🍥｜anime

[tired quarry]

i can't seem to find Gishikiv1

[lyric zealot]

Thank you Abe

[tired quarry]

thanks

anyone know a model that would produce results like this?

With those nice colors and depth of field?

[lyric zealot]

I’m thinking models from ctd might be able to achieve that

[tired quarry]

what is that haha 😅

[lyric zealot]

If you want, you can try some of her models

Closertodeath

I think if you search that with huggingface in google, you can probably find her profile

[tired quarry]

ill try

[lyric zealot]

But this also probably requires a lot of Lora

[tired quarry]

what's lora

[lyric zealot]

Lora is a concept that can be loaded into other models to “teach” them

[untold glacier]

Ok, color me impressed with Bing's DALL-E 3 integration:

[tired quarry]

ive never seen that interface in my life

[tired quarry]

how would i go about doing that

[silk narwhal]

LoRA stands for Low-Rank Adaptation. It allows you to use low-rank adaptation technology to quickly fine-tune diffusion models. To put it in simple terms, the LoRA training model makes it easier to train Stable Diffusion on different concepts, such as characters or a specific style. These trained models then can be exported and used by others in their own generations.

Stable Diffusion models have been gaining popularity in the field of machine learning for their ability to generate high-quality images and text. However, one major drawback of these models is their large file size, making it difficult for users to maintain a collection on their personal computers. This is where LoRA comes in as a training technique to fine-tune Stable Diffusion models while maintaining manageable file sizes.

LoRA models are small Stable Diffusion models that apply smaller changes to standard checkpoint models, resulting in a reduced file size of 2-500 MBs, much smaller than checkpoint files. LoRA offers a good trade-off between file size and training power, making them an attractive solution for users who have an extensive collection of models.

[lyric zealot]

Basically if you want to generate a character, but the model you used don’t know the character, then you can download a Lora someone made of that character and use it

[untold glacier]

Thank you chatgpt

[silk narwhal]

😄

[tired quarry]

i just downloaded automatic111

[lyric zealot]

You can find Lora at civitai

Beware, there’s a lot of nsfw stuff there

[untold glacier]

LoRAs are network extensions for Stable Diffusion. They're a type of model that extends into a certain concept or idea that other people have trained. They help your existing model reach into more difficult to gen concepts, or styles.

[silk narwhal]

Simply use <Lora:name of the lora you added to the Lora folder>

[lyric zealot]

Okay, since everyone is around now to help, imma go back to bed

[silk narwhal]

gnight Smol

[lyric zealot]

[untold glacier]

Hello, and Goodbye/G'night Smol!

[zinc flax]

Have a nice rest

[tired quarry]

this is so overwhelming

one step at a time

[lyric zealot]

Hello and goodnight everyone!

[tired quarry]

how do i download these models https://huggingface.co/closertodeath

[lyric zealot]

Take it slowly, it’s not gonna go away

[lyric zealot]

You are at her profile page, go into one of the ctdmixs

And there should be a lot of files

File ending with safetensors/ckpt are most likely model

Unless it’s really small, like less than 1gb, then it’s likely Lora

You almost always wanna use file with safetensor tho

[untold glacier]

You should understand what you're downloading first. I'd start with Civitai.com, as it's easier to download and see what you're getting. I'd just make sure your browsing mode at the top right corner of the screen is on safe, to avoid most of the... more unwholesome elements of the site.

[tired quarry]

except there's like 20 safetensors

[lyric zealot]

Safetensor file will be “safe” from executable malware

[lyric zealot]

Yup, so you can pick one to try it out

[untold glacier]

Huggingface really isn't the best place to start when you're still trying to figure out how everything works.

[lyric zealot]

That’s true

Maybe save ctd profile as bookmark for now, then go to civitai and find models

Find something that you kinda like to test it out and get used to it

[untold glacier]

Yep. Start slow like that.

[lyric zealot]

Before using huggingface and stuff

[untold glacier]

Experiment. See how things work with your existing workflow and when you change things. Always keep the same seed number when testing, so you can see the difference your changes are making.

[lyric zealot]

@untold glacier if you are gonna stick around, teach about the vae as well

No idea tbh, I think most of her stuff are good, just need to get used to the prompting style

[tired quarry]

so i think im problem...

does my prompt suck

[untold glacier]

You need a VAE, like Smol was saying. Lemme pull one up.

[tired quarry]

i usually midjourney

thanks

[untold glacier]

https://civitai.com/models/23906/kl-f8-anime2

[lyric zealot]

This is already pretty good

[tired quarry]

where do i put the downloaded file?

[untold glacier]

In your VAE folder. Within your SD installation folder > Models.

[tired quarry]

for reference this is what im used to..

[untold glacier]

Then you'll need to select it within Auto's interface.

[lyric zealot]

Well, you will be able to achieve something similar soon

[tired quarry]

hopefully lol

you guys seem so knowledgable, im curious to see what your generations look like

[untold glacier]

Abe...

[lyric zealot]

In technical terms, there’s actually vae for all models. VAE is needed to convert those noise to pixel, or something like that, not sure the correct terms for that

[silk narwhal]

You got this ❤️

[untold glacier]

If you post multiple images, can you please post them in a single post? We're trying to talk at the same time, lol.

[lyric zealot]

Having knowledge doesn’t mean good generation

[untold glacier]

They can vary, as many of us like to change, try, and experiment with new stuff all the time.

[tired quarry]

anyone know why im getting a grey thing now

[unreal ridge]

tf do you mean, your gens are great

[untold glacier]

What have you changed since last gen?

[lyric zealot]

Did you interrupt it?

[tired quarry]

i just added the vae in the folder and restarted it

[tired quarry]

nope.. i ran it twice to see if it was a bug or smt

[untold glacier]

This takes me back... I seem to remember this meant something early in my days of genning. Lemme try and remember.

[lyric zealot]

I mean, it’s not bad but can be better if I’m not lazy appreciate it tho

[silk narwhal]

do this

[lyric zealot]

What’s your gpu?

[tired quarry]

3060

[lyric zealot]

Then you don’t need medvram

[tired quarry]

whats that

[lyric zealot]

You can just use xformers

[tired quarry]

ok

oh ok ok

[untold glacier]

Yeah, I agree with the commandlines above, aside from the --medvram and the --opt-split-attention.

[lyric zealot]

^

[untold glacier]

--opt-split-attention should be on by default (not require a command line) for quite a few versions now.

[tired quarry]

thanks for helping me btw guys

[lyric zealot]

Really needed is just xformers, other optimisation methods can be toggled on and off in settings when webui is up

[lyric zealot]

Except for xformers

[untold glacier]

Somehow my eyes glazed over the no half vae. Lol

DON'T put no half vae.

[tired quarry]

so the same thing happened...

[unreal ridge]

It's a private one, but I'm using some LoRAs tho

[untold glacier]

Yeah, there's a reason for it. Lemme try and find what I remember.

[tired quarry]

okk

[lyric zealot]

I don’t really remember this issue so

[untold glacier]

Ok, bad news. You're going to have to delete your venv folder and have it reinstall the guts by starting it again.

[unreal ridge]

https://civitai.com/models/24584?modelVersionId=29405 (1.0)
https://tusiart.com/models/605586978534138789

[lyric zealot]

[tired quarry]

great! i love that so much

[lyric zealot]

[unreal ridge]

[untold glacier]

Hahahaha... this is one time that I can feel the sarcasm through the internet.

[lyric zealot]

Okay, good luck and have fun

Imma go sleep fr fr

[untold glacier]

G'night again Smol!

[unreal ridge]

wait, u need to register now?

Oh well... You need a chinese number now... so...

Guess I can't use tusiart anymore

[untold glacier]

Does it need to be a Chinese one?

[unreal ridge]

Yes

[untold glacier]

[unreal ridge]

It fucking sucks

[untold glacier]

Imma try with my aMeRiCaN one.

[unreal ridge]

so many good shit in there locked behind a chinese number

[untold glacier]

You were right

WAIT

I just forgot the 1

[unreal ridge]

?

[tired quarry]

ITS STILL GREY

[untold glacier]

Well... I'm waiting for a text... but I probably won't get one.

[tired quarry]

I DID THAT FOR NOTHING

[untold glacier]

It's reinstalled already?

[tired quarry]

yeah

[untold glacier]

[unreal ridge]

Are you using the detail lora?

[unreal ridge]

I even tried using those public chinese numbers to see if I get lucky and nothing

[untold glacier]

Well, I'm not finding anything on the web. My next suggestion would be to start a new installation with a different folder, re-getting everything how you did before, whether redownloading or pulling from github.

It obviously worked for you at the start, so something changed after that.

[unreal ridge]

https://tusiart.com/models/609957850803901222 I really wanted to download this

[untold glacier]

Isn't that this...? https://civitai.com/models/101402/movie-style-filter

[unreal ridge]

[untold glacier]

It's even in the discord description of the link, lol.

[silk narwhal]

all of them are on civit i believe ..not under the same name

[untold glacier]

Yeah I think the other one was too.

[unreal ridge]

Not reading moment

[untold glacier]

That LoRA does look nice though.

[unreal ridge]

I hope it doesn't look like shit when I try it

[untold glacier]

Hmm, judging from the preview images it seems a bit overtrained. But it might still be useful at lower weights.

[unreal ridge]

[untold glacier]

Lol

Suspicions confirmed.

Probably need it at 0.6 or less to mitigate the overtraining.

[tired quarry]

@untold glacier turns out it was a problem with the checkpoint

[untold glacier]

Fantastic news.

I recommend Aurora, if you know booru tags. Or DarkAlfa if you want a more natural language prompting.

[tired quarry]

let's take the gens one step further. do you know how to remove the grain there

[untold glacier]

Are you using the VAE?

[tired quarry]

yes

[untold glacier]

Did I link the wrong VAE?

What model are you using?

[unreal ridge]

maybe he didn't select the vae?

[untold glacier]

Pretty sure that's the right VAE. Are you sure you're using it??

[tired quarry]

oh no it was js a bad model

[untold glacier]

And you for sure selected the VAE to use within the interface itself?

[tired quarry]

you had to do that?

[untold glacier]

Lol

[unreal ridge]

yes

[untold glacier]

Yep

[tired quarry]

🌚

where do you do that

[untold glacier]

I customized mine, so I can't remember. @unreal ridge ? Can you point it out?

[silk narwhal]

settings>VAE

[unreal ridge]

same

mine is at the top

[untold glacier]

Darn lol

[unreal ridge]

can't find on settings

[untold glacier]

Yeah, it's not there

[unreal ridge]

It's on settings... somewhere

[unreal ridge]

ye

SD VAE

select the one you downloaded

[tired quarry]

oh ok

[untold glacier]

Huh. So it just plain removes it from the settings if you put it at the top of the interface.

Nice album cover.

[tired quarry]

im getting better and better generations

[unreal ridge]

U using hires?

[tired quarry]

started from this

[tired quarry]

whats that

[silk narwhal]

it downscales the generation and then upscales it again for better quality

[tired quarry]

how can i check if that's enabled

[silk narwhal]

set steps higher than 0

[unreal ridge]

I use it at 15 steps/0.45~0.52 denoising

[untold glacier]

Also, try starting with this in your negative prompt area. Any further negative prompting you want try adding it in the area before the break:

BREAK
(disfigured, unclear, indistinct:1.3), (bad anatomy:1.3), misplaced hand, misplaced foot, (text:1.3), (signature:1.3), (title:1.3)```

[unreal ridge]

I don't remember where I downloaded my upscaler

[tired quarry]

alr let me try thta

[untold glacier]

Additionally (sorry, I know this is all a lot to start with) download this and put it in your Models > ESRGAN folder: https://openmodeldb.info/models/4x-Fatal-Anime

[silk narwhal]

i always download directly in the ESRGAN folder 😄

[tired quarry]

how do i fix this

[silk narwhal]

lower resolution used

[tired quarry]

i used 150 sampling steps 🙂

[unreal ridge]

U don't need to go that high

[tired quarry]

i wanna see the max potential

[topaz lichen]

Most of the samplers converge around 30.

[untold glacier]

Steps doesn't have anything to do with resolution, also, I don't recommend higher than 50 steps for anything, lol.

[silk narwhal]

32 steps is max for most models 😄

[untold glacier]

I'd say 50, but it really depends.

[topaz lichen]

Anything past convergence is a waste of time.

Yeah 50 it WILL have converged.

[untold glacier]

A good starting sampler is DPM++ 2M Karras at around 20-50 steps.

[topaz lichen]

Yeah any of the DPM samplers actually converge a bit earlier if I remember right.

25 steps is a good number for a DPM sampler.

[topaz lichen]

Also lower your res or use CPU offloading.

[untold glacier]

If you still run into memory issues, then at that point I would put medvram in the commandline arguments, and then restart SD.

As long as you're not trying to hires upscale at absurd levels.

[silk narwhal]

using vectorscope extension

[untold glacier]

I think he has the idea by now, lol.

[unreal ridge]

[silk narwhal]

Local AI image generation ..lots of experimentation and headaches .. but once you get it it just works

[untold glacier]

Oh right, I forgot to turn on Adetailer.

[tired quarry]

this still happens at 50 steps

[untold glacier]

Are you using hires fix?

[tired quarry]

yes

[untold glacier]

At what scale?

[tired quarry]

2x

[untold glacier]

Then, at that point I'd say: #🍥｜anime message

[silk narwhal]

weird solution but try enabling or disabling if ON gpu acceleration on your browser

[untold glacier]

Oh, and ensure you don't have anything else taking up your GPU memory. Firefox, other things, etc.

[tired quarry]

what's sd

[untold glacier]

Stable Diffusion

[tired quarry]

how od i check that

[silk narwhal]

switch to Opera GX 😉 not having any issues with that on my gtx 1070

[earnest orbit]

doggo

[topaz lichen]

I'm ComfyUI Pilled

[earnest orbit]

ive not used auto1111 in two months now

[unreal ridge]

My brain hurts whenever I look at comfyUI

[topaz lichen]

I liked the idea of Fooocus refiner technique but you can't use any LoRAs or anything else with it anyways.

[untold glacier]

Open Task Manager. Go to Details tab. Right click in an empty area where the sections are and click "Select Columns", then add the Dedicated GPU memory and Shared GPU memory columns.

[tired quarry]

im so close to that perfect render

why is it blurry though

[silk narwhal]

still generating?

preview is low quality till done

[tired quarry]

there's no select collumns for me (im on win 11 btw)

[tired quarry]

its done

[untold glacier]

Erm, firstly close the Refiner tab. Then if you want to use latent upscaling, then use Latent (nearest-exact).

[tired quarry]

idk what latent upscaling is

it was enabled on default

[untold glacier]

That's where it stretches the latent image (the image that it has in mind before generating it out) to fit the upscale. I'd recommend usually using an upscaling model like I linked earlier though.

I also recommend setting it to 0.4 or 0.5 denoise.

[tired quarry]

this is the progress so far:

[untold glacier]

Wow. Getting much better!

[silk narwhal]

looks good

vectorscope extension is freaky 😄

[untold glacier]

You have to make sure your task manager is showing you everything. I believe it's in some sort of simple mode by default.

Something's grabbing her hair though

[unreal ridge]

Man, can't think of any prompts

[untold glacier]

M I K U

[unreal ridge]

I'll just prompt anime girls with no background

[silk narwhal]

grab/modify anything from https://docs.qq.com/doc/DWHl3am5Zb05QbGVs

[unreal ridge]

I'll just include miku in some of the prompts I have saved

[untold glacier]

These prompts are ridiculous and artless.

[tired quarry]

i got this again

[untold glacier]

#🍥｜anime message

[tired quarry]

i did that already

[silk narwhal]

what settings are you using?

[tired quarry]

@silk narwhal

[silk narwhal]

did you add --medvram to the webui-user.bat commandline args yet (and restarted the webui)

[unreal ridge]

Moment

[untold glacier]

As Kuromi said, did you restart it?

[tired quarry]

yes

[untold glacier]

Well... then the only thing left to do is to get the Tiled Diffusion/VAE extension and set the settings lower.

[untold glacier]

Go to Extensions Tab > Available > and look for Multidiffusion Upscaler.

[silk narwhal]

Fooocus might be a bit better to use (has better optimizations for 4GB and 6GB VRAM cards

[untold glacier]

Could be I've not used it.

[silk narwhal]

needs more passes 🙂

[tired quarry]

i have 16gb of ram though

[untold glacier]

I love Auto1111's interface and workflow the best though.

[silk narwhal]

VRAM is what your GPU uses, RAM is seperate

[tired quarry]

how do i increase that then

[unreal ridge]

Goofy legs

[untold glacier]

Just download more VRAM.

[tired quarry]

HAHAH

[untold glacier]

I'm kidding. You'll need a new GPU.

But you could just follow what I said above, with the extension.

That will split the gen into tiles for the upscale.

[silk narwhal]

🙂

[tired quarry]

i have to restart sd after every gen to get rid of that

how do i clear the memory

[untold glacier]

Just so you know, this is what it looked like for me to gen something with similar settings to yours. It usually always takes 6gb to gen something at that size.

On Auto's build of SD anyway.

[untold glacier]

As far as I know, you clear the memory by restarting SD.
I haven't heard of any other way.

[silk narwhal]

https://github.com/IgorMundstein/WinMemoryCleaner/ this might help ..but not sure as it is not for VRAM

[untold glacier]

Uhhhh. Yeah I wouldn't use that.

[tired quarry]

how do i change from automatic111

i js checked and i have 6gb of vram

[untold glacier]

If you want to move to something like ComfyUI or others, I won't be of much help to you. All of these are different builds for SD.

It's a lot more complicated to figure out, but here's ComfyUI: https://github.com/comfyanonymous/ComfyUI

That depends on the size of the gen.

13-14.

[silk narwhal]

https://github.com/lllyasviel/Fooocus#download

one click install pretty much

[untold glacier]

Wait, that's for 640x640

Wait, no it's not.

Or maybe it is. It's hard to read.

[silk narwhal]

640 yea

😦

[untold glacier]

https://tenor.com/view/sad-anime-boy-gif-23701094

[untold glacier]

I was looking at it at first and thinking, "Well that's not too bad..." before seeing that the "s" and "it" were switched around, lol.

@supple raptor You seen this? https://developer.nvidia.com/blog/unlock-faster-image-generation-in-stable-diffusion-web-ui-with-nvidia-tensorrt/

[silk narwhal]

it is what is powering text to 3D scene experiments yea

[untold glacier]

TensorRT Extension for Stable Diffusion Web UI

https://nvidia.custhelp.com/app/answers/detail/a_id/5487

[silk narwhal]

4096 px frames in < 2ms on RTX 40xx and TensorRT cards

[unreal ridge]

Oh damn, 640x640 is pretty good

[wispy canopy]

i'm building engines for all resolutions i use xD

and because of that, this thing now exists

[unreal ridge]

[silk narwhal]

tasty

[unreal ridge]

wtf did I just watch

[silk narwhal]

blame ai video generation 😄

[wispy canopy]

xD

that's... why

WHY?!

what did humanity do to you, for you to punish us with... THAT

[untold glacier]

He is your creation. Do you not love him?

[wispy canopy]

i love all my children

[silk narwhal]

with a shotgun to the face

[wispy canopy]

all equally, just some abit more equal than others

i mmean...

[silk narwhal]

this next gen is coming out naicu

5 minutes -_-

[wispy canopy]

ai based animation?

[silk narwhal]

Pika Labs Ai image to video animation yea

this failed gen reminds me of system shock

[wispy canopy]

at least she's covering up her uhh... private parts

[silk narwhal]

too many upscale steps :/

[wispy canopy]

kinda overcooked much

[sonic portal]

hot

or would that be cold in this case?

[untold glacier]

Lol

[sonic portal]

hullo

i did a thing

[sonic portal]

the good model finally downloaded

[sonic portal]

why is cat in Venezuela?

[wispy canopy]

which one?

[sonic portal]

first one

[untold glacier]

[wispy canopy]

well, i did specify post-apocalyptic... i suppose the model has some thoughts about what resembles it

what makes it venezuelan, actually?

[earnest orbit]

Could not not understand why my upscaled pictures looked like someone washed them with soap, turned out custom VAE encoder kinda important sometimes X.x

[sonic portal]

it was just a joke cus the country isnt in a good place

would Chicago have been more accurate?

[wispy canopy]

this feels like lyricsprompting

[silk narwhal]

its not 😄

[topaz lichen]

she made herself as an android and now they've fallen in love (narcissism)

[sonic portal]

fires are dangerous

they should put them out

[silk narwhal]

better version ❤️

[untold glacier]

K Loki

[wispy canopy]

wildcarded the animals 😄

[sonic portal]

thats hot

[wispy canopy]

worked pretty well on jolteon... slowpoke however...

[untold glacier]

She's certainly on a mission.

[wispy canopy]

tactical panty-flashbang

[silk narwhal]

bit overcooked ...

[elfin bolt]

her trash bin is under attack

[wispy canopy]

pokemon!

[silk narwhal]

zoom out and credits 🙂

[sonic portal]

👀

hmmmmmmmm

[stable crane]

lmao

[stable crane]

oh cool

[untold glacier]

Difference between using TensorRT Unet and not. It was a pain to set up but seems to work now. Its limitations are many though, and I can't seem to build a Unet model that's large enough to use for a Hires. fix.

With and without Unet.

[wispy canopy]

dis spearrow

[wispy canopy]

unet?

the tensorrt engine stuff?

[untold glacier]

Yep. #🍥｜anime message

#🍥｜anime message

[wispy canopy]

i'm running on tensorrt right now

can't live without it

2x speed increase

[untold glacier]

You use it with Hires. fix?

[wispy canopy]

yeah, built 2x single models

using a mmodel that supports both is incredibly slow

(in my case)

but building the 2048 model took me all 24gb of vram i have

[untold glacier]

The RAM I used balooned to 40gb with python and the thing just froze when I tried to build a unet with it.

[wispy canopy]

yeah, it's SLOW and takes up huge amounts of resources

[untold glacier]

And I was just trying for a 1024x1536 hires model.

[untold glacier]

How slow is slow though? Did it freeze on you? Use your RAM in large amounts?

[wispy canopy]

it felt like freezing up... i think it was occupied for like 30 minutes

[stable crane]

same, someone mentioned to try tiled diffusion but it seemed like even if the tiles should fit the TRT, it didnt' work

[wispy canopy]

i mean... 4k image in 40 seconds is really worth it tho

are you using xformers?

[untold glacier]

Yep.

[wispy canopy]

for pure prompters like me, it's a god send

[untold glacier]

Sure, brag it up

[wispy canopy]

hmmm

1024x1024 did work?

[stable crane]

what's the non-TRT time?

was the key just building a 2048x2048 TRT?

[wispy canopy]

yeah, built a 1024x1024 model and a similar 2048x2048 trt

[stable crane]

does it still switch TRTs partway thru?

[wispy canopy]

switching models totally worth it on sdxl

yeah, it swithes on highres

[stable crane]

switching models
TRTs you mean?

[wispy canopy]

the single resolution model FAR outspeeds the dual resolution engines

yeah, TRT's

[untold glacier]

You use static or dynamic models?

[stable crane]

oh the static one

[wispy canopy]

dynamic -> else i can't make an engine that supports 75 to 450 tokens

but keep it on single res only

[wispy canopy]

WORTH.

thats over 2x speedup

@untold glacier this one suicided it's run on your pc?

[untold glacier]

I only tried a static one of those, not dynamic, and only with 150 tokens.

I'm going to make a few small sizes for aurora first before switching to trying to make a larger BreakdomainAnime one.

[wispy canopy]

do it like this

[untold glacier]

[wispy canopy]

keep optimal to 75, and only change max prompt -> this way it'll support 75-whatever tokens

if you're on HDD -> perhaps double resolution models, but if on SSD -> try making a model for every size, i feel that's WAY faster in inference

(for sdxl)

[untold glacier]

Most definitely an SSD.

Yeah, it almost seems like it would make SDXL viable on auto's build.

[wispy canopy]

i mean, 40 seconds for a 50 steps restart, 15 steps highres?

[untold glacier]

It's too bad I can't queue up these model builds, heh.

[wispy canopy]

1024x1024 -> 2048x2048

that makes sdxl very viably actually

[untold glacier]

What GPU you running?

[wispy canopy]

... yeah... about that

3090

i might have some vram to spare

[untold glacier]

Lol

Maybe just a little

3080's have less Tensor cores to use, so I'll see less benefit from it. Still, it's quicker.

[wispy canopy]

my pc has never had so many upgrades since sd/llm bullshit came around

[untold glacier]

[wispy canopy]

tensorrt is way quicker

https://tenor.com/view/money-crying-woody-harrelson-wiping-tears-with-money-gif-6945518

well, this one is interesting

[stable crane]

this is 2048?

[wispy canopy]

that's fro 1024x1536 only

but that makes sure you can run short and long prompts

[stable crane]

so 512x768 for highresfix?

[wispy canopy]

if you want to make it support both, you'd have to change it up a bit

[stable crane]

both what?

[wispy canopy]

both resolution

[stable crane]

oh

like, the 512x768 part, and the 2x highres part?

[wispy canopy]

yeah

[stable crane]

gotcha

have you tried tiled diffusion at all with it?

[wispy canopy]

sd1.5 might benefit from both in the same engine

[stable crane]

you'd think the tiles might fit into a smaller TRT

[wispy canopy]

sd1.5 is fster, so loading a new engine wont'win out

[stable crane]

I found the switching time of TRTs (even on SSD) to negate the perf gains

[wispy canopy]

but for sdxl, a specific engine per engine feels a lot faster

what's the switching time? mine feels nearly instant

[untold glacier]

[wispy canopy]

i mean... the full 2048 image is done in 40 seconds, including the engine switch

[wispy canopy]

solution -> no more loras

[untold glacier]

Imagine making a model for all 2k+ LoRAs at multiple resolutions.

[untold glacier]

https://tenor.com/view/anime-anifid-anime-damage-mad-angry-gif-17502176

[wispy canopy]

yeah... for lora abusers this tech is just totally useless

not just useless, garbage tier even

makes inspecting songs for songmaster material a lot easier tho

[untold glacier]

Why the hate

[wispy canopy]

prompting masterrace!!!!

[silk narwhal]

moving all lora's,models,vae's to stability matrix so i can use all of it in Fooocus/comfyUI and Auto1111 without duplicate all over my drive

[wispy canopy]

oh wait... that's literally just me 😢

[wispy canopy]

symlinking everythign?

[untold glacier]

https://tenor.com/view/jo-jo-gang-beat-up-anime-kick-gif-17369990

[wispy canopy]

saw on reddit sd.next had some updates... anyone tested it?

hmmm... this feels like a pretty huge fail for songmaster... only a single interesting result

[untold glacier]

Aside from using LoRAs though, it's always fun to push the boundaries of what a model is capable of. I've done that for Aurora and DarkAlfa. Composable diffusion is an underrated force.

[wispy canopy]

i think regional prompter supports lora's too

[untold glacier]

Also, why put the optimal tokens at 75 and not somewhere in-between?

[wispy canopy]

because it changes the minimum tokens with it -> that's what it did after i tested it

[untold glacier]

Ah

[wispy canopy]

and if your engine generation takes 30 minutes, it's a very costly bug to have

and most prompts are 75 or less tokens

[untold glacier]

ғᴏʀ ʏᴏᴜ

[untold glacier]

Most of mine are probably around 150 or less these days, since I use BREAK for everything.

[wispy canopy]

they should be for you too -> unless you're staying close to what you already did

[untold glacier]

Use BREAK. Reap its sweet benefits.

[stable crane]

I still don't get BREAK

[wispy canopy]

dude... i BREATHE the BREAK

[silk narwhal]

we all need a BREAK sometime

[stable crane]

brought to you by kit-kat

[untold glacier]

That increases the tokens though right? It's padding after each break to the next token limit.

[wispy canopy]

yes, what break actually does is pad the current group to 75 tokens

so that prompt automatically is 225 tokens

[stable crane]

how does this help me tho

[wispy canopy]

seperate style from subject and composition

[untold glacier]

Oops. Sorry, got talking with Eface lol. Lemme get you what I said yesterday.

[wispy canopy]

for me -> a HUGE thing. for anime model users, a lot less (i'm using a generic non-anime model)

[untold glacier]

#🍥｜anime message
#🍥｜anime message

Oh hey Nubby.

[stable crane]

oh interesting

ty @untold glacier

[supple raptor]

tokens have stronger influence the closer to the beginning of a block of 75 tokens they are and their effect cascades down so token 1 will have effect on something like tokens 2-20, then 2 will effect 3-21, etc. They have minimal effect on each other accross separate blocks of tokens so token 1 will have little effect on tokens 76-150.

[wispy canopy]

the word(token) blue on token position 1 in group 3 would like have a word with you... -> this is why BREAK is a good thing (hint: the image was... BLUE)

[silk narwhal]

So the most important word should be the first out of 75 , gotcha

[wispy canopy]

this is a good songmaster prompt!

[supple raptor]

Yep, the first 15 tokens or so are where all of the most important words should be and try to separate differently colored things throughout the block, usually with the ones that are inherently the strongest towards the end.

[silk narwhal]

https://mojim.com/usy255151x1x1.htm try for this song 🙂

[wispy canopy]

is generating ^^

[untold glacier]

You see that TensorRT Nvidia SD Unet model stuff yet?

It's a pain, but it's crazy.

[wispy canopy]

all depending on model too -> if a model responds very strong to specific word, it's generally not smart putting them in the first token position, best fixed by delaying their introduction until later [tokens:0.2]

[supple raptor]

Yeah, I gave up on setting it up because it was just giving errors and I'm not going to store and switch between 10M different ONNX models for the different resolutions, models, and loras that I want to use.

[silk narwhal]

wished windows file copy/move was as fast as most game installers ... why does a 140 GB game install faster than copying a 30 GB folder on the same drive

[wispy canopy]

(it's not nipples)

[untold glacier]

Other people had the errors too, it was a not too difficult fix according to here: https://github.com/NVIDIA/Stable-Diffusion-WebUI-TensorRT/issues/12#issuecomment-1768564187

*Fixed my link

After I had the extension installed, I just cmd lined:

pip uninstall tensorrt

pip cache purge

pip install --pre --extra-index-url https://pypi.nvidia.com tensorrt==9.0.1.post11.dev4

pip uninstall -y nvidia-cudnn-cu11```

[wispy canopy]

setting up tensorrt was a pain for real

[supple raptor]

I also have only 16GB RAM so even making the size that I would want would be a huge pain.

[wispy canopy]

the extesion is pretty fuckn buggy to set up

[silk narwhal]

this is the gens i got with those lyrics 😄

[supple raptor]

I usually gen at 544x960 with 2x hires fix so I'd need a 1088x1920 engine to run my hires fix.

[untold glacier]

Yeah, it froze up on me for a simple 1024x1536 hires fix model build earlier. But maybe I didn't let it run for long enough.

[wispy canopy]

gonna remove the backseat part tho -> it's horrible for ALL generation that have it

THERE IS TWO

no wonder i got it all the time

[wispy canopy]

i'm using my verser script

Tangled up in knots, feeling lost in this uncertainty. Frustrated, caught in twists of this life complexity

| Suffocating in the grip of my shattered dreams. Chasing rainbows, drowning in shattered moonbeams Tears streaming down, I'm searching for reasons why
| I'm breaking free from these chains. No matter how hard it gets I gotta keep on moving You can't take my life away I just wanna be free Please, set me free Let me be me
| Barely survive in this ocean of inconsistency. I'm the emotional storm, longing for simplicity Find more lyrics at
| Thrown myself in the backseat since forever. I'm drowning in the ocean of my own frustration In the darkness, I lost my way Those laughs keep echoing
| Set me free
| Tears streaming down, I'm searching for reasons why. No one to blame but me, myself and I
}```

[untold glacier]

Hey what starting res do you use for 16x9 with 1.x SD? I've alternated between a few, but since I'm trying to build a Unet model I need to narrow it down to one.
I'm leaning towards 800x448

[silk narwhal]

ah

[wispy canopy]

fixed lyrics 😮 they go into songmaster!

[untold glacier]

Ah, it looks like there are still stricter resolution limitations for TensorRT.

[supple raptor]

yep, needs multiples of 64 in both directions

[silk narwhal]

Just wait till we get consumer 64 GB VRAM TensorRT GPU's 😄

[wispy canopy]

yes. that consumer part was important

i mean..

[supple raptor]

544x960/1088x1920 is going to be the best option.

[silk narwhal]

put the gpu outside because it will act like a house heater on its own

[wispy canopy]

-> my house heating strategy for coming 'winter'

[stable crane]

that's 1 small blessing

[wispy canopy]

this looks like a fun prompt!

[untold glacier]

If it's a multiple of 64, it goes from 512 to 576. I'd have to find another set.

[supple raptor]

oh right. I have the dumb

[untold glacier]

No worries, I did the same thing lol.

[silk narwhal]

512px tiling 🙂

[supple raptor]

that's probably why I got an error when I set it up the first time. setup went smoothly and then I tried to make that engine.

[wispy canopy]

am i just that lazy clamping to multiples of 256 in ui.json?

[untold glacier]

That bottom right pic on the left side: in D&D when you cast Melf's Minute Meteors.

[untold glacier]

Ah, could be.

[supple raptor]

then after it broke my auto-complete extension, I uninstalled it and had to rebuild venv and I haven't been able to get it to install again since.

[wispy canopy]

second run on this prompt. damn, this one can run solo!

[stable crane]

I wish the clipboard handling in A1111 was more reliable

maybe it's cause I'm using webp that I can't copy-paste from txt2img to img2img

and you can't drag and drop from 1 website into A1111

[wispy canopy]

this is whatever is on your clipboard indeed

copying a webp is not the same as copying a png or jpg

[stable crane]

I imagine they just didn't implement a paste handler for webp?

because you can drag a webp from FS just fine

[untold glacier]

Alright. Wish me luck, I'm trying for a 576x1024 build.

[supple raptor]

you can though, at least with PNG. I do that all the time to grab PNG info from discord images.

[wispy canopy]

good luck! i hope it succeeds

[stable crane]

hm can you drag one of my imgs?

[wispy canopy]

never got that to work

[supple raptor]

yep

[wispy canopy]

the lower left tho... that one vibes with me hard

[stable crane]

hm

I can't in firefox

are you dragging from discord app or discord website? I'm on website

[wispy canopy]

same for me, on firefox

[supple raptor]

Gradio hates browsers that aren't Chrome

[supple raptor]

click image, open in browser, drag from that browser into webui.

[wispy canopy]

🥳 SONGMASTER is over 500 lines!

[untold glacier]

At least it's progressing now

[wispy canopy]

i had this too -> it's going

[untold glacier]

Somehow, it's not really using my GPU like it should though...

[wispy canopy]

it deferred some tstuff to ram

[stable crane]

I should give TRT another chance

[wispy canopy]

it's going to be slow from now on -> but it will keep running

[stable crane]

because I can use lora in txt2img, then TRT + raw model in img2img to upscale it

[wispy canopy]

you should... the speed is just insane... it just comes with so many restrictions 😦

[stable crane]

I liked the speed but I was put off by lora and TRT switching

[supple raptor]

I think TRT engine building is handled on CPU because of the very high RAM requirements. Unrealistic for most consumers to use it if it had to happen in VRAM.

[stable crane]

but my current upscaling for these recent images has been to img2img without lora so they don't blur the image as much

so if I can upscale those faster, win

[wispy canopy]

it tries some of the things in vram, but falls back to ram if you run out mildly

[stable crane]

before

[wispy canopy]

this is a songmaster fail. DENIED

[untold glacier]

Well it's done. That was a lot quicker than I though it would be, even if it's only 576x1080.

Uh

What the heck did I just say? Lol 1024

[wispy canopy]

got a nice upscale engine now?

[gloomy scroll]

I think it's firefox that hates clipboard

[untold glacier]

It was a starting resolution.

[wispy canopy]

good fucking luck on the final resolution

[untold glacier]

It's a minimum for 16x9.

[wispy canopy]

if that one already crapped out your vram

[untold glacier]

Well... at least I have 64gb of RAM.

I've been able to run a 65b LLM, albeit slowly, heh.

[wispy canopy]

i feel this is ajust a tiny bit unrealsitic, i mean, those are 40kg of pauldrons...

[stable crane]

I need to get back into local llms

[wispy canopy]

totally unrealistic

[supple raptor]

Gradio has lots of problems in non-Chrome browsers. I've even seen plenty of problems on chromium browsers.

[stable crane]

haha

[gloomy scroll]

Haven't rly tried myself but I know that firefox clipboard doesn't include image data

[stable crane]

wym? I regularly paste images from clipboard into discord

and from firefox -> r-click copy -> back into discord

[gloomy scroll]

I used hours in work project to get paste into firefox working but file data didn't exists anywhere

[wispy canopy]

-> copies the actual png/jpg to your clipboard

[stable crane]

yeah sometimes converts file types on me

[wispy canopy]

lower left ❤️

[wispy canopy]

full prompt for the variations, or just the lower left one? XD

[gloomy scroll]

Left bot

[wispy canopy]

BREAK Where we go, what we find, what we will leave behind is. How we choose to define the journey within our mind Where we go, what we find, what we will leave behind is How we choose to define the journey within our mind```

[gloomy scroll]

And now it goes to never used prompts list

[wispy canopy]

it certainly hit my songmmaster list ^^

songmaster goes very well with prompt alternation

[your actual prompt | songmaster bullshit]

gives some very nice flavour

and, i get to listen to new songs too!

[wispy canopy]

songmaster fail, nice crow tho

one of these prompt parts might have a very strong effect...

[wispy canopy]

the question is ... WHY is she blushing that hard...

[supple raptor]

she just realized I was following her

[supple raptor]

https://civitai.com/models/110244?modelVersionId=118849

It'll look pretty different on most other models. Probably my favorite LoRA.

[supple raptor]

tbh, I haven't bothered to set up adetailer in comfy so I'm not sure what it's supposed to look like.

[silk narwhal]

this is fine,right

[dawn jasper]

We love the hair for both the women. 😊

[untold glacier]

Photographed mid-blink

[cyan burrow]

So how has the SD scene progressed while I was gone for a few months?

[supple raptor]

aside from new models and loras, there's a couple of new extensions that claim to improve quality (FreeU mainly), and TensorRT came out which can give a big speed boost if you don't mind dealing with a lot of downsides.

[cyan burrow]

what downsides?

[supple raptor]

You have to build an "engine" which is an ONNX model that's only compatible with pre-specified resolutions and prompt lengths so you probably need several for each model, lora, resolution, and prompt length combination you want to use and realistically 2 for each of those combinations if you want to use Hires Fix. If you want to use LoRA you have to build them into the model at the weight you want to use them. The possible resolutions are much more restrictive than using Pytorch now days. It takes a lot of ram to build the "engines". Each "engine" is almost 2GB for 1.5 based models. It doesn't work with a lot of really useful extensions at all. And it's a pain to install.

But hey, you can get almost double the generation speed on a 4090 (increases are less the slower your card is)

[silk narwhal]

aka ... first release SD hassles for next gen cards 😄

[supple raptor]

yep, pretty much

[cyan burrow]

ah seems like a lot more hassle than a passing amateur like myself could deal with

thanks for the information

[supple raptor]

This extension is kinda neat though. https://github.com/ljleb/sd-webui-freeu.git

[silk narwhal]

no idea, how to use it 😄

[supple raptor]

me neither I just turned it on and used some settings someone else shared for a while

[cyan burrow]

I'm rather curious about stable audio and any AI speech modulators/t2speech that are open source

[cyan burrow]

though I guess having FL Studio and a midi keyboard of good quality I could just make my own sounds

and a ton of plugins of virtual synths

I just need to get autotune or whatever is the best one now

[silk narwhal]

synplant or something is getting traction recently

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

[cyan burrow]

does that work on voices?

I mean I want to not only generate music, but I want to generate speech, or experiment with generating singing

so I need something that does speech

[silk narwhal]

it is a more advanced Vocaloid synth so .. i guess it would do speech (singing)

[cyan burrow]

ah

can it do voice sampling? to learn a voice and make a model

[silk narwhal]

as far as i understand it samples the audio input and breaks it all the way down to the tones that is needed to make that sound

[cyan burrow]

okay so there's quite a learning curve

is it free?

[silk narwhal]

free trial / $149

[cyan burrow]

cause I was asking about open source

[silk narwhal]

https://soniccharge.com/synplant

[cyan burrow]

something that'd be free

[silk narwhal]

https://www.bespokesynth.com maybe?

[cyan burrow]

I think you're misunderstanding what I'm asking for here. I'm asking for human voice speech generation

in open source

cause the paid options are real expensive

[silk narwhal]

not sure any are released publicly yet... meta was planning to i believe

[cyan burrow]

hmmm...I thought there were at least one or two that I don't remember the name of. That deal in both text2speech and speech2speech

[silk narwhal]

https://github.com/suno-ai/bark

is kinda meh though

[native halo]

like RVC?

[cyan burrow]

I dunno

[native halo]

just search rvc on youtube and you will see it can also do TTS

[cyan burrow]

but can it do speech2speech?

[silk narwhal]

rvc-v2 yea is pretty good

[native halo]

yea thats where its better

[silk narwhal]

https://github.com/Tiger14n/RVC-GUI/releases

[native halo]

this webui does tts with rvc https://github.com/litagin02/rvc-tts-webui

[cyan burrow]

can it train voices too somewhere?

[native halo]

yes it can train

[cyan burrow]

nice

thanks

https://github.com/WadRex/RVCompact looking at this one now

Im guessing RVC V2 refers to the model

with the interface still being the interface

[silk narwhal]

yes

[cyan burrow]

https://github.com/Mangio621/Mangio-RVC-Fork this might be better

[silk narwhal]

Mangio is decent yea

i use this mostly but my pc can barely handle it in realtime: https://github.com/w-okada/voice-changer

a 20xx+ graphics card definitely recommended

[cyan burrow]

does it exist in English?

[silk narwhal]

yea.is english too

https://github.com/w-okada/voice-changer/blob/master/README_en.md

[cyan burrow]

cool, thanks again

[silk narwhal]

when these two meet..things get hot

[silk narwhal]

from fire to water

[native halo]

the miku hatsune

[silk narwhal]

g' night all

[nova remnant]

@untold glacier hello

[keen hollow]

evenin'

[untold glacier]

[wispy canopy]

gmorning

[untold glacier]

[wispy canopy]

2 minutes render and have freeu, or 1 minute render, and not have freeu

i'm very inclined to say -> freeu quality upgrade way better than tensorrt speed

[untold glacier]

The first image is better. Is that with FreeU?

[wispy canopy]

yeah

[untold glacier]

I just haven't seen a marked improvement with Anime models.

[wispy canopy]

... it's not 😛

lets make it a bit more fair comparison, to compensate for the tensorrt speed, i'll halve the steps freeu gets

if it's still better quality, i'm sold

[untold glacier]

Let me rephrase. I haven't seen an improvement anything like the images you show up there, with the anime models that I've used.

[wispy canopy]

i suppose freeu stabilized my stupid prompts a lot 😮

[untold glacier]

Have you tried it with different models?

[wispy canopy]

just downloaded it, so taking it for a spin

[untold glacier]

[wispy canopy]

have only been using brightprotonuke for like a month now

serves me for all purposes

[untold glacier]

Guess I should try with DarkAlfa to see if that's any different. I've only tried it on Booru-based anime models thus far. It works better on Aurora than BreakdomainAnime, but it's not much and it's not better most of the time. Or even half the time.

[wispy canopy]

your milage may vary ^^

[nova remnant]

eface

[wispy canopy]

tensorrt (no freeu) 1:20 VS freeu low steps 1:24 VS freeu high steps 2:14

i feel freeu gets a LOT of the smaller details right

[untold glacier]

Hmm. The implement I'm using to try FreeU doesn't have step options.

That could be the difference.

[wispy canopy]

nah, just sampling steps

to make up for tensorrt's speed, i just halved the amount of steps to get comparable speed

[untold glacier]

Interestingly, there's a large difference in FreeU vs not with my gens. It's not subtle changes.

Some of them almost look like completely different seeds.

[wispy canopy]

yeah, had that happen a couple of times too

the interpretation just being vastly different, however, better adhering to prompt imho

[untold glacier]

For instance: No FreeU

FreeU

With DarkAlfa.

Certainly harder to tell if it's better or not when the entire image changes, lol.

[wispy canopy]

same seeds?

[untold glacier]

Yep. No variation.

[wispy canopy]

wow, some are completely different

[untold glacier]

[wispy canopy]

non vs freeu also different outputs

[untold glacier]

I dunno. Her torso didn't turn into a head covering up half the screen lol. She's still in the same position, head turned left with her eyes closed. Backless dress, etc.

[bitter latch]

dare I ask what Freeu is?

[wispy canopy]

yeah, i suppose fair

[untold glacier]

https://github.com/ljleb/sd-webui-freeu

And actual info here: https://github.com/ChenyangSi/FreeU

[wispy canopy]

sampler extension that seems to improve output quality

[untold glacier]

That's what they claim. But I haven't seen it with any of my gens yet 😛

[wispy canopy]

i see it tho

[untold glacier]

BUT I DON'T

[wispy canopy]

small details get resolved better

try without loras 😛

[untold glacier]

I know what you're saying. I've been looking for it myself.

[untold glacier]

I haven't been using any

[wispy canopy]

goooooooood

[bitter latch]

ahhh ok

[untold glacier]

It's PART OF THE TESTING.

[wispy canopy]

but i love whacky output nonetheless ❤️

[bitter latch]

Pon YOOOOOO

[untold glacier]

Lemme ask ya, what's your settings? Maybe that's my problem.

[wispy canopy]

just press the recommendations and the ✅

that's what i did

[untold glacier]

Maybe I'm crazy... but I'm already seeing an improvement from the recent gens.

I'm probably crazy.

[bitter latch]

if you are here.. if not .. soon will be 😄

crazy that is

😄

[wispy canopy]

nope, it's certainly improvement

[untold glacier]

Well, there are aspects of that pic that I like more in the first image, but I get what you're saying. I'm just wondering if it's a model-specific thing; where it won't really improve anime models. But I'm still testing.

[bitter latch]

cd tuner at 0 for details

[wispy canopy]

it missed a part of the prompt in the first image, flat color, so i feel it adheres a lot better

[untold glacier]

Ah, I getcha.

[wispy canopy]

and the shadow seems a bit better thought out

and again, better resolved details on the grass

[untold glacier]

So far after the adjustments to get it (I think) closer to your settings I'm liking things more. It's difficult to know if the settings were changed correctly though, as my interface is different, heh.

[wispy canopy]

i wonder why

[bitter latch]

cd tuner at 0.5

[wispy canopy]

i'm using this one https://github.com/ljleb/sd-webui-freeu

[bitter latch]

hmm not much diff

[wispy canopy]

fixed pole in background, better spacing on wall left, hair ornament less blurry, same for arm, actual crease in the skirt

not a lot, but all these small details get resolved better

[untold glacier]

Lol

[bitter latch]

cd tuner 5 . umm

hmm to much

[wispy canopy]

now if only tensorrt could work with extension 😢

[bitter latch]

tries it with adetailer , hi-res fix and few other setting change, cd tuner 1

[untold glacier]

I dunno man. Look at the presentation, the cohesion of the hands, and the less-baked nature:
No FreeU.

FreeU

[bitter latch]

oh boy tis failed.. umm (meaning mine)

https://tenor.com/bbDlz.gif

[untold glacier]

It could be that the prompt is bad, and so the FreeU version is getting closer to a bad prompt. I'll have to try a different one.

[bitter latch]

grinz evilly and hits a button

and a few more, changes prompt.. and waits

blinks

blinks

BLINKS

[nova remnant]

@untold glacier what are some good anime models one that makes good, not weirdly shaped bodies

[bitter latch]

MOOO

[untold glacier]

If you want Booru-tag prompting, Aurora. If you want more natural language prompting, DarkAlfa.
https://civitai.com/models/40199/aurora
https://civitai.com/models/64843?modelVersionId=74900

[nova remnant]

ree

I don't have those

I can't download them for this service

[bitter latch]

leaves this here

[wispy canopy]

a LOT of regional prompting and dumb stuff going on (as always with my prompts) freeu hard winning this one

[untold glacier]

I'm giving you 16 comparisons at a time, and you're giving me one back lol. I just don't see it yet.

[wispy canopy]

because that takes 8 minutes

[untold glacier]

Fair

[native halo]

https://tenor.com/view/ez-noob-skill-issue-simply-a-difference-in-skill-gif-20780203

[wispy canopy]

my pc is burning non-stop

[dawn jasper]

She is strong.

[untold glacier]

[nova remnant]

is milk on the menu

[nova remnant]

[native halo]

straight from the source

[bitter latch]

MOO

[untold glacier]

No FreeU

FreeU

It certainly darkens the images a fair amount.

[bitter latch]

oh I had the seed set. .DOH!

[native halo]

i wanna eat that peach

[bitter latch]

and transformation complete?

my adhd brain went from one to the next and o...k

[nova remnant]

[bitter latch]

takes away a word, adds a word

presses button

[wispy canopy]

it's called addiction

[untold glacier]

When you plan for things to get darker, it's a bit better to work with. I'm still not convinced though... I'll have to give it time and letting the community figure out what settings are best so I can compare then and with more models.
I'm heading off for the night. Goodnight peeps!

[wispy canopy]

sleep well

[bitter latch]

I SHOULD be in bed but... sigh

[wispy canopy]

@untold glacier here's a grid for ya ❤️

[wispy canopy]

basically, i feel coherence and small details are improved pretty much

[bitter latch]

seed 8675309

[untold glacier]

Yep. A definite improvement with nearly all.

Might be the method of prompting. I'm using mostly exclusively booru-tags with mine.

[wispy canopy]

booru tags are very restrictive tho

[untold glacier]

Restrictive, but focused with the proper booru-prompting anime models like Aurora.

[wispy canopy]

i can't work with them, i need my models to be free 😛

[bitter latch]

I just use RANDOM Things 😄

[untold glacier]

Hence darkalfa. And things actually looked better using it with that. With the brief test I did anyway.

[wispy canopy]

using booru tags on darkalfa? 😮

[untold glacier]

Shhhhh

[wispy canopy]

NOOOOOOOO WHAT HAVE THEY DONE TO YOU MY BABY

[untold glacier]

[bitter latch]

umm--> bedtime hoodie, Yuki Miku, HUGW bobble head, unreal engine render , 8k, 3.5D

DOH!

[nova remnant]

just tried a more medieval prompt

with denim shorts in it

🤦

[native halo]

where is it i dont see it

[nova remnant]

It's dead

[untold glacier]

Yep. Definitely better following the prompt and fixing issues with DarkAlfa, at least with the small amount of additional testing here.

Ok. I really am going to bed now. G'night peeps.

[wispy canopy]

yeah, fixed up a lot of small things, doesn't it?

gnight @untold glacier

[untold glacier]

[wispy canopy]

why does it look like she is sneaking?

[nova remnant]

She doesn't want anyone to notice that she is a fox

[nova remnant]

[bitter latch]

hre ya go

[nova remnant]

I find it funny how detailed SD makes armpits

x3

[wispy canopy]

heh

[bitter latch]

ok this one came out great

almost

[zinc flax]

AI secretly abbreviates to armpit illustrator

[zinc flax]

Scary

[nova remnant]

types armpit in the prompt

👍

[zinc flax]

arms behind head, sleeveless

[native halo]

(giant,massive,gigantic,gargantuan armpit:4.5)

[bitter latch]

umm.. wht?

[nova remnant]

Everything seems fine

[bitter latch]

huh.. muku? what?

interesting

Little Red RIding Muku

[nova remnant]

@native halo rate this moon

[native halo]

i love that massive moon

[zinc flax]

Very important pose

[native halo]

what a friendly creature he even created a bonfire so we can rest🤠

[bitter latch]

me thinks something went wrong

ok off to bed. nite

[nova remnant]

👍

[nova remnant]

more moons

@native halo Are smaller moons also relevant ?

Enough

[native halo]

beautiful round perfect circunference moon

[near elbow]

just "a 1980s manga magazine" 🙂

adding "gundam"

[wispy canopy]

👈

[near elbow]

too bad about the right arm