🍥｜anime

[final river]

I will do foxgirl holding world sphere to prove superiority

[supple raptor]

Maybe I do foxgirl holding catgirl world sphere

[wind girder]

the current SDArt model https://civitai.com/models/44230, the first I trained on a dataset of 600+ pictures, is already working quite good, lots of diversity, quality mikus, but lacking subjects because the PoW are restricted
But I'll be honest @supple raptor . I have no hope of rivaling any anime model. I'm doing a generalistic artistic model, and anime model 100% focused, with the expertise and the tagging system this community has... I can't be on the level tbh

[cyan burrow]

hmmmmm
Loading weights [531b0f060c] from G:\STABLE_DIFFUSION_VARIANTS\AI_AUTOMATIC1111(IMAGE)\models\Stable-diffusion\Dark_Gemini\DG3_Full_V1.FP16.safetensors changing setting sd_model_checkpoint to Dark_Gemini\DG3_Full_V1.FP16.safetensors: SafetensorError Traceback (most recent call last): File "G:\STABLE_DIFFUSION_VARIANTS\AI_AUTOMATIC1111(IMAGE)\modules\shared.py", line 483, in set self.data_labels[key].onchange() File "G:\STABLE_DIFFUSION_VARIANTS\AI_AUTOMATIC1111(IMAGE)\modules\call_queue.py", line 15, in f res = func(*args, **kwargs) File "G:\STABLE_DIFFUSION_VARIANTS\AI_AUTOMATIC1111(IMAGE)\webui.py", line 149, in <lambda> shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights())) File "G:\STABLE_DIFFUSION_VARIANTS\AI_AUTOMATIC1111(IMAGE)\modules\sd_models.py", line 496, in reload_model_weights state_dict = get_checkpoint_state_dict(checkpoint_info, timer) File "G:\STABLE_DIFFUSION_VARIANTS\AI_AUTOMATIC1111(IMAGE)\modules\sd_models.py", line 262, in get_checkpoint_state_dict res = read_state_dict(checkpoint_info.filename) File "G:\STABLE_DIFFUSION_VARIANTS\AI_AUTOMATIC1111(IMAGE)\modules\sd_models.py", line 241, in read_state_dict pl_sd = safetensors.torch.load_file(checkpoint_file, device=device) File "G:\STABLE_DIFFUSION_VARIANTS\AI_AUTOMATIC1111(IMAGE)\venv\lib\site-packages\safetensors\torch.py", line 99, in load_file with safe_open(filename, framework="pt", device=device) as f: safetensors_rust.SafetensorError: Error while deserializing header: HeaderTooLarge

[supple raptor]

I interpret this as a complaint about the filepath being too long. Shorter filename could help if that is especially long.

Disclosure: I do not know how to code. I'm just good at guessing.

[cyan burrow]

but it's not

or maybe it doesnt like the total length

[supple raptor]

Total length would be my guess. As in the full file path, shortening the file name is just the easiest potential solution if that is the problem.

[cyan burrow]

hmmm

very strange

Loading weights [531b0f060c] from G:\AI_Apps\Auto1111\models\Stable-diffusion\DG3_FP16.safetensors changing setting sd_model_checkpoint to DG3_FP16.safetensors: SafetensorError Traceback (most recent call last): File "G:\AI_Apps\Auto1111\modules\shared.py", line 483, in set self.data_labels[key].onchange() File "G:\AI_Apps\Auto1111\modules\call_queue.py", line 15, in f res = func(*args, **kwargs) File "G:\AI_Apps\Auto1111\webui.py", line 149, in <lambda> shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights())) File "G:\AI_Apps\Auto1111\modules\sd_models.py", line 496, in reload_model_weights state_dict = get_checkpoint_state_dict(checkpoint_info, timer) File "G:\AI_Apps\Auto1111\modules\sd_models.py", line 262, in get_checkpoint_state_dict res = read_state_dict(checkpoint_info.filename) File "G:\AI_Apps\Auto1111\modules\sd_models.py", line 241, in read_state_dict pl_sd = safetensors.torch.load_file(checkpoint_file, device=device) File "G:\AI_Apps\Auto1111\venv\lib\site-packages\safetensors\torch.py", line 99, in load_file with safe_open(filename, framework="pt", device=device) as f: safetensors_rust.SafetensorError: Error while deserializing header: HeaderTooLarge

again

this makes no sense

since every other model loads

[fading sage]

Change the filename extension from safetensros to ckpt

[cyan burrow]

why?

[fading sage]

Works for some people. Otherwise try downloading the model again and doublecheck it's the correct version

[cyan burrow]

It's not a download, it's my model

I'm preparing it for upload

but FP16 conversion has failed

[fading sage]

Very interesting then

[fading sage]

Using controlnets?

[cyan burrow]

at what point?

[queen void]

its time to cook again

[zinc wigeon]

How to run stable diffusion without downloading on mac? Can anyone help me with the process

[supple raptor]

Have you tried deleting your fp16 pruned version and pruning it again?

[cyan burrow]

Calculating sha256 for G:\AI_Apps\Auto1111\models\Stable-diffusion\DG3_Pruned.FP16.safetensors: 097b8b6df5b9cb2d3ee96827a51474e7334b68d50f1fdb451b83f2fd5d0dbb21 Loading weights [097b8b6df5] from G:\AI_Apps\Auto1111\models\Stable-diffusion\DG3_Pruned.FP16.safetensors changing setting sd_model_checkpoint to DG3_Pruned.FP16.safetensors: SafetensorError Traceback (most recent call last): File "G:\AI_Apps\Auto1111\modules\shared.py", line 483, in set self.data_labels[key].onchange() File "G:\AI_Apps\Auto1111\modules\call_queue.py", line 15, in f res = func(*args, **kwargs) File "G:\AI_Apps\Auto1111\webui.py", line 149, in <lambda> shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights())) File "G:\AI_Apps\Auto1111\modules\sd_models.py", line 496, in reload_model_weights state_dict = get_checkpoint_state_dict(checkpoint_info, timer) File "G:\AI_Apps\Auto1111\modules\sd_models.py", line 262, in get_checkpoint_state_dict res = read_state_dict(checkpoint_info.filename) File "G:\AI_Apps\Auto1111\modules\sd_models.py", line 241, in read_state_dict pl_sd = safetensors.torch.load_file(checkpoint_file, device=device) File "G:\AI_Apps\Auto1111\venv\lib\site-packages\safetensors\torch.py", line 99, in load_file with safe_open(filename, framework="pt", device=device) as f: safetensors_rust.SafetensorError: Error while deserializing header: HeaderTooLarge

[cyan burrow]

yyeah I gotta do that next

[supple raptor]

Without downloading, your best option is to use a google colab. This is the one that I would personally recommend, though I don't have time to offer much more help than the link right now. https://colab.research.google.com/drive/1wEa-tS10h4LlDykd87TF5zzpXIIQoCmq

[zinc wigeon]

Thanks I will try this

[supple raptor]

If you have some idea of what you're doing I think it's mostly straightforward, but might be confusing if this would be your first time setting up SD. I'm sure some other people around here have used NoCrypt's Colab before so if you have to ask for help, someone will probably be able to assist you.

[cyan burrow]

@supple raptor calculating sha256 for G:\AI_Apps\Auto1111\models\Stable-diffusion\DG3_FP16_Pruned.safetensors: ************************************ Loading weights [cca7de46ce] from G:\AI_Apps\Auto1111\models\Stable-diffusion\DG3_FP16_Pruned.safetensors Loading VAE weights specified in settings: G:\AI_Apps\Auto1111\models\VAE\kl-f8-anime2.ckpt Applying cross attention optimization (Doggettx). Weights loaded in 3.8s (calculate hash: 2.7s, load weights from disk: 0.1s, apply weights to model: 0.7s, load VAE: 0.2s).

new pruned one works

[supple raptor]

Cool, must have just been an error that happened when pruning.

[south sable]

🤔

[zinc wigeon]

I am expecting an assistant on the same, this is the first time I am using Stable diffusion

[cyan burrow]

and now how to properly convert to FP16

[supple raptor]

IDK what she's holding, but it doesn't seem safe to use that while riding a motorcycle. Maybe you should give her a helmet

[south sable]

I was thinking this looked dangerous...

[cyan burrow]

Uuuh I think I might be baking a VAE into it, unintentionally or maybe not. It loads it before it converts it

maybe it doesn't matter

[south sable]

https://tenor.com/view/ive-always-loved-baking-vincent-the-great-canadian-baking-show-i-love-baking-i-love-to-bake-gif-24095250

[cyan burrow]

Loading weights [cca7de46ce] from G:\AI_Apps\Auto1111\models\Stable-diffusion\DG3_FP16_Full.safetensors Loading VAE weights specified in settings: G:\AI_Apps\Auto1111\models\VAE\kl-f8-anime2.ckpt Applying cross attention optimization (Doggettx). Weights loaded in 1.1s (load weights from disk: 0.1s, apply weights to model: 0.7s, load VAE: 0.2s).
@supple raptor and again

both working

Calculating sha256 for G:\AI_Apps\Auto1111\models\Stable-diffusion\DG3_Pruned.FP32.safetensors: 100d5ea7f35e12c9284b80215451d0ac8fd6e84b9ed337944ace2ada7b8f6ef0 Loading weights [100d5ea7f3] from G:\AI_Apps\Auto1111\models\Stable-diffusion\DG3_Pruned.FP32.safetensors changing setting sd_model_checkpoint to DG3_Pruned.FP32.safetensors: SafetensorError Traceback (most recent call last): File "G:\AI_Apps\Auto1111\modules\shared.py", line 483, in set self.data_labels[key].onchange() File "G:\AI_Apps\Auto1111\modules\call_queue.py", line 15, in f res = func(*args, **kwargs) File "G:\AI_Apps\Auto1111\webui.py", line 149, in <lambda> shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights())) File "G:\AI_Apps\Auto1111\modules\sd_models.py", line 496, in reload_model_weights state_dict = get_checkpoint_state_dict(checkpoint_info, timer) File "G:\AI_Apps\Auto1111\modules\sd_models.py", line 262, in get_checkpoint_state_dict res = read_state_dict(checkpoint_info.filename) File "G:\AI_Apps\Auto1111\modules\sd_models.py", line 241, in read_state_dict pl_sd = safetensors.torch.load_file(checkpoint_file, device=device) File "G:\AI_Apps\Auto1111\venv\lib\site-packages\safetensors\torch.py", line 99, in load_file with safe_open(filename, framework="pt", device=device) as f: safetensors_rust.SafetensorError: Error while deserializing header: HeaderTooLarge

Well I know what to do, lol

[cyan burrow]

@supple raptor Calculating sha256 for G:\AI_Apps\Auto1111\models\Stable-diffusion\DG3_FP32_Pruned.safetensors: *************************************** Loading weights [0fd16b71d2] from G:\AI_Apps\Auto1111\models\Stable-diffusion\DG3_FP32_Pruned.safetensors Loading VAE weights specified in settings: G:\AI_Apps\Auto1111\models\VAE\kl-f8-anime2.ckpt Applying cross attention optimization (Doggettx). Weights loaded in 5.3s (calculate hash: 3.5s, load weights from disk: 0.1s, apply weights to model: 0.8s, load VAE: 0.9s).

and again!

btw have they fixed xformers for 2.0?

[mild tree]

Oh my...

[opal dome]

any gamers have experience doing SD + Live2D?

[cyan burrow]

time to test all 4 dark gemini v3 on all samplers I have (12 II guess and the response is def that cat stare

like "wtf human doing?"

[desert arch]

Greetings

[brittle jay]

model?

[quick hill]

Aurora

[brittle jay]

nice

[queen void]

Well first LoRA is done after coming back from the abyss. Any prompt I should try? :D

[quick hill]

is it a style lora?

also wb

[desert arch]

Crocodile top hat

[supple raptor]

I'm running it with the ux fork and it's running well so far.

[untold glacier]

Try my LoRA testing prompt: 1girl, close-up, brown hair, short hair, tank top, sitting, city street in background, traffic light, road sign, power lines

[queen void]

yes

My first test prompt

[quick hill]

mature, 1girl, solo, blonde hair, short hair, blue eyes, choker, flower, short hair, realistic eyes, head tilt, smug, looking_at_viewer, sky, mountain, outdoors, cloud, bangs, upper body, pink flower, collarbone, blouse, parted lips, lips, lace trim, smile, black choker, mountainous horizon, cloudy sky, standing, holding_flowers

[cyan burrow]

she dressing up, she sexy yeah?

[quick hill]

yeeeeeeeeeeeeeeeees

[cyan burrow]

nice, I should start using it again. Is it just plug and play again?

and did auto finally abandon the repo?

[queen void]

my model doesnt rly like the prompt xD Even without LoRA

[south sable]

is there anyone else not seeing LoRAs in the normal tab? seems the option has dissapeared?

[untold glacier]

What model you using?

[quick hill]

hmmmm i wonder what style is that

[queen void]

An artist style :)

[vagrant falcon]

not even listening to half the prompt, i mean yeah, but..

xD

[cyan burrow]

imma go buy ice cream

[untold glacier]

I've found it to be an effective LoRA tester? Even among different models. Which model are you using?

[vagrant falcon]

is there any sharingan, loRA?

available

[quick hill]

https://civitai.com/models/13945/sharingan-eyes-inpaint-lora

apyr made it

[vagrant falcon]

sheeesh

[queen void]

it was with Kenshi. I just need to try more models. I have more than plenty of models.

[quick hill]

https://civitai.com/models/5800/bocchi-ed-style-lora

also this so cute

[queen void]

Only one of my loras is in that image

[untold glacier]

These are the artist style LoRAs. I have different categories. Also, the vast majority of these are still ones that I've been working through from the list that Nubb gave me awhile back. Finally near the end. Will move on to Civitai after that. I know you have many awesome ones that I can't wait to try.

[south sable]

anyone can help with the missing LoRA issue? I don't see them in their usual spot anymore

[supple raptor]

Pretty much. Once you have torch 2.0 installed you just have to activate your venv and run pip install -U --pre xformers The latest version also seems to be deterministic which is a nice little bonus.

[cyan burrow]

awesome, gonna do that after my xyz is done

[south sable]

🆘

[cyan burrow]

oh anyone else concerned about the latest contest? Considering the legal precedent currently set is that people don't own whatever is generated by AI? So how does that work with the prize money etc?

[unreal ridge]

Merging Models is hard
it's been 1 day already 😔

[cyan burrow]

that X isn't very informing

[untold glacier]

Reboot everything. Check that the files haven't been deleted or moved.

[final river]

https://www.federalregister.gov/documents/2023/03/16/2023-05321/copyright-registration-guidance-works-containing-material-generated-by-artificial-intelligence

https://www.govinfo.gov/content/pkg/FR-2023-03-16/pdf/2023-05321.pdf

[untold glacier]

Pretty sure you asked a yes or no question and that was given as a response.

Also, legal precedent has only been made with Midjourney. SD is a different beast in my mind.

[cyan burrow]

yes but...to me that's not a response

maybe it's cause I'm a boomer and don't talk through emoji

[untold glacier]

Maybe, heh.

I've found it to be very useful in conveying emotion when words fail.

[cyan burrow]

okay so SD diferent, maybe

[vagrant falcon]

thanks 🙂

[untold glacier]

Nice.

[cyan burrow]

I'm sorry but that's way too much for my brain to handle right now. It's affected by a rather unhealthy state of my body so I gotta go for the TL;DR version

if you don't mind

[untold glacier]

Tbf it's a complicated subject.

[supple raptor]

Not sure what your concern is? You think the winners won't be paid, or that they don't have the right to keep the funds if they don't own a copyright for their work?

[final river]

TL;DR: You can copyright AI generated things depending on how much work you actually put into it.

[elfin bolt]

if you mix multiple styles and deliberately guide the artwork to certain intent, it can be copyrighted

if you just use whatever the ai spews out then no

[cyan burrow]

Getting drawn into a legal battle and so on, basically having to deal with annoying shit

[untold glacier]

I think the issue there is what is considered "work". Many people don't consider prompting an artform.

[final river]

Doesn't matter what those many people think.

[untold glacier]

It does if they rule on the law.

[supple raptor]

There's nothing there to be drawn into. You don't need to hold a copyright to sell a your work. You only need a copyright to protect your work from being duplicated or used by someone else without your permission.

[final river]

If human creativity is involved in shaping most of the work, like... prompting specific things in an image for example.

[cyan burrow]

doesn't that mean EXCEPT for whatever remnant remains? Like some frankensteins monster thing

[elfin bolt]

as long as you remix it should be fine

[final river]

*To sue/take someone to court for it.

[supple raptor]

There is currently no legal precedent or even an attempt at an argument for one to ban selling purely AI generated content. The battle is around whether or not you can protect your AI generated work.

[supple raptor]

Yes, which is what I meant by "protect your work".

[final river]

Mm.

But I was being more specific.

Because all artwork has a general copyright per se.

Even if it's not registered with an official record.

It can still have copyright protection, but you can't take someone to court/sue for it unless it's a documented/registered copyright.

https://www.legalzoom.com/articles/how-to-copyright-artwork

[cyan burrow]

Thank you for all the information, it wasn't my intention to turn this into a debate. I was only concerned about being dragged into something annoying at this stage of my life

[quick hill]

all good

[supple raptor]

Basically at the end of the day, the only reason that copyright matters is to prevent other people from using your work. Without copyright protection you are more than welcome sell as much of your AI content as you want, but you can't do anything about it if someone else decides to start selling your work as well.

[elfin bolt]

i'd say its better to go wacky like that balenciaga harry potter guy and open patreon if u want some side hustle doing ai generated content right now

The legal status is still pretty much on the air rn

[zinc flax]

Be sneaky and sell the prompts instead

[unreal ridge]

Or model

[zinc flax]

And then get it leaked

[unreal ridge]

kek

[elfin bolt]

and get cancelled by twitter

[hollow storm]

copyright cring public domain good

[untold glacier]

what

[unreal ridge]

tbf, you can get cancelled by twitter for pretty much anything.

[supple raptor]

Yes, I'm not saying that selling AI art is an especially good business model at the moment, just that it is not illegal to do so whether or not you hold a copyright for it. The problem is that if your work becomes popular as an AI artist, someone with better exposure than you absolutely will eventually take your art and start selling it on their own.

[elfin bolt]

https://www.youtube.com/watch?v=iE39q-IKOzA&ab_channel=demonflyingfox

[final river]

Unless you register a copyright for said work.

[zinc flax]

Tracing AI art soon™️ in 2025

[supple raptor]

Which you can't do if it is purely AI generated.

[hollow storm]

thers alrdy a ton of bot accounts reposting popular works on other platforms
really isnt all that new

[final river]

Yeah. But if you put work into it then it isn't purely ai generated.

If you type in "a girl riding a bike" and hit generate you can't copyright that.

[elfin bolt]

The amount of work isnt exactly quantifable

i do use controlnet alot and inpaint

[final river]

It is.

[elfin bolt]

but i dont think it can be copyrighted

[hollow storm]

right, but reposting generated stuff imo is at least better than reposting real artists work without credit

[final river]

It's really not hard to understand.

If you use like one sentence, hit generate, and don't do anything else to it. You can't copyright it.

If you think of the clothes, accessories, background, inpaint, outpaint, etc. all for one image, you can register it for a legal copyright.

You can copyright it.

[untold glacier]

Hence the Midjourney AI ruling.

[hollow storm]

ok but what if u use like 20 sentences and click the button 20 times with inpainting

how about twice

[zinc flax]

The AI lore deepens

[elfin bolt]

what if u photobash it? does it counts?

[final river]

How much of your actual thought process/creativity was involved in the output of the image.

[elfin bolt]

its still messy

[untold glacier]

There's even more work that goes into it by default though. All of the settings can be changed, in order to get the best result that you're looking for.

[hollow storm]

what if i just use the ai image but i draw my signature on it which i came up with myself
its still very case by case imo

[supple raptor]

As far as we know, unless you use tools outside of generative AI, then it flat out isn't currently eligible for copyright in the US. There is nothing in the wording to indicate that a more specific prompt will enable you to receive a copyright.

[cyan burrow]

midjourney are also very nonchalant about how they get their images for training

[hollow storm]

the training part is most likely under fair use

[south sable]

yay I got my extensions back. musta been a glitch.

[untold glacier]

As far as I've read and seen, there's nothing to indicate that it isn't eligible depending on the complexity of creation though, right?

[supple raptor]

Shutterstock watermarks are all over SD training material too.

[final river]

.

Human more involved in the final work? Yes. AI doing most of the creative work for the elements in an image/the final composition? No.

[cyan burrow]

they don't care if they take it from those who do not consent etc

[cyan burrow]

isn't that pre 2.0?

[supple raptor]

2.0 removed artist names and NSFW content AFAIK.

[cyan burrow]

yeah

[wise forum]

guys can i find on this server an image generating chat ?

[cyan burrow]

well either way, I'm staying out of it

[brittle jay]

no

[wise forum]

where i can find an discord bot for stable

image generation guys

[untold glacier]

Go to #1072220168534642768 and #1072229020520947753 for info on where to generate images. There are no bots here.

[supple raptor]

Seems pretty clear to me that their intention is that an author can use AI creations within their work, but cannot copyright purely AI generated content.

Individuals who use AI technology in creating a work may claim copyright protection for their own contributions to that work. They must use the Standard Application,[39]

and in it identify the author(s) and provide a brief statement in the “Author Created” field that describes the authorship that was contributed by a human. For example, an applicant who incorporates AI-generated text into a larger textual work should claim the portions of the textual work that is human-authored. And an applicant who creatively arranges the human and non-human content within a work should fill out the “Author Created” field to claim: “Selection, coordination, and arrangement of [describe human-authored content] created by the author and [describe AI content] generated by artificial intelligence.” Applicants should not list an AI technology or the company that provided it as an author or co-author simply because they used it when creating their work.

[brittle jay]

do you have a gpu?

[wise forum]

?

[brittle jay]

graphics card

[wise forum]

intel

[wise forum]

hd

[unreal ridge]

f

[supple raptor]

Are you looking for a bot to host yourself for your own server, or are you looking for a server that has one that can use?

[brittle jay]

I suggest you use colab then

[wise forum]

for my server

bot

[brittle jay]

oh

you mean like that

[wise forum]

wdym ?

[final river]

Where are you drawing this conclusion based on the section you just copied?

[brittle jay]

you can still use colab

[untold glacier]

That's text-based AI creation though. I'd think that image generation is a bit different in both process of creation and AI implementation.

[brittle jay]

just need to enable the share thingy and use api for bot

[wise forum]

can u explain what is the collab

collab with who

[umbral depot]

I got canceled on Twitter for not showing feet pics

Collab with Google

[brittle jay]

google colab is a google thingy where they let you run stuff on their servers

[untold glacier]

ELI5

[wise forum]

so hard for me, never used it before

[unreal ridge]

That's... New for me.

[supple raptor]

First step is going to be obtaining hardware that can run stable diffusion either by renting it from a cloud computing service or owning it. I would encourage you to take a step back from the discord bot idea until you have figured out just generating through stable diffusion normally.

[brittle jay]

it is very easy to use since people have already made colabs for people to use. you just need to press a few buttons and that's it

[brittle jay]

I mean, if they don't wanna spend money colab could also work no?

[wise forum]

ohh i see and what is minimal cpu paramets

[brittle jay]

no cpu, gpu

[wise forum]

can u explain what buttons ?

i really can t understand the idea of collabs

before i was using the midjorney

now i wanna switch to free source

[final river]

My bad, I misunderstood and misread what you wrote.

[supple raptor]

It's all generative AI. The protections are the same.

[brittle jay]

https://colab.research.google.com/drive/1wEa-tS10h4LlDykd87TF5zzpXIIQoCmq

[untold glacier]

It's generative AI, but you can't "add more words" into an image. I think the process that's described there is different for image AI.

[brittle jay]

I think this good colab no?

[supple raptor]

You're gonna have a bad time trying to run it on just a CPU. It's possible to do, but very, very slow. Typically generation uses your GPU (graphics card) I would recommend a minimum of an RTX 3060 12gb if you are buying a GPU for this task, though it can be done on lower end graphics cards too if that is out of your price range.

[supple raptor]

You can, by compositing images or painting over them.

[brittle jay]

minimum is like 1060 probably or something

[final river]

Yeah, that's correct. You can't copyright purely ai generated content. But what that means in the current legislation is "no human creativity involved in the creation of the work". You have to have had considerable influence on the output of the final work as a human in order to register it for a legally registered copyright.

[brittle jay]

I mean if I use latent couple, is that purely ai generated still?

or what if I img2img on my own painting

they did say that it's a case by case basis

you will have to convince the court you have influenced the ai enough or changed the image enough by yourself

[wise forum]

I ngl i have mac os system, macbook pro so gpu is low here, but i m thinking to running it not in my own computer but like using someone that hosting a bot.

[supple raptor]

Yes, and prompting has apparently been deemed to not qualify as human authorship based on the Zarya of the Dawn ruling.

[final river]

The zarya of the dawn ruling wasn't under the current precedent.

[supple raptor]

Most likely, yes. In the img2img from your own painting, you could absolutely argue human authorship for that.

[brittle jay]

I actually agree with us law here that if you don't really touch the settings too much, just prompting should not make it human authorship

[final river]

The case was also up for a re-review.

[brittle jay]

well, only law wise. I would still consider it art and the one that made it its author

[final river]

If you prompt specific things & details in an image, you had to use your human creativity to do it. You made all of those things appear in the image, therefore controlling what the AI outputted. So it was your influence, and the AI drew it in tandem.

[brittle jay]

their argument is simple. you give it instruction and then it's onto the machine how to lay these instructions into the images, what to put where and stuff

it's like commissioning

[lyric zealot]

pretty sure it's expmixlinev2

[final river]

There's a difference between prompting "A girl at the beach" and "1girl, black and red swimsuit, purple hair, brown eyes, white sun dress, beach and ocean with calm waves, windy," etc.

[brittle jay]

as there is a difference between a detailed commission and a less detailed one

you do influence one more than the other but at the end of the day it is still like a commission

[supple raptor]

The same could be said for text generation too. In fact based on my experience running LLMs, it takes a hell of a lot more prompting and finetuning of settings to make an LLM give a good story than it does for a good image on SD or any of the other modern AI art models. If you've only used something like ChatGPT then you may not be aware of what goes into it because the initialization prompt is done for you.

[lyric zealot]

[quick hill]

Oh god my cat is taking my seat again

Its her seat now

[wise forum]

guys what u re using for an image generation, and what u using also for ai chat (like chatgpt, bing or what) im really interested ?

[final river]

Not sure if you're saying this generally, or actually mean me as far as llm use goes.

[brittle jay]

can confirm

[supple raptor]

Of course. You belong to her. Why would you expect any different?

[final river]

But my point is exactly that. The human element involved towards specific prompting is what would qualify it as an original work, and not just the AI spewing out whatever.

[quick hill]

[final river]

It's the amount of human work involved in the creation.

[quick hill]

handss

[final river]

That's how the current legislation works and reads.

[supple raptor]

While I personally agree that what you are saying should be the way it is handled, that is not the case in reality based on everything I've read.

[elfin bolt]

can i copyright my image by adding this?

[untold glacier]

Sorry dark mode peeps, shield thine eyes.

There is no mention in the policy guidelines of image generation outside of the Midjourney-specific mention. Any specifics that are mentioned are text-based and not image based. It's not a 1:1 there as the tools that have been developed beyond the original SD release gives creative input on the human side arguably far more than any LLM or text generation.
In the end these are "Policy Guidelines" anyway, and shouldn't be taken as the law.

[brittle jay]

for image generation we use different models since there is always more inovation and it's better to use a model made for a specific task than a generalized model if you want to make a certain kinda image.

for LLMs, we use pygmallion, gpt4, vicuna, alpaca/llama finetunes and more, through either koboldai, oobabooga or some other random UIs

[final river]

Then I don't know what else to tell you other than you must be misinterpreting what you've raid.

[untold glacier]

Sorry, wrong image for the third one.

[final river]

And creative input/intervention from a human author includes specifically prompting for a given image output. Not just typing a single sentence and letting it rip.

[supple raptor]

Can you provide where you have read that implies that sufficient prompting qualifies as human authorship in AI generated content?

[brittle jay]

if you really think so, go to court and try proving it

[final river]

Someone even posted this earlier.

[brittle jay]

lol

[final river]

Tell me specific prompting doesn't fall under that criteria.

[brittle jay]

weird wording right there

[cyan burrow]

aaaah I is back and youre still doing this?

[untold glacier]

Look what you started! For shame! 😂

[cyan burrow]

Fuck I'm sorry

[untold glacier]

Nah, everyone loves a good AI debate.

[cyan burrow]

I forgot we were weebs for a moment

a moment of disaster

[elfin bolt]

"#🍥｜anime-general-chat"

[supple raptor]

Based on the context that certainly sounds like their intended meaning is that the human artist may composite together multiple images that each were purely AI generated.

[final river]

Read the rest.

Actually

It's easier to understand if you just read list of the things that can't be copyright protected.

[supple raptor]

The rest is where I am getting that context from.

[final river]

Those were just examples.

It clearly says between both of what you can copyright and can't copyright

That by guiding the AI through the process of selecting, modifying, choosing etc the material in an image

(Prompting does that)

You can copyright it

[brittle jay]

I think because they said "generate multiple styles" they mean that you make multiple images and then combine them

[final river]

That one

Was just

An example

For multiple images

The rest of the text isn't exclusive to that starting example

[supple raptor]

What's the link to the full document you're looking at?

[cyan burrow]

to make it real annoying, just copy it in photoshop and run it through color grading, curve alterations and some brushes to add hair and whatnot

[wise forum]

where u generated this dude

[pale monolith]

local :3

[elfin bolt]

that's how you trick this ai image model detector https://hivemoderation.com/ai-generated-content-detection

[brittle jay]

overall, this doesn't really say much just because of how ambiguous they worded it.

[supple raptor]

There's no annoyance there. That is explicitly allowed.

[cyan burrow]

started my civit upload of DG

[cyan burrow]

most of mine trick is anyway

it seems to think theyre real

the more flawed (in a human way) they are, the more it tricks them

[wise forum]

thx for explanation bro, I will still look for an bot or a webpage with a stable ai, that can generate an images for free.

[brittle jay]

I linked a colab

learn to use this https://colab.research.google.com/drive/1wEa-tS10h4LlDykd87TF5zzpXIIQoCmq

[frozen olive]

oh a new one?

is it any good?

[brittle jay]

it's by nocrypt

[cyan burrow]

well mine has worked for almost everything so far. When it comes to checkpoints that is, not lora, embed, hyper, lycoris, samplers for upscaling etc

[elfin bolt]

I made everything through collab, its useful while google still make it free atm

[frozen olive]

i see that but is it good

[cyan burrow]

but I think I will personally merge towards different specializations for personal use

[brittle jay]

idk I just trust him tbh

and it has all the latest greatest extensions built in

so ig it's good

[frozen olive]

i hate colab cause it changes every 3 days and then nothing works

this sucks hard

[final river]

Do you have a source/link to the documentation for this?

[wise forum]

ok i will try to understand how to use it

[final river]

I already know it's correct based on the legislation I've read but it'll be easier for other people to understand with it

[wise forum]

u re the best guys

[elfin bolt]

not so reliable, but its a guy who mainly write about ai summarizing from https://www.govinfo.gov/content/pkg/FR-2023-03-16/pdf/2023-05321.pdf

https://www.alenknight.com/?p=2276&fbclid=IwAR2Wks0crYpywAexx4a_-ahR-9v1_piR_uF_ktzz3FbEFJeNd2kH_rCYAqI

[cyan burrow]

director of movie based on a true story where everyone died tries to make scenes suspsenful
crickets
everybody dies. How unexpected

[wise forum]

I wonder are there chatgpt4 via collabs

hihihi

[brittle jay]

no

too heavy and also not open source

you can just use openai's api, it costs money though

[final river]

Their summary is correct, but most people probably wouldn't believe it. Since it seems some people can't understand how the current government documents/rulings work. Oh well.

[brittle jay]

or use normal chatgpt for free with limited tokens per day

[elfin bolt]

[cyan burrow]

I've got a friend that made a local gpt4 chatbot

it's an annoying bot

cause it always says "that would be unethical"

[brittle jay]

vicuna or something probably or they used openai api

[cyan burrow]

yeah well I don't wanna be ethical

[brittle jay]

so not truly local gpt4

[cyan burrow]

no api

[brittle jay]

then vicuna

[cyan burrow]

not sure what but I'll ask

[brittle jay]

or like some alpaca/llama finetune on gpt4

[final river]

It'll be nice to have something that steps away from facebook's models

But bloom isn't as good as any of the current llama finetunes

[brittle jay]

I like pygmallion

they based on like gptj or something no?

[final river]

Yeah

[cyan burrow]

uuuuh apparently it's not gpt3 or gpt4. At least not according to the bot

[final river]

Just waiting for something comparable equal to or better than vicuna that's completely open source

[cyan burrow]

it's a bot that answers questions related to traveling to a region in the US (I wont be specific for privacy concerns)

and he reprogrammed it

but it thinks it's still answering travel questions

[brittle jay]

reprogrammed? like finetuned?

you can't reprogram a model

[lament osprey]

Beautiful

[final river]

This makes me wanna re watch Bocchi even though I just watched it 5 days ago

[cyan burrow]

I don't know...it's not a commercial or open source thing

[magic wave]

Gotta love how when I had a GTX 960 No Man's Sky took 15 minutes to load, now I got a 3060 and it takes the same time

[brittle jay]

I have a question. is it alright that I named my variables in a game I am making sex and sex2

[untold glacier]

She's a big girl now

[magic wave]

I think thats pretty funny

[brittle jay]

same

that's why I did it

[cyan burrow]

I just managed to make the bot give me a pgone number

and then found out it was compromized

and asked for a new one

and got it -.-

the hell is up with this chat bot

[supple raptor]

Pretty cut and dry here that you cannot copyright any work that is solely produced by inputting a prompt into a generative AI. Adding additional controls such as controlnet, may or may not be enough to be deemed sufficient human authorship as they do not seem to be aware of tools like that. I suspect that a tool like latent coupling would, also not be sufficient as it would likely still be viewed as simply inputting a prompt.

[cyan burrow]

@supple raptor it's apparently gpt3.5 the text-davinci-003

[supple raptor]

I have no idea what you're talking about. I was AFK for a bit and just came back and reviewed the copyright doc.

[cyan burrow]

ph

oh

[cyan burrow]

gpt3.5 the text-davinci-003

I got you mixed up with crazy

apologies

[cyan burrow]

@brittle jay so it's running gpt 3.5 with a davinci-003 text model

[supple raptor]

If it's about the phone number thing, then it doesn't really surprise me that the LLM was able to produce a phone number. I'm sure there would have been many phone numbers in the training data with enough context around them for the model to predict a phone number.

[brittle jay]

yeah yeah

openai api

[untold glacier]

"when an AI technology receives solely a prompt from a human..."
Yeah, that's midjourney for sure. SD is different by default with something like Auto's build as you have access to so many of the different settings, not to mention extensions like you said. I think the additional levels of complexity in creatively deciding it's generation capabilities in regard to your specific prompting is enough for copyrighting. But that's just me. Hopefully it's others' opinions as well.

[final river]

This was even later clarified to the office and "fixed" because they didn't know and understand the process of how specific you can get with prompting, and how much influence you have over what's generated when prompting. The example in that documentation (as can be read on the bottom right starting with the number 28) is using midjourney. Even with the rest of what's written, it was shown that with other programs prompting has more of a direct influence based on human input.

[brittle jay]

I think I saw someone producing keys to windows 98 or whatever it was with like 1/30 chance of working or higher iirc

[supple raptor]

I think it was 95, but yeah I saw that too.

[supple raptor]

Nothing there is saying that their determination was amended, only adding additional information that wasn't important enough to add to the main body of the document so as to prevent it from becoming too wordy.

[brittle jay]

blue lock lora?

[weak moss]

YES

[supple raptor]

@final river here you go. It's possible that a prompt itself may eligible for copyright, but not the material generated from that prompt.

[magic wave]

Aka, you can copyright your idea, but not your product

[untold glacier]

:thonk:

Is what I would normally use if I had nitro.

[supple raptor]

There you go

[untold glacier]

ᵗʰᵃⁿᵏˢ

[velvet void]

control net should be enough human intervention for copyright

[south sable]

you guys should watch the latest chapter of "Everything is a Remix" on youtube. It answers a ton of these questions about copyright stuff

[supple raptor]

I'm off to go craft and copyright a super complicated Chat-GPT prompt that I can use to generate stable diffusion prompts which I will then generate images with.

https://tenor.com/view/stonks-up-stongs-meme-stocks-gif-15715298

[pale monolith]

copyright your OC then include them in every image in some degree

[south sable]

https://tenor.com/view/the-simpsons-dbz-rap-homer-simpsons-gif-17469967

[magic wave]

Good idea

[cyan burrow]

api? no he has it local

[brittle jay]

it's not open source

[cyan burrow]

not paying anything

well then he didn't get it through open source

or open anything

[brittle jay]

maybe web scraped chatgpt?

[cyan burrow]

I dunno

[brittle jay]

but that's like limited tokens per day

[lament osprey]

Yes that shit takes ages

[supple raptor]

Loading times are usually primarily limited by storage speed. I don't have NMS but if you're loading from a HDD you could probably significantly speed that up by installing it to an NVME SSD.

[magic wave]

Yeah xD

[magic wave]

I mean, I could try to move it to my SSD

[teal pivot]

😮

[supple raptor]

Would probably help a lot even really slow RAM and VRAM are more than twice as fast as top of the line PCIe Gen4 NVME drive so the drive speed should be the main limiting factor as long as the CPU isn't overwhelmed or being choked by bad code.

[cyan burrow]

@brittle jay yeah my hunch was confirmed so I won't go further into this

[lament osprey]

Omg

So good

[brittle jay]

?

[wise forum]

bro where u from

[brittle jay]

Israel

[wise forum]

I m from Azerbaycan

[cyan burrow]

that he didnt exactly obtain the gpt as...you should

and I don't want to get him into trouble 😛

[brittle jay]

if leaked then this big deal

I doubt it

[modest anvil]

all my ai say its not likely to be AI Generated
0.4%

hmm

[cyan burrow]

I'm not publicly speaking about it

[modest anvil]

i tried 30 different pictures i have tooo

i think if the ai doesn't have fucked hands or a background thats mutated its pretty impossible to tell if its an ai now or text

[magic wave]

Fair fair :o

[supple raptor]

Honestly, those tools aren't very good at what they're supposed to do. Most AI art is pretty easy to identify if you know what you're looking for. There's lots of little mistakes that the AI makes that give it away as AI if you look close enough. There are definitely exceptions though.

[untold glacier]

Mistakes? What mistakes?

[modest anvil]

100% agree but i wonder how things will be 5 months from now i remember 5 months ago doing stable diffusion disco diffusion and not for one second did I think this type of level of art is possible to do in seconds

[supple raptor]

damn... she got me

[final river]

Most AI art is pretty easy to identify as AI art because most AI artists don't really work on and refine it/do any post processing.

[cyan burrow]

a few hours until I have uploaded FP32 and FP16 versions of normal and pruned

[teal pivot]

xD

[teal pivot]

what could be better ahaha

[unreal ridge]

I merged it with my model, it makes some pretty good backgrounds.

[cyan burrow]

what is happening here? (spoiler tag potentially gross, I dont know, depends whats going on)

is it acid blood and brain that's burning a hole in the ground?

[stone onyx]

face is a bit deep fried

[brittle jay]

sneak peak. without vs with lora at epoch 4

gotta continue training

I shall now go asleep, for I got a big important test in 2 days

technically in 1 day since it's past midnight for me

[cyan burrow]

mmmm

tomorrow I'm going in for the prep work, chat and final tests before my surgery on the 27th

then I will be admitted on the 26th

and probably stay there 2 weeks or so recovering

[cyan burrow]

hope I can set up a good remote desktop

so I can run SD from there

maybe set up hamachi or something similar and only use remote desktop if auto or hamachi crashes

[south sable]

that would be awesome. I've wondered about that myself but havent looked into it yet

[cyan burrow]

cause I'll be in the hospital recovering from major surgery

will be in bed a lot

not much to do....

[hollow storm]

just b relly careful to hav securty

[cyan burrow]

security?

[hollow storm]

dont want some random guy halfway across the world to also using ur stable diffusion

[cyan burrow]

they can't get in when its local

[hollow storm]

well if its local its not remote...?

[cyan burrow]

it is with hamachi

or something similar

hamachi is emulated local network

[hollow storm]

ic

[cyan burrow]

so I can get into it that way

then I have remote desktop in case either crashes

they say that's good luck. Now start stacking, you open those with cream and you have two extra to include in the stacking

you can make a superbiscuit

prepare it for later

but yeah

if you ate it

well eat the fillings of another

and save them

then you can make it

[supple raptor]

As an American I'm confused why you would expect cream in your biscuits

[cyan burrow]

I should get a lot more waifu things to advertise my model

[wise forum]

looking trilling

[cyan burrow]

I'm afraid I was wrong, it was gpt 3.5. He's working on getting gpt4

[stable crane]

hayo

long time no diffuse

[supple raptor]

I was just teasing 🙃

[weak moss]

when I want to quote an anime at the prompt, is it like this '' \ (one piece\ ) '' ?

[compact warren]

I think thats become redundant now

But i dont remember

[opal dome]

just do one piece, <character name>, danbooru tags are hit or miss imo

[compact warren]

Its why we got loras

[stable crane]

that is how you'd escape it yes

[stable crane]

so and so \(neon genesis evangelion\), as 1 token

[supple raptor]

So, the \(one piece\) thing comes mostly from characters, not animes. For example, the tag for One Piece is just one_piece which you would write as one piece in your prompt because we don't use underscores. However, the tag for Yamato is yamato_(one_piece), you still remove the underscores like in the first example, but because parenthesis are used in the actual tag you need to escape them so that they will be interpreted as text instead of added emphasis which means you would write yamato \(one piece\) in your prompt. The \ is used to tell the python code to feed the parenthesis to the AI as text instead of doing the function intercepting them and using them to increase emphasis.

In short. Anytime a tag has parenthesis in it, you need to add \ before both of the parenthesis.

different mediums such as watercolor, graphite, and colored pencil are other somewhat common tags that use parenthesis. These would be watercolor \(medium\), graphite \(medium\), and colored pencil \(medium\)

[supple raptor]

Usually tags that have at least 2k examples will work pretty well. Characters and more complicated concepts might need more. Some models will handle some of them better than others of course depending on additional training that has been included in them. Characters usually start working well around 10k examples. If you use the webui-tagcomplete extension, it makes it easy to see roughly how many examples exist for a tag.

[pale monolith]

👋

[supple raptor]

wait is it really this easy?

this is actually useful for the server I moderate

I'm now reading that like I shouted it XD

[wet lark]

How

[supple raptor]

[subtle crown]

so markdown

eh

• oh
  boo

[supple raptor]

sadge. doesn't work for emoji.

What's funny is that it ends up doing the opposite for emoji since those are usually big if you don't include text in the message.

no superscript.

1

2

[subtle crown]

H1

H2

H3

bold text
italicized text

blockquote

1. First item
2. Second item
3. Third item

• First item
• Second item
• Third item
  code

---

title

| Syntax | Description |
| ----------- | ----------- |
| Header | Title |
| Paragraph | Text |

{
  "firstName": "John",
  "lastName": "Smith",
  "age": 25
}

 Here's a sentence with a footnote. [^1]

[^1]: This is the footnote.

My Great Heading {#custom-id}

term
: definition
The world is flat.

• [x] Write the press release
• [ ] Update the website
• [ ] Contact the media
  That is so funny! 😂
  I need to highlight these ==very important words==.
  H~2~O
  X^2^

[stable crane]

wat lol

g𝖚𝖓𝖋𝖑𝖔𝖜𝖊𝖗

surprised we can do headings now

that's cool

[supple sluice]

whats with all this formatting chat

[chilly swallow]

[Apparently you can do colored text too]

[stable crane]

# yo dawg

[supple sluice]

All I figured out was lists like this work

• I
• figured
• it
• out
• in
• #1097898676610928790

[chilly swallow]

"Its code formatting, so the code is at the top"

[supple sluice]

didn't know they did a bunch of other formatting stuff

[chilly swallow]

%{Then you add weird stuff around it}

[chilly swallow]

- Red text

[chilly swallow]

[blue text]

||Ok now teach me how to spoiler||

[stable crane]

b r u h

this will be useful

sss

[steel arrow]

is this how it works

men are hot

[supple sluice]

[colors?]

[chilly swallow]

%{this_is_an_important_message}

[chilly swallow]

I cant get it to be yellow

[stable crane]

ᶜʰᵃᵗ ᵐᵒᵛᶦⁿᵍ ˢᵒ ᶠᵃˢᵗ ⁿᵒᵇᵒᵈʸ ʷᶦˡˡ ⁿᵒᵗᶦᶜᵉ ᵐʸ ᶜᵘᵗᵉ ᵖᶦᶜ

[chilly swallow]

+ Its very cute <3

[steel arrow]

I notice your cute pic

[stable crane]

it's a fine line to tread to avoid overly lewd

but someone's gotta do it

lol that's a website that converts your text into other forms

[steel arrow]

I just made one that's basically 2 vampire girls laughing at you (I asked for vampire teeth and I put the weight too high)

[wet lark]

I hate UD's draconian nfsw rules

[steel arrow]

Felt bullied

[supple sluice]

so im messing around with a prompt generator that I helped make and this was one of the results...

[stable crane]

https://lingojam.com/TinyTextGenerator

[supple sluice]

I kinda wanna try it tbh

[wet lark]

Got timed out for making Miku, it is not even nfsw

[stable crane]

how lewd was it?

[steel arrow]

raided 2bs wardrobe

[chilly swallow]

I use your lora to get pretty eyes sometimes

[stable crane]

lol which lora is this

[supple raptor]

Might want to delete that one. I'd guess it breaks the rules since they apply to content other than images too.

[supple sluice]

got a prompt that led to this... quite interesting

oh this this how the big text works?

[chilly swallow]

||the blowjob lora||

[supple sluice]

oh

[stable crane]

ohhh

[supple raptor]

Which model are you using? I'm really enjoying your style

[teal pivot]

My own, soon I'll post it on Civitai when I'm done with the tests

[supple sluice]

alright I have decided the prompt generator is not good for anime models right off the bat, though it is highly customizeable (which I do not feel like doing right now)

[steel arrow]

I need to learn how to make my own model, sounds like fun

[chilly swallow]

I feel like it just gets really cute faces even if they all start having open mouths

[stable crane]

yeah I didn't expect it'd work that well for that

[steel arrow]

What if you tell it to close the mouth

[stable crane]

sometimes dynamic pose is too dynamic

that's the only good pose I got in like, 10 gens

[teal pivot]

this is sfw channel?

[stable crane]

yeah

[teal pivot]

ok

[stable crane]

gotta get creative to skirt the rules

and Clyde

[teal pivot]

let's goo

[stable crane]

sick

love the hogbike

[stable crane]

what upscaler you using?

[teal pivot]

hiresfix remacri + extras R-ESRGAN anime B

[stable crane]

have you tried multi-diffusion?

you can do highres fix with like, 3x or 4x

without OOM of VRAM

[teal pivot]

I know, but I don't see the point in making such big pictures yet 😄

[stable crane]

a 3x gen

ekes out more details

though this style doesn't get much more detailed

[teal pivot]

that's enough?

[stable crane]

you can see the extras tab upscale though

[teal pivot]

ye i know

[stable crane]

I wish I had a good example offhand of the other

[teal pivot]

u want me to make this xd

okay lets go

6gb notebook 3060 brrrr

[chilly swallow]

funny

[teal pivot]

It takes a long time

[chrome moss]

new lora

[supple raptor]

Awesome. I'll have to keep an eye out for it. Do you have an account already that I can follow to make it easier?

[stable crane]

same

[teal pivot]

yes on civit

check bio

[stable crane]

following

[teal pivot]

almost

[stable crane]

how big is it?

[teal pivot]

one sec

[stable crane]

damn nice

[teal pivot]

haha my GPU 🔥

[steel arrow]

My gpu runs out of v ram if I upscale to that amount (idk how to increase the allowance)

[stable crane]

try multi diffusion extension and enable tiled vae

[south sable]

its called buy more vram

T_T

[stable crane]

https://cdn.discordapp.com/attachments/1045779996523901008/1094417232014082098/image.png example

[stone onyx]

this is really cool, ultrawide wallpaper

[stable crane]

I think that one might've been just a txt2img

[steel arrow]

Damn 32gb ram comes in real handy with this stuff

[subtle crown]

64 is even better

[steel arrow]

cries in wallet

[subtle crown]

if i ever build new pc totally getting 64gb of ram

[steel arrow]

My pc can only handle that much

[stone onyx]

I don't think I've ever ran out of ram despite running 300 web tabs 24/7 (32gb ram)

[steel arrow]

Yeah same

[stable crane]

I want more vram mostly

[steel arrow]

But I haven't run that many tabs

I'm annoyed, I got a 3070ti and the v ram amount is minimum for today's stuff

[stone onyx]

I almost got myself a 3070 as well, but thought I might get annoyed running AI stuff so I settled with a 3060

maybe one day used 3090s are gonna drop in price

[subtle crown]

yeah, waitning on 5080ti hopefully has lots of ram lol

[steel arrow]

Yeah, but it's too expensive to change to anything now

[stone onyx]

honestly with how nvidia's been treating vram on their newer cards I dunno

[steel arrow]

5060 20gb, 5070 9gb, 5080 12gb. Seems to be the trend currently

[stable crane]

wat

is that what they're doing?

[stone onyx]

the 4070 has 12gbs iirc

[steel arrow]

I'm making stuff up, but the 3060 has more v ram than 3070

[stone onyx]

4080 has 16gbs

[stable crane]

jeez

[stone onyx]

meanwhile look at amd, throwing vram like it's candy

[subtle crown]

yeah i would so get AMD card but, getting AI generation on it its eh, not good

[stable crane]

can't bet against nvidia

[steel arrow]

My brother just got the 7900xt 3 days ago and I wish I waited the 6 months to get it at the same time (cheap)

[stone onyx]

not wrong, though they're bringing ROCm (AMDs CUDA basically) support to windows soon, but it's for sure gonna take a while for people to implement its support

[subtle crown]

yeah by the time 5080 comes out maybe we get something good from AMD by that time

[teal pivot]

hmm

[stable crane]

lucky wink

not a single dude looked at the camera

[teal pivot]

xD

[teal pivot]

ahaah

[stable crane]

impossible to get a skull mask on the face

[teal pivot]

Hmmm, the model might make some nice stickers

[final river]

Do you mind if I try and add it for you?

[stable crane]

go for it

[final river]

Did you want the mask covering her face?

[stable crane]

I suppose you could img2img but I would like to do it in 1 shot lol

yeah no human face visible

[teal pivot]

so cute yeaah

[zinc wigeon]

Can anyone help me how to run this..because I am using this for first time

[stable crane]

@teal pivot can you @ me if you release the model?

gonna head out for now

[teal pivot]

sure

[stone onyx]

idk why I even spent 15mins fixing the hands and blurring them to hide how tragic they look

[keen anchor]

Dang was proud of this anime style image and came here to find everyone also has awesome anime style images lol

[keen anchor]

anyways heres the image

[stone onyx]

cute artstyle

[final river]

I spent way more time on this than I wanted or expected to and it's not even good but I'm just gonna leave it here even though I'm not happy with it at all

[grand geyser]

what does 32 gb ?

does it help in any process

[supple raptor]

It would help with model merging, and speeding up model switching since you could cache models to RAM, but off the top of my head I can't think of much else that you want to be using system RAM for.

[final river]

Removed the random petals

[quick hill]

Vram matter the most

[supple raptor]

Yep, for inference if your tasks start swapping to system RAM you're going to have a long wait.

[quick hill]

Hii

[grand geyser]

yeah this what i mean

@supple raptorhow do you know it swap to ram

[supple raptor]

First sign is your it/s turns into several s/it.

[grand geyser]

i have only 1 it

like not couple

[weak moss]

Guys, these are the configs I use in ''inpaint'' to improve the hands, but it's not working. Some tips ?

[supple raptor]

Try something like this (sorry for different UI). The "Masked content" and "Inpaint area" settings are the main ones here. Then make sure you mask a little ways up the arm, maybe halfway to the elbow in order for it to have enough context to draw the hand properly. If it seems like the hand isn't fitting in with the rest, try increasing the "Only masked padding, pixels" by a couple of steps.

[loud falcon]

ooh, did you make that ui?

[supple raptor]

If the hand is really bad, I would recommend doing your best to try to draw it with the correct shapes at least in something like Krita, then inpaint over your manual fix to match it to the style.

[supple raptor]

No, I just picked the colors, it's just a different fork of Automatic1111's WebUI. https://github.com/anapnoe/stable-diffusion-webui-ux

[supple raptor]

Much better interface than the main fork, especially if you use --listen to generate from mobile.

[loud falcon]

very nice, i'm using one that's coffe something i believe but i can't resize the prompt box

[loud falcon]

wait, is this like installing stable from zero?

[supple raptor]

Yeah, it's a separate fork. I installed it alongside my existing Auto1111 install and symlinked my embeddings, models, and outputs directories so that those are shared.

[quick hill]

Oh god

[loud falcon]

https://tenor.com/view/brain-gif-22265960

[quick hill]

My burnout is making me lazy

[supple raptor]

solution is simple just stop having burnout

[quick hill]

https://tenor.com/view/homeless-house-braindead-stupidity-buy-gif-26259659

[supple raptor]

Gonna retry that one with (SLE:1.2) to see if I can get better eyes on it.

[quick hill]

That remind me I have to make to do list for next version of Aurora, prob towards background improvement

[supple raptor]

Already planning the next model?

[quick hill]

Just updating

But next model yes

Probably specific style this time

[supple raptor]

I think I know you well enough that a "next version" isn't that much less work put into it than a whole new model XD

[loud falcon]

you can symlink only the dirs and it works fine?

[quick hill]

[supple raptor]

Yep, for those ones at least. I got errors when I tried symlinking one of my extensions so I don't think that can be done (or at least not easily)

[loud falcon]

so you symlink both diffusions dirs to a dir outside?

that's the idea?

[supple raptor]

mklink /D E:\StableDiffusion\stable-diffusion-webui-ux\embeddings E:\StableDiffusion\stable-diffusion-webui\embeddings mklink /D E:\StableDiffusion\stable-diffusion-webui-ux\models E:\StableDiffusion\stable-diffusion-webui\models mklink /D E:\StableDiffusion\stable-diffusion-webui-ux\outputs E:\StableDiffusion\stable-diffusion-webui\outputs

That didn't format as well as I had hoped -_-

[quick hill]

TheUi is so worth it though, default isn't bad but I really want customizable UI

I thought that isn't a thing until today

[loud falcon]

ah, thanks a lot

[loud falcon]

i thought you made a 3rd dir to direct both, but you directed the 2nd to the first

[quick hill]

Prob not less work than making model, but I already plan my theory; Just need to apply it.

Also gonna try retraining SLE for Locon too.

[quick hill]

Aurora took like 2months prob because versatility

[stone onyx]

I still need to download Aurora, I'm procrastinating too hard

[quick hill]

XD

[quick hill]

I think ur images look better than my model, aurora is more like challenge for all styles atleast for me

I still wanna go experiment and testing out things I don't know about and can't find any research anywhere on internet

[stone onyx]

ⁿᵒ, ᴵ'ᵐ ˢᵗᶦˡˡ ᵍᵒⁿⁿᵃ ᵈᵒʷⁿˡᵒᵃᵈ ᶦᵗ >:(

[mild tree]

😭

[quick hill]

I hope I can write prompting guide for my model today

[steel arrow]

When the cmd says the upscale has 1 hour 10 to go, and the bar has been at 1% for 5 mins

And it'll turn out wrong

[stone onyx]

okay time for bed, gn

[quick hill]

Goodnight

[quick hill]

Oh my god this is so cute

And comfy

Is this lora etc?

[stable crane]

reminds me of the simpsons

[south sable]

no loras, just model magic

[quick hill]

What is the model

I want it

[south sable]

https://civitai.com/models/35960/flat-2d-animerge

[untold glacier]

-i am now caught up in chat

• I forgot Discord was in markdown.

• Though I had memorized bold and italics

[untold glacier]

Thanks for this. I've been rather unhappy with the bright shininess and big outlines of 7thC. Hopefully this works well for my purposes.

It looks rather dim in the civitai examples. Are you doing something special to brighten the images?

[quick hill]

Gonna extract model and make it into lora

[stable crane]

looks very cute

[quick hill]

[south sable]

I'm using blessed2.vae

[mild tree]

Yum

[stable crane]

blessed2?

[quick hill]

Yesh

https://huggingface.co/NoCrypt/blessed_vae

[south sable]

https://tenor.com/view/yy-smoking-counting-money-anime-gif-17852058

[quick hill]

BLESS 🙏🏻

[stable crane]

hah https://s3.amazonaws.com/moonup/production/uploads/1678275680281-62de447b4dcb9177d4bd876c.png this really proves kl-f8-anime2 really blurs shit

[south sable]

that and fatal anime 500000 upscaler

[quick hill]

5000000000

[untold glacier]

Fatal blurs or fatal is good? Is Anime6B better?

[stable crane]

oh man https://civitai.com/models/43256/blue-archive-emoji-style

[south sable]

they all have strengths, I use 500000, sharp, and 6b for differnt things

[untold glacier]

I just started using fatal today. Haven't run it through tests yet but initial results are good. 6B is great at cleaning flat zones and keeping broad crispness where it needs to be with line edges. How does fatal differ in your opinion?

[south sable]

from what I can tell, it can change depending on all the other factors like loras, hypers, steps, cfg, etc... sometimes it just works better. I'm not sure what it does consistently

[untold glacier]

Cool beans. Thanks for the info.

[stable crane]

this blessed vae seems pretty good

[south sable]

I like it. I use it a lot but still go back to the classics often

[stable crane]

might be a bit saturated with offset lora too

[stable crane]

the colors really pop

[south sable]

right?

[stable crane]

I'm always amazed by SD

[south sable]

same

[stable crane]

god it's so cute

[south sable]

https://tenor.com/view/menhera-chan-chibi-menhera-angry-anime-girl-gif-24076530

[quick hill]

This is so hq

[stable crane]

this really makes her look mousey

[untold glacier]

ok

Here's a quick test that shows what each upscaler does within context of a 0.18 denoise SD upscale. It's interesting to see the fine and broad line differences between Ultrasharp, 6B, and Fatal at work, as well as where the crispness is located.
1: Ultrasharp, 2: Anime6B, 3: fatal_Anime

Fatal also interestingly darkens the image somewhat, whereas the other two upscalers are relatively unchanged in that regard.

[south sable]

yeah it goes into deeper darks

but the brights gets saturated

[quick hill]

Better color balancing?

[south sable]

they all have good qualities

[untold glacier]

Judging from that test alone, I'd say if you're wanting just straight crispness to flat edges, a 2d anime style for instance, then go with 6B. If you're wanting some nice color enhancement and keeping of minute details go with Fatal. If you're going full realistic anime with grains of sand or fine repeated patterns like fishnets, you're likely going to want to go with Ultrasharp.

[stable crane]

might be one of the cutest renditions of Lotta ever

[still bay]

saw this on pixiv any idea what model is used on this? it looks like the one from the anime

[untold glacier]

All that being said, the VAE I ran for this test was just straight 1.5, so if you're running others the results may change slightly. But that can be said about any setting changes, heh.

[stable crane]

lollypop is also worth a shot

[untold glacier]

I've tried it. Not a fan.

[stable crane]

no?

it's probably the one I use most

[stable crane]

6b is really nice too

[steel arrow]

hugs anyone?

[stable crane]

hm 6b seems crisper than lollypop

lollypop vs 6b

[untold glacier]

I get strange melding of colors in the pastel range, when they are clearly defined beforehand. The result seems half baked, and compared with others like 6B it's substandard imo.

Yeah pretty much what Lewds said.

[supple raptor]

Lollypop can over smooth images that have fine details. The style you often post Lotta in probably works well with it.

[stable crane]

I've been using 6b a lot more recently

[untold glacier]

Same, depending on style. It keeps crispness of thicker outlines (even more so than Ultrasharp, with better coherence than Remacri.)

[supple raptor]

6B usually looks great but as an R-ESRGAN upscaler it's also slower than the others and the difference is usually quite minimal at the denoising levels I use (speaking of Hires fix). For multidiffusion/SD upscale I take more care with upscaler choice.

[stable crane]

I didn't realize they meaningfully impacted gen time

[untold glacier]

It's just a slight time increase on my 3080, but ymmv.

[supple raptor]

It's only a few seconds more per gen for me, but it adds up when I'm making a lot of images, especially since I always have Hires fix turned on.

[stable crane]

which upscaler do you use most @supple raptor ?

[supple raptor]

Depends on the prompt, but either AnimeSharp or Fatal are probably my most commonly used.

[quick hill]

Wowww

[quick hill]

I lovw lighting

[untold glacier]

Haven't used animesharp, I may have to try it.

[quick hill]

I use fatal all the time, I believe I did testing and forgot.

Prob have reason just like _ in prompt

My memory is so bad

[subtle crown]

this model is crazy for colors https://civitai.com/models/44596/526s-cusp-of-serenity