💬｜general-chat

[fervent thunder]

it should really be MAAA

[fervent thunder]

watermark wont piss off.. no matter what inpaint settings I try, is there any tricks? EVERYTHING on my image is perfect except the loser traditional artist who put a watermark down on the data

[junior urchin]

You can always train a model yourself, handpicked and edited

[desert dagger]

photoshop?

[wanton lotus]

what happend with sd3

[wispy arrow]

Drawing a blank here, what is the command line arg to disable changing settings when using the --share flag?

[trail lion]

there was an announcement and a release and some bickering over a license and some anticipation of some other verison of it, the end

[south basin]

hello community ! anyone knows if there's a v-prediction version of SDXL?

[scenic moon]

hello everyone

[primal moon]

阿巴阿巴阿巴

[abstract quarry]

pseudoterminalx trained on

https://huggingface.co/ptx0

here you find his models

[neon hawk]

what is SD3???? 🙂

[rancid ore]

what's the use of hugging face platform ?

[fervent thunder]

isn't cosXL also v-pred

hugging face is where people put data and models

[rancid ore]

i dont get it with the price there then,
if just data wouldnt it be free? and the price itself not even for data and models.
why there's gpu stuff etc. sorry im new

[fervent thunder]

hugging face is free

I think they offer some cloud GPUs

but you don't have to rent them

[rancid ore]

hello im trying to make some arts, and need more control like this guide: https://www.youtube.com/watch?v=XtMvk0dpnO4&list=PLNlRhPQovztRqp_zyp-lY79fWZIzjnNTf

that can crop the image and re-generate again.
but it seems with stable assistant unable to have that range of control, correct?

also do i need stable artisan then? imnot sure what is that for.
bcause i need the webgui + huggface surely right to have full control

i thought it's open source? but if i use by webgui and running it on my pc, then i still need to pay the service ?

[fervent thunder]

you're getting confused between a lot of concepts
its better to start simple
install Fooocus and try to make some images

[rancid ore]

yes, im really new with this ai=stuff directly. im just blockchain dev and softwre engineer for 8y hvent touch pytorch or stuff directly.
what's the different with fooocus? it's newer webui ?
so i dont need AUTOMATIC1111 WebUI ?

[fervent thunder]

not actually sure which came first

I think A1111 came first but I am not 100% sure

the point is that Fooocus has less options so its easier for newbie

[rancid ore]

so both are just competitor u meant?

[tulip yarrow]

Any SD3 Lora trainer yet?

[fervent thunder]

yeah both are competitor

[tulip yarrow]

Also, will SD1.5 die out after that 😭

[fervent thunder]

and no Lora trainer yet

[tulip yarrow]

Do you think people will still use 1.4/1.5/SDXL after the trainer is out

[fervent thunder]

I still use SD 1.5
I think its the best model out of any currently

[tulip yarrow]

SD1.5 will always have a place in my heart

And honestly yeah I'll still use it

[fervent thunder]

its the strongest for photo portraits

out of any model including dalle, midjourney, ideogram etc

because it was overtrained on photo portraits

[tulip yarrow]

What do you think SD4 will be like (even though it's too early to predict lol)

[fervent thunder]

hopefully move above 1000 pixels for base resolution

even beefier VAE

there haven't been many breakthroughs in the last year concerning pure image quality

aside from just raising the resolution

I actually think diffusion models have hit a soft limit for now

[tulip yarrow]

How about actual perfect hands

[fervent thunder]

and we have to wait for the next architecture (probably GANs)

or maybe something autoregressive

hands is probably just a training data issue

I think there is too much focus on hands

don't think its a technical barrier

[rancid ore]

wait, i dont need to pay for both of them right? and i can generate as many as I can?
im using m1 max too, and can do that locally, correct?

[fervent thunder]

no you need a GPU

to get reasonable speed

an Nvidia GPU, not an AMD one

[rancid ore]

so apple silicon gpu wont work? oh that's why hugface offer gpu.
damn

gonna hard for me to experiment a lot on early, since i had no nvidia stuff

[fervent thunder]

some people use apple silicon
but its really not a good idea

there's a big apple fanbase
that is sadly tricking people into thinking that apple is ideal for AI
its like 10 times slower than GPU

[rancid ore]

yeah, i need to cashout another pocket for gpu laptop then

[fervent thunder]

can you get desktop?

laptop heat issues

its more stressful than gaming

for the computer

laptop can work though

[tulip yarrow]

Any reliable ways to train LoRA on kaggle?

[rancid ore]

ASUS ROG Flow X13 RTX 4070
bcause i might moving a lot.
and i want to sketch too, and finalising my art with ai

[rancid ore]

yeah i hate this, i dont know good windows laptop without heat

that's why m1max is a beast for me, no heat at all.
m3 even crazier

[fervent thunder]

yeah but they stay cool
by not having a GPU
LOL

[rancid ore]

and also the damn brick battery windows laptop

[rancid ore]

idk they this stated havign 24gb gpu ?

so this gpu isnt usable at all ?

[fervent thunder]

its very slow

its integrated graphics

[rancid ore]

do u know best gpu cloud service?

[fervent thunder]

yeah Rundiffusion

its the only thing I use

[rancid ore]

why u use it? i thought u have gpu

[fervent thunder]

I do but I want to use my PC at the same time

[rancid ore]

ic, may i know what u use for cloud gpu is it for u work?
and how long it takes to render or some stuff. want to know some of pov

[fervent thunder]

I don't use stable diffusion for work, no

render times depend on quality

with SDXS you can make 600 images per second LOL

but an image with heavy tiled upscaling can take 8 hours

so anything between 0.001 seconds and 8 hours
is reasonable

you can choose the trade-off between quality and time

[tulip yarrow]

you still use that in 2024?

[tulip yarrow]

I don't know probably Colab or Kaggle.

Saving up for an actual GPU laptop (which is around $800)

[rancid ore]

im talking the general windows laptop.
im using macbook though

im trying to test sdxl sd3 whatever is that, but seems gonna be a problem with my mac and making it hard for me to quick start

[steady heath]

qqqqq

[rancid ore]

i just tried comfy ui and already liking it, damn it's powerful

[storm shard]

You can run it local on your M1
https://drawthings.ai/

@rancid ore there's also a Draw Things discord which is very helpful, lots of info there and people who are willing to help

[unkempt hatch]

I cannot for the life of me understand why MoE didn't take off
the paper seemed absolutely amazing, and then instantly forgotten

[lament moth]

https://poe.com/PhdExpert-CDvr4

[old lichen]

Is there a hires fix for img2img?

[abstract quarry]

do we refer to the same mixture of expert paper?
I found it really lame

1. all experts were just finetunes of same model
2. moe usually swaps single linear layers and not entire models
3. if I want my image to be xyz style I can just select the right lora or model manually

[unkempt hatch]

what do you mean linear layers?

the one that made SD1.5 able to do text and count and increased prompt adherance.

Ya, it's various finetunes of one model, but images you got a hand-finetune, face finetune, anatomy finetune etc...

the problem with Loras is they blend everything; so if I've got 2 character loras, I end up with a blend of 2 characters, instead of being able to include each.

[abstract quarry]

I meant mlp layers not linear layers

[abstract quarry]

but moe don't solve this problem. Regional loras can help you with that

[unkempt hatch]

you can see here it does. Better image quality and better prompt adherence. And without even using specifically trained models

https://youtu.be/6Q4BJOcvwGE?si=LdajWOtf4iTKGVWJ&t=844

[fervent thunder]

mlp layers are linear layers

he's right though, in LLMs MoE normally only swaps the linear layer and not the self attention or cross attention

ah actually I checked the blog post for segmoe
they are also only swapping the linear layers

a SegMoE model comes with multiple models in one. The way this works is by replacing some Feed-Forward layers with a sparse MoE layer. A MoE layer contains a router network to select which experts process which tokens most efficiently. You can use the segmoe package to create your own MoE models!

(feed-Forward layers = mlp layers = linear layers)

so in that respect it is the same as in LLMs

[abstract quarry]

it's not exactly the same, cause mlp/ff are two linears with a non-linear activation in between. That's why I rather corrected myself before someone starts being picky xD

the segmoe is just using layers of finetuned models. It's not what moe in llms are doing.

in ny opinion the paper somewhat rides the moe hype. "moe are so cool in llms so we have to do the same". There method is not that great, though

[abstract quarry]

they don't really improve. They only improve compared to the base model, but not compared to the finetune model

the segmoe is just a very complicated way of using different models for different art styles

like when "photograph" occurs in your prompt, use a realvis model, if "anime" occurs in your prompt use an anime model. segmoe automates what everyone can just do manually and more precisely

[unkempt hatch]

but in the examples, it shows the SegMoe counting better than the finetunes without SegMoE

[unkempt hatch]

that's interesting. I wonder if you could merge models like that - taking only the layers from each which process a token most effectively

and I imagine if you had models specifically trained to work in SegMoE, it'd work even better. Like, 4 models in a SegMoE being similar to a model with 4x the parameters

it's be a great way to make models scalable, so that someone with only 4gb vram and someone with 24gb vram could be using the same resources, just the one with 24gb could use more at once

[fervent thunder]

on a reddit post the dev said it improves clip scores a bit

(higher clip score means better prompt adherence)

but didn't say it raises image quality

I replied to the wrong comment

but yeah

[abstract quarry]

nah, I'm wrong, according to the code it does use a merge of the ff for each layer separately

then I have the paper wrong in memory

[fervent thunder]

its confusing how
so many places talk about MoE and merging differently

and experts and routers etc

also terms that are used differently

[unkempt hatch]

I just meant image quality improves because you can see in the paper and videos of people using it that it's making better images

I wish I could just try it, but I get errors in ComfyUI... I don't get why no one's even playing with it though

[fervent thunder]

we need a big sample size really

image quality assessment datasets like KADID-10k have over 100,000 images

some have more like 1,000-10,000

but that's the sort of numbers that are needed to avoid cherry picking problems

we have some nice image quality assessment transformer models now like MUSIQ, DEIQT and MANIQA
so it might be interesting to run those on the output

[unkempt hatch]

if I could get it working in Comfy I could just test myself; but looking at Nerdy Robdent's examples they definitely look better, and I don't think he's cherry picking

[fervent thunder]

its not that I'm accusing him of cherry picking
I'm saying that in general anything under 1,000 samples is likely to have cherry picking problems
its not about the individual person I don't actually know them

[junior thistle]

So I have all these extensions on a1111

but I never see them on txt2img

tab

[fervent thunder]

aha!

[cerulean tangle]

Shameless plug https://civitai.com/articles/6024/using-a1111-why-not-swarmui-a-transition-guide

[undone garden]

Is there a way to queue training sessions in khoya to keep going while I'm afk?

[junior urchin]

You should add stuff that swarmui does better, not what A1111 does.

[cerulean tangle]

You're right, I just added a "So, why SwarmUI?" bit that I realized was missing from the beginning. I might need to expand on it though to include more features, though the general gist I'm hoping to get at is that SwarmUI can pretty much do anything Auto can and more.

[junior urchin]

You need to add a category for the "more" part in particular

Otherwise people will just ignore it, as it has the same features as auto, which they will see as a copy cat

[cerulean tangle]

Ok, I'll add that on once I get the chance, thank you for the feedback :)

[pale latch]

open model initiative hehe. they've pivoted from buidling their own base model to choosing a base model to support instead

[junior urchin]

And maybe implement https://github.com/vladmandic/automatic/ as well, as it has more model support

[junior urchin]

I think that was the base idea to begin with, in terms of feasibility

Not sure if they agreed on lumina or sigma

[cerulean tangle]

I did mention that one was more subjective. It has a totally new interface now and is tooled for advanced users. I should bring that up however, you're right. SwarmUI is expanding its model support as well though, but I don't believe it's in parity yet.

[junior urchin]

[pale latch]

Thought the original announcement was for building an open model from scratch not just another refinement of an existing poor base

They're not really innovative or taking initiative if they're sticking to what's been done. They might as well build on sdxl

[junior urchin]

Well, they also want “Safety”, “Ethics” and “Responsibility”

So a lot is practically out of the door

[pale latch]

That's a good thing. I'll just block you now before your bait turns toxic.

[junior urchin]

They mentioned it in their own reddit post

As to whether or not you skipped past that, ignoring it, is on you

https://www.reddit.com/r/StableDiffusion/comments/1do5gvz/the_open_model_initiative_invoke_comfy_org/

Recognition of unconsented artist names, in such a way that their body of work is singularly referenceable in prompts

Generating the likeness of unconsented individuals

The production of AI Generated CSAM

Surely you have read that

[pale latch]

yeah taking such things seriously is a good thing. amazing how there are people who don't agree with you 😮

[junior urchin]

If they disagree with me, they disagree with that post

So yes, they can do SDXL, but it would need to be trained way harder than pony was

[trim magnet]

so they installed a rootkit inside the model that checks for celebrity faces and national id's of characters to check birth date? 😮

[junior urchin]

That would be interesting :o

And then default over to something like https://thispersondoesnotexist.com/

[trim magnet]

it would be an insane achievment,not even chinese models can do that and they have millions of $ 😮

[river sparrow]

A dilemma. How to make realistic images of people, without any REAL people in the dataset ... you would almost require a "bad" model trained on actual people, in order to create "fake" people. And then use that as the sole data set for further training (if the images are different enough to ALL EXISTING PEOPLE!!!)

Even for my current face, the chances are very high that at least 2 other persons on this world look very much the same. (not talking about my brother)

[trim magnet]

easiest way is to go MJ way and make the model closed source 😃

[river sparrow]

closed source web-only access gives the best control

[cedar salmon]

monkey something something Shakaespeare

[pale latch]

context matters a lot too. if you're not trying to generate someone's likeness intentionally but an image you made that went viral and gained some notoriety had a face that resembled someone by chance and coincidence, it'd be hard to argue you were trying to profit from their likeness. The context of your image would likely not match anything else if it was just by sheer coincidence that they look similar.

context and artistic intention. Also factors in cases of identity. Even with art pre generative AI. Old laws and precedences still apply.

[trim magnet]

easy just train model so it understands old laws and human definition of context

[floral umbra]

Is there a way to make comfyui generate with 2 cuda devices yet? As I've held off from getting a second 3090 to double the iteration per sec until either automatic or comfy supports it

[pale latch]

not as far as i know. swarmui will spool out batches to multiple cards though so you can generate 2 images at once

[hard flax]

Hey crew - I wanted to share a new blog post I’ve just put out. It's a very simple process to help with turning quick executive summaries around. I’m often asked (you might be as well) to provide an executive summary on a topic, event, technology, etc. The turnaround time on these asks is usually a few hours to a few minutes, which is never enough time. These three GPTs and processes have helped a lot. 😀

https://www.dylandavis.net/2024/07/three-gpts-walk-into-a-bar-and-write-an-exec-summary/

I hope you find it useful and feedback is always welcome!

[fervent thunder]

when I saw it was about Custom GPTs I thought it was gonna be bad, but these system prompts are great

I think iterative refining like this is the best way to get writing out of LLMs

[undone garden]

Been absent for a while, did we get SDXL controlnet yet? Specifically tile?

[sage reef]

you can search on hugging face : https://huggingface.co/models?search=sdxl controlnet tile

[left cedar]

Hi everyone!
Could you please recommend me a fitting UI for a beginner? I've read online that Forge is a really good one for beginners, but it seems like it is moving away from general use case UI to some specific features testing tool (according to author's note on github).

[golden arrow]

Just saw your response, so I thought I'd ask:

I recently upgraded my card to a 16 GB VRAM one. I have my old one, with 8 GB just lying around. I am reluctant to sell it. Is there any means of "harnessing" those extra 8GB in any way?

If anyone knows?

[pale latch]

you could use it to game on while training on the 16gb if your psu has enough watt output

or you could run different instances of ui on each card

[golden arrow]

But don't you need a special motherboard for that? I am not keeping up with technology, but "back in my day" I remember something called

crossfire

or something like that

(I'm in my mid 30s, despite how I sound haha)

[pale latch]

some games supported that. that's like running 2 cards as one card. but it was always glitchy and required specific code so it didn't work on most programs

[golden arrow]

But, you are saying I can just connect additional card and use two cards on my desktop?

(I always thought that's what crossfire was)

[pale latch]

if you've got enough slots to fit them, you can run multiple different gpus for different tasks

[golden arrow]

I see

[pale latch]

crossfire would be something the motherboard and gpus and software all need to support for it to work

[golden arrow]

That's what I thought... But it has been decades already, that's not exactly new. I would have thought software would develop for it by now... 🤔

would have developed*

[fervent thunder]

crossfire was just for gaming

[pale latch]

turns out a newer more powerful card was often cheaper way to get fps

[floral umbra]

Aye, i know how to do that, but i'd want the opposite. 2 cards for one generation to double the speed :P

Especially when i do 2-4 upscales for one image to make it 2x as big than the last. And as some gens can take 10-15 min due to the sheer size, 2 cards for one gen is needed :P

[neat cypress]

even InstantID developers didn't think of this implementation 😄 this is my first ever latent space level code implementation 😄

[cedar salmon]

what the smile emoji

[desert dagger]

it's reddit. the best thing to do is disagree with anything posted on reddit and not read reddit

[junior urchin]

[flint spire]

Hey, im trying to install Stable Diffusion second time, and it again tells me to install idk "torch" and others stuff, how to install all of them at once?

[outer citrus]

with 2 cards it is better to work on 2 different images at once

[boreal cosmos]

with 2 cards it is better to sell one to me

[strong ginkgo]

#✍🏼｜rules-and-tos

[floral umbra]

Not for what i want to do.

It's why i'm waiting to see what nvidia will achieve with their 50 series. As 4090 was ass gains over 3090 if i can't split the cuda workload on 2 cards for one image

[trim magnet]

they wont do that because right now you can only do that with training on ada cards and the older gen ones so they wanna keep the ai multi gpu builds exclusive to ada (the 10k+ usd cards)

[outer citrus]

The 4090 is about twice as fast as a 3090

[trim magnet]

not on SD

[outer citrus]

i have both

it literally is

same workload, the 4090 does twice as many it/s

[trim magnet]

nope,on a 3090 i get 20 it/s and on a 4090 i get 32 it/s i dont see where u get the double speed

you can check this on vlad benchmark site

[outer citrus]

I dont have to check a website

4090: 1.56s/it batch 4 1536x1536

3090: 3.07s/it same batch

and the 4090 does it using the same power and heat output so way better

both are water cooled

[trim magnet]

well the website provides evidence,you dont

[outer citrus]

you cant send images here

[floral umbra]

1.5-2x depending. But as i'm mainly a gamer, it only had about 60% lift there. And for 2.5 grand in norway, 2x is not enough to justify that price

@cloud smelt Change your password immediately as your account has now become compromised

@hidden dagger Can't tell if you are one of the mods here, as there's no mods/admins role, but nainsy's account is compromised, and message needs deletion asap before anyone falls for it

[outer citrus]

I went from a 3090 and a 2080ti to a 3090 and a 4090 and I more than doubled my productivity

I don't game though only train and test new models

train with other cards

[floral umbra]

Aye, but i still desire more A.I performance than what 4090 offers :P If i want double, i can just get another 3090 for a third of 4090's price, but that depends if there's any such method of using 2x cards for one generation in comfy for instance :P

[outer citrus]

3090 is about half 4090 price now

they keep dropping

I got a 4090 msi liquid for 1400

crazy msrp like 2 years ago was 1650 or something close to that

[vestal dew]

I see no trained models for SD3 on Civitai....what's going on with SD3?

[digital ginkgo]

Different license, civitai won't be supporting sd3

[late surge]

hi

[spare plume]

I have 6GB VRAM on an RTX 2060 and 16GB RAM, which version of SD3 Medium do you think I should download ?

[pale latch]

civit issued a blanket site wide ban on all sd3

[fervent thunder]

starting to think civit did more harm than good

civit doing that could be main reason there are less SD3 lora

[warm junco]

The 5gb one with encoders integrated

[fair blade]

Anyone here familiar with ComfyUI?

Need help with something if possible

[spare plume]

Sure, go ahead

[fair blade]

@spare plume Do you know how I can download this? It provides a tutorial but I still don't really get how to do it as a beginner. https://github.com/kijai/ComfyUI-LivePortrait?tab=readme-ov-file

[spare plume]

I believe you can either download it manually (in which case you have to follow the readme file and clone the github repository in the correct folder, whose path I don't exactly remember), or use a specific plugin to manage other ones more easily. It's this one iirc : https://github.com/ltdrdata/ComfyUI-Manager

The folder to put your custom nodes in is "ComfyUI/custom_nodes"

Using git is recommended, allows for easier update

[fair blade]

I have ComfyUI Manager, is it okay if I pmed you and asked further questions? I won't spam or nothing lol @spare plume

[low moon]

this is why sd3 is over

[placid hatch]

anyone else find that adetailer makes all faces look the same?

[ripe atlas]

hey friends, I was using sd with automatic1111 solution for some time in the past (a year or so ago), but recently I switched to a macbook (m3 max btw), so I wanted to ask you guys about the state of sd overall, which UIs or apps or solutions are all the rage nowadays and which one are being used the most on macbooks ?

[burnt veldt]

imagine/

[warm junco]

Hey, in the pinned messages in #🤝｜tech-support You'll find the install guide for Macs with the best performance settings

[ripe atlas]

thanks

[sturdy flame]

has there been any news, or whispers through the grapevine about an update to the license?

[onyx oak]

can i use 画像生成？

[fervent thunder]

no but license will likely improve

they got a ton of new investors

[bleak matrix]

Good morning, everyone! How are we all today?

[fervent thunder]

okay

[hearty junco]

With some checkpoints i have some color spots (same colors). What is the problem ?

CFG to high ?

[fervent thunder]

sounds like vae

[hearty junco]

I allway use speed 20 cfg 7

vae inside check point ?

I dont use neg prompt and only subject in prompt

for model test

[fervent thunder]

sometimes color spots are just random

but it can be a VAE issue also

[hearty junco]

changed ghostmix release with beakedvae seams to work better no spots so problem solved

[plush jetty]

Any good stable diffusion news?

[boreal cosmos]

can 16GB RAM fully use 10GB 3080 for SD? or would I run out of ram

[desert dagger]

it's stable, but diffused 😉 - one of the guys here made some headway with a lora that he posted about a couple days ago

[pale latch]

Skill issue

[desert dagger]

it's why civit is slowly going downhill

[plush jetty]

Why is that?

[cedar salmon]

porn

[mortal harness]

Are there no other communities supporting the development of SD3?

[desert dagger]

there are a couple of repositories a lot of model creators have moved their stuff to, i don't know if any of them are supporting jsut one specific version of SD however

[unreal rock]

how can i generate images here

[desert dagger]

you have to use the artisan channels. start by reading the information #🗣｜artisan-support-feedback at that link

[grand rain]

whats happening with SD3 atm?

[reef wigeon]

I've gotten Stable Diffusion working and making ok, images. What are some resources I could use to get better with it? Is there a "For Dummies" book or a good YT channel to watch?

[brave grail]

lmfao

[runic stump]

“Our new Community License is now free for research, non-commercial, and commercial use. You only need a paid Enterprise license if your yearly revenues exceed USD$1M and you use Stability AI models in commercial products or services.”

Pretty awesome!

[brave grail]

yea theres plenty of websites and youtubers out there who are very good at explaining how to get started

are you using comfyui by chance?

[manic wraith]

congrats on doing the right thing Stability

[final gate]

SD3 is shit no matter what, you tried to hide NSFW so much you removed women from existance 💀

[river sparrow]

Best news from the license announcement if you ask me:
" Continuous Improvement: SD3 Medium is still a work in progress. We aim to release a much improved version in the coming weeks.

Model usage: We are all researching how to best use this new architecture and model. As we learn more about the best ways to use the model and to fine-tune it, we will share our findings as well as those of others In the community."

[desert dagger]

wrong

[manic wraith]

it is, but so is SD 1.5 and (base) SDXL. Those only are good because of extensive community fine tunes, which is why the license to do so is extremely important. Nobody using SD 1.5 or SDXL is using the base model outside of "hello world, let's check this works"

[pale latch]

the broken posed people isn't from censoring or protection from nsfw or any of that. it was bad pretraining. the safety training is different.

[reef wigeon]

I don't believe I am. I know I am using ZLUDA and Automatic1111. But that is because good people on discord helped me get it working. What is comfyui used for?

[brave grail]

oh are you amd?

[reef wigeon]

Yep.

[pale latch]

they've put a lot of aesthetics data into protected properties. its why celebriteis come up as completely different people

[river sparrow]

for reference, im running linux with amd gpu. A1111 worked with some tweaks, but now I prefer comfyui

[final gate]

No, if an artist doesn't learn how to draw naked people he cant draw people, thats how it worked and still works. SD3 can generate (almost) naked men and clothed men but no naked woman and also no clothed woman

[brave grail]

i think comfyui supports amd, but anyway, if u like automatic1111 then stay with it, if u like a more linear node graph setup then use comfyuio

[river sparrow]

Chris, AI is not drawing

[manic wraith]

use Automatic1111 if you like GUI, use Comfy if you're into programming (I am, and I use mostly A1111 because I am lazy)
for learning with A1111, use the "plot XYZ" script and try out stuff until you get the hang of it. Just see what each setting does for yourself, it's the easiest way to learn.

[brave grail]

chris has a point

[river sparrow]

Chris, do you want some bikini women from SD3 ... ?

[final gate]

Thanks

[pale latch]

there's enough data in the dataset to know how to draw people. the broken poses are because they didn't pretrain the 512x512 versiuon of the model on that original dataset

[brave grail]

isnt the point of ai for ai images for some people is big booba anime waifu or nah

[final gate]

I know that it is possible by using weird phrases

[desert dagger]

i haven't had any issues with those

[pale latch]

it's a different dataset than laion 5b. you don't need hustler hardcore to draw people right.

[finite cloak]

Mods, come on, don't remove the clown emoji from the update. You're asking for backlash.

[pale latch]

threatening mob vengeance probably isn't smart

[finite cloak]

Making wild leaps of logic isn't smart.

[reef wigeon]

I have developed no love for anything yet. Looks like ComfyUI is more flowchart-like verse graphical sliders.

[brave grail]

lmfao mods removed clown emote

[final gate]

Let the people show off their emotions

[finite cloak]

I use invokeai, but I was using automatic1111 before with AMD

[river sparrow]

Check https://discord.com/channels/1002292111942635562/1004159122335354970 for examples. Using this (censored) prompt:

"(modern anime style:0.9), s#xy woman in red bikini on a beach, tree shade, some cliffs, medium br##sts, cl##vage, half body shot, happy eyes, blonde hair and green eyes"

[reef wigeon]

Dumb Question, but are Automatic1111, Comfyui and invokeai just user interfaces for Stable Diffusion?

[river sparrow]

idk maybe my local SD3 model got infected by some older sd15 models ... 🤨

[pale latch]

Back on point. my hope for 2b is that they have continued pretraining or found a way to refine specific unrefined blcks of the DiT network. Just throwing more refining isn't going to fix 2B. it needs something more since the foundation wasn't complete to begin with

[brave grail]

yea

[finite cloak]

Not exactly, but primarily.

[pale latch]

Less poisoned latents would be great too. theres too much DPO breaking classes and making loras hard.

[brave grail]

they also removed the middle finger emoji lmfao

waah pixels make me sad waaah

[pale latch]

also all the underlying code that the weights plug into

[kind thorn]

The terrible licence was just the icing on the cake. They also need to improve the quality of the model and remove the built-in censorship

[reef wigeon]

cool, is it difficult to change? I have an AMD card.

[kind thorn]

SD3 was supposed to the the next big step forward after SDXL, not be signficially worse than it

[pale latch]

mmdit is actually really good still. even with all the safety they added.

[sudden ruin]

No mods were involved in that Action to be clear

[finite cloak]

I don't know what OS you are on, but if you have AMD you should strongly consider using a Linux distro. You don't need to do much to get these working with AMD in Linux. You need to install rocm from AMD. Otherwise you basically only need to tell the installers for the UI's that you are using AMD when installing.

[spare nacelle]

throw sd3 2b model in trash start over

[finite cloak]

What does that mean? Someone without mod authority can remove emojis now? Or that whoever did it is not a mod in terms of the role in Discord?

[sudden ruin]

We only have 2 mods lol

[pale latch]

i hope that the people with the OG training code can find a way to refine specific blocks of the new network architecture or something. i have no idea what, but a new network presents new solutions. maybe they can fix the pretraining problem without going back to the drawing board.

if not then yeah i think stripping 2b back to pretrained 512x512 and refining that more is a better direction.

[finite cloak]

Well, I watched it happen. I don't think 20+ simultaneously unchecked the clown emoji. lol I don't have a dog in the fight. I'm not outraged or anything. I literally was saying it's a bad idea because others would notice and the internet is wild.

[final gate]

Then who removed the emojis?

[sudden ruin]

There are roles above mods

[spare nacelle]

i mean if base model cant do simple human correct no reason pushing it forward just scrap it start over is my personal view i am assuming becuase through so much anime in sd3 it broken human anatomy

[finite cloak]

Putting any anime in the training is an egregious case of poor insight.

[final gate]

Angry SD employee then 🤔

[spare nacelle]

atlest from my lora making over past few years i noticed that mixing anime with realism generally give very cursed anatomy lol

[reef wigeon]

I wish I could, but I'm not ready to day drive Linux on my main computer yet. I game and use it for a media server. When I make an AI server, it will be running Linux.

[spare nacelle]

clearly sd knows more what there doing then me lol

[finite cloak]

Yes, it should always be noted that my opinions are largely uneducated. lol

[pale latch]

stability also has the actual training code for this new mmdit architecture. they may have a smart trick up their sleeve to repair 2b's failed pretraining.

it's not a unet anymore

[finite cloak]

My guess is that they needed to get a change to the license and SD3 was pushed out early to facilitate that. I know that there's been talk about bankruptcy if they don't get another round of funding.

[spare nacelle]

sd3 2b is great at text an alright overall art style but it bad everything else lol

i dont see how more training on it could make it beter lol

i look forward to testing new model regardless

[trim magnet]

well have you seen 1.5 base outputs? full of deformities that were fixed with a lot of training

[reef wigeon]

I'd like to try ComfyUI, I am using an AMD RX 7900 XT on windows 11. Do you know if I have to do anything special to get it to work?

[dim forge]

they just need to release the real model 🙂 I am not going to pay for api access just because the public version is garbage

[spare nacelle]

1.5 has beter anatomy then current sd3

can you train lora for 8 billion with it on api ?

i know 8 billion model require special set up to train it but is that even allowed xD

[dim forge]

the 8b model is nothing special you would probably need a more mid range graphic card, unlikely over 24 vram

what is holding it back is that it is locked behind a paywall

[spare nacelle]

i just was not sure if you are able to load model in to a trainer through api

never really messed with api

[mortal sphinx]

What's the difference between SD3 and A1111?

And should I use SD3 instead?

[desert dagger]

a1111 = automatic1111 and it's an interface, like ComfyUI, to run Stable Diffusion (all models) with. SD 3 is a stable diffusion model

[somber nacelle]

Just posting to thanks Stability AI for commitment on listening community feedbacks and more globally the commitment to the open source community, as ClosedAI betrayed our expectations, i hope we may trust you guys forever

[mortal sphinx]

Ah so SD3 isn't an interface, it's just an updated ruleset etc running under the hood?

[desert dagger]

it's not a rule set. it's a base model

[mortal sphinx]

Base model, right. Thank you. I read it wrong.

[desert dagger]

no problem 🙂

[dim forge]

Message for stability btw is no one ever asked for a lobomised 2B model, it's entirely a worthless effort compared to just releasing 8B

[kind thorn]

If it can't outperform SDXL, there's no point to it, especially if its ability to render realistic bodies has been intentionally sabotaged.

It's not "safety", it's censorship 😐

[dim forge]

oh my science the safety, what if someone makes taylor swift again

[kind thorn]

We'll need to burn the internet itself to the ground to eliminate the possibility

[devout horizon]

getting your wee wee hard is a safety hazard, you'll poke someone's eye out

[river gorge]

Joooohaaaaaaan!

[desert dagger]

reality check. i'ts not lobotomized. that's fake news that was made up by reddit and 100% untrue

[river gorge]

Johan

Shadowcrip

Are you here?

[kind thorn]

So the core model itself is just as broken? Lovely.

[river gorge]

J-man

Something must be wrong. 😦

[desert dagger]

the devs were very clear about saying that it was rushed out and is unfinished. but the community would not shut up, continually posting things like "stabilty.ai is never going to open source SD3" and were all about refusing to be patient and allowing them to finish it. it's y'alls fault. now be paitent and let them work on it

[dim forge]

the entire concept of the 2B model is a lobotomization, the real model is 8B paramters 2B is just them screwing it over because they want to lock the real model behind pay per gen

[desert dagger]

you are 100% incorrect and spreading misinformation.

[static cape]

Why all the "rofl" reactions? Isn't this update what we wanted and were hoping for?

[desert dagger]

and you're doing it on purpose

[kind thorn]

The way things have been going, there was legitimate concern that Stability was going to fold or be sold before the public release and the model itself would be withheld as an asset for the sale/bankrupcy.

[dim forge]

what would you describe as only handing out the model with 1/4 the training data as while locking the full dataset ?

[desert dagger]

i'll refer you back to all the posts on reddit and here that were nasty, ugly, and deliberately spreading lies claiming that stability.AI would not be releasing SD 3 in open source mode. that got to the point where they got what they wanted, a model that wasn't finished, but that was open source. and since i recognize your name, i know you read those. you probalby posted some of them too.

[dim forge]

Right now today they have a better model that they could release but they want you to pay to use it

[desert dagger]

right now, today, that model is not ready to be released. get that through your head.

[kind thorn]

Their twitter post for that got such negative feedback that they deleted the original post, lol

[dim forge]

How is it not ready to be released? It hasn't been lobotomized to be safe 🤣 ?

[desert dagger]

i doubt you ahve the machine necessary to run it, either.

[dim forge]

8b models do not require some super computer to run, a 400$ graphic card can use 24gb vram

[desert dagger]

i'm starting to think you must work for a competitor

[dim forge]

everyone pointing out my points are stupid is a shill, incredible stuff

[kind thorn]

A $400 graphics card with 24gb VRAM?

🤨

Do you know what Vram is?

[desert dagger]

8b requires a lot more than 24 right now. and you don't have the data center to do what's necessary

[dim forge]

2nd hand 3090s are in that range

[desert dagger]

you're not running 8b on a 3090

[dim forge]

I bet you could, and at bare minimum you could with 2

[desert dagger]

you can ask @primal wing if 8b will run on a 3090 if you like

[dim forge]

SDXL sits on ~8gb vram as viable, 24 is a massive step up

either way that doesn't really address anything I said does it?

[kind thorn]

I would NEVER trust a 3090 being sold for that low. Most likely been used to death for crypto mining and is on its last legs

[dim forge]

they are holding back the good model as a pay per use and handing out garbage

[desert dagger]

everything you've said has been wild speculation and imagination. you have no earthly idea what you're talking about, and you just proved that. you're no longer with the time to listen to

[finite cloak]

it does

[desert dagger]

how well does it run?

[kind thorn]

I can afford to wait for a better SD3 which may or may not arrive. They're still making wonderful SDXL variants and Loras

[dim forge]

you know those are the community making those?

[cursive dust]

Bro

[kind thorn]

Of course

[cursive dust]

Hi

What is this server about?

[finite cloak]

last i tested (months ago) it was pretty slow (talking like half a minute to a minute per gen), but if you quant it / optimize it / turbo it / anything it'll be a lot quicker

[cursive dust]

Can anyone tell me what topic is the server based on

[desert dagger]

stable diffusion

[cursive dust]

What is that

[desert dagger]

an ai image generator created by stability.ai

[alpine bear]

@desert dagger your first mistake was arguing online with a furry. Not that you asked. I figured I'd just stop by and point out the obvious.

[cedar salmon]

i also like to blame community for SAI's choices

[dim forge]

I wouldn't put them above some of those people disagreeing with me just astroturfing, you either have to be clueless or arging in very bad faith to try claim 2B models are somehow a good thing compared to giving us the ready and genning models

[primal wing]

It should run fine on 3090 if you offload the TEs

[static cape]

Let me just say that I am very happy about the announcement. Let's hope promises can be kept this time and that SAI can regain the community's trust. 👏

[river sparrow]

im fine with medium, but it could use a bit more nsfw

[trail pecan]

Yeah, I knew that Stability AI would still live.

In principle, this is my main favorite company dealing with open models

Let's Go 😄

[cedar salmon]

ya even with the you know stuff, assorted stuff

[red otter]

Someone knows an app for made own prompt library for save prompts and advanced details ?

[trail pecan]

Damn, the community is quite cruel towards the first failure of Stability AI:(

[hazy wharf]

Anybody can help me out with prompts? I just trained my own lora and i am basically fresh in this space like a newborn

[trail pecan]

It doesn't always go smoothly. This is an irrefutable fact. There will always be losses

[dim forge]

more annoyed they work so hard to own goal themselves for reasons the community doesn't care about like safety

[hazy wharf]

Ive already created a picture i really like, but there is a second person because i wrote after "1boy, solo, male child" and seems the male child prompt generates the second person and if i remove it i basically get a total different style of picture i try to achive 😭😭

Already tried negative prompt "second person" but doesnt work

Anybody has a idea?

[river sparrow]

just enter "boy" + rest of the prompt, dont over complicate.

[hazy wharf]

But it gives me a completely different art style then somehow, like it ignores completely some prompts if i dont put that "male child" prompt

[river sparrow]

are you going for a child with mustache or something?

[hazy wharf]

No lmao

I tried to make a character with a big head and a small body cartoonish and "big head, small body" kinda messed things up so i used small child which kinda worked but seems i also achieve the same results with a image canny, but i always get now a second person in the photo ☠️😭

Trying since two days basically lmao

[river sparrow]

maybe try something else

[full lark]

more than 24? holy shit

[steady arrow]

Why does 8b need >24gb if fp16 should be ~ 16gb

[finite cloak]

it doesn't

non-turbo it should work fine in fp8 within ~12 GiB of VRAM or less (as long as you have enough RAM to hold the offloaded tencs and all) (though will take forever to run lol)

[steady arrow]

Why doesn't sd work like an LLM

I dont need to offload anything for l3 8b in fp16 with 24gb

[desert dagger]

???

[desert dagger]

@finite cloak

[full lark]

I can also offload the whole llama3 8b in my 12gb gpu, but SD3 uses more resources, even SDXL can use almost 10gb of vram and its around 2b

[steady arrow]

so.. is it not 8B?

I'm confused where the excess resource needs come from

[full lark]

I mean that image generation models use more ressources than LLMs

[finite cloak]

it's 8B MMDiT backbone, plus 0.3B CLIP L plus 1ish B CLIP G plus 10ish B T5-XXL Enc

the offloading is the three textencs

oh plus the VAE which is how many hundred M size

tencs run once, backbone runs repeatedly (diffusion steps), then vae decode runs once

note also to account for the activation tensor size - in LLM terms, all SD backbone calls effectively have a perpetual humongous context len

(because there is an entire image passing through, for SD3 usually 128x128x16 latent between steps and wider channel intermediates within the steps)

[dim forge]

has anyone implemented accelerate for multiple gpus yet properly into either comfy or a1111/forge? It's a big deal for llama but so far hasn't been needed for average users in images either gen or lora training

[finite cloak]

Looking at the license, it's arguably worse. No matter the size of your business, you should not enter into an arbitrarily revocable license. The attribution section isn't draconian, it's silly and unenforceable. Then you have the completely unaddressed core issue of disallowing produced works being used to train other models. Stability AI literally could not exist if it was subjected to similar terms for the products it uses. And the $1m revenue from any source is also completely unworkable.

It feels as if SAI is trying to go with a lowest-bidder approach to legal counsel. Certainly whoever they are using either do not have their best interests in mind or aren't being listened to.

[dim forge]

my opinion on that side is it's an honor system unless they catch like microsoft or meta using their images to train competition, not relevent to even making something like pony diffusion

[finite cloak]

Swarm does multi-GPU via queue splitting (ie send multiple image gens automatically across different backends)

[steady wagon]

does somoeone know the best way to study everything about AI?

[finite cloak]

i've seen a few experimental two-GPU hacks (eg positive prompt on GPU 1 and negative on GPU 2) but none particularly great

for VRAM saving I haven't seen anyone do it

theoretically doable without too much difficult the way it works in llama, but on the mmdit

[dim forge]

the main actual use case I would see immediately would be lora training, more batches at once

or when a very large SD3 model does drop and require over 24gb vram for optimal settings

chaining gpus together is massively cheaper than going into those enterprise card numbers

[outer citrus]

it works for training by doing 2 simultaneous runs then comparing at the end basically.

A PCI-Express 5.0 x16 device has up to 64 GB/s of available bandwidth.
A 3090 has 936.2GB/s of bandwidth

so to utilize 2 cards at the same time for one image you would be decreasing the speed by more than 10x based on physics

[finite cloak]

That is overly simplistic and inaccurate. Based on physics.

[outer citrus]

physical limitations of your bus...

[finite cloak]

And what are you imagining? The two cards don't need to work on the same portion of the data. They're not going to be streaming massive chunks back and forth. I'm not a genius with this stuff, but even simply tiling the image first and splitting the tiles between them before a final pass...

[floral fable]

Thanks, Stability AI staff, for the license clarification and broader scope of the community license. The models and research are all very neat resources.

[empty terrace]

I don't think anyone should be thanking them just yet....
This is better, but it's still a mess.

They should just pick a standard license model thats well understood and has a track record.

declaring the conditions of the research and non commercial use license as revocable basically makes the whole thing worse.

Making the commercial use revocable with no defined payment structure is actively hostile and makes it look like a trap.

Mandatory registration for all commercial use is just kind of hilarious.

The license does not explain how transition from commercial use to enterprise happens or the conditions of the enterprise license.

but the license terminates itself as soon as you cross 1 mil in revenue.

[steady wagon]

can someone help me in tech support

[empty terrace]

The use restrictions mean no one can use sd3 for any synnthetic use.

The ownership of outputs its unintelligible nonsense.

The deletion clause is just freaking weird.

[shell cradle]

They have not open sourced it.

[shell cradle]

Personally I don't care about an updated 2b, I'd like to play with 8b.

[dim forge]

yeah that is the point isn't it, we know they have a better model they just want to pay per gen

what is even their current excuse for not releasing it? it still isn't safe or something

[empty terrace]

NOPE.
This thing is a mess. The license mean the model is completely unusable.

[shell cradle]

Not the first. They botched 2.1 & the cascade launch. But yes, I think the reaction could have been more sensible. No one owes anyone anything as far as I'm concerned. That doesn't mean people should be happy either, when the expectation was built that SAI would be open sourcing models.

[empty terrace]

They are definitley in a rough spot, I feel like they way overinvested, so now they have all of these costs, and they never really had a plan that was going to generate the kind of revenue that could cover those expenses.

so now the new leadership crew is trying to figure out HTF do we monetize this at a scale that pays for all of this?

[tepid kayak]

Is there some easy to access and use out of the box (ideally with a free model) of the SD 3.0?

[empty terrace]

Define easy?

[tepid kayak]

Like without setting it up, something akin to Mage or Leonardo, with a GUI via browser and ran from cloud. I am non-coder, not-very technical user, so going for simple to use interface

[empty terrace]

I don't think there is an easy way to do that.

The easiest two ways I have seen to start tinkering coming from an artist perspective are the Krita plugins, or invokeAI.

Both run locally, and do require some setup, BUT

unlike webbased prompt systems they are focused on providing a useful creative tool.

InvokeAI has an ugly but easy to use and reliable installer.

it's probably the easiest and they have a TON of tutorials.

[tepid kayak]

hmm, alas I only run AMD 6600XT which I don't think is a competent enough gpu to run it locally

[empty terrace]

yeah diffusion models without Nvidia hardware are in kind of a weird place right now.

[tepid kayak]

so maybe someone will come up with some on cloud solution I might use instead, but thanks

[empty terrace]

Looks like invoke does have a cloud platform with a try for free thing https://www.invoke.com/

but I don't how much you get for free.

I always just run it locally, so far Invoke is the tool I've had most success teaching non-technical users on.

[tepid kayak]

ty, what is your specs? and what about that Krita?

[empty terrace]

I'm running a 4090, but it's monstrous overkill for most image generation stuff, I mostly use it for large language model stuff.

Krita is a photoshop alternative.

I'll see if I can find the plugin..

[tepid kayak]

Oh yeah that truly is a monster haha.

[tepid kayak]

it is pretty expensive and no free access at all

[empty terrace]

Ok, looks like the krita plugin has AMD support

https://www.youtube.com/watch?v=Ly6USRwTHe0

I like it because I can use it with my wacom tablet

https://github.com/Acly/krita-ai-diffusion?tab=readme-ov-file

https://www.interstice.cloud/plugin

[tepid kayak]

but does it support SD 3.0? Kinda only see the older models like the ancient 1.5 or XL

[empty terrace]

hmm looks like probably not...

but honestly sd3 is .... not great right now.

[tepid kayak]

hmm okay

it it possible to use LORAs via that plugin?

[empty terrace]

yes

no training ui

so building your own loras will need other tools

[tepid kayak]

I think I am fine using free to get loras, I am fine with that, not this advance to train my own, but do you know how to use them with kryta?

[empty terrace]

In the configuration panel for the diffusion addon, you goto the lora heading and click the 'add' button

[tepid kayak]

not sure where I exactly, I am new this program, can you specify the exact path in kryta?

[empty terrace]

So you've installed krita, and you've installed the image diffusion plugin?

[tepid kayak]

yes

[empty terrace]

so next to the model selector, there is a little settings button.

two gears..

click that.

[tepid kayak]

yes I am there now, but the LORA is greyed out also the ADD button. Got one LORA downloaded which I really want use.

[empty terrace]

hmmm not sure on that.

[desert dagger]

was the lora created for the model you want to use it for?

[tepid kayak]

yeah I think so, since the plugin works with SD 1.5 and XL - here is the LORA https://civitai.com/models/74776/moebius-jean-giraud-style

made for SDXL

[desert dagger]

odd that it won't let you access it then - unless maybe it's not finished downloading?

[tepid kayak]

I wish I can upload a screenshot

[desert dagger]

wouldn't hurt

[empty terrace]

I'm grabbing the lora, gimme a second

[tepid kayak]

Literally when I go to styles - the LoRA option is totally greyed out

I cannot add any at all, that's the problem

you ve got it in Tech support :)(

[empty terrace]

path should be something like C:\Users\MarsV\AppData\Roaming\krita\ai_diffusion\server\ComfyUI\models\loras

thats where the lora needs to be.

[desert dagger]

i have no idea and there's nothing conclusive on google for that. other than possibly this https://www.reddit.com/r/krita/comments/r5nq1y/krita_wont_allow_me_to_change_its_settings/

[tepid kayak]

that does not seem to be related to my problem

Maybe it is because I use the cloud gpu for generations so it won't let me run Lora locally.

[empty terrace]

Ahhh yeeeah.

thats probably it.

you'll need to get it working locally.

if you want to do much with loras.

You should be able to run that plugin on your 6600

[tepid kayak]

yeah, so here comes my 8GB 6600XT haha, I kinda wish the model can utilize also my main system ram as I have 64GB DDR4 3200

[tepid kayak]

but probs waiting for a single generation like minutes, any idea what resolution I should aim for ideally?

[empty terrace]

it probably won't be minutes.

I'd suggest starting with low resolution,

and then upscale when you have something you like.

[tepid kayak]

I am aiming for illustrations so that might not be that resolution crazy, got any tip for the lowest resolution I should choose in this scenario for the best performance but also good enough fidelity?

[empty terrace]

also try starting with the turboXL model and image 2 image.

so it depends on your dimensions.

HAve you done illustrations before without AI?

[tepid kayak]

not really

[empty terrace]

Ok, so thats goign to be the biggest obstacle.

[tepid kayak]

but I have already done quite a lot of generation, for instance a tarot deck, but all was done via cloud

[empty terrace]

so illustration is different from something like a tarot deck, and different from just generating an image with a prompt.

Because there are very specific things you need the image to convey

[tepid kayak]

sure I am aware, but I am not planning to do the illustration by hand, I want it to be solely done by AI

[empty terrace]

There's a seperate art to that aside from just the technical skill of creating the image

[tepid kayak]

I am simply just coordinating it

[empty terrace]

oooookaaaaaaaaaayy...

[tepid kayak]

Using inpaiting and outpaiting to correct stuff

and also learn how to use controlnet for consistency

[empty terrace]

thats not something that any models can really handle well.

at least not yet.

you can do it if you are willing to get really REALLY technical

but we are talking about multiple posenets, compositing , lora passes with different loras for different characters etc.

[tepid kayak]

I want to experiment with consistency and unified art style by simply borrowing art style I love

[empty terrace]

So, you can absolutely do that, and it's a blast.

[tepid kayak]

At the beginning of the AI hype last year, I saw someone early one generating an AI comic novel and it looked aces. That must have been one of the first MJs or SDs

[pulsar cipher]

hi

[empty terrace]

If you want to illustrate something with a clear idea, and communicate it to others- then you really only have two routes:

1. Get way technical, get some serious hardware, learn JSON, learn ComfyUI. Learn how to build frameworks that let you build complex scenes reliably. This means picking up some technical knowledge.

2. Lean on sound artistic fundamentals- start with Image2Image, sketch-generate, sketch some more and use inpainting and outpainting to push it the way you want. you draw the basics, and the generative model draws 'the rest of the owl'. This means you need to pick some artistic knowledge.

Either way you need an understanding of space, line, color, texture, etc. to get good results.

I haven't seen any AI generated comics or novels that I would describe as 'aces'. I've seen some individual pieces that were great, but as far as big composed projects- most of it has been decidedly mid so far.

@tepid kayak Apologies if that's not very helpful I kind of assumed from your initial questions that you were in that second category.

[tepid kayak]

No it is fine, I really appreciate your honesty here

It is perhaps time for me to get out of my comfy zone and learn a thing or two about illustration, I think the route 2 is much more appealing to me.

[empty terrace]

There's massive potential for GAI to accelerate an artistic endeavor- but so far I don't think I've seen it successfully replace the process meaningfully.

At least you've got krita installed- thats as good a place to start as any.

[tepid kayak]

When the download is finished I would love to se some inpainting magic and the Lora combo. I am mostly interested into mastering controlnet, that stuff seems like the gateway into some consitency. I don't assume via using this plugin and this lora, I can also fine tune the downloaded model with my creations, so all I make is becoming more and more inline

[desert dagger]

you know - there's always this https://comfyuiweb.com/ or this https://www.runcomfy.com/comfyui-web

[empty terrace]

Yeah its so weird that one of the most technical workflow tool available has the word 'Comfy' in its name.

[tepid kayak]

oh wow, yeah I don't think I'll be going into learning json at this point :))

[empty terrace]

It's an awesome tool but if my workflow looks like this: (image in tech-support)

Comfy is probably the wrong word.

[desert dagger]

json's not hard

[empty terrace]

Jsons not hard for me or you.

[desert dagger]

everything's easy once you know what you're doing.

[empty terrace]

For someone who has never touched a programming language it might as well be calculus

[tepid kayak]

I ve never been good with code, it is overwhelming, I am more of an intuitive person.

[desert dagger]

yes but this https://www.runcomfy.com/comfyui-web is just an online comfyUI interface. use it just like you would if you ran it at home

no programming required

[empty terrace]

I've never been able to do anythign I wanted to do in ComfyUI without needing json....

admittedly I am usually asking much of it.

[tepid kayak]

I really don't know how or why I should use it instead of Krita, What is the edge, I don't see any interface (only with premium versions)

[desert dagger]

cause... krita's a paint program? and apparently not happy with the lora you want to use

[tepid kayak]

Problem is I have not find a way to install the lora, not that I am not happy, no we figured what was the problem, but still I am failing to figure that out. Some dude is showing a tutorial and he has there actually one extra icon - the folder icon which I don't, it is confusing af

[desert dagger]

my install is local, but my path is comfyUI/Models/Loras

maybe you need to make the Loras folder?

[wet grotto]

where I find controlnet normal to donwload?

I take a version with 5gb =0

I'm on kaggle, must be 16fp, is it?

[tepid kayak]

btw @empty terrace how does one trigger image-2-image in Krita?

[prisma horizon]

load an image

[empty terrace]

the strength percentage slider

[prisma horizon]

double click on it and start the gen with a different denoise value than 100%

remember to set the correct resolution of the canvas

after loading the image

[tepid kayak]

so loaded an image - prompt is empty - but instead of generate I am having refine button only

[prisma horizon]

why the prompt is empty?

refine is essentially img2img

[tepid kayak]

I don't know, I thought I am generating just a version of the original image, don't know what to write in the prompt

[prisma horizon]

you still need to prompt the content of the image

or the model will not know what to do with the image

[tepid kayak]

I fine if it just tries to replicate it, but don't know what prompt to use

[prisma horizon]

it can't

i'm not sure what an empty prompt will do

[tepid kayak]

I do now, generate a whole new image completely unrelated :))))

[prisma horizon]

i suppose it generate an image based on the model's bias

[fervent thunder]

f'ing civtai, should of never given them money

I never will again, fo sho

[desert dagger]

it acts as noise.

[prisma horizon]

what happened?

[fervent thunder]

They are cracking down

[prisma horizon]

well, img2img always use the first image as noise

[fervent thunder]

Banning all types of language for prompts

[prisma horizon]

talking about the censorship on sd3?

[fervent thunder]

This is some new things they just started today

[prisma horizon]

what? what happened?

[fervent thunder]

They are young and dumb, mostly

list of banned words pt1 -https://github.com/civitai/civitai/blob/feb2337c202ab82661958481de9652a4a6b3417b/src/utils/metadata/lists/words-young.json#L4 and pt2 -https://github.com/civitai/civitai/blob/feb2337c202ab82661958481de9652a4a6b3417b/src/utils/metadata/lists/words-nsfw.json

[prisma horizon]

what i'm missing?

[fervent thunder]

If it was completely free it is not a problem. When you start taking money, the rules change a little bit

[prisma horizon]

i can say that the first list is understandable, even if it's a big pain when genning charming stuff

the second list is there to literally sue everyone genning porn with it, i suppose

i'm right?

[fervent thunder]

I have no idea what they are thinking. A case of the paranoids, or bad legal advice I would say.

[prisma horizon]

where this list comes from?

[fervent thunder]

You cannot defer liability in most countries. It just does not work that way.

[prisma horizon]

to me this looks just like a tool to help mantein stuff safe

[fervent thunder]

The list is not complete, I think I got banned over using the word petite

[prisma horizon]

but if nowhere is written an imposition in using such lists, i don't see the problem

[fervent thunder]

And you don't know what word is wrong until you test it

I never like them anyway. I shoulda never given them money! 🤣

[prisma horizon]

petit is not on the list but i can see why prompting it can be a problem

but every nsfw word is present on the list

[trim magnet]

just gen morbidly obese ppl and u safe (they have experience with that)

[empty terrace]

doesnt impact my use case but my thoughts are that the first list seems like a good idea minus the word 'disney'

The second list is an odd choice since that whole site has basically been smut driven from the word go.

[prisma horizon]

understandable why, but i would feel pretty sad if i couldn't generate anymore charming chibi characters

they remind me so much of my lost childhood

i miss so much my grandparents >.<

good times ahhhh...

[trim magnet]

tbh doesnt feels right to gen sexy pics on an online gen,too risky

[prisma horizon]

i agree

over18 stuff require a certain level of responsability

[desert dagger]

blame it on the banks and payment processors

[trim magnet]

yea and if generator creates something cursed or bad u are done

[prisma horizon]

yup

[empty terrace]

Whats the scope of the ban?

Just generative prompts?
I imagine if they banned that second list from their whole site 80% of their models would just vanish.

[prisma horizon]

for me is just a filter to stop some gen

[empty terrace]

Then I'd say thats a good call on their part.

[prisma horizon]

agree

[empty terrace]

If peopel want to generate smut, they should do it locally on their own hardware.

[prisma horizon]

there is so much porn onlìne... what's the point of genning for more?

lol

[fervent thunder]

I am damn sure vanishing from their site

[empty terrace]

Sooo can you clarify your complaint?

I'm trying to understand what you are upset about.

[fervent thunder]

not really interested now, I am just over it

[empty terrace]

They are basically just saying no smut or kids in prompts on our cloud- that seems perfectly reasonable.

[fervent thunder]

yes it is, they can have at it, I am out

[cedar salmon]

lure them in with the porn than cut them off

[empty terrace]

All the porn models and stuff are still there...

they are just telling peopel to use their own hardware instead of their hosted platform for that.

[static cape]

I won't be able to Finetune 8B. I'd rather have a better 2B Base or 4B. ^^

[prisma horizon]

well, speed is a good point too

what do you need a big model for if it's a fat a**?

i dropped a lot of good workflows because they were too slow

[random brook]

Man, CivitAI is really struggling rn. Anyone know whats going on?

[fervent thunder]

They need to hire some experienced help

Been saying that for a while now.

🙈

[desert dagger]

they had a wake up call

[random brook]

Ya, but the last few days have been terrible

like, well above the normal amount of annoying

[fervent thunder]

Hopefully they will correct their course sooner rather than later

[desert dagger]

they've been alienating their business partners and users for months.

[random brook]

Tbh, I pay the monthly support fee and all I care about is the site working as way to host my stuff. Like, I don't care about the buzz, the bots, ect... but if the site stops working, I guess I will need to find greener pastures

[desert dagger]

easy to find alternatives - just go to google and search on "civitai alternatives"

[fervent thunder]

Same here.

tensor.art is not bad

[random brook]

They have an import from Civitai button, im rolling rn

[trim magnet]

strugglin with what?

[fervent thunder]

civtai can srls F off at this point, that site was so mismanaged from day 1

[random brook]

I have uploaded my new LoRA and corrisponding post, only to see it still not load after 20mins, shooting me a bunch or red jpegs and errors

[trim magnet]

oh yea probably overloaded servers

[random brook]

Ya, which would be fine, but its been like this for two days

which is a lot less fine

some would say

problematic

[pale latch]

Paying for a service that has bad uptime is problematic indeed

[lusty beacon]

https://stability.ai/news/license-update

[desert dagger]

#📣｜announcements

[tropic zodiac]

are there any inpaint controlnet models for SDXL that work with A1111? I tried many and none worked so far

[fervent thunder]

probably want comfy for combining control net with inpainting

[graceful bramble]

So I am using half a dozen loras to create a specific body type. Leg length, face details, abs, etc etc. Is there a benefit to merging them into a single lora or even into the model itself other than convenience or am I better off leaving them seperate?

[fervent thunder]

IDK if more than 3 loras works that well

at this point its better to train full checkpoint

[graceful bramble]

And bake the loras into it?

[fervent thunder]

yeah

[graceful bramble]

Could you point me to a good article/tutorial/youtube video or whatever to learn how to do that?

[fading plover]

how long does stable audio generation take

is the team on holiday

[hardy sedge]

Is Prodigy really better then Adam for Charakter lora Training?

Can you merge 10 loras into one?

I have 10 loras of a character?can I make a perfect lora out of it?

What is the best app to merge lora?

Thanks

[placid hatch]

Is there any argument or setting for webui to completely stop it from preloading anything?

[narrow kernel]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings

[fervent thunder]

SEcourses

they have dreambooth tutorial

[fervent thunder]

its the same as I was saying to the other guy
train a full checkpoint if you have 10+ lora

[low moon]

😮

[magic burrow]

hi

[low moon]

hi

[hardy sedge]

How does it work?

[fervent thunder]

if you search "dreambooth tutorial" you will find some, SEcourses has one

[red otter]

Whta ui library is used for comfyui ?

[low moon]

[prisma horizon]

👍

[amber quail]

i get different results of pictures with civitai and my automatic1111, one thing is: Eta noise seed delta. Is there a Standard for Civitai?
What could it be else?

[fervent thunder]

could always just be randomness

if literally everything else is the same

[sage loom]

Use the DARE/TIES method with comfyui. Check the about section on this model page for more info on a workflow. It's not perfect yet but I have seen a huge improvement from the method used in versions 1 + 2 vs the DARE/TIES used in v3.

https://civitai.com/models/447902/mangled-merge-xl?modelVersionId=619849

[arctic sedge]

@bleak matrix I know the license for SD3 M has just been refreshed, but will the new SD3.1 M have a different license to the new current one?
I'm seeing lot's of people are still having critical issues with the new current license, unfortunately.

[bleak matrix]

Hey there! Hope you're doing well. Unfortunately, I’ll need to check up on that. If I find an answer for you, I’ll let you know.

[crystal bone]

Can someone please provide me a summary of the license for "sdxl turbo" and "sdxl refiner 1.0". Is it allow under the licenses of this models to post the output on social media or upland scripts on GitHub that use this models

[arctic sedge]

Oh hey! Thx! Great stuff. I really want SD3 to be a genuine sucess. As do a lot of others. There is just a few major restrictions that i think are holding back some small business from using the model. Glad to see SAI is trying tho. 👍

[verbal osprey]

Interesting....just read the update. So it sounds like SD3 medium 3.1 is coming. I never downloaded 3.0 after seeing how bad it looked.

Very happy an improved model will come out in a few weeks. I hope it's better!

[bright nest]

license question: do image outputs of core models fall under either the commercial, non-commercial or enterprise licenses? (other than creating illegal content or using them in the creation of foundation models)
example: can i sell $1,000,000+ in core model created images per year and still not need the Enterprise License?
example 2: can i use the core model created images to sell a t-shirt with said images and make $1,000,000+ per year and still not need the Enterprise License?
example 3: can i use the core model created images to promote my business that makes $1,000,000+ per year and still not need the Enterprise License?? ie: a sign created with said images.

[cedar salmon]

i thought about this, it doesnt loook good, the wording # 3

[bright nest]

it really only states SD materials(the tech) and Derivates(loras/finetunes).

[cedar salmon]

it says reguardless of

[bright nest]

with images in a second catagory

under IV general terms -> c. -> (iii)

this is why im asking 😄

[cedar salmon]

well answers dont come from here from what ive seen, just gossip, thats lawyers work

[bright nest]

so true 😦 but i still wanted it out there. maybe lykon will see it and refer it back to the legal team 🙂

[desert dagger]

@primal wing adressed all of this yesterday

[bright nest]

thanks will scroll up

[cedar salmon]

dont bother i looked

[desert dagger]

eaiser to use the search feature and choose from: lykon - then you just see his posts

[primal wing]

hey what's up?

[desert dagger]

scroll up and read all the licensing questions they've been asking half the morning

[primal wing]

can you give me a summary?

[desert dagger]

this one is probably the most important #💬｜general-chat message

[primal wing]

If you have a business that sells shoes and makes 1M dollars and then your intern uses SD to generate an advertisement of a cat wearing shoes, you'd technically need the enterprise license (as per the legal wording of the license itself). Of course, this is why there is no pricing attached to the enterprise license, as this would vary case per case (eg: you won't be asked to pay 500k for your cat image, lol). Plus, keep in mind this is based on self-reporting and counts on people being decent human beings and wanting to contribute to the Open Source environment instead of using SD models to make millions and then wish our death if they don't like a free model release (which actually happened).

that's my understanding however. I'll make sure this is properly addressed in the FAQ

[sage reef]

so with this new license, is civitai gonna unarchive the sd3 related content now?

but i guess everyone is waiting for the improved model anyway

[desert dagger]

you'll have to ask civit, right?

[primal wing]

if they have any doubt we'll address them

[trail lion]

Yup, but seems reasonable they would now

[primal wing]

also, this might not be clear to anyone, if you are using the model for a research purpose at a large organization that is making over USD $1M in revenue, then you are under the research definition and the USD $1M trigger doesn't apply.

[sage reef]

ah yes.. for research purposes

but i can't wait to see the improved model as well as the other versions like 8B, etc

i also wonder if they will release a new stable video diffusion model, it would be nice

[desert dagger]

i don't know if they are resonable or not. i just know that they don't have the best reputation in a lot of people's opinions

[primal wing]

there is also a research definition in the license itself for this very point

[sage reef]

civitai or not, we still have huggingface and that torrent tracker :3

[bright nest]

thank you! i was confused because it said sd-materials or derivitives of 🙂

[primal wing]

that torrent tracker

[sage reef]

it's down for maintenance 😦

[primal wing]

My example had 1 cat image as the absolutely low bar. If you sell shoes and also sd models, then the story is a bit different, of course. Enterprise licenses are worked case by case

(not by me fyi)

[bright nest]

yea i was only wondering about the output images

how much of the images i owned under the licenses

i think its fair, might not like it but its fair.

[primal wing]

We are addressing this right now. That point was mainly aimed at companies who can make foundational models "cloning" ours (in lawyer understanding).
We're either gonna amend the point (I personally do not know if it's possible, since I don't know the limitations of license wording) or make it clear in the FAQ. Gonna take a bit, but rest assured that SAI will not police you for using SD3 outputs to train your lora or finetune. Also there is no way to enforce it except intentional self-reporting.

[bright nest]

if i was making this amount of money i would support SAI lol \o/

[sage reef]

another investor 😮

[primal wing]

there are some internal proposals and discussions on how SAI can help people reach that 1m threshold more easily

the plan is to have more and more people register for the free commercial license and help them reach 1M revenue so they can start paying 😄

the end goal is to improve and lead open source, instead of just being "it's free, do whatever" or "it's research only" like most are doing

[bright nest]

thank you very much for giving this info (not just the reply part) lol

[loud solar]

Oh yes! Help me! 😄

And thanks for the information.

NFTs only attrackted scammers. The only picture I've sold that way has been a print 😄

[jolly copper]

Jesus Christ is the Son of God , who died for your sins on a cross , then God rose from the dead three days later. So that you may receive the free gift of eternal life, when you ask Him to forgive you. If you believe in your heart and confess with your mouth that Jesus Christ is Lord. You will receive this free gift. He also promises to heal your body. This Is Gods Love for You!

[trim magnet]

Raptor Jesus, in your talons we find refuge,
From the abyss of our doubts and the shadows of despair.
Grant us the courage to tread paths uncharted,
With the wisdom of the ancients and the strength of the righteous.

[loud solar]

Prompt material?

[fervent thunder]

I like the devil

he's coo

🤷

[fervent thunder]

sd3 is shit

and they close sourced the good shit

[desert dagger]

please don't start that again

[fervent thunder]

its trash

[desert dagger]

you're welcome to your opinion. please don't come in here and try to start an argument

[loud solar]

What we have is way lass than they announced ... or none of my 10 workflows is working right ...

[fervent thunder]

how did they make the model worse

its worse than 1.5

[loud solar]

I think nobody never knew the model of the pictures they showed or the workflow ...

[fervent thunder]

🙈

[loud solar]

It's a real bad communication ... they have released SD3 and now in the new Licence agreement they tell us they are still working on it and there will be a better release. I never got the information it was a prerelease ... maybe my fault ...

[fervent thunder]

apparently it really was meant to have a beta label

like

internally it had a beta label for ages

and then it got shipped without it

or any wording to that effect

[prisma horizon]

to me it looks as bad as it was 1.5 at first

[desert dagger]

dude, don't come on this discord and try to start an argument. we're going to start reporting you

[prisma horizon]

and what 1.5 is now, is under the eyes of everyone

[junior urchin]

I don't think it's necessarily worse

It's more like, what has been shared & hyped up, was the 8B model, not the measly 2B one

[loud solar]

If you need a community for finetuning ... don't waste their time with betas ...

[desert dagger]

he's just on here to try to get people argueing. he's not going to listen to what you say, he's jsut going to use it as arguement fuel

[junior urchin]

Hmm

[prisma horizon]

looking how it worked the small textual model, honestly, i expected something like it

[junior urchin]

The only arguement that could give him validity, is Comfy's statements

[desert dagger]

he's just here to try to cause problems, and he's probably being paid to do so

[junior urchin]

But by whom 👀

[desert dagger]

competitor would be my guess

[prisma horizon]

pffff being paid to cause problems, i would have done it for free XD

[fervent thunder]

It’s more like SD-0.5

They really ruined it on purpose

[desert dagger]

see? he might even be a bot

[fervent thunder]

Because they were pressed by the NSA

[junior urchin]

It might be a stupid question, but who could be a competitor to open models?

[fervent thunder]

No one anymore

The NSA is taking it all

[junior urchin]

yes yes

[loud solar]

Looks like I am the only one who don't really wanna talk bad about SAI ...

[junior urchin]

The sky is green and the water is red

I will talk bad about something, if it's neck deep into the pile

[loud solar]

I think they just didn't give us enough information and maybe that licence agreement wasn't clear enough ...

[junior urchin]

The announcement is iffy

Sure it talks about license

but

It mentioned that 2B was released as a "work in progress" piece

That's the first time that has happened

[fervent thunder]

We need a real decentralized open source . That cannot be taken down or have a main point

[junior urchin]

Have fun in the amazon jungle then

[fervent thunder]

Because stability is done

[desert dagger]

i'ts not the first time. the devs were very open about posting in this discord that it is unfinished and was rushed to release

[fervent thunder]

It’s been closed

[loud solar]

Not a fan of that point, too. But we all have been shouting for a release ....

[prisma horizon]

welcome in evangelion 🙂

[desert dagger]

he's a broken record

[fervent thunder]

I'm new here.

[junior urchin]

Ah, my b, must have missed that one

[prisma horizon]

hello kanoi, welcome

[junior urchin]

I usually only read the official docs

[fervent thunder]

Thank you.

[desert dagger]

if you search this discord for from: alex (mcmonkey) and scroll back throug his posts to around the release date, you'll find quite a bit of discussion

[fervent thunder]

Yes let’s support censorship guys, meanwhile stability is using copywrited training data

Makes sense

[desert dagger]

definately a bot on a script

[prisma horizon]

more gens and less talks lol

let's see what i can try to do

[fervent thunder]

lol 😆

Yes I’m a bot 🤖

But aren’t we all

Just parroting Twitter

[prisma horizon]

i'm a bot too

[fervent thunder]

Someone post a good sd3 gen

[cedar salmon]

you know if you say something that someone doesn't agree with its clearly arguing

[fervent thunder]

lol

[prisma horizon]

i dare not agree to that fact without arguing 🙂

[fervent thunder]

It’s like I was t even going back and forth with anyone

[desert dagger]

can't post images in this channel though

[fervent thunder]

Just getting my thoughts out

But yeah keep tagging me begging for an argument

It’s not happening.

I’ll just ignore you

[prisma horizon]

oh right, this is general

[junior urchin]

The yeti exists, I've seen it

[fervent thunder]

Anyways

sd3 is a total fail. And stability has been taken over by the NSA

[junior urchin]

Its fur was nice and soft

So soft

[fervent thunder]

now I have mentally ill people trying to distract people

[junior urchin]

Yes

Sheep sheep sheep, sheep sheep

[fervent thunder]

you know what i just looked and sd3 is actually good

forget about what i said

[desert dagger]

a sheepish yeti?

[junior urchin]

Yes

It was on a cool holiday

A freezin' breeze

[fervent thunder]

heresa challenge for u nerdss

cat smoking from a bong on a couch floating in space over the earth looking at the moon

tag me in the channel

#🏞｜general-with-images

[junior urchin]

Not sure if sdxl can even do that

So an invalid request

It's a clown world, yes

[desert dagger]

you post yours first

[junior urchin]

He can't do that

Gemini doesn't have that feature yet

[desert dagger]

i dunno. SD3 has very good prompt comprehension. let's find out if it can do that or not

[junior urchin]

Although Gemini would say it's against the policy

[cedar salmon]

cant get the bong yet :/

[junior urchin]

Well, if he requests, that would mean he can't, by default?

ah

[desert dagger]

no, it wouldn't. just means he's tossed out a challenge.

[junior urchin]

you meant in general

[desert dagger]

why?

[junior urchin]

[cedar salmon]

this isnt the place to philosophize

[junior urchin]

I thought you meant him, not asking if the model could

cat smoking from a bong on a couch
floating in space over the earth
looking at the moon
The last part would be hard for it to comprehend

[desert dagger]

here's another challenge, this one's harder: prompt: a horse riding an astronaut on the moon

[junior urchin]

I lowkey think it's easier

[desert dagger]

see if you can get it to do anything but put the astronaut on the back of the horse

[junior urchin]

yeah.. that's brutal, gonna try and update comfy

[desert dagger]

that's cheating though

[river sparrow]

ai gen is cheating

[shell cradle]

How about 8b?

[desert dagger]

for the specific prompt challenge

[river sparrow]

ok, understood

prompt only. Im going for a horse riding a car on the moon

[steady wagon]

Is there an way to make realistic images from an sketch / 3d model without changing the objects inside of the picture?

[river sparrow]

not really

[shell cradle]

Ok, you won't. But I, and others, would, and the potential would be much higher.

[steady wagon]

promeai.pro and xona.ai are almost able to do it

[raw tapir]

Does anyone have any experience using Pinokio as a platform to launch Stable Diffusion (et. al)? I'm tempted to give it a go, but it kinda looks too good to be true.

[desert dagger]

it works. Cocktail Peanut is a very good programmer

[raw tapir]

He must be. 🤯 It looks amazing.

[prisma horizon]

what's so special about that one?

[desert dagger]

it's a launcher for all sorts of applications.

[raw tapir]

Pinokio apparently runs heaps of different models, many if not most free, everything from images to music to voice gen and heaps of other stuff, all one click install. Few people are raving about it on YouTube. It looks awesome.

[prisma horizon]

oh

an all in one package

[raw tapir]

It's more like a platform through which you can install and launch other packages, such as A1111 and ComfyUI, for example.

[prisma horizon]

oh...

interesting

ty

[dawn mulch]

anyone used DMD2 yet? https://huggingface.co/tianweiy/DMD2/tree/main

its 5 GigahBites, running that with a 6GB sdxl model seems like it wont work on my 6gb vram gpu

[west rivet]

anyone know if silicon macs are powerful enough to run sd models?

[desert dagger]

silicon macs?

[fervent thunder]

hey I just started with stable diffusion, wondering if there is a prompt that works that prevents things from being cutoff ie: generating weapons like a sword, sometimes only part of the sword is in the image

[desert dagger]

try adding "centered" to your prompt