]

I learned something tragic today: A1111 has a controlnet API endpoint that returns filtered data based on control types (like in the UI).
The tragedy is that gdamned Forge doesn’t have it

[karmic cedar]

What if one day there’s a new img2img mode where the AI interprets the content of two images using T5-level attention and combines them into one

Like they have this sort of using things like instant ID and controlnets but…

Something more fundamental.

[karmic cedar]

https://videogigagan.github.io

Omg Adobe

[soft iron]

😭

[fervent thunder]

😆

[sage reef]

https://tenor.com/view/gimme-what-about-bob-bill-murry-i-need-gif-19552065

[pearl ocean]

no crying!

[placid elm]

Hello, is there any free web based SD UI?

[pastel lynx]

Checkout on HuggingFace there might be some

Hey, anyone know some nodes or workflows to do inpainting in ComfyUI in real time ?

[slate orchid]

Want to check if an Image is an Advertisement? Try this: https://github.com/chitradrishti/adlike

[pearl ocean]

how realistic can an image possibly get

[hasty hornet]

That probably should be in #🌶｜off-topic

[lunar pendant]

Hi

[pearl ocean]

are there any other sites to download models from other than civitai??

[pastel lynx]

hugging face

Someone does know where to put the api key to activate SD3 ?

[pearl ocean]

other than that lol

[pastel lynx]

i guess hugging face, civitai and github are the most trusted platform where you can get models

obviously any website that could host files can have models but do it at your own risk

[lethal jay]

Hey guys Im getting really horrible images in SDXL Turbo using ComfyUI how do i solve it

[pearl ocean]

use Forge, and epicrealismXL

[static cape]

That looks to good to be true.

That being said... how can we get that local? xD

Maybe Topaz can figure something out on that base and use it for VideoAI 4 or something?

[pearl ocean]

I hear some people who use SD are extremely obsessed with Anime

[pine fiber]

i hope we can something like that for anime

old anime that is only preserved in like 360p lol

[vale steeple]

you may be surprised how many people use SD for NSFW purpose

[meager kayak]

Hey guys, do you know if exist a good ControlNet inpaint model?

[trail lion]

I dont think for XL, but there is a good one for 1.5 models

[charred mesa]

for XL there are like foocus models that make inpainting with SDXL a breeze

[trail lion]

really, never tried fooooooooooooocus

[charred mesa]

https://github.com/Acly/comfyui-inpaint-nodes it doesn't have to be foocus

you can use these models with comfyui as well https://huggingface.co/lllyasviel/fooocus_inpaint/tree/main

@bleak matrix could you please destroy this person

I am not upset

[bleak matrix]

Please read the #✍🏼｜rules-and-tos for more information about server rules. Thank you!

[charred mesa]

thanks

[rich kestrel]

depends. if human, then pretty damn close

[visual briar]

Any news on when or if the bots might come back? Status hasn’t been updated in over 2 months.

[hot vortex]

Hi everyone. What lora would you recommend for anime real like pictures? I'd like to make one of higher quality. Today I experimented a bit, it took a few hours to fix hands etc. and upscale it, but I'd like something unique. I see a lot of pretty pictures but don't know which loras could be used for this job. preferrably SDXL (yes, it somehow works on 4 GB VRAM)

[sudden ruin]

Check #🍥｜anime and see what others do and ask around maybe

[hot vortex]

thanks for the suggestion, I'd like something of a good q uality because it takes like 3 to 4 hours to upscale the picture into a really big reslution

wait, I did more than 4k resolution, that's why it took like 5 hours

[shy flicker]

Sd 3 will be released on hugging face the same way sdxl

Was

Right

[low moon]

I wodner if there is like really bizarre weird undeground models only shared privately or some weird torrents lol....

[loud solar]

I bet we don't wanna know ...

Imagine a model based on police crime scene photos ...

[rich kestrel]

i mean

most 1.5 models can already do cheese pizza

so there really wouldnt be a point

never blame the tools, always blame the perpetrators

[sage reef]

if only there was a technology that can capture footage from our dreams and then we could train models to generate some weird dream pics (or videos) LOL

[low moon]

Or anything medical.

Or pizza without pizza

[loud solar]

Pizzaface Freddy?

[sage reef]

a good model is one that can generate a boneless pizza as well as pineapple on pizza

[karmic brook]

this ^ let's veer away from inflammatory topics here in the server

[rich kestrel]

kohya ss

prodigy scheduler

50-100 repetitions

good tagging

and 2500-7500 steps. boom ur done

u can do both

if u got big vram u can do dreambooth

if not u can do lora

I recommend both but for most folks lora is enough

16 should be plenty

[dusty coyote]

Open Source Stable LM Chat App : https://github.com/nerve-sparks/iris_android

[mortal delta]

I ordered some pc parts from china and i must say i dont like waiting, its been 2 and a half days and im so bored.

[sage reef]

i ordered sd3 weights and i dont like waiting as well

[rich kestrel]

lora can be used on different models

dreambooth is a model finetuned to your art, so less flexible in that regard

[mortal delta]

Thats pretty relatable...

[rich kestrel]

I think XL is both dreamboothable and lorable but I wouldnt know if the same specifics apply since I dont use anything other than 1.5

for that ur better off asking in #🔧｜finetune

[sage reef]

why do you have a Toyota if a Ferrari exists?

[oblique elk]

I prefer good refined models of 1.5 sd. You got plenty to chose from as they are easier to train and merge. So one of my workflows is using one model (eg anime style) for the initial image (denoise about 0.9), followed by an upscale and tile control net to regain details. For the tile I use others models like realistic ones. So you get interesting results. Anime clothes, poses and face but with a realistic touch. It is fast and if you do not like how the first image turns out you can stop and retry. Without the need of performance intensive calculations.

[sage reef]

@rich kestrel btw, have you tried other things than koyha, like OneTrainer for example? cause i want to train a lora, and will start with OneTrainer, but wondering if there are any major objective differences or is it more just personal preferences.

[regal snow]

I'm here to toot my own horn real quick. I'm a bottom of the barrel, fake python coder.
Yet in a few hours time, I managed to implement ControlNet's 'control_types' API endpoint into Forge.

Which is missing probably because they deprioritized it then never got around to re-adding it

I'll be shooting a PR to Forge, where it shall languish for eternity

In the meantime, it's truly an essential endpoint for anyone using ControlNet via API

[tribal sequoia]

I'm curious about something but I'm not sure how to search for this. Tell me if you know which search terms to use for this? You know how you use positive prompts to move the AI toward certain concepts and negative prompts to move the AI away from certain concepts? I want to see how far apart the AI considers various concepts. Like if I tell it "negative prompt: ugly, cartoon", the result isn't necessarily going to be closer to "beautiful, realistic", right?

[regal snow]

someone will probably have the right answer, but the model probably has weights for cartoon and by putting it in the negative prompt those weights are lowered? Not necessarily increasing the weights for something else?

[tribal sequoia]

yeah nobody's going to be able to summarize in one or two sentences the entirety of how the model interprets prompts. That's why I want a way to visualize what it's thinking.

[rich kestrel]

1.5 is vastly superior

matter of opinion ofc

but check out my model senblend 1.10 link in my bio and judge for yourself

in the right hands, 1.5 is indestructible, but it isnt very n00b friendly, you gotta know how to use your tools to make it good.

when it comes to lora, kohya_ss is probably the best bet, I havent tried onetrainer but if it works for u go for it

Lmk if you want my preferred settings file for prodigy and/or if you want my guide for dreambooth (do not use kohya for dreambooth, we use a1111 for that)

[rich kestrel]

its not magic g its just tags

there is a small dataset that has been tagged with "ugly" or cartoon and the AI knows what that looks like