💬｜general-chat

[fervent thunder]

actually a good idea.

whats the best sampling method for hyper photo realistic?

[desert dagger]

that totaly depends on the model you're using, actually

[fervent thunder]

... you mean version?

[desert dagger]

i would add the terms "cinematic" and "photographic" to your prompt, and AVOID the terms 'photoreal' 'photorealistic' and 'photorealism' - those terms are painting terms and the AI knows they are. it'll give you a painted look

[fervent thunder]

ahhhhhhhhh

[desert dagger]

yes or if you're using a checkpoint for one of the base models. or you're using loras

[fervent thunder]

yeah Im watching a video right now on this to get a hang of it.

Runs amazingly well on my computer I am kinda surprised.

Can't believe I was about to pay 20 bucks a month for this hahaa

[desert dagger]

also you might add HDR (but NOT HDRI) to your prompt - those are both photographic effects that the AI is very familiar with the look off. Avoid stuff like 8k resolution, though. it's not a camera, you can't prompt it to make something high resolution

as you're just getting started, find out what the AI thinks about by default for your terms by using just one term as the propmt and generating a few times

[fervent thunder]

is there a way to get it to pay attention to my requests more lol. like I asked it

full body cinematic undead horror abomination monster with a massive, cancerous and gangrenous right arm and bulky legs with glowing red eyes

Eyes arent glowing red (I did 8 images), arms are not what I asked for. Pretty creepy stuff tho

[desert dagger]

put the "glowing red eyes" at the start of the prompt so they have the most weight - and remember you're talking to a computer. think like it does

[fervent thunder]

I juust learned about CFG scale too

[fervent thunder]

I got another one:

I keep asking for a sword and it keeps giving me knives/daggers. If I put "knife dagger" in negative prompt, it just gives me handles (no blade) or leather straplike things that christian grey probably had in his room 🙈

Whats a good prompt to get a proper sword?

[raw tapir]

maybe try something like 'longsword' or 'great sword' or "a bloody big sword"

[west rivet]

apple silicon

ive got an m3 mac with 18gb of ram

[fervent thunder]

i tried longsword that worked better

I wonder what prompt I'd need for a long longsword

[desert dagger]

there is no such thing. are you trying to draw a final fantasy sword?

[fervent thunder]

there is such a thing, they are called greatswords

[desert dagger]

a greatsword isn't just a long longsword. a long sword blade is thinner, the haft is different

[fervent thunder]

stop making me look uncultured 😤

[desert dagger]

too much time in the SCA 😉

[fervent thunder]

dunno what sca is

[desert dagger]

an orginization full of people that run around pretending that it's the middle ages

[fervent thunder]

like a HEMA club?

or just RP?

[desert dagger]

https://www.sca.org/

[fervent thunder]

so... both? I guess haha

theres a HEMA branch like 20 minutes from my house

Keep thinking about it. My inner roman wants to come out and play

[desert dagger]

you should go and visit

[fervent thunder]

I should and I might have to

[narrow kernel]

Do you think the auto captioner, that generates the image dataset captions to be used for SD model training, was trained to know the difference? That's one of the biggest things you have to ask yourself when facing issues like this.

[desert dagger]

yes, actually

[narrow kernel]

I'll try running some broadswords and longsword through the half dozen VLMs I have on my machine then to see then

Vlms that are far more capable than the older ones used for things like the lion dataset

[fervent thunder]

it always seems to want to make swords shorter

like sword = dagger. longsword = sword.

[desert dagger]

and a few others as well

[fervent thunder]

apparently SDXL needs lora for weapons really

which is fair

can see why they were not put into the training data en masse

[desert dagger]

yeah, that's probably the best solution. i'm not sure anyone's created one

[fervent thunder]

its better to train your own loras anyway

standards of lora are too chaotic

[desert dagger]

auto1111 has added SD3 support

[nova lily]

Even if the model and Lora are compatible, what if they ask for different clips skips?

[vast ingot]

The new license details feel quite better than before. Thanks S.AI on that course correct !
Personally, I've been using SD on my twitch channel as part of my bot, drawing pics for the viewers to spread knowledge and hype about this tech. Since the channel "makes money" (about 15€ in 6 months yeahhhhhhh), I would have had to upgrade to a paid license that was costing a lot too much to be viable. So quite happy to be able to upgrade my models and stay on the free tier

[fervent thunder]

for the most part, don't use models that skip clip layers

if you have to, to get the effect that you want, then that's fine
but in general if you are skipping layers of your deep learning model you are making it weaker

can be a fun experiment for variety though

[wanton barn]

Hello, my text2img with control net is incredibly slow. Usually it only takes me a few seconds to create a regular image, but with control net open-pose, the generation takes up to 25 minutes which is incredibly long. Is there anything i could do?

[narrow kernel]

Watch your vram and shared memory in the task manager on the GPU section. It sounds like you're out of vram and it's shuffling calculations with system memory, which is exponentially slower

[wanton barn]

does that mean if i have 6 gb vram and in the task manager my shared memory works 6+/12gb , it shuffles calculations with system memory,

[junior urchin]

What about regular ram 👀

My regular ram is my bottleneck

And I got 16 gb + 12 gb vram

[warm junco]

Increase the Windows Pagefile

[storm shard]

You can run all models from 1.4 to SD3 on that with
https://drawthings.ai/

[junior urchin]

What about recommended mb amount?

It's currently 32115

The top amount of committed ram I've seen, on my own pc, is 52gb

[warm junco]

Oh if its on 32 then leave it on system managed

Make sure its only enabled for the C drive

[junior urchin]

Yeah, system managed c drive, then 12000-16000 for E drive (where auto1111 is located)

[warm junco]

Is E an SSD?

[junior urchin]

Both are, yes, both are beyond 1500/1500gb

[warm junco]

Ok

Normaly you don't need to create a pagefile for E

One should be enough on the fastest drive

[junior urchin]

gonna try and put a stick in it and experiment then

[warm junco]

Also dont forget to restart then

[junior urchin]

Switched it around to 24576 - 49152, no more ram bottlenecking

ya

Staying at around 12400 mb

[warm junco]

Perfect ^^

[junior urchin]

I was pretty much just waiting for the new AMD to drop, so I could upgrade my system, as my screen had blacked out twice, for some reason (which I expect was the ram issue), so thank you ❤️

[narrow kernel]

If you go above your 6gb of vram, then yes, the shared memory is your system memory and it's painfully slow for large numbers of parallel tensor calculations. You never want shared memory to be used at all. Look into options in whatever diffuser program you're using and check to see if they model offloading options(moves models and parts of models to system ram when not being actively used for computation).

[warm junco]

No problem, what's your GPU?

[narrow kernel]

Oh and another thing to note is that sometimes, things will work okay if usage spills into the shared memory, but you can't always rely on it since it's hard to know for sure if 100% of whatever currently used model is within the VRAM and that none of it is in the shared memory. But again, the safest bet is to always ensure that you stay entirely within the VRAM.

[junior urchin]

4070ti

ryzen 5 3600 cpu + 16 gb 2133mhz

mobo prevents higher

[warm junco]

Great card, make sure you use --xformers with it and if sdxl is slow use --medvram-sdxl

A AM4 board. The best you could get would be the r7 5800X3D

[junior urchin]

I do use xformers, but medvram has more issue for me, than without

[warm junco]

Oh ok

[junior urchin]

It's like the it/s takes a 2 iteration hit for no apparent issue

[warm junco]

With --medvram or --medvram-sdxl ?

[junior urchin]

either

yup

I see people blessing medvram

I'm like, wat, why do you bless poor performance

[narrow kernel]

oh you're using a1111

[junior urchin]

yeah

[narrow kernel]

no wonder why you're having problems lol

[warm junco]

--medvram-sdxl can improve the sdxl performance (with 8-11gb vram)
And --medvram only helps with 6gb vram or lower

[junior urchin]

Haven't figured out how comfy can adjust the -- stuff

ahh

[narrow kernel]

medvram and medvram sdxl are identical if youre using sdxl models. the medvramsdxl just only turns on if it's an sdxl model

the plain medvram just applies to all models and not just sdxl

[warm junco]

True

[junior urchin]

set COMMANDLINE_ARGS=--no-half-vae --xformers
That's what I use :P

[narrow kernel]

try some of the wrapper ones that have an a1111 like interface like stableswarmui

[junior urchin]

What about vlad :o?

[narrow kernel]

i highly recommend stableswarmui if you're wanting a simple frontend. a1111 was really good when it came out, but they've been extremely slow at updating and have some pretty archaic memory control

[junior urchin]

Yeah, I feel that one

[narrow kernel]

correction https://github.com/mcmonkeyprojects/SwarmUI

forgot he forked it after the fallout with stabilityai

[junior urchin]

That's why I've been contemplating switching over to vlad instead + more model support

For models with Sigma

I will probably use swarm, if it gets the support

[narrow kernel]

comfyui has it if you the use the extramodels addon

[junior urchin]

oh

[narrow kernel]

oh and also, swarmui can load comfy workflows, so you can set up a simple pixartsigma workflow(it's like 8 nodes) and then use the regular UI to run it

not 100% sure how well it works with custom nodes because i only use comfy these days, but i'd assume it would probably work

[junior urchin]

ah, I'll consider that next time

[lusty beacon]

Can I use it A1111 with SD3

[karmic bear]

Prob

Check the A1111 GitHub page for the change log

I believe they mention SD3 there

[warm junco]

With the next stable release update

[undone marsh]

Does anyone know how to mask a live stream video based on a prompt? Any link or suggestion would be really appreciated

[finite cloak]

pixart sigma is supported semi-natively https://github.com/mcmonkeyprojects/SwarmUI/blob/master/docs/Model Support.md#pixart-sigma

[junior urchin]

👀

I hope it works better than the regular comfy then
Nope, NETSDK1045 error like on launch, again

[broken smelt]

id look up vids of turbo models doing that, have seen a few vids showcasing that and other live stuff that you'd want fast inference for

[undone marsh]

Thanks @broken smelt Any resources or links where I can get info about it?

[broken smelt]

probly just google search for turbo models live stream, etc. the videos I've seen were all spread out across different social media

[undone marsh]

Ok, will look into it. From now I had no success that is why I asked here. I did my research before

[feral pike]

can someone please explain to me whats the difference between 'refiner' and 'refiner vae fix here'?
https://civitai.com/models/101055?modelVersionId=128080

[wise ginkgo]

I thought amd gpus sucked at SD /SD3?

why r ppl making posts about amd hype?

I havent' read one person recommend getting an amd gpu for sd

[feral pike]

i have amd gpu

only issue is the skill issue , GPUs are not the issue

[fervent thunder]

git gud

!!!

[west rivet]

thank you

[warm junco]

The refiner VAE fix has a different VAE in it which doesn't has the issue the first sdxl VAE had

[empty prism]

Friends, I have RTX 3050 VRAM 4 laptop, does it work, does it work Stabel Diffusion

[opaque needle]

Hello, I have a small question. I'm using RevAnimated for my realistics images. I tried Pony SDXL but I can't achieve anything realistic and most of my LORAs doesn't works at all on it. Is there a way to achieve realistic with Pony SDXL ?

[steel mural]

Does someone know why stable diffusion are not starting on pappersapce these days?

[humble sapphire]

Hello, I'm looking for a person/team who is an EXPERT in stable diffusion (and NLU). The mission would be remunerated.

[nova lily]

For example if your already generating at beyond 1K, you never have to worry about using hiresfix or SD upscale right?

or is there some value in it maybe the denoise process can add some fine detail?

[carmine edge]

hi

[cedar salmon]

exactly this, or if you want higher res

[lean laurel]

Hey guys, is the failed difusion channel gone?

[cedar salmon]

they nuked a bunch of image channels only anime survived, "low use"

[fervent thunder]

because in anime, as we all know, everyone is over 18

[nova lily]

I gotcha, The problem you have to get your prompt on point first as you dont want to be upscaling lackluster stuff. And I have trouble getting perfectly content with prompts.

[meager stump]

Nothing has been shared yet

[desert dagger]

you dont want to be upscaling lackluster stuff< why?

[fervent thunder]

This is my current standard method that I use for every image:

1. Start with latent shrinking methods like Kohya Deep Shrink or Hi-diffusion as the base, to boost the diffusion model resolution by 1.5x
2. Combine this with SAG or PAG where possible (compatibility can be tricky)
3. Take into photoshop and use very soft brushes to paint on white point, black point, exposure, contrast, gamma curve, colour adjustments and sharpening
4. First pass 4x upscale with either DAT2, HAT-L, ATD or RGT transformer upscale, or CCSR diffusion upscale
5. Repeat photoshop step
6. Second pass 1.5x upscale with SUPIR diffusion upscale
7. Repeat photoshop step

Would recommend this over hiresfix or SD upscale

It avoids any of the issues like tiling problems that SD upscale can have

The important part is that its only 1.5x upscale for Deep Shrink and SUPIR, yet 4x upscale for transformer or CCSR

This is because transformer or CCSR are less likely to mess up the image (trade-off is that they add less details)

This method gets you to 8k from SDXL, or 12k if you bump the SUPIR multiplier from 1.5x to 2x

[nova lily]

i mean like when your first starting working on a prompt, your generation probably wont be so good right away so no need to upscale.

[nova lily]

Itd be great if I knew what any of that stuff was. apart from photosoip haha

Im gonna screenshot your flow for future refence though appreciate the post.
With regards to upscaling though we gotta make sure the detail is even present before mega upsclaing right. all up to original generation

[fervent thunder]

no you don't need detail to be present in the original image necessarily
the upscalers can invent a decent amount of detail on their own

This is how I would rank them in terms of adding detail:

1. deep shrink 2. tiled upscale 3. SUPIR 4. CCSR 5. transformer

In terms of keeping the image consistent, the ranking is the opposite:

1. transformer 2. CCSR 3. SUPIR 4. tiled upscale 5. deep shrink

so if you are restoring old photos, a great option would be transformer into CCSR

but if you really want changes then deep shrink plus tiled upscale into SUPIR would be better

my method posted above was a hybrid

you can also change the trade-off between details and consistency by changing the upscale multipliers

so for example I had SUPIR at 1.5x but the transformer at 4x and you could flip these multipliers if you want (would not necessarily recommend it but you could do it)

[desert dagger]

why is why you do not use the AI you are generating with to do the upscaler

[fervent thunder]

sorry could you rephrase the question I am not sure what you are asking

[desert dagger]

i'm not. that was supposed to be "which is why" but autocorrect. you don't want to use the AI you generate with to upscale as that is exactly what they do. they add and change details. you want to use an AI that is just going to read the image and make an identical copy of it but larger and higher resolution

[fervent thunder]

it depends. some people seem to want their upscaler to change a lot

some people seem to call that enhancer rather than upscaler though

like look at the popularity of magnific.ai

if you want the upscale to be as consistent and conservative as possible then I think a single transformer upscale is best

[desert dagger]

capcut has a very good image upscaler that just upscales.

[fervent thunder]

ah thanks I will take a look

[desert dagger]

it's on their magic tools page. you'll need a free account, but it's free, and you even get a couple gig of storage for free. just remember to clear it out after you've finished a project

[fervent thunder]

I had a look, it looks like a strong model
this is using diffusion though
its less conservative than a transformer

[desert dagger]

yes, but it does a very nice job

as does the video upscaler there

[fervent thunder]

its a potential SUPIR competitor yeah

it depends on what you are looking for
because that Capcut model doing at least 4x diffusion upscale will change the image a lot more than my method

as my method limits the diffusion step to 1.5x

and does the bigger 4x step using transformer

[desert dagger]

might work with it some. i've upscaled a lot there, and haven't had anything change

[fervent thunder]

it can be subtle, anything stronger than linear interpolation will have some changes
bare in mind it depends on the upscale multiplier
for something like a 4x upscale it is a lot easier to do it with less changes
my workflow was for 9x or 12 upscale, which normally breaks most models including SUPIR

[desert dagger]

i'm a graphic artist. i haven't seen any changes. not even subtle

[fervent thunder]

well we are doing no-reference super-resolution so there is no way to know for sure
because we don't have a reference image at the higher resolution to compare to
the image in our head of what the higher resolution image is meant to look like is also "hallucinated" by our brain

linear interpolation is the only method that just takes the original pixels and resizes them without changes

[desert dagger]

you still might want to run some detailed tests through it

[fervent thunder]

sadly our upscaling models (and our image generation models like SDXL) are like 100 times better than our image quality assessment models

but yeah we can just run lots of example tests and then go by human preference

Capcut it a private company so it would not suprise me if they beat SUPIR

because closed source gets better funding

I will try some tests with deep shrink -> ATD -> Capcut and see how it does

[desert dagger]

they're your competition, so jut a suggestion to be aware of what they can do

[fervent thunder]

no I think this is a different category of upscale, not direct competition
because its in a very different spot in the trade-off between consistency and detail

Capcut is a competitor to doing a single pass with SUPIR

these are the competitors to my workflow:
https://github.com/Phhofm/models/tree/main/SUPIR```` https://old.reddit.com/r/StableDiffusion/comments/1b50sp0/ccsr_vs_supir_upscale_comparison_portrait/```
there is a third one, a reddit post that combined 2 non-tiling upscalers with tiling upscale

[opal hedge]

Hey folks, has there been any news on sd3 8b release

[desert dagger]

no. and there won't be for a long time

[fervent thunder]

sadly no news

[snow blaze]

Is there a way to generate imgs in private chat\channel? (not using artisan 1-5 when everybody see the results) maybe some subscription needed?

[nova lily]

When "Looking at viewer" isn't complying, im sure you could use some photoshop trickery to "aim" the gaze. Does anyone do that?

[fervent thunder]

well you could move the eye in photoshop I guess

but that would not help head angle

which is likely the bigger issue

sometimes you just have to cherry pick out of like 200 generations
before upscale

[desert dagger]

you could go to the website

[narrow kernel]

Think about the captioning that went into the image set. I highly doubt autocaptions would say "looking at viewer" when captioning things like portraits. They would say something like "looking forward" or "looking toward the camera" or "looking directly at the camera" and so on. Also, depending on the captioner, they might only make mention of the directionality if it deviates from a normal "portrait," since probably 95% of portrait photos have the subject looking directly at the camera.

yeah i just tried a handful of VLMs and they seem to all use wording involving the "camera"

lemme try on some traditional art and see how they word it

yeah seems like even for traditional art, they are using the same "at the camera" or "forward" style wording. but i have seen a couple "viewer" references with the bigger chat style VLMs, if you have them really spell it out with a ton of words. but autocaptioning used for datasets aren't writing five paragraph essays about the image

[graceful bramble]

Anyone have an idea for how I could get a persons eyes closer together with pony? Eyes close together or close set eyes do nothing.

[fervent thunder]

omg its a tiger playing soccer

that would be totally unfair irl

[rugged warren]

hi everyone! Anyone knows about a spanish comunity channel or something to make questions about stable diffusion in Spanish?

[fervent thunder]

There are probably some spanish speaking people in this Discord.

[rugged warren]

Perfect, anyway I ask about this channel. I need to know a little about how to take a door design and generate different rooms with the same door that I have

I usually make a mask for the door and create prompts for living rooms and rooms with img2img, but many times I simply place the door on the wall of a dining room or on top of some furniture...

[bleak matrix]

Good morning, everyone! How are we all doing this beautiful day?

[fervent thunder]

Think about the captioning that went into the image set. I highly doubt autocaptions would say "looking at viewer
you would have thought so but some models do respond to this
likely due to the manual captioning

[austere cosmos]

Hey everyone! I’ve summarized Radek Osmulski's "Meta Learning"

https://riteshrm.github.io/posts/Meta Learning By Radek Osmulski Chapter Wise Summary Points/

Would love your feedback and thoughts!

[fervent thunder]

not sure is its worth ever doing summaries ever again
because LLMs can do this now

[primal basalt]

Hello, is there an effective way to prevent the SD of sensitive images from being illegal?
Any complement? Or does the positive prompt have to be passed through AI to identify more intentional orders?
I will pay whoever advises me on this issue

[trim magnet]

what

[primal basalt]

dm?

[trim magnet]

no ty

[desert dagger]

i doubt they say that, either.

[sudden tiger]

Hello! I generated a chair using stable diffusion, but now I would like to change the color of the chair to black, do you know how I can go about it? THANKS

[pale latch]

nope. if you have images to protect, make sure they're behind a paywall and you don't distribute high quality versions of them on the public internet. be prepared to defend your copyright if someone posts what is clearly derivative work

same old ways of protecting your copyright still exist

[pale latch]

WD tends to create "looking at viewer" captions. They're also trained into most sdxl and sd15 refines because they use a lot of booru style tagging, which is where looking at viewer comes from. https://safebooru.org/index.php?page=post&s=list&tags=looking_at_viewer

[fervent thunder]

hello

[steel hull]

Hi

[fervent thunder]

i am blind so i am not able to say much but i am interested in science.

[desert dagger]

https://research.nvidia.com/labs/dir/jedi/ this looks promising

[fervent thunder]

i look promising.thank you very much.

[pale latch]

still waiting to see the backlash for blocking the clown emoji. wonder when it's gonna happen.

[unkempt hatch]

Interesting. So it needs to be trained on 512x512 before 1024x1024?

And so if I do a fine-tune, might it help to do an epoch of 512x512 and then epochs of 1024x1024 with the same images?

[pale latch]

base model training is a lot different from refining an existing model. pretraining usually is lower resolution

[unkempt hatch]

But what's the benefit of that lower resolution?

[pale latch]

i have no idea about what would help now. i don't think a 512 pass on the final refined model would help so much

[pale latch]

its part of the process of building foundational weights

[nova lily]

Yes you can get away with it. But not very high resolutions you wont be able. https://github.com/CS1o/Stable-Diffusion-Info/wiki/Installation-Guides#nvidia-automatic1111-webui-stable-diffusion-webui

[unkempt hatch]

But could it be useful in a fine-tune introducing new concepts?

[pale latch]

i think if it did, we would've seen articles about that already

[pulsar venture]

Hi, there is someone that could help me with some SD extensions? I need help for LLUL and composable lora

[chrome herald]

they might have concatenated booru tags after the rich captions at least for the anime images in SD3, like this is sd3 just on "1girl, from side, autumn leaves, solo, full body, 2d, anime, looking at viewer": #🆕｜sd3 message

[river sparrow]

a list of simple 2 or 3 word groups, between commas, seems to work, no issue, for sd3

[umbral shale]

curious - has sai mentioned in any papers that they used auto-captioners for their data? my understanding was that they only used the scraped <img> alt tags on the internet for 1.5 and xl (haven't had a chance to dive into the sd3 paper yet so might be outdated)

[narrow kernel]

yeah, sure, for old ass models they used a lot of tagging like that. for newer models, they don't. things like cogvlm, florence2, llava 1.5/1.5, moondream2, etc, all tend to lean toward what i was saying though. as for them adding in manual captions like that, that's a very likely hell naw lol... peppering in even tens of thousands of manually captioned images into a sea of tens of millions isn't going to do anything. what they could do though would be to manually link concept phrases together so that things like "looking at the camera" and "looking at the viewer" would essentially map to the same spaces internally. you can't hardlink the actual word though like camera and viewer because then what if you want an actual camera in the photo.

[narrow kernel]

a lot of that would just be from crappy captioning scraped from sites that had the tags for the image, but they definitely could have added in a bunch of DB tags into the mix like i was just saying, by manually linking some of the common concepts

[narrow kernel]

yes, they use things like LAION and other image datasets, which are mostly tagged by AI, mixed with some manual captions, mixed with scraped meta data from the sites the images were taken from. as for SD3, they also used cogvlm

[pale latch]

people are still going to refine with booru tags

[narrow kernel]

sure, just don't expect it to keep working with newer and newer models

[pale latch]

it will, because people will keep training the image boards

[narrow kernel]

sure, some will

[fervent thunder]

I have no earthly idea why I would put millions, or billions of dollars into ai trained on booru tags, I just dont get that

It's just...dumb

mainstream it

[trim magnet]

NAIv3 uses booru tags and its not an old model

[fervent thunder]

I just do not understand how ai competes without using natural language models

[pale latch]

base models? sure ok. maybe not. community refines, 100%. i had no idea what booru tagging was before i started with ai, but then i looked into it and wow, its established. image board culture going back a long time.

People won't simply stop relying on those datasets and visual language

[trim magnet]

because the training images are pulled from booru sites which uses those tags so converting those tags to natural language would be a waste of time/money

[pale latch]

not to mention, booru tagging is an established vocabulary tied to image descriptions

[trim magnet]

yea if it works it works

[fervent thunder]

If you are going to potentally make billions of dollars, how is it a waste of money? I do not understand.

[trim magnet]

because ppl who make anime images already use booru tags so they are familiar with them plus spending a few million dollars to convert tags versus not spending those millions,easy choice

[narrow kernel]

well one of the leading reasons why is because the average reading/writing level for most people on the planet is no higher than that of a 10-12 year old. a lot of people just don't have the literary skills or vocabulary to prompt correctly, so things like DB tags took over and have remained. oh and also because of token limits, but that's not really as much of an issue any more with newer models. even though i'm dunking on DB tags, there's nothing wrong with throwing in some lazy short single word tags in an image

[fervent thunder]

who cares about the anime people? This is where I get lost

I see Microsoft investing billions into mainstream and think, "they get it"

Why can't you do multiple models

[trim magnet]

NAI cares,even bing cares considering they included anime in their training

[pale latch]

and so long as weights are availabel to train, the motivated few will train the booru image sets

[narrow kernel]

sure

[trail lion]

just based on the amount of content I see on civit, I doubt it's few

[fervent thunder]

I will never understand all of this. It is beyond me.

[pale latch]

lol anime nerds are not so much a minority as they were when revenge of the nerds came out

[narrow kernel]

even civit is a tiny echo-chamber vs how many people actually use AI for image creation

[pale latch]

the nerds are people's bosses since dotcom boom

in cases especially with tech, nerds are often ahead of the curve and drive adoption trends

[narrow kernel]

well either way, between all the bad tagging and the flooding of ai content, model collapse will happen at some point. now whether or not that happens before laws and regulations restricting the training/usage of generative AI get enacted, is a different story

[fervent thunder]

Will be interesting to see the laws

once people get triggered

[pale latch]

we got ways of dealing with data restrictions. just mail hard drives to each other

[narrow kernel]

and if you get caught, you get charged

[pale latch]

microsoft got these glass sheets they laser etch data into too. stays good for million years or something

[narrow kernel]

like selling contraband or weapons illegally

[pale latch]

such an underground data economy will thrive

[fervent thunder]

Kids are already getting addicted to ai bots

Society will collapse. Nothing will be real. Mcdonalds everywhere will have functioning ice cream machines again tho.

[trim magnet]

they are addicted to phones already

[fervent thunder]

I know right. But so are old people.

[narrow kernel]

just as you're addicted to using a computer

[fervent thunder]

Everyone is addicted to the interweb!

🙈

[narrow kernel]

just as boomers were addicted to watching tv

everyone gets addicted to something

[pale latch]

vulnerable and lonely people were already addicted to sex chat lines. it sucks that the bots are marketing at kids now a days, like directly at the kids. its a bet pedo'y

[trim magnet]

that one is hard to police because your app would have to ask for some sort of ID to use it and when you ask that ppl start complaining about data privacy and idk if google allows apps that ask for id to be on playstore

[pale latch]

i dont think sex bots are going to get much of a market penetration at all. same as there is for existing sex bots.

most people will just the cheap easy route . ol palmela handerson

[fervent thunder]

I just lost a friend to chat bots

[pale latch]

chat bots might shrink the market for phone sex lines. not such a bad thing.

[fervent thunder]

She was..a nice person lol

[trim magnet]

id pay for cyborg pamela anderson

[pale latch]

until you learn that you're the one who has to 'clean' it

[fervent thunder]

oh yes, cyborg brad pitt

[pale latch]

lonely people will exist. addictive personalities will exist. i just don't think ai chat bots will drastically change anything. it'll just shift where that market potential goes

[fervent thunder]

I tried character.ai just long enough for the bot to piss me off

hmmm

[pale latch]

ai dungeon was fun for a bit, but it just gives you what you want to hear. not a very crafty GM

[fervent thunder]

I forgot about that one

[pale latch]

it has really bad world consistency too

would love one of these models hooked up to proper game state management

[fervent thunder]

I just want a good game of dnd!

🎲

[light skiff]

can someone please recommend a SDXL model/lora to convert isometric pixel-art into realistic looking with high detail art?

[fervent thunder]

try zavy maybe

[light skiff]

thanks, i tried zavy lora, that works better than my previous attempts

[balmy moss]

Anyone use the ChatVision app?

[hollow notch]

Wtf I asked for help and no one answered my ticket just closed it

[desert dagger]

did you post in the #🤝｜tech-support channel?

[hollow notch]

Nope but I did now

[dawn heron]

Yo how you guys generate images that look nice?

[quaint pawn]

The key is to be specific in the description or prompt.

[placid hatch]

what 1.5 models make good refiners for anime and illustrative style content on an sdxl or pony base?

[fervent thunder]

not sure but try Ghostmix and Rev animated

Hey

[gray ferry]

怎么使用图生图

[stray notch]

Hi all 🙂 Since it looks like mov2mov extension is dead - is there something similar that could be use for "face replacement" in gifs/movies? Doing it via img2img (each frame) is not consistent and there are glitches so I am looking for something that could make it simpler and more reliable

[warm junco]

Roop-unleashed can do this.
Its a seperate tool but works great for faceswap videos

[stray notch]

thank you! this looks great

[gleaming topaz]

Does anyone here have experience merging models?

[stark tusk]

So you are going to experiment on kids?

[fervent thunder]

🤨

[light skiff]

I use the "vary (subtle)" feature in Fooocus to alter an Input image, it converts to the new style pretty well however the model also changes the composition too much and some details are missing. How can I avoid that?

[warm junco]

hey, you can try to install the zluda variant of SD, a guide for that is in #🤝｜tech-support
it supports the rx580 when a fix is applied. but make sure you have the gpu with 8gb vram

[desert dagger]

i have a suggestion for you - rather than posting this to random discords full of people that have never worked in a classroom or taught students, go to the people doing the work, the teachers, and let THEM tell you how they want their classrooms to be revived. teachers put a LOT of work into their classes, so if you want your project to be useful to them, rather than ignored or laughed about, go to them

[tender turtle]

Does anyone have experience improving domain-specific inpainting? I'm working on a comfyui workflow for inpainting environments around products, but SD always wants to paint more of the product around the edges of the given image. I've tried using an edge controlnet with a line drawn around the inpainting mask, but this doesn't seem to be effective.

[knotty dove]

Hi

[fervent thunder]

I think a Lora is what you are looking for
another option is style transfer using IP adapter

a more powerful option would be to train a full checkpoint model

if you use comfyUI then you can use many different IP adapters together rather than just one, which can help

[quasi olive]

主題是烘焙蛋糕餅乾的icon 主體是一個蛋糕 由上往下看的視角 蛋糕是四分之一塊 缺了一角 缺的那一角要畫上餅乾 畫風文青風 像是蠟筆花出來的感覺 高質感 平面或是3d的icon

[fervent thunder]

[bleak drift]

One statement I agree with against AI images is that they are ruining google images

Becauzs now it is filled with not even good ai art that I think sea art just uploads

And other sites seem to do it too

[gleaming garnet]

Hi all - not sure if I’m in the right place but hoping I am.

Is there a way to set up stream diffusion on TD with a Mac?
Everything I am finding it’s only on windows

https://youtu.be/bcZXlhy7KDE

I’m looking at building something like this for some new physical art works, that I hoping to run in TD and make data reactive

[unkempt hatch]

Adetailer doesn't vae encode and just works at the full resolution.

Doesn't that mean using adetailer is like having the best vae decoder ever? Shouldn't it be way more detailed than anything, including SD3?

[untold loom]

How channel to generate images ?

[verbal scroll]

a detailer IMO is just a instant impaint option there is no "magic" to it

[untold loom]

This is not free ?

[river sparrow]

1 get a gpu
2 download free stable diffusion software
3 download some models (checkpoints)
4 generate images for free (read: energy costs + costs of having a computer)

[nova lily]

Wait do you actually have to put space after camas? Or has that just been me being OCD Grammar Greg this whole time?

[wary belfry]

generate an image with and one without while keeping the same seed and you will find out 😄 But using xformers slightly alters your image, even with the exact same settings

[desert dagger]

every ascii character carries data. the AI learns three things when it trains - it learns what words look like, it learns concepts, it learns relationships between words. make the experiment. use this prompt: an apple <--- and use the seed of 11. generate that. then change the prompt to: an,apple. generate. change the prompt to: an, apple. generate. change the prompt to: an. apple. <- now compare what changed (subtle or not) on all of those. then try something like: an apple sitting on the table. an,apple, sittingonthe. table

[silk gulch]

for anyone on mac, what is the fastest ui?

[low moon]

buying windows

[sudden fjord]

How do I get stable diffusion

I got the license

[silk gulch]

https://huggingface.co/stabilityai/stable-diffusion-3-medium

[modest lodge]

1

[sudden fjord]

Where do I paste that code it gives me for diffuser’s pipeline

[silk gulch]

a python file?

[junior thistle]

Does having the checkpoints and web uis in a hdd instead of a SSD make a difference?

[valid comet]

用stable diffusion

[thorn ether]

just out of curiosity anyone running stable diff on a amd rx6800?

[graceful bramble]

Anyone know a good token for a thinner face or bringing the eyes closer together or changing the nose size on pony?

[vital gazelle]

Iam new to stable diffusion, what model should i download?

[warm junco]

Should work with the AMD Zluda guide from #🤝｜tech-support
(In the pinned messages there)

[thorn ether]

i was just checking your stuff out! i cant wait to learn all about it and give it a shot if i have any questions i will see ya there! thanks for everything

[warm junco]

Np, yea let me know if any step isnt clear enough 🙂

[thorn ether]

oh i promise i will see ya soon very. new to everything spent the last 5 days banging my head at a wall with anaconda guidehttps://www.stablediffusiontutorials.com/2024/01/run-stable-diffusion-on-amd-gpu.html ended with errors that i couldnt find a fix for

[warm junco]

Oh please not that guide xD

[thorn ether]

yeahhhhhhh.......

[warm junco]

And get rid of anaconda before starting my guides

[thorn ether]

deleting everything as we speak

[gray vortex]

#artisan-faq

[silk gulch]

does anyone know what the fastest mac ui is

[novel hearth]

hello

[pulsar venture]

Hi, when Dora will be available on automatic1111?

[bleak matrix]

Good morning, everyone! Hope you are all doing well!

[obtuse badge]

Is it possible to host the basic automatic 1111 using vercel.com or wix.com?

I just wanna use the inpainting feature

[trail lion]

I'm not the developer but I see a commit related to that in the dev branch so there's hope, assuming you meant the finetuning thing, and not the cartoon character

[pulsar venture]

I'm just a normal user, so my question is when i can use the dora for txt2img for anime characters?

[trail lion]

I understood your question, I dont know...but It seems like it's coming

[pine fiber]

did anyone try Anole yet?

[alpine bear]

[toxic plank]

Hey do we have some promting experts here? If so feel free to shoot me a DM, I am facing some issues with my prompt. I use the "ultra" model.

[nova lily]

When you use the gold star highres fix button, it uses settings from the highres section?

thank you

I used it a few time yesterday and think I see what why it's good. I didn't realize it can potentially turn lack luster images into improvements. skin texture wise and face wise.

What im wondering is ill try it later if I can still get enhancements while making the scale factor very miniscule.

[novel axle]

hi there, im working on a project that takes in a user's image, and then uses ai to gradually age / deteriorate the user (and their background). im new to ai dev tools and was wondering if there are any recommended apis or models that are good at image alteration (stable diffusion or otherwise). the ones ive tried are not always great at maintaining visual consistency between the images

[fervent thunder]

I think most people just do 2x

[rustic nova]

Hi everyone.I have a question on Comfy-portable. Can anyone help me how to fix the error?
Or is there no discussion on Comfy in this community?

[loud solar]

Feel welcome!

[rustic nova]

Thank you. Then I'll write to tech support.

[loud solar]

It's not official Comfy here but many of us try to help ...

[fervent thunder]

[hardy sedge]

Is there a way to train loras specifically for pony checkpoints? At pony checkpoints, loras don't look so accurate.

[craggy kindle]

I think it depends on the pony checkpoint you use. Ive gotten good pics with Pony xl v6

[hardy sedge]

But a character lora looks much more realistic at the normal sdxl checkpoint. With pony you don't recognize the character so clearly

[candid epoch]

@hardy sedge
Are you implying your eyes can't handle recognizing a character without their trusty pony? Maybe it's time for a prescription check-up before you mistake your friends for strangers next!

[desert dagger]

you might want to ask the people that specialize in training for pony...

[visual lagoon]

has civitai still not lifted its ban on SD3 stuff?

[pale latch]

thats correct

they're now invested in competing with stability too, with the omi

[visual lagoon]

what's omi

and that's dumb, since SD has cleared up sd3 stuff civitai should let people upload finetunes

[pale latch]

Open model initiative. They're investing into supporting a model with no commercial restrictions.

We may never see civit AI allow sd3 content

[desert dagger]

let's hope they're not using models created by other people then

[pale latch]

Their plan is to find one on a libre license and refine it

MIT or Apache or something

[cedar salmon]

sounds like misinformation

[low moon]

tsk tsk without emad there is no civit, now they compete with the emad... this omi model better be good then

[low moon]

no but they will train on images generated by othe rmodels XD

[pale latch]

civit havne't just not got around to unbanning sd3 content yet. the fact is it's still banned given the new license. they're planning their business strategy is why it hasn't been unbanned yet. all we can do is speculate on that strategy which includes an investment in OMI

[low moon]

tbh the future is personal models, in a few years everone will train their own model at home with datasets they make thats the relatively near future

[low moon]

omg can u imagine civit also becoming like sai down the line when investor pressure kicks in.... all of a sudden you'll need tokens and u can only dl 3 models a day and then there is paid plan etc 😄

[pale latch]

i dont want to imagine but they are a business that venture capitalists are trying to build and it isn't cash positive yet. so there's a lot of unknowns

[low moon]

the onyl way to be 100% safe elon musk must buy civitai

[acoustic dagger]

Why do my images look awful?

[low moon]

coz its an unrefined model

ur trying sd3?

[pale latch]

i could see schmidy and parker doing something ballsy like acquiring civit as part of their investment group

[acoustic dagger]

What models are refined?

[low moon]

SD15, SDXL and Pony

but beware with Pony

[acoustic dagger]

I tried SDXL

Awful

[low moon]

that should give you OK images

noooo

[acoustic dagger]

Yeah they're decent some of them

[low moon]

whats your workflow?

[acoustic dagger]

Default

[low moon]

in comfy?

[acoustic dagger]

Yes

[low moon]

which model u tried?

[acoustic dagger]

SDXL

And SD1.5

[low moon]

oh wait ar eu using the basemodels?

[acoustic dagger]

Yes

[low moon]

no wonder

[acoustic dagger]

So what should I do?

[low moon]

go to civitai and look for the lastest "Juggernaut" model for SDXL and "realisticstockphoto" for SD15.

zavychroma is good for SDXL finetunes also

and epicphotogasm for sd15

when u search look for stuff that wa supdated past 2-3 months

[acoustic dagger]

Is it true you shouldn't run stable diffusion XL on 512x512

[low moon]

yes use 1024/1024

[acoustic dagger]

That was the mistake I was making

[low moon]

probably

[acoustic dagger]

Also

How can I make 512x512 images

Do I need a specific model?

You know

Ones that actually look good

[low moon]

if u must u can resize it in comfui

make it 1024 and resize to 512

[pale latch]

some sdxl refines are better at 512x512 , but if thats all you want then just use sd15 honestly

[icy solar]

Hold my beer: Dell Inspiron with NVIDIA GeForce 1050 mobile + Fluently XL + LoRAs. Code: https://github.com/InServiceOfX/InServiceOfX/blob/master/PythonLibraries/HuggingFace/MoreDiffusers/morediffusers/Applications/terminal_only_finite_loop_main_with_loras.py

[nova lily]

Does anyone use tiled diffusion, anyone know why im getting subtle circle shapes all over the image.

[alpine bear]

Have you tried using 'subtle circle shapes all over the image' in your negative prompt

[halcyon wasp]

which channel can make pic

[low moon]

artisan

[nova zodiac]

Ive heard of a 960 4gb running a sd1.5 model but wasnt much faster than a modern cpu lol

[acoustic dagger]

Images are coming out terribly how do you optimise Stable Diffusion to make good images?

[high wigeon]

Hi there!

Is it possible to find out which Stable Diffusion model the bot is using to generate images? Additionally, can I get information on what model was used on specific dates, such as in August 2022? This is for a research project I am working on in college. Thanks!

[fair crystal]

Hi everyone, I'm having some questions about writing prompts for Stable Diffusion 3 (SD3), especially considering its new features like natural language and longer prompts.
Here are my specific questions:

Prompt Length: Is there an ideal prompt length in terms of tokens? Does going beyond a certain number negatively affect the results? (For example, does SD3 struggle to remember all the instructions if the prompt gets too long?)

Prompt Order: Does the order of elements in the prompt matter? For instance, should I prioritize mentioning camera angles or the style of the final image first?

Character Placement: When using SD3 to depict multiple characters in a scene (e.g., a red fox near the camera on the left and a yellow lion far away in the distance on the right), what's the best approach for describing them? See my options below
Option A: Describe each character individually, then their placement after.
Option B: Describe one character's appearance and position first, then the second character's details after.
Option C: Describe the overall scene composition/layout first, then details about each character's appearance.
Option D: Another way?

By clarifying these points, I hope to improve my SD3 prompting and achieve better results.
Thanks in advance for any insights!

[feral pike]

hiii

[onyx axle]

[fossil rune]

hey guys im having trouble installing ReActor can anyone help me?

[uncut tundra]

Hello, I'm having a pretty serious problem. I've been generating images with Automatic1111 for months, using SD 1.5, and I have an RTX 2060 Super. Approximately every 40 minutes of generating images, the screen would suddenly go black, and after restarting, everything would be fine. The temperature would rise to around 70 degrees Celsius. However, two days ago, I started having a different issue where I can't generate anything at all. Every time I try to generate an image, the screen goes black. I've even tried formatting the PC, but the problem persists. With the default settings, I can't get it to work.

This issue started happening after using commands like --xformers. I don't remember exactly which command caused it, but this is the information I have. Can anyone help me? Did I break something? For now, I can still play games without any problems.

Also, does anyone know of a good quality-to-price service I could use to run Automatic1111 remotely (in the cloud) with my setup? It needs to support custom LoRAs, NSFW LoRAs, checkpoints, etc.

[fervent thunder]

🆕 🆓 🆒 🆗 🆙

[sturdy sandal]

Hello

[trail lion]

Runpod would let you setup anything you want, and there are options to keep the image you build, which costs more than an instance that gets destroyed every time. There's other services like rundiffusion, but I haven't used it and don't know if it supports all your extensions

[warm junco]

You should use --xformers --medvram --no-half-vae and nothing more as launch args in webui-user.bat

[modern moth]

Guys ik this is very off topic and shit and ik it's probably not allowed here but I'm in deep huge shit

I need to sell my valo account so if interested pls DM

[ember tide]

Hello guys

I needed some guidance

I want to build a webapp that lets you faceswap,upscale and transform images using custom models like the ones available on civiai

I have experimented with them locally using automatic1111

and wanted to know how do I proceed with building my own image generation tool with above features?

I looked into diffusers library but it's in python and I only know Javascript

I am fairy new in learning programming have made some little projects till now so if anyone can help me out it would be good

[light skiff]

gpu heavy tasks on a server will be really expensive, but if you want to use machine learning tools it would be useful to learn python

[ember tide]

I only want it as a project

[light skiff]

you can also look into automatic1111 or fooocus source code to see how they load the models and predict images

[ember tide]

It seems to advanced for me rn

Can I not just learn Python like how things work and use diffusers Library?

[light skiff]

personally, i would start with a smaller project and train a simple pytorch model

but you can of course use the diffusers library

[vital agate]

You can use Express or NestJS to build a middleware layer and directly call the Stable Diffusion API from there

[graceful jackal]

Hey, need someone's help with gen. that has stable installed on a local machine

Long story short im away at work and will be home in like a month. Need someone with "stable" running locally (non XL, can be makeayo). Had been bored and wrote a prompt in my spare time. Wanted to see how well it performs, but it needs to been locally since its long/weighted and i wont get the same results with cloud based gen. (like never). Side note its nsfw(nothing hardcore or lolicon). I would send both positive and negative privately. Thank you

[visual lagoon]

@fossil rune lol why did u dm me

bit sus

idk u

but if you're rly having troubles with installing reactor for comfyui

https://www.youtube.com/watch?v=vCCVxGtCyho&

Follow this

It worked for me

[viral forge]

Hi, I just recently started using Deforum Stable Diffusion and I was hoping someone with more experience may be able to point me to some techniques / parameters that would be best for evoking clear color transitions. I'm trying to make an abstract video to music where for the first 300 frames the entire landscape is black and white in ink brush texture, but then on the beat drop, starts to integrate more colors. When I tried this, the colors would show up briefly for a frame or two, but then be drowned out by the original black and white color scheme (I assume to keep the cohesion of the video). I'm currently using the API version to make deforum calls if that's relevant. I wanted to try a version where I set color_coherence to None, but would love any additional insight. Thanks!

[undone garden]

Why does it take a year and a half to interrupt the generation on auto1111?

Why can't it be easy

[ember tide]

R u sure it's not your hardware

[crude marlin]

Is there any way to use this for free or a similar alternative? I just checked this Discord since January, and now it's not free all of a sudden 😦
I don't like Bing and that's the only good option I'm seeing that is free so far..

I'm looking for something in a web, not to download.

[echo kite]

huggingface spaces

[crude marlin]

Thats the one I have to download rigth? I only have a chromebook and I dont even know to install those things so..

[echo kite]

no you don't have to download it

[crude marlin]

ohh

[echo kite]

idk if I can send links here but just google SDXL huggingface space

not the one with the model download and model card

but huggingface space

(not space as in spacebar)

it is what I used until I get SD back working locally again

[crude marlin]

alr thank you so much I'll look into it!

[signal violet]

Can anyone tell me how to make x-ray in promt without using LoRA or other plugins?

[pale latch]

just prompt for a medical xray image

[lyric orchid]

hi

[nova lily]

Learned about tiled diffusion/tiled vae yesterday.
Is there anything out there the makes makes Hires fix faster or optimizes it?

[meager wharf]

It’s partly a VRAM thing, partly a A1111 thing

[nova lily]

its simply like a train. After chugging hard it needs to ramp down.

[pale latch]

a new gpu

though seroiusly maybe the kohya hires fix? since it'll fix in fewer steps

[nova lily]

okay ill check out what kohya is. I just wanted to make sure there's not some obvious MUST USE extenstion or stuff I may not know of

[trail lion]

"I want to make it less filling, but tastes great" 😄

I dont get this demand to make things faster, like what's that all about? are you on an assembly line? if your perfect photo takes 2mins vs 1 min, what is the drama here?

for commercial it makes sense maybe

[fervent thunder]

I generate at minimum 200 generations in one batch

so it really adds up

I have a cherry picking ratio of 0.5%
so for every 200 images I generate I throw away 199 and pick one

that's before any upscaling

[fiery magnet]

does anyone know a discord bot that allows you to provide an image set, and allow people to sort it?

[static drum]

hi

[shell tendon]

for every 200 images i generate i generate another 200k and run out of hard drive space

[fervent thunder]

LOL

these models are so random that I don't rly understand why people don't do big batches

aside from time and cost

[desert dagger]

i just stick all my images into channels on my discord and dont' worry about how many i make

[fervent thunder]

storage is finally cheap anyway

[narrow kernel]

I don't save most of what I generate. Like per 100 images I generate, I might save 1. While I enjoy making stuff with diffusion, the creations are of almost zero value to me. So I don't endlessly hoard stuff I'm never even going to look at a second time

But then again, maybe that's because I do traditional/digital art and photography

[fervent thunder]

I don't save either

I should start making a system probably

[vital agate]

Hi,I would like to know what kind of mask image I should upload for inpaint sketch when using the API.

[fervent thunder]

The image passed to this parameter should be a black and white image that represents, at any pixel, the strength of inpainting based on how dark or light the given pixel is. Completely black pixels represent no inpainting strength while completely white pixels represent maximum strength.

In the event the mask is a different size than the image parameter, it will be automatically resized.

Alpha Channel Support

If you don't provide an explicit mask, one will be derived from the alpha channel of the image parameter. Transparent pixels will be inpainted while opaque pixels will be preserved.

In the event an image with an alpha channel is provided along with a mask, the mask will take precedence.```

does that help?

go in photoshop, make a new layer
paint the shape that you want
then make it just black and white

send mask from that layer to Stability AI

[vital agate]

Thank you, buddy. However, I believe that the black and white mask only applies to the inpaint mode, and it does not carry color information.

I tried using a grayscale image, but it still didn't work.

[narrow kernel]

it probably assumes to use the A channel of an RGBA image

[fervent thunder]

you can go to channels in photoshop and paint in the alpha channel

to do it that way

[narrow kernel]

so youll need to save it as a 32bit png or some format that handles all four channels

[fervent thunder]

not sure if it wants pure black and white or if it takes greyscale

there are different types of inpainting

it might just want pure black and white

the only greyscale inpainting I did required the Differential Diffusion tool

but this might not be the default

[vital agate]

Thanks for your reply, mate. I'll give it a shot right away.

[narrow kernel]

yeah i dont know what type of inpainting they use, but diff-diff is pretty much all you need now, so it definitely is best to have some blends in there, rather than a binary mask. plus, even if they dont use diff-diff, it just rounds up or down to 0 or 1 from <127<

[fervent thunder]

I really love the blending rather than just old binary inpainting

[vital agate]

      for (let i = 0; i < imageDataContent.length; i += 4) {

        if (imageDataContent[i + 3] < 128) {
          imageDataContent[i] = 0; // R
          imageDataContent[i + 1] = 0; // G
          imageDataContent[i + 2] = 0; // B
          imageDataContent[i + 3] = 255; // A
        } else {
          imageDataContent[i] = 255; // R
          imageDataContent[i + 1] = 255; // G
          imageDataContent[i + 2] = 255; // B
          imageDataContent[i + 3] = 255; // A
        }
      }

This is the way I convert masked images.

[fervent thunder]

why not python lol

[narrow kernel]

yeah i only use diff-diff these days, havent touched a single inpaint model since. (old inpainting was binary masking and even if you tried to blur it, it didn't really work well, don't think it even did anything internally though)

[vital agate]

i can't python 😦

lol

[narrow kernel]

of course typical inpaint setups would try to do some post process blending though between the base image and the inpainted part though, but that's why it always looks bad

[narrow kernel]

if you can javascript, you can python

[vital agate]

I understand that their syntax is similar, and Python is even more flexible in some ways. However, personally, I'm not a big fan of Python's writing style. Nonetheless, I still believe that Python is a great language.

[fervent thunder]

its mostly that
AI stuff is almost entirely in python

so now its kinda a bigger deal

[narrow kernel]

well as neon said, if you want to do anything in the AI ecosystem, you'll need to get used to python because it's 99% python

[vital agate]

I can understand some Python code, but when it comes to neural network algorithms and machine learning, I'm completely lost. It's just too difficult for me, lol.

[fervent thunder]

its best to start with classical methods

linear regression, decision trees etc

[cloud sundial]

Hi guys, i am new to Stable diffusion and was just wondering when generating txt2img, what the max resolution is (not hires-fix) for width and height? I'm trying to make sure i'm using sdxl, is the width/height slider supposed to be only 2048?

[fervent thunder]

there's no exact number

but as the latent increases in size, it becomes more likely that the composition will get strange

there are things you can use to reduce this such as hidiffusion, deep shrink and control nets

if you just want a resolution that is most likely to work well then 1024x1024 is best

[fair crystal]

I followed some Python courses online to get up to speed, but a lot of the terms used in machine learning are pretty overwhelming indeed. I did create some stuff with the Diffusers pipeline on a cloud service but would like to learn more.
Do you have any free online resources where I can ease into learning these principles better?

[fervent thunder]

do you have postgrad level math and stats or do you need a course that also goes through the math?

[fair crystal]

I def don't have next level math skills

[fervent thunder]

Hello, when I download an XL Lora model file, I place it in the models/Lora location, but when I open stable diffusion, the XL Lora files do not appear in the Lora section. What is the reason for this? Do I need to place it in a different location?

[fervent thunder]

maybe start with an Introduction to Econometrics textbook
that's a decent way to learn regression

[fervent thunder]

is this A1111?

@fervent thunder webui_forge

[narrow kernel]

Lol you don't need it unless you're trying to create a whole new model architecture. For run of the mill use of models, simple algebra is enough. Ever used an equation like calculating the volume of a sphere? It's algebraic, but is derived from calculus. Even needing to explicitly know how to calculate things like regression aren't needed when there are libraries for that that you just plug your numbers into

[fervent thunder]

oh I agree to a good extent
its just that I would recommend a completely different sort of book to someone who already had the stats background

[narrow kernel]

point is, he doesn't need to know some ultra high calc level math to understand a lot of the math behind models. but yeah, some basic knowledge of statistical terms is good to know since all AI is based on curve fitting. there are far simpler crash courses for getting up to speed

[fervent thunder]

@narrow kernel help me ❤️

[narrow kernel]

the two biggest concepts to understand in the "ai" world are curve fitting and backpropagation. it's literally what it sounds like, you work your way from the end to the start and take a bunch of derivatives to determine how far to tweak all the N-dimensional tensors on the next run through(if you're training a model)

backpropagation is literally ALL derivatives pretty much. need a meme of the "always has been" astronauts. what is a derivative? going from 3d to 2d, 5d to 4d, Nd to N-1d, it helps you determine the slope of whatever n-dimensional shape you're on

[fervent thunder]

I don't know forge sorry

@fervent thunder ❤️ np

[narrow kernel]

have a 3d contour graph and want to know the instantaneous slope at any point? take a derivative at that point. what does that look like? a 2d plane touching the 3d contour

so why do you need that for training AI? because otherwise you'd be at the mercy of brute force trying every single combination of the million and one dimensions

so when you try to fit a curve on some N-dimensional data and it doesn't fit correctly, you need to tweak a bunch of knobs, but what knobs do you tweak? some are dependent on each other and by tweaking this one, you lower that one's effect and so on. so when you use back propagation, you can estimate, within a very decent level, how much to tweak all the knobs and which directions to turn them(based on the slopes) to head toward "fit," based on how much error there was. and when you set a max amount it's allowed to shift the values per step, that's literally your learning rate

so once you understand the basics of that, it should start making a bit more sense

[runic marten]

Hi

@narrow kernel Hi

@fervent thunder Hi

[narrow kernel]

but again, knowing all the math behind it isn't as important as simply just understanding the gross overview of the various concepts and the loops. the math itself is highschool level on a per component basis(some of it gets a little hairy, but we work with PCs, not pen and paper)

[fervent thunder]

someone out there has probably worked out a "minimum viable product"
for how much math is needed at minimum

[narrow kernel]

mostly just algebra. using the volume of a sphere equation i brought up, the version you use is just a simplified calculus equation. the calc version is V = integral from -r to r of pi(r^2-x^2) dx. but when you do the math, it simplifies down to 4/3pi*r^3 like you're used to using. do you need to know how to derive that? nah, you just use the algebraic version of it. unless you're reinventing the wheel, you don't really need to know the hard shit. just use what people have already figured out. this is mostly how humanity works and what set us apart from the rest of the animal kingdom: we can stockpile knowledge and pass it on.

do you really think most of these researchers are actually reinventing some wheel? nah, they are just pulling already known "tools" out of the toolbox and piecing them together. knowing which tool is right for the job is far more important.

but anyways, thanks for coming to my tedtalk lol

[uncut beacon]

hii

[fervent thunder]

DO NOT click on the banana

🤔

[fair crystal]

Thanks for the help guys. I'll start with the OpenCV Python course for now and will find out what types of math knowledge I need to brush up on

[uncut beacon]

so how to get started with stable diffusion in python?

[trail lion]

Well you would choose which interface appeals to you, there are at least 5 to choose from at this point, then you would install it

[uncut beacon]

elaborate iam completely new

[wise ginkgo]

what's the best gpu to use out of these gpus?: 3090, 4070 Ti S, 7900 xtx?

[lusty beacon]

Is there a free creative image upscaler

[bleak matrix]

Good afternoon, everyone!

How are you all doing?

[frosty bluff]

Good and you?

[bleak matrix]

Pretty good! It's a cooler day than it's been the past few days! That tropical weather was quite the fun time, to be sure

Just working on some things; what are you up to?

[frosty bluff]

Planning to teach a course on AI to my co-workers.

What tropical weather? From Beryl? You were in the path?

[low moon]

Friday.

[bleak matrix]

Oh? What's it centered on?

And yes!

I got to experience all that wonderous...heat. And...humidity.

Love the rain, though.

I'd rather be baked in 120 in the desert, in the shade of an oak tree, y'kow? Dry heat vs humidity, 110% on the dry heat, thanks!

[frosty bluff]

They want me to teach what it has taken me years of research to learn. I'm going to start with prompt engineering. I think if I start with neurosymbolic AI and horn clauses, they will become lost. 😂

Humid conditions are what I am accustomed to. We typically are impacted by at least 4 hurricanes each year. I do not know dry heat. I only know humid and soggy heat.

[bleak matrix]

laughs Well, at least they might learn some new words, eh!? Give them something to think about, to ponder on, don't you think?

Ach! Teaching is a skill--and you have to know your audience! That can be pretty hard. I certainly struggle with it, so, you're certainly not the only one~! 💖

oof! It's subtropical here, and the soggy and I? Well, I have Sjogrens (and other autoimmune stuff,) so my body + humidity are on a scale of 1-no. I rather think soggy and I aren't friends at all, and I've been repeating that phrase to myself all week, bahahaha! Everything around me is just...soggy.

[frosty bluff]

hahahaha I agree. it is a skill. A skill that is not in my repertoire. I teach computers, not people. Computers have been my audience. People...not so much. My take on it is...I am self-taught. There is a ton of information out there. They should also teach themselves. If you have any experience in teaching basic AI to humans, I'm all ears!

[bleak matrix]

People have been self taught for centuries; I consider it an advantage.

It is easier to learn with other people, though--and you're providing that to other people, which is great! Accelerate the learning all around!

In terms of teaching people, I think the basic principles of teaching apply, regardless of the subject.

Be yourself, regardless of who you are teaching. By genuinely connecting with your audience, having a plan, formulating the kinds of questions they might ask, reflecting with the audience (asking questions to see what they are understanding,) you'll find yourself becoming better at teaching.

That back and forth between student and teacher is essential, so the basic principles of human interaction will always apply when it comes to learning.

Visual + audio + writing (taking notes,) and incorporating these things into your learning/lesson will automatically make it much easier for your students to understand exactly what they're doing.

This can be done by doing object lessons (applying what by associating color with words,) and the sound of your voice + the subject, etc.

Prompting questions based on these, and asking people to relay the things they've learned during the lesson, and at the end, will help them remember the information you taught them.

Providing feedback is also part of the puzzle, as well as taking it. If you can keep yourself open to improvement, you'll always become a better teacher--no doubt about that!

[pale latch]

I've always learned that grade school and high school isn't teaching you anything important other than to give you the skills to learn and teach yourself.

the man pages have always been integral to this stuff. it used to be a matter of pride in hacker culture to have read the darkest manual pages from the most gigantic reference material. hackers movie even referenced this

https://youtu.be/4U9MI0u2VIE

[frosty bluff]

Object Lessons - how does this work with teaching prompt engineering?

[uncut beacon]

pls help

[bleak matrix]

You could do it in a wide variety of ways.

For example, bring in something easy to prompt. I don't know what country you're from, but go into a grocery store. Pick up something that's easy to prompt, like food.

Bring it. Then talk about its color. An orange might be a good example.

Talk about the differences between the color orange, prompting an orange, and using an orange color palette, as well as warm colors (like sunset colors that have orange in it.)

You might also talk about textures, like, an orange, or the SIZE of an orange, in comparison to distance, perspective, angle, etc.

There's all different kinds of things you can do. It's really just up to you.

Then pair that with the language.

[pale latch]

Hard to teach about prompt engineering since it's not engineering at all. It's a lot of voodoo

practical experience helps with it more than studying methodologies . its' good to have a framework for your explorations and develop a personal prompting style though

https://blog.fal.ai/auraflow/ this is a very strong model release but they did the same thing stability did with 3. didn't release training code and are leaving it to diffusers team to do that for the community.

gone are the days like the sdxl beta release where they had training code implemented everywhere for the beta release

[bleak matrix]

I certainly agree that it's good to have your own style; that's a part of expression, and part of the creative process. It's part of the fun, too! Being able to create is really where it's at for me, since I'm an artist, and will always be one.

[pale latch]

i've never liked the term prompt engineer since prompting seems like paint to me. wouldn't call bob ross or other oil painters a paint engineer. Engineers typically have very practical approaches to a problem

they don't find happy accidents

[low moon]

They find calculated accidents

[pale latch]

those are plans. when thye have accidents its disatrous

[low moon]

All major discoveries have been accidents

people are idiots

look at the leader of the free world

clowns

[pale latch]

i'm not sure about that. manhatten project seemed pretty determined to get the result they got

[low moon]

were all clowns we have no clue whats going on

you dont even know whats in front of yourr face

you know nothing

you dont know if youll wake up tomorrow

[pale latch]

you ok?

[low moon]

very much so

🙂

[pale latch]

because you just wnet on an insult tirade unprompted and i'm not sure why

[low moon]

not an insult my man

its the truth

its liberating

;

😉

[pale latch]

yeah i feel like you're lashing out here. this doesn't seem like a winky winky face moment

[low moon]

even tho it is :/

online moment

oh well

no one is face to face anymore so more and more gets lost intranslation

[pale latch]

well i mean, you just called me a clown who doesnt' even know whats in front of my face, but i guess if i say that's offensive i'm just thin skinned or something? Really don't know what you're trying to imply at all and why calling me a moronic clown is just fun an games.

i'll drop it . this got stupid for sure. was sort of out of left field.

[low moon]

well if u scroll up i never called you a clown

i said we are all deluded clown thinkign we know anything- includign engineeres and uhm logical people and whatever - we dont know nothign and most discoveries were mad eby accident

is what i meant

anyway

[pale latch]

if this were face to face and you said that shit, there'd be more than words being exchanged

[opal hedge]

sd3 8b wen

[low moon]

2 weeks

[opal hedge]

Is this cope or did SAI announce something

[low moon]

ah icey cold is a fighter

youd punch a lady in the face?

😄

[low moon]

ah we never know

[opal hedge]

😥

[pale latch]

ladys don't talk that way

[low moon]

oh we do now

were equal remember

[pale latch]

oh you thought i meant women. no i meant ladies.

[low moon]

Ask disney

[pale latch]

just like if a guy was acting like a goon he'd never be called a gentleman

[low moon]

Girl that says it as it is is a lady 🙂

Contradict one word i said. do you know youll wake up tomorrow for sure? wer enot most modern discoveries made by accident amd isnt the leader of the free world a clown?

[pale latch]

The blue led was created with years of dedicated research towards the goal.

[low moon]

objective empirical truth- wer eliving in a $hitshow

[rare berry]

Nuh uh

[pale latch]

subjective opinions are rarely objective truths

politics have always been a shit show you're just tuning in now

[low moon]

not this much tho

[pale latch]

history begs to differ

[rare berry]

We used to be peasants

Under an all powerful king

[pale latch]

Ever read anything written by churchill? holy bombastic.

[low moon]

well and now were all homeless and bitter under the whims of commerce and geriatric grandpa

at leats were not living in any country ending in "stan"amirite

[pale latch]

i'm canadian

[low moon]

😄

[rare berry]

You're not homeless bruh you're on discord

[low moon]

well

[pale latch]

oh wow. true colors shining through

[low moon]

i dont own my home is what i mean

[rare berry]

Muslim countries are extremely shit wdym lmao

[low moon]

i dont mean "muslim

[rare berry]

Wanna talk about a shitshow

[low moon]

but those countries are bad

objectively bad

like the craziest dictator is there in one of them uzbekistan i think

and the taliban

[pale latch]

thatsbait.gif

[rare berry]

Extremely violent area and the way they treat women is literally the worst in the world

[low moon]

yeap

[pale latch]

has to do more with economics. UAE doesn't have high crime rates

[low moon]

yes as logn as ur loaded you are ahuman being

[ruby zenith]

lol

[low moon]

otherwise you ar enothing in uae

[rare berry]

I'm talking about the Middle East in general

[ruby zenith]

ure prompt engineer not an artist

[pale latch]

i know you're generalizing. it's obvious.

[ruby zenith]

unless u physically do stuff

without ai

[low moon]

who cares about the real world tho soon we cna have our own VR heavens amirite

[rare berry]

Lmao I'm generalizing because that's 90% of the Middle East

[low moon]

lets go 😄

What will your VR AI heaven look like?

[rare berry]

I would live in the matrix

[ruby zenith]

comparing it to bob ross

wonderful, just can;t

id say you can call yourself ai artist

but not an artist

[pale latch]

i'm not sure who are more reprehensible. the thinly veiled racists or the right out there "no true scotsman" type people. i suppose both groups are worth being prejudice towards.

[low moon]

Only happy accidents.

Anyway, lady whistles Wizard! You're the popcorn man, have at it. I am off to South Korea. Coz it's Friday.

[ruby zenith]

sad af

[low moon]

annyeong!

Soju > sake

[bleak matrix]

I'm having a little break; a nice snack

Strawberries and bananas

Let's all be nice to each other, eh?

[low moon]

As long as no pineapples end up on pizzas I am ok with that

🍕

[loud solar]

We are always nice ... the most of us ^^

[nova lily]

You guys ever start generating one day, using the same model, the same prompt setup that was great another day and your just getting crap and your like something must be broken

[low moon]

all the time

[loud solar]

It happens ...

[rough needle]

does anyone know why comfyUI works so much faster than a1111?

like its atleast 15x faster for me

[low moon]

me 2 lol

[rough needle]

like i love the speed on comfyUI but i LOVE the controll on a1111

[low moon]

A1111 is no very optmised

Foge was better

and even Fooooocus

Forge

oh but you have way more contorl i cmfyui!

🤯

[rough needle]

is there like an inpaint like on a1111 on comfyUI? i havent found one i like yet.

[low moon]

I havent use dit but I know there is

[rough needle]

the one i use now uses SAM which im not really a fan of

[low moon]

Yes Sam gave me issues with hand detection XD

[rough needle]

and its tedious

having to click every where, i like to draw where i want it to redoo

[low moon]

ohhh you just want plain inpaint

[rough needle]

yea

[pale latch]

i hate doing custom masks in comfyui. you either use a model to automatically create the masks for you, or you hand paint them in another program, save them, then import the file into comfyui

[rough needle]

cuz i make so many different pictures

[pale latch]

i had a canvas node installed at one time that sort of worked and let me draw masks, but it wasn't that great and when i updated it broke so i just removed it

[low moon]

Well you can always inpaint in photoshop any mask you want and loa dit into comfy

if you dont liek the automatic selectors

i only make my masks externally for the light model

ic light that is

[rough needle]

i dont have access to photoshop sadly

to expensive atm

[low moon]

there are alternatives

krita?

[trail lion]

or gimp

I actually prefer gimp over photoshop tbh

[fervent thunder]

darktable is good

[pale latch]

https://www.getpaint.net/download.html always a fan favorite

[pine fiber]

did u guys see that

llama3-405b will be able to generate images

which means its 100x bigger than sdxl

[narrow kernel]

Krita

Use it, it works great

[pale latch]

i'd love to read about that if you got a link

[pine fiber]

it was reported by theinformation

[pale latch]

ah. paywalled

[mortal delta]

random question but do any of you pay for ai tools or do you try to do everything local?

[ornate flame]

there's a lot of both

[primal shuttle]

Dammit. Why does Juggernaut X insist on making pantsless hermaphrodites. I am specifying pants. I want pants.

Nprompting NSFW, too.

[primal shuttle]

Do you remember back when you could buy commercial programs and run them locally? Pepperidge Farm remembers.

[mortal delta]

can you name some examples are we talking about software that is pay once or subscriptions?

[primal shuttle]

Runway never gets updated, but I've found it essential for my workflow. Their outpainting is both garbage and the best out there. I regularly outpaint with runway and then reprocess through stable to blend it.

And the new Gen 3 TXT2VID is uncannily realistic.

[mortal delta]

dosnet runway have a credit system? i heard some platforms you pay a certain amount for unlimited use or commercial use.

[primal shuttle]

Yes. Their best option is steeply priced at 90/mo but I still shell out for it.

And with that option you get unlimited txt2vid.

and IMG2vid.

[mortal delta]

jeez that alot of money.

[primal shuttle]

Be aware, too, that they are very thenthitive individuals on that team. Highly censored, very prudish, still worth the cash.

Seriously, tho. Where are muh pants!

[mortal delta]

i try my best so far to use free tools but they dont look as a good as what ive seen other use with paid tools, someday im afraid i might have to pay.

[primal shuttle]

I still refuse to use anything paid from stability.

[mortal delta]

like your asking for just pants or a person with pants?

[primal shuttle]

I am asking for a person with pants and it's giving them a prominent and slitted... member... anyways. Deeply disturbing stuff.

It's giving my character buttless chaps. Those don't look good on anyone.

[mortal delta]

thats something... im kinda nooby but im sure someone else could help.

[pale latch]

is that sensitive with a lisp? Why you speaking in code ?

[primal shuttle]

Yeth.

[pale latch]

because they're from barcelona?

[primal shuttle]

Don't make fun of my lithp. Ith embarathing.

[pale latch]

people don't type lisps. that's why it's still spelled barcelona

[primal shuttle]

Not when I do it, it's not.

[mortal delta]

i have dyslexia and autism so that affects all forms of spelling for me.

[pale latch]

hmm. i don't believe you have one and still think you meant it as some coded phrase

[primal shuttle]

OMG ACTUALLY SAME. We're twinsies.

[mortal delta]

i dont know how to reply to that, but can we all move on or something?

[primal shuttle]

Pants. I got pants. With a wide fleshtone zipper flap. There is progress.

[novel stump]

Is it possible to invite the stable diffusion bot into my own discord and use it there?

[desert dagger]

if you use the API, you can make a discord bot that calls it

[novel stump]

wonder if I could put one in a kubernete

[clever vault]

is stable diffusion 3 a model?

[pale latch]

its a group of models . 2b parameter model is out right now. the 8b parameter model is on the api

[low moon]

just put "nudity"in the neg prompt and everything will be fine XD

[lyric jasper]

How to generate image in this server?

[low moon]

artisan

[uncut beacon]

Hi
Iam back

Again

[devout salmon]

Hello, I am Millie

[uncut beacon]

Ok

I need a gpu to run stable diffusion?

[hushed lynx]

hi im quite new here im looking for help maybe someone will be able to help me. I want to generate a image showing one of judo technique i have a photo of this technique i tryed open pose but it wont fork for 2 people. What im doing wrong ?

[deft quartz]

whats the promot to generate images please

[deep salmon]

hi, what can i do to edit my pic in ps2 style?

[grim holly]

Prompt:-Create highly detailed hyper realistic A digital art piece of a spider man holding a bouquet of flowers in front of the Taj Mahal agra.ratio of 9:16 high resolution, clearer potrait,with black and white background and model is colour full with good lighting glow

[stoic spire]

今天的刮刮乐中奖号码

[void jasper]

is there any AI tool/software/addon that scans anime images and calculate the cup sizes?

[outer crescent]

#🆕｜sd3 message

[outer crescent]

well you can use blip to interrogate an image and i guess if you fed it through an llm it would be able to make an educated guess like teacup, mug, espresso cup

[void jasper]

🤣 thats not what i mean with cup.. i meant brah sizes. like A , B , C , D, E , F ... (that should be safe for work, cause its just math -> returning a string)

you made my day, i print your message and put it in a frame on my wall

[outer crescent]

Go with god. 🙏

but seriously, you can interrogate images, but it can only really return what it is trained on and i doubt people who draw animes write specifics like that in for context

[void jasper]

na somehow the stable diffusion if you do the deboru scan can detect "small breast, medium breast ...." just need it more in detail 😄

i am honestly supriced i cant find anything liek that online

[outer crescent]

ok, despite your question being silly and inane, im now wondering if theres an equivilent to lora for BLIP.

OH! llava has!

so you actually could train a lora for llava that would teach cup size. all the cups, americano, cappucino, ceremonial.

[void jasper]

just need a smart person who code an ai, that returns body part meassurements 😄

[outer crescent]

https://medium.com/ubiai-nlp/how-to-fine-tune-llava-on-your-custom-dataset-aca118a90bc3 here, if it gives purpose to your life

[void jasper]

could be helpfull in daily life too like shoe sizes

but yeah my pc is to weak to train anything with ai

only have 4gb vram

😢

[outer crescent]

colab would probably let you, theres probably a little jupyter book somewhere.

[void jasper]

but i cant believe no one ever had the idea to make an api/ai that returns body part sizes. that has to exist somewhere already.

so i dont have to reinvent the wheel. just find it

and use that api then

[outer crescent]

lazy as heck good buddy. you asked an absurd question, and i somehow managed to give you the process to manifest such a silly thing. you just dont wanna do work. i say do it, then you can have a little project and be forever known as the teacup person, friend to all

[void jasper]

yeah, you gave me a complicated answer, i dont know how to do or even start.

i expected something simpler to be honest

[outer crescent]

can i have something that doesnt exist?
yes, here is how it can exist.
unacceptable! i want the thing that doesnt exist to already exist.

that's where we're at. 😄 ❤️

[void jasper]

i asked reddit, lets see

if it doesnt exist, for sure, then maybe can start working on it and become rich :3

[outer crescent]

rich, rich from all the open source money you make creating loras

😉

[mossy river]

What's the best non-pony XL model for generating characters?

In y'all's opinions

[golden cypress]

hi everyone, ive tried to find the answer elsewhere but i have to ask for help. Im trying to generate a D&D character portrait. I have specifics for different pieces of clothing and features but it ignores some of them in the prompt. Ive tried taking one generated image and fixing the missing features, but it always comes out worse. Is there a way to build "pieces" and them put them together? For example, the prompt words im trying to get together are: muscular orc, turquoise skin, beard, bald head, green tunic with a popped collar, black-and-white pinstripe pants. TIA if anyone could guide my playing-around or tell me what to google

[fossil rune]

guys is there such thing as the ReActor but for short form videos?

[jade hazel]

1

one

[fervent thunder]

won

[desert dagger]

what is a 'popped collar"?

[golden cypress]

Vertical, not folded down

[desert dagger]

so how many images in the AI's data training set do you think it saw labeled with 'popped collar' and how many with "upturned collar"? And it knows that orcs are green, it's going to be obstinate about that

[novel raft]

hey

i am just starting ot get into stable diffusion, still figuring out what to download and how to use it

[desert dagger]

you might want to make sure your machine will run it first

[novel raft]

how can i check?

[desert dagger]

what sort of machine do you have?

[marble coral]

I'm a beginner and I'm playing with A1111, and I downloaded this XL model https://civitai.com/models/133005/juggernaut-xl to try to generate text. the letters are mostly fine, but the words are always nonsense

what's the best way to fix the text? I found some suggestion online about using image editing software like Canva, but you have to pay for them. Ideally I could have something automate in A1111 itself

[novel raft]

AMD Ryzen 5 5600X 6-Core Processor, 32.0 GB ram, AMD Radeon RX 6600 gpu

[desert dagger]

use SD 3 - the other versions of stable don't have the comprehension necessary for good text

[desert dagger]

okay, start by going to google and searching on "run stable diffusion on AMD radeon" and start reading through the hits

[marble coral]

but I saw that even SD3 doesn't get the text 100% right (https://stable-diffusion-art.com/stable-diffusion-3-local/#Text_generation)

is there no easy way to fix incorrect text?

[trim magnet]

on sdxl theres the harrlogos lora you can get decent text but its not perfect,on sd3 u can also get better text gens but is also not perfect,even dalle is not perfect text so u either gen lots of images til u get a good one or fix bad text on a img editor

[desert dagger]

sd3 2b (medium) which is what's currently release doesn't do the best job, but it does a decent job if you just use the t5xxl text encoder instead of using all 3 of them and keep the text fairly simple

[novel raft]

i am trying to check if my gpu can run it but im getting mixed messages

[desert dagger]

https://opendata.blender.org/ this is a benchmark site. granted, blender put it up, but it's a good benchmark site regardless. download the benchmark tool and run it on your machine.

[marble coral]

@trim magnet @desert dagger thanks! I'll play a bit more with sd3. I also found a paid online service that supposedly fixes text, I'll see what works best

[novel raft]

running it hoq do i know what is a good score?

[desert dagger]

it's eaiser, in my opinion, to just add the text with photoshop in post production.

[desert dagger]

read through the benchmark site.

[undone garden]

Do you guys have resources to point me to to diversify my prompts on subsequent generations without having to micro manage every single one? Like, stringing them together reliably with minimal to moderate variations

[slender vault]

sd on amd kinda sucks. you need A1111-ML
if you want "okay" performance you need to run linux so u can utilize rocm

[slender vault]

controlnet is gonna be your friend for this

[warm junco]

your gpu can run SD, you need to follow my Automatic1111 AMD Zluda guide to get the best performance out of it on windows

[latent stump]

does SD run on OpenCL?

has anyone tried dopencl or vopencl to "merge" multiple servers into one by running an remote OpenCL instance? I wonder if it would increase the performance or degrate it.

3 years ago I tried both of them but not for SD, one would memory leak so I had to restart it daily, forgot which one... I also had troubles with some libs supporting only 64 GPUs too.

[tepid flower]

Anyone come up with a good prompt for chatGPT/Claude to improve upon AI image prompts?

[rigid hedge]

hey, when training a LoRa for SD1.5 would you recommend that dataset is exclusively 512,512 or doesnt it matter anymore. I usually generate in 512,768 or vice versa. I know kohya_ss uses the buckets so it should be okej to use mixed ratios right?

[desert dagger]

that's what i'd do

[spice zenith]

"/imagine prompt: A delicate baby pink rose with soft petals, with a sparkling diamond emerging from its center, representing elegance and strength"

[novel raft]

@warm junco @slender vault alright so just download 1111?

[placid hatch]

Are there any good open pose editors that aren't part of an automatic1111 extension?

[slender vault]

cs is gonna be more helpful than me since i bothered him for months to get all my knowledge.

[warm junco]

you find my install guide in the pinned messages of #🤝｜tech-support

[rigid hedge]

anyone tried perplexity.ai , best god damn thing since sliced garlic bread

[fervent thunder]

ad?

[rigid hedge]

oh no, just a friendly tip. I recently discovered it and I love it 🙂

[zenith summit]

Just wondering, if anyone else has had this issue. When I use the 1.5 LCM loRA with the correct settings its slower than a standard generation. Thats odd right?

[pale latch]

bucketing was invented for sd1.4 by novelai and it works great. just make sure that you have few as possible buckets. if there are buckets with only 1 image in them, try to crop them so they fit into antoher bucket

[ashen flower]

Anyone know if you can get ControlNet webui extension working on amd gpus? (I've managed to get stable diffusion 1.5 working with zluda)

[trail lion]

yes, I run controlnet on xl and previously on 1.5. I'm a linux user, but that shouldnt really matter if you have inference working already in windows

[desert dagger]

auto1111 or comfy?

did you put it in the right directory? has it finished downloading all the way?

and do you have an SDXL model loaded?

you can't load a lora that's for SDXL if you don't have an SDXL model

https://civitai.com/models/122822/crystal-clear-xl this one's good

[rigid hedge]

crystal clear is fantastic, agree

[desert dagger]

sure. basically you put a set of images together, you then use a trainer (or write one). However it's quite a bit more complicated than that. I would suggest you start by watching some youtube tutorials on how to train loras for stable diffusion

okay. deep breath. i use this one. it will train a LoRA to be used with SDXL https://replicate.com/zylim0702/sdxl-lora-customize-training pull it up, take a look at it, and then you're welcome to DM me if you'd like me to walk you through using it

i can only walk you through the one I've used

[short cipher]

hamburguer

[frail rock]

Oi

[neon hawk]

🙂

[gleaming sage]

Hey

[stone oxide]

One message removed from a suspended account.

One message removed from a suspended account.

[night leaf]

Cómo puedo acceder al bot de Stable Diffusion?

[stone oxide]

One message removed from a suspended account.

[mellow geyser]

Hello!Can you please tell me what is the best way to swap faces nowadays?

[fervent thunder]

hello!!

Ive spent like the last 500 hours addicted to SD!!

I have amd gpu and training loras seems like it will be a technical challenge, I tried already and installing requirements breaks rocm

anyone here make loras? I am really hoping to see a dark syde phil lora, so I can really share with the world where my mind can go

[hidden dagger]

I don't study nor use Lora. I mostly spend time on how to use SD model for every possible downstram task in medical imaging

SD model is like a great part of a lego house in DL

[warm junco]

Hey, you can train Loras on AMD with OneTrainer

[latent stump]

not an ad - it is good tbh, but huggingface.co/chat is great too imo. they're both rather popular

[fervent thunder]

is that on github

[fervent thunder]

does anyone here make models?

or is there a better discord to find those things

[fervent thunder]

that looks great, thank you

whats the legal implications of turning a public figure like a youtube streamer into a lora

[fervent thunder]

Minimum to none in the United States, as far as the courts are concerned.

It's nearly impossible to win a case like that in reality. But there are often other factors outside of court that apply.

[fervent thunder]

win as in suing me?

well they sure got my reddit account banned for making kamala harris twerking

I was accused of sharing someones lewd without their knowledge, so now im wondering if im having my creativity suppressed due to political reasons or if im in the wrong

[desert dagger]

same as if you used their likeness without their permission for anything else. they can take you to court, and probalby sue you for everything you have.

[fervent thunder]

I would look at terms of service violations in social media. Legality is a much more gray area with these things

for context kamala harris said she was "out in these streets" for voters

[fervent thunder]

great, so maintaining poverty will keep me honest just as the great philosophers have said

if they want to shut me up they will have to go the other way, bribery lol

[desert dagger]

ask yourself "would I like it if someone used MY face without my permission?" - go get their permission

[fervent thunder]

if I was a public figure I would understand parody

its protected

[desert dagger]

sure. but would you like it?

[fervent thunder]

As a former attorney I would advise you not to get your legal advice on the internet ya know

im not just making lewd images of her to fatansized. It was clothed

[fervent thunder]

if I accept the advice does it mean you now represent me