🍥｜anime

[steady grail]

now i need seconds even to detail upscaled faces

[rich tartan]

That's perfect anatomy ngl

[steady grail]

the same set up of yesterday, but i added white... oh, doves, they are doves

[native halo]

heres another

[steady grail]

i don't know if i can call this image holy, or cursed

maybe an abandoned chapel

[rich tartan]

is it bc of lora or model?

[native halo]

both

[steady grail]

it's weird, just addng the birds, the model transformed the colored lights in them!

from this to that

[nova remnant]

@wooden coral @dull jackal bad link in chat

[quick hill]

[wooden coral]

[steady grail]

i wonder what happens if i randomize the noise of the upscaler

[steady grail]

how the?.....

[steady grail]

i absolutely love comfy

it's a constant source of surprises

a part of my want to cover known characters in atipical misterys, but the dominant one just want to push the absolute chaos out of ai generation XD

the next thing i implement, are wild cards

[nova remnant]

https://tenor.com/view/bunny-bnuuy-steals-your-bunny-steals-your-bnuuy-gif-26209043

@gritty spruce @wooden coral @dull jackal another one

[ornate echo]

no can has bunbun

[dull jackal]

But...what if there's...nothing to click!?

[steady grail]

lol

[dull jackal]

Santa's givin' that person coal

[dull jackal]

This is so charming, esp with the bunny

bunny 💖

[steady grail]

yup lol

i should gen bunnys too

[ornate echo]

It's based on this girl I know. She has this bunny she loves.

[dull jackal]

Also, yeah, Comfy is so great!

I always wanted to have a bunny. I had a bff who had one when I was a kid, so I've always loved them. They're so soft and adorabibs, so how can you not?

[ornate echo]

They have their downsides, but they're cute.

[steady grail]

i had 2, but they were big, and chewed all my pc cables, so, i gave them back

not a thing for me

i prefer my cats, that constantly drop vases and cups on the floor, but i resolved with plastic once lol

[dull jackal]

haha, yeah, small animals will be out for your cables, for sure! Cats, dogs, and rats.

[dull jackal]

Well, you know it's a rite of passage when your cat takes something and whaps it to the floor.

[steady grail]

no more glass or porcellain for me XD

well, they are cats. 10 times better than lara croft XD

[dull jackal]

You'll get no disagreement from me

I love animals

[steady grail]

me too ❤️

[dull jackal]

haha, high five!

[steady grail]

[dull jackal]

What's your fave!?

[steady grail]

i'm a cat guys, always grown cats from childhood

they once i find on the streets

every time the time takes them away from me, it's a trauma, but even that it's love

what they game me, will always be part of me, now and ever

[dull jackal]

It's always tough when that happens, but I will always love all the animals I've owned.

Gorgeous! Love all the traditional art aspects/blending going on.

[steady grail]

this workflow is insane. i don't know what's going on myself

[dull jackal]

Tell me about it.

[steady grail]

i'll be fast showing you a screen XD

[dull jackal]

sweet!

[dull jackal]

cute!

[zinc flax]

A different kind of bunny lol

[steady grail]

and this is still an incomplete project

some nodes may be connected mistakenly

and some settings could potentially even damage less performing pc

and if my pc is in a mid low range

[vital raptor]

but what does it do

[steady grail]

lcm and coadapter to set a pose and depth to the image

than i pass it on 3 ksampler, 1 first starts the work in exponential, the second in karras, the third is connect to a new seed and the gen with a small upscale, to ad detail

the result get depth, over imposed images, like the birds in this case and absolutely not prompted details, that gives more complexity to the image

i even added concat, when i add single details that gives even more randomness to the end result

in this case, in concat i added birds and the gen spawned birds everywhere

maybe i should add regional on concat to add details in some precise zones

and use wild cards on the other 2 cocat to make things even more messier

in the end, i just have upscaler and face adetailer

on the top right, ipadapter

[steady grail]

cute lol

adding reginal prompting on the concats, influenced by wild cards

i can't wait to assemble it

i love comfy. i can't do such things on auto

[dull jackal]

I like it!

Sorry, had to do real life for a bit

[steady grail]

ah no problem, i'm generating in multy tasking, when i'm reading and listening

time is never enough

the true madness, is in setting every node

as if connecting them is not enough

[dull jackal]

I can see the various different aspects in your image---it's certainly a lot of effort

but I like the overall output

and the depth to the background

And again, c'est la vie, but....FAREWELL

[steady grail]

interesting

[rich tartan]

what tf

optimization +10mil

[steady grail]

it's never enough

when i finish with existing nodes, i will study how to develope nodes myself XD

[rich tartan]

Am i using somehow a bad model or sth.. I only get bugged anatomy(literally)

What are yall using for fixing extra libs/legs/hands/fingers?

[steady grail]

negative prompts?

usually i fix them manually

when i feel it lol

[rich tartan]

uh, wait. Currently on win 10. Gonna need some time as im unpacking something

[nova remnant]

Can you send an image with info that you generated

[steady grail]

face adetailing this will be hell

i knew, adetailer is capturing every single face lol

[untold glacier]

Well... that's not a bad thing here. It looks great.

It's doing what it was made to do. Don't take that away from it, lol.

[steady grail]

uahahah

damn, my obsession on precision

i don't like the end result 😦

[untold glacier]

Well, zooming in, it looks like you don't really have a high denoise on face detailer. Put that up to 0.4-0.5 and have the steps at 12 or so. That should fix the problem for the faces that it's actually identifying correctly.

[steady grail]

i lost the old lady on the right

[steady grail]

now, she became a yung nun 😦

let's try with more denoise

[unreal ridge]

old lady?

[rich tartan]

lowres, polar lowres, bad anatomy, bad face, bad hands, bad body, bad feet, bad proportions, {bad leg}, {more legs}, worst quality, low quality, normal quality, gross proportions, blurry, poorly drawn, text,error, missing fingers, missing arms, missing legs, short legs, extra digit, 2girls, long skirt, out of frame, only_upper body, only_lower body, [petite], low background, distorted perspective, dynamic angle, dynamic pose, nsfw,

now on ubuntu

[untold glacier]

Oh the horrors

[steady grail]

eheheh

oh, the higher denoise is working

[untold glacier]

I mean, it looks like you just need a simple full upscale first before you start inpainting faces.

[steady grail]

for the old lady, well, it's a miracle

[rich tartan]

and an example of it:

or this

or here the face and the hand

[steady grail]

oh, i can set an expression wild card on adetailer

that would be funny XD

[rich tartan]

that's literally insane, i don't see much of fuck ups

[unreal ridge]

I wish I could understand comfy metadata

[nova remnant]

oh it's comfy

[unreal ridge]

isn't 12 cfg kind of high?

[rich tartan]

not really, the first gen needs it ig

getting upscaled by model and then downscaled latent for more details

[unreal ridge]

my man yumeshiro writing the whole bible

👀

[rich tartan]

soo.. what you guys using and not getting ana faults?

[unreal ridge]

This

The only thing I change depending of what i'm going for is the resolution/num of steps

[untold glacier]

How Do I Get Good Hands with a 1.x Anime Model?

Ah yes, a question as old as the time of AI itself.

If you're using 1.x: First, you need a killer model; one that's extremely cohesive like Aurora or DarkAlfa. Then you need to prompt for the model. If it's a booru-tagging model like Aurora, try to use as little natural-language prompting as possible (describing the subject or scene in sentences are a no.) If it's a natural-language model like DarkAlfa, use short descriptive sentences combined with booru tags to help it along.

Next, a decent negative embed will help it along in the direction that you're wanting it to go; something like negative_hand works well, if subtly. An additional general negative like AuroraNegative or EasyNegative also helps.

Additionally, a solid negative prompt that has as little as you can while still taking care of potential issues is the goal. Here's what I use as a starting neg prompt in Auto1111, but I often cut it down a bit further before genning.

BREAK
(disfigured, unclear, indistinct:1.3), ugly hands, extra arm, extra hand, split arm, missing finger, extra finger, three fingers, four fingers, six fingers, merged fingers, (bad anatomy:1.3), misplaced hand, misplaced foot, (text:1.3), (signature:1.3), (title:1.3)```

Final tips: 
- Use the CFG you want, but know that the lower you get, the less cohesive, with more potential for bad hands and indistinct features. 
- Use an amount of steps that works for the sampler that you're using. 20 is a good baseline, but if you want more details then 40 is good as well for samplers like 2M Karras. If you ever use any that end in Exponential, then increase the steps a bit, as it requires more.
- Check network extensions. Sometimes it's a bad LoRA or incohesive embed that can mess things up.
- Be aware of your settings as a whole. There are many moving parts that can affect things.

[unreal ridge]

Also, keep in mind that pretty much any kind of weapon comes out looking like shit if you're not on sdxl

😔

[untold glacier]

Negative Hand embed: #🍥｜anime message
Aurora and EasyNegative: #🍥｜anime message

[rich tartan]

i just saw on civitai that some people on the model i use, use badhandv4, should i also do that?

or only one at time?

[untold glacier]

Additional guides of mine here.

• **How to Upscale with the Most Detail and Least Problems: **
  #🍥｜anime message

• Help I Can't Increase My Contrast when Generating!
  #🍥｜anime message

• What 1.x Anime Model Should I Use??
  #🍥｜anime message

[untold glacier]

Don't use badhandv4. Just use the negative_hand that I linked. And never use more than one bad hand embed.

[rich tartan]

alr thanks, i'll try out

hmm, i just generated some interesting image, it's just hiding the hands.. (No embedding currently, my internet is busy with another model)

[steady grail]

i need to set better the adetailer

gotta go

see ya

[untold glacier]

[unreal ridge]

Man, I just realized that I like horny cartoon-ish models

[nova remnant]

welcome to the reality of entire humanity

[rich tartan]

can i add just the prompt in comfyui or how do i include the neghand?

[unreal ridge]

I feel like it's easier to make mature looking characters in them

🤔

[untold glacier]

If you have the file the same name as what's in my neg prompt then it'll work. Just make sure you've restarted your ComfyUI after you've put it in the embeddings folder.

[brittle rampart]

you do it!!! bravo!!!

[untold glacier]

Oh, and if you're using comfy, you'll need to concatenate the negative prompt clip conditioning instead of using BREAK.

[rich tartan]

um, what does that mean?

[untold glacier]

You'll need something like this. If you search "concat", you'll find the node. It essentially combines two CLIP Text Encode conditionings into one.

If you don't already know, you can double click empty space in the workspace to manually search for a node to place.

[rich tartan]

oh alr, now i understood. Im not using any breaks tho

or ym i should do it?

[untold glacier]

Right, you'd put what's before the BREAK in one and what's after the BREAK in the other.

[rich tartan]

Ig it worked, does comfy also give meta data out? if yes how'd i read them?

wait nvm, it clearly worked

not perfect, but much better as double/mixed up hands

[unreal ridge]

it definitely improved 👍

[rich tartan]

{{{{{{{lens flare}}}}}}} what exactly do those {}? Do they make the weight go up or down?

[nova remnant]

https://www.aidemos.info/what-are-parentheses-and-brackets-in-stable-diffusion/

[untold glacier]

DON'T use {} use () to increase attention.

I recommend not using silly amounts of brackets though.

[quick hill]

Neither

[untold glacier]

Just use something like (black and white:1.3) or (skinny:0.5) to increase or decrease weighting.

[quick hill]

(((((((((())))))))))) for curse image

[untold glacier]

Hahahaha

[rich tartan]

it makes my eye ache to read the prompt, does this mean it would be 1.7 because of 7 {?

[untold glacier]

Eh, something close.

That's just a stupid amount though.

[rich tartan]

**{} is NAI’s “implementation” of () **what is nai and what does that exactly mean?

[untold glacier]

NovelAI diffusion. Don't worry about it--it's not a local thing that we have access to with SD.

[quick hill]

Novel AI, probably wayyyyy back

[rich tartan]

oh, alr

[brittle rampart]

I'm sorry for interfering. But what about this syntax - [blue hair:medium:1.25], for example

what mean - medium:1.25

[nova remnant]

general purpose with weights would usually require a value between 1.0-3.0 max on the incremental sides and you would rarely go below 0.1

[untold glacier]

That's prompt editing. https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing

In short, it allows you to change that part of the prompt over time, from something to something else.

[untold glacier]

When prompting, I generally don't recommend going above 1.5 in weighting. If you are, then it's likely that the problem is elsewhere.

[rich tartan]

alright.. now my generation is fkd up

[untold glacier]

What did you do

[rich tartan]

changed all the {} to weight

[quick hill]

lol..

[untold glacier]

Show me your prompt

[quick hill]

Model issue

[untold glacier]

Yep. Or a network extension.

[nova remnant]

general idea is that (test) is actually a weight of 1 and there are cases of (((test))) just a general outline of how the weight works, but yes going above 1.5 is not necessary

[rich tartan]

the one up there i left with {} as i was a little double confused

[untold glacier]

Holy cow. Ok, get rid of all the {}. Make them disappear.

Not the parentheses. The brackets.

Not (), but {}

Anything that's over 1.4, bring it down to 1.4.

Or even down to 1.3.

[rich tartan]

like this?
{shimmer hair:1.4}

[untold glacier]

Change that to (shimmer hair:1.3)

And the same for the rest that are 1.4 and over.

[quick hill]

I forgot Network Extension even a thing...

[untold glacier]

You're gonna make me cry

[quick hill]

The hypernetwork

LOL

👴

[rich tartan]

still weird

[untold glacier]

Copy the text of the prompt in here. Put three of ` before and after the text.

[rich tartan]

(best quality:1.1), (masterpiece:1.2), (highres:1.1), extremely detailed girl, solo, sharp focus,(flower shop:1.3), (cinematiclighting:1.2), (character:1.3), (1 girl:1.3), solo, (beautiful plain pink t shirt:1.3), pants, standing, light angry, closed mouth, beautiful detailed eyes, blue eyes, (sharp focus:1.3), (masterpiece illustration:1.3), (medium shot:1.3), shiny hair, short hair, blonde hair, (messy spiky hair:2), (shimmer hair:1.3), glowing hair, (iridescent silver hair:1.1),(hair (houseki no kuni):1.3), disdain, science fiction, (lens flare abuse:1.3), glowing light, bloom, black wear, holding deathly alien weapon, war ready, cyber city

[quick hill]

So many weights

[untold glacier]

That actually doesn't look too bad, if still an eyesore.

[untold glacier]

I know

[rich tartan]

i'll try to load back my old prompt, wait a sec

[untold glacier]

Wait, why the heck is the image so big lol

[rich tartan]

wdym

[quick hill]

wdym

[nova remnant]

fyi the general idea of using weight is not to use them just for writing prompts but only where attention is required that might be missing in your render

[quick hill]

no hires, img too wide etc?

[rich tartan]

upscaling

[untold glacier]

Alright, well I've loaded your workflow up in my Comfy. Lemme see what I can do.

[rich tartan]

now it works again, with those {{{}}}

[untold glacier]

Well, does comfy even accept brackets as a attention modifier?

[quick hill]

Probably due to this reason @rich tartan

[untold glacier]

If it doesn't then it's not even changing the weights.

[quick hill]

Tell ChatGPT to remove all brackets and paste it in ComfyUI?

Maybe removing all weights and start adding on what's needed for the image

[untold glacier]

I loaded up Aurora and am checking the gen now.

Yep. Horrors, lol.

Ah

This is probably at least part of the issue "(messy spiky hair:2)"

[rich tartan]

yh, you found out?

[rich tartan]

hm, lemme try again

[untold glacier]

Hold on, lol. Still fixing. There's something up with the flow

[rich tartan]

kk

[untold glacier]

Who wrote this prompt... it's making my eyes hurt

[rich tartan]

haha, mine too actually.

{{{{{{{{{{{{{{{{{{{{#

[untold glacier]

Ok now we're getting somewhere. Let me change a couple more things.

[nova remnant]

best quality, masterpiece, hires, extremely detailed girl, solo, sharp focus, flower shop, cinematic lighting, character, 1girl, solo, beautiful plain pink t-shirt, pants, standing, light angry, closed mouth, beautiful detailed eyes, blue eyes, sharp focus, masterpiece illustration, medium shot, shiny hair, short hair, blonde hair, messy spiky hair, shimmer hair, glowing hair, iridescent silver hair, disdain, science fiction, lens flare abuse, glowing light, bloom, black wear, holding deathly alien weapon, war ready, cyber city

[rich tartan]

yh works also

thanks, but i wonder what messed it up..

weighting can mess it up so hard?

[quick hill]

Yes

Try using insane amount of weights

[nova remnant]

btw taken from your prompt you could save this as a template for generic anime prompts ...

best quality, masterpiece, highres, extremely detailed, sharp focus, cinematic lighting, illustration

[rich tartan]

howd you guys read the metadata actually?

[quick hill]

TIL Illustration is generic prompt

[rich tartan]

just put it into auto1111?

[quick hill]

yes

[nova remnant]

i copied your paste and edited them in my editor to remove brackets

but you could also read metadata with PNG Info tab on a1111

[rich tartan]

yh, i thought it wouldn't be compitable with comfy actually

[untold glacier]

Ok. You also didn't have a VAE in. Which was a big problem.

Or... maybe it's built into your model.

[rich tartan]

it changed the deign a little, it's more of a face photo now

[rich tartan]

built in

[quick hill]

Every model have VAE inside

[quick hill]

Otherwise it will just fucked from what I understand

blackscreen

[rich tartan]

can't even gen image without vae

[steady grail]

no, some models need an esternal vae

[nova remnant]

good catch, im removing illustration from that list

[steady grail]

what a splendor

[quick hill]

Was wondering since model I use "Illustration" turns into 2D illustrate, but I guess those are difficult to find.

[steady grail]

don't you see that fantastic smily face? XD

[quick hill]

Something I noticed awhile ago

[nova remnant]

yeah its logical not to put that word into generic template

and i switch a lot between realistic and anime

[rich tartan]

ig, the design also improved by removing those {{}}, it listens more to the prompts

[untold glacier]

Alright. I'm using a different model, but the prompt results look much better now. Try it with your model and see if it works. Just open the image in a web browser, download it, then drag it into your comfy to adopt the workflow. I tried to keep as much of the original as I could, and I'd really like to concatenate your positive prompt like I did the negative one but...

[steady grail]

stunning colors

[untold glacier]

Oh, and you'll have to reattach your model VAE lines to the original spots.

[brittle rampart]

under mask

[rich tartan]

dam, this now looks like russian to me xD

[quick hill]

oh mother russia

🇷🇺

[rich tartan]

interesting design it got..

but dpm/karras somehow destroy the look of the chars

[untold glacier]

Uh... not sure what you mean by that. DDIM is a very... outdated sampler method these days.

If you want a softer look, then try euler A or DPM 2 Ancestral

Just be sure to up the steps to at least 28 if you use Euler A.

[quick hill]

I haven't use that sampler like 2 years ago

[untold glacier]

Yeah

[rich tartan]

it looks more detailed with ddim and simple

[quick hill]

I disable it on my A1111

[rich tartan]

The eyes most times suffer

[quick hill]

Gacha upscaler?

:idk

[untold glacier]

I'm not seeing the 'more detail' part, but it does look cleaner. Try Euler A. That'll likely be what you're looking for.

[rich tartan]

im actualy a PLMS user, but comfy doesn't have it, or i couldn't find it

[rich tartan]

yh, i'll try that

[untold glacier]

Euler A with 28 steps on mine.

[rich tartan]

ig that's much worse

[untold glacier]

Oh, and scheduler set to None

Well, in the end, you should gen with what you're happy with and what looks best to you. If you like how DDIM works with your setup, then I'm certainly not going to tell you that's wrong.

[rich tartan]

But why is it outdated tho?

ddim on the upper and euler_a the bottom one

same gen time

[untold glacier]

Because more efficient and better resulting samplers have come out since then.

DPM 2M Karras is still the go-to for most.

DPM 2M or 3M SDE Exponential is another that some have moved to.

Restart, UniPC, etc.

They all have their quirks and ideal step counts.

And for some, they just like the 'look' of the sampler in their particular setup.

[quick hill]

Also most people who uploaded the model doesn't test DDIM etc. Results are vary too, no? @untold glacier

[untold glacier]

Results can vary by so much Model, steps, CFG, etc. etc.

[rich tartan]

uni_pc is also good if i remember correctly from auto1111

[untold glacier]

I remember liking Uni_pc_bh2 in Comfy when I used it a few months back.

[rich tartan]

i'll try, the best dpm i found was the adaptive one, takes very long tho

[untold glacier]

Whatever you do, just have fun doing it. Don't be pressured by others to change your workflow if you like the way things are going. Just try to keep an open mind and don't stop learning! That's the best thing about generative AI: it's always changing and evolving as new techs and solutions come out.

[rich tartan]

yh, i'll try out

thanks for the information tho

[steady grail]

do what he says and you will end with a monstrosity like mine

[rich tartan]

and the prompt fixing, it really did a massive change to easier prompting

[quick hill]

The (((((((()))))))) is really funny lol, I swear I saw it in Civit everywhere

[steady grail]

it just confuses the gen

they try to force something, usually, failing miserably

[quick hill]

Esp the NSFW gen ((((big backpain))))

[rich tartan]

i also love heun for the environments, but the characters looks broken

[steady grail]

like u prompt a glass full of ((((((((water)))))))))) and you will get the sea

[untold glacier]

LOL

[rich tartan]

hmmm wait

[rich tartan]

is it possible to generate with heun the environment and put the character that is generated with another sampler into it?

[rich tartan]

But would need an transparent background ig

[untold glacier]

Not easily. You'd have to automatically inpaint it with a Detailer.

[rich tartan]

can adetailer fix the face for heun?

[steady grail]

i remember there was this detector that could distinguish people and objects

[untold glacier]

Essentially you'd gen the image first, character and all, and then it comes back in with a Detailer inpainting the character with different settings.

[quick hill]

ComfyUI prob gonna convince me to do inpaint and mess with things like these in future..

[steady grail]

using it and ip adapter, maybe

[quick hill]

I swear I just go TXT2IMG and never look back

[rich tartan]

yeah, uni pc is actually really good

[rich tartan]

ig that's gonna break my brain at the end

[untold glacier]

Be breakable. Mold your mind to...

the p o s s i b i l i t i e s

[quick hill]

Ooga uunga

[untold glacier]

But to do any of the detailer stuff you'll need the impact pack https://github.com/ltdrdata/ComfyUI-Impact-Pack

[quick hill]

My dirty mind seeing "SEGS" nodes..

[rich tartan]

those nodes gonna extend out of my screen into my nightmare, bet

[steady grail]

if i'm not wrong, segs was that detector i was talking about

[steady grail]

charming

[rich tartan]

transparent/white background i would wish, but seems like nah

auto inpaint gonna be a mountain to crack with this

[steady grail]

you could use a very rare color for bg and than delete that color with another app, to get a transparant bg

[rich tartan]

i may want to train and use some loras first, never really used some. Any recommendations and what loras take effect on?

[steady grail]

i personally use digitalpainting, epicrealismhelper, hyperdetailer

all your gen are really gorgeous

[steady grail]

your style is very peculiar and unique

still not perfect, but very cool

[brazen rapids]

perfect to what

[steady grail]

some plastic effect is still stuck on the image, but the realistic, vintage style is very cool

[steady grail]

bitcoin spam too.....

[unreal ridge]

poor dude probably got hacked 💀

O7

[steady grail]

yup...

[brittle rampart]

I would reduce the strength a JPFilmGrain little.

[brazen rapids]

no, me like grain

[untold glacier]

I definitely get FFVII:AC original dvd release end credits vibes from this.

With "Calling" by Kyosuke Himuro

Before they had to put Gerard Way in there and remove that whole section for the Bluray release.

[rich tartan]

Getting better

Finally a state, that i can work with

[nova remnant]

What could ever go wrong if you add ((((cursed))))

[near elbow]

ahoi

wot?

pseudo 3d 🙂

[rich tartan]

bright asf

it's really getting good

[steady grail]

this is generated? how?

it's cute af

[rich tartan]

hmm, if i give more to the background lora, the character suffers, if i give more to the character, the background suffers.. What do you guys to in such situations? Any ideas?

[steady grail]

wow

[steady grail]

that's just a question of model

i knew neko-girl could fly in space

[rich tartan]

i guess yes, without the lora that supports the same style as the model, the character breaks or not even getting drawn.. Can't really balance them without something breaking or getting left out

[steady grail]

some models are just very good with waifu, but inconsistent with bg, some other are good with bg, but fails in generating excellent waifus

[steady grail]

to maintein the form of the character, maybe using control net may help

[rich tartan]

character is now more detailed vs background is now more detailed

[rich tartan]

hm, that sounds good, is controlNet available for comfy?

[steady grail]

yep

[rich tartan]

alr thank you

[unreal ridge]

Perfection

[rich tartan]

Finally managed to get some good emotions

[nova remnant]

@wispy canopy a quick nudge since you are working on a new merge and if you want to consider turbo ... https://civitai.com/models/215418

[nova remnant]

Buttons.. must press them

[unreal ridge]

damn, how tf did you notice that lol

had to check it 5 times to see it

[nova remnant]

Haha it's very common for AI to place buttons and like armor stuff over there

[nova remnant]

@nova remnant knows all the secrets of Ai

[tribal mango]

hey guys merry chirstmas fools

[nova remnant]

sdxl textures are too blockish

[nova remnant]

having to download multiple XL models is becoming an issue for me, especially when each XL models can cover all of the styles

and trying to restrict a XL model to do one specific style is a waste of checkpoint file itself

only kind of difference you'd see with XL models are those that are well trained and those that aren't

[vital raptor]

I don't think that's an XL thing, 1.5 has the same thing. Most models are fairly generalist and require loras to push a specific style or use. Any heavy leaning one way or the other is specifically from the training and merges in it and has nothing to do with the architecture. Realistically the only difference between SD 1.5, SD 2, and SDXL are the base resolutions used for training the original model. Which can also be changed by training

The only reason SDXL has "more detail" or "higher quality" is because it has 4x the pixel information in it's training, but that doesn't always mean it's better or actually good quality

Realistically, most people won't have any real benefit using SDXL over 1.5, especially if they aren't taking advantage of the larger resolution, because that's the only real advantage it has. On top of that, SDXL is really slow in comparison on lower end hardware. 1.5 is also vastly more mature as far as availability of model types, loras, embeddings, etc, and can be trained on normal low end consumer hardware. Training something for SDXL takes much more hardware and time, and it's more finicky because of the extra pixel data so it's hard to make training even work properly

Not to mention 1.5 models are also 2.5-4.1GB usually, my smallest SDXL model is 6.7GB

[supple raptor]

I mean, XL also has both the same CLIP model from 1.5 as well as OpenCLIP which is a larger text encoder. The UNET shape is also different think, but I don't know very much about that.

[vital raptor]

Well, there are multiple technical differences, but as far as practical use it doesn't matter as much. The text encoder changes how prompting works, but also people are training and using the models with the same tag mess from 1.5 so it doesn't really matter on heavily trained models

If you were comparing just base 1.5 with base XL, there's massively obvious differences in how they function with prompts and outputs

With sdxl I tend to prompt more "real english", but throw in some tags when i need to push a weight on something as it seems to work best that way. 1.5 was just word salad

[supple raptor]

That's because CLIP is pretty shit text encoder that we've just gotten decent at bending to our will when combined with the grossly overfitted models that we use

OpenCLIP is quite a bit better in that respect.

[supple raptor]

But, tbh the SDXL anime models that I've used are really not great yet. They have little existing character knowledge, and are largely just as overtrained if not more-so than the 1.5 models that we're used to.

[vital raptor]

I think that mostly has to do with the novelAI thing, we don't have a leaked model to merge into everything with XL

unless someone puts the effort and time/money into training a solid base anime model, it can only get so far

but anime specifically also has the benefit of not really needing XL to begin with, it doesn't have as much detail as a realistic image so it can just be produced at a smaller resolution and upscaled

upscaling a photo is a lot more difficult and benefits a lot more from denser pixel data from the start

my problem so far with most of the XL anime models is that the CLIP is destroyed with all the random merges people are doing with no actual training on anything, so they have horrible prompt coherence and disfigured bodies all the time

I'm still trying to figure out why models like bluepencil have so many merges, it's almost like they think merging actually adds all the data into the existing model so now it contains everything they merged

[near elbow]

yes, SDXL based model, most of the prompt is in the filename - must be generated since the text doesn't make any sense 🙂

[steady grail]

lol

[near elbow]

yeah, this is with a fully NSFW capable model and it tends to substitute buttons for... other things 🙂

[near elbow]

having a better prompt understanding (like SDXL) would benefit anime as well, although people have become used to booru tags there anyway

[native halo]

it would benefit us if it could create the anime characters we are typing

[steady grail]

maybe, but most of the people can't even properly run sdxl <.<

[vital raptor]

and even fewer can train it

[steady grail]

maybe at a buisiness lvl, would make sense

[nova remnant]

no chance, sdxl can't do it

this is a grid lock

[vital raptor]

it can do it, it might just not be trained in a current model

[nova remnant]

sdxl has certain advantage over realistic look but not w/o block like render

[native halo]

gib money ,no longer a gridlock

[nova remnant]

right

this is how great ideas fall

$$$$

also as you already know sd1.5 and sdxl are different in their versatility

by design sd1.5 can be trained for specific task that are focused on that task itself but sdxl by design have much wider scope of doing things across different styles, and i dont get why we have to download multiple xl models when one can do all those tasks given its properly trained

total technological mess up

[vital raptor]

who said you had to download multiple models?

sdxl can be trained for specific things too, it's not generalist because it has to be, it's being trained that way

[native halo]

u have to if u want to change the style

[nova remnant]

you can do all the styles with one xl model

[vital raptor]

ok, but you have to change 1.5 models to change the style too, or use loras

[nova remnant]

its how xl is designed in its core

[native halo]

yea i can just use a 1.5 style lora on a single 1.5 model

[vital raptor]

if xl 'can do everything' then why are you getting multiple models is the question

[nova remnant]

thats cause sd1.5 has certain limitations

devs will tell you if you want decent anime its 1.5

[vital raptor]

my roommate has been playing with the realcartoon 1.5 model and it's so far been more flexible than most of the XL models i've tested

he doesn't do anime though

I don't care what people tell you, I've found multiple anime models that will do what I want

but what I want and what you want is not the same thing

fox girl with white hair wearing a long dress dancing

exactly what i asked for, it does the job 🤷‍♂️

I see a lot of people complain about this sucks or that sucks, but no one with actual comparisons, just vague "I can't get what I want"

[native halo]

cant really compare something that doesnt exist

[vital raptor]

but, like I said earlier, if 1.5 does what you want and you can't get that from XL, then don't use XL, it's a really simple solution

especially since XL is horrible for performance, even on high end hardware

[vital raptor]

tbh, I really want to see an example of "XL can't do this" that was made in 1.5 just to see what magically impossible task is being asked of XL. I'd be willing to bet it's a limitation of the available trained models/loras, or at least the few tested, and not a limitation of the architecture

[supple raptor]

You'd be correct. The architecture is better in every way on paper. But having a better architecture is utterly meaningless unless the models actually leverage it. Take SD2.x for example. On paper the architecture is higher resolution with higher parameter count than SD1.x. In reality the text encoder was absolute garbage and the training left a lot to be desired so it consistently was outperformed by SD1.x models until it was basically forgotten.

[nova remnant]

emad is busy with new business strategy to profit, dont think we can expect much from sdxl

[supple raptor]

The fact is that SDXL is more difficult to train both due to hardware requirements and the hyperparameters being harder to dial in (and likely captioning being far more important) than SD1.5 is which will slow progress on finetunes significantly.

[nova remnant]

sdxl is now into 6 mo. since its release

[supple raptor]

And SD1.x models didn't really start to get good until around 6 months after the initial release even with the huge boon of the NAI leak helping it along.

[rich tartan]

Were there only a few models available at the time? Or why was nai model such "hit"?

[vital raptor]

nai was heavily trained by professionals on a massive high quality dataset and was not intended to be freely available

but once it was, it got merged into everything else vastly improving the models

[zinc flax]

It kickstarted the whole anime model lol

Without the leak I doubt AI would be as popular

[vital raptor]

at the very least it wouldn't have produced such high quality so quickly

[zinc flax]

It started with anything v3 iirc

[native halo]

yes without nai 1.5 anime models would suck like 2.1

[rich tartan]

Interesting, i also started to get know the ai with nai, as i thought first this would be Standard sd

[native halo]

and with recent nai sdxl 3.0 looks like nai are the ones who know how to train models properly

and they also have the money so 💸

[zinc flax]

Yeah, nai is the best anime model now

[vital raptor]

that's the ultimate problem, those guys do that as a job and have access to the hardware to make it happen, no one is going to dedicate the time and money to training that kind of model for the good of the free public

[native halo]

i think waifudiffusion tried to do that but the result is not the same

(they dont have money that nai has)

[vital raptor]

on the other hand, even if I wanted to spend the time and money to make some kind of amazing anime model, i'd have to know what's wrong with the current ones but no one can seem to explain it

[rich tartan]

With how many est. pics do they train the model?

[zinc flax]

Probably million lol

[native halo]

on nai 1.5 i think it was 5million pics

[supple raptor]

Novel AI is a company with huge amounts of money who were able to leverage a massive amount of compute along with a ton of GPU hours to make a model that was very good at what they wanted, then someone burned a zero day exploit to gain access to the private repository and leaked the model. Since then, virtually every decent SD1.x model that's been released has contained some amount of the leaked NAI model.

[supple raptor]

estimated to be at least the entirety of the danbooru database, likely with more added from other sources.

[rich tartan]

Dam, That's actualy a heavy numb, i thought about 100k or something

[zinc flax]

It's still scary how good nai v3 is

It's been only a year since the first ever nai 1.5 right

[supple raptor]

NAI leak was Oct of 2022 and it had been around for at least a couple months on their site before then.

[zinc flax]

From anime girl eating ramen the weird way to copying over 7k artist styles

[nova remnant]

look i got the hand right ... (less than five fingers, more than five fingers, missing fingers, extra fingers)

[zinc flax]

waving prompt hack

[nova remnant]

i actually had waving in prompt

[vital raptor]

no negatives

anime girl with long brown hair wearing a christmas sweater waving, standing in the snow at night, street lamps, from side,

[rich tartan]

is that xl model?

[nova remnant]

btw bluepencil v2.9.0 coloring and brightness are too harsh compared to v2.0

yes it is

and with bluepencil v2.0

[rich tartan]

xl models literally always have such clean colors ig, also looks much more realistic

and btw, did you guys try out to use two vaes at once?

[rich tartan]

I wonder what happens if i mix a vae into latent image and higres with that

[rich tartan]

Interesting, mixing heun into ddim gives some more accessories without messing the character up

[supple raptor]

I'm not sure what you mean, VAE is either converting latents to pixels or pixels to latents.

[rich tartan]

oh yeah right, i was in the mind mixing with some preset image or sum

instead of the pre-gen one

finally, heun also gives background with good character instead of void or messed up face/focus..

uh, what is the difference between approx and none?

[nova remnant]

https://huggingface.co/dataautogpt3/OpenDalleV1.1

[vital raptor]

sounds cool, but if you look at the previous version page it's just a merge of albedo, juggernaut and mergeheaven, and it can't even do 'anime girl' properly

[steady grail]

very suggestive

[native halo]

shush 🤫

[steady grail]

cute

[steady grail]

chitanda rules

[steady grail]

they reminds so much the original

[nova remnant]

takes her with me

[near elbow]

what are you missing in SDXL? for my taste, that's anime-ish enough (just "a cute anime cyber ninja, cherry blossoms zen" with SDXL anime style)

and for old school, just add "((flat 2d cel shaded studio ghibli style)) a cute anime cyber ninja, cherry blossoms zen"

[vital raptor]

me? nothing, that's what I was wondering, people keep complaining about sdxl being bad at anime or "not as good" or whatever, but I think it's fine

[near elbow]

right. it does everything anime I need...
the only thing the SDXL models i looked at fail with is kinky stuff (no novelAI leak for SDXL 🙂 )

[vital raptor]

yeah, I've been trying to get a good nsfw merge figured out since that's the only real failing

[near elbow]

well, it does more adult stuff nicely (this is still SFW i think, just for illustration)

up to full and detailed nudity, anime and real, but i'm not posting that here 😉

[near elbow]

"(watercolor) ((flat 2d cel shaded studio ghibli style)) a cute anime cyber ninja, sitting on a bench in a beautiful flower garden"

all just SDXL, nothing special, no LoRA

less flat

still more 3d

[rich tartan]

dam, why is kohya_ss such a pain in the a** somehow i can't get it working, as it tells me cuda dependencies not found or sth

dreambooth ooms me instandly, can't run it

[silk narwhal]

What GPU do you use?

[rich tartan]

rtx 4070

[silk narwhal]

on a desktop? ..any other gpu's (internal maybe that is getting used)

[rich tartan]

yh ubuntu, i don't have any other gpu in my pc(also no igpu)

[silk narwhal]

ah ..linux, don't think kohya_ss works proper on linux, try going over all the steps again else ask in #tech-support

https://github.com/bmaltais/kohya_ss?tab=readme-ov-file#linux-and-macos

[bitter latch]

may or may not be legal in some states

may or may not be legal in some states for a xmas gift

[rich tartan]

thanks, i did, but im currently trying to compile a package myself, that's somehow not working

[bitter latch]

and the manga of the anime

[rich tartan]

and it says 6gb vram users can run this thing

ooooooooooooooooooooooooh

my maximum bucket res was at 2048

meanwhile it should be 512

alr

[silk narwhal]

yea 2048 needs at least 12 GB 😄

[rich tartan]

i even crash with oom at 1024 weird

(just tested)

[silk narwhal]

ask in #🤝｜tech-support i would say

[rich tartan]

oh wait, i had batch size at 10

are those settings good for lora?

1 epoch, 100steps pic

[supple raptor]

for training I prefer this over the regular sd-scripts repo. gives a nice gui wrapper that you can use but also installs sd-scripts which you can run normally if you ever want/need to for any reason. https://github.com/derrian-distro/LoRA_Easy_Training_Scripts

[native halo]

i dont know any anime that looks like this

thats way too slow for a 3070

[nova remnant]

people are splt in two views, anything goes sketch anime and actual perfection of anime.

[native halo]

doesnt look like a sketch either,looks like a cg render

the checkpoint that looks like a good anime style is the neko one

[nova remnant]

all i want is pretty faces

🙂

and thats the only beef i have with sdxl

[native halo]

are u using linux?

[nova remnant]

i have broken out of convention with canvas size

sd1.5 with 640x800 and sdxl 816x1024

those are also 4:5 ratio

[native halo]

a 3060 gets 4it per second on XL at 1024x1024 so a 3070 should get like 5 or 6

the 1.5 looks normal

7 to 8 it/s

oh yea somethings wrong then

how much ram do u have?

then idk maybe try with another nvidia driver version

no,older versions

536.23

yea try that before downgrading drivers

also sometimes slow speed is due to wrong webui installation

[native halo]

yea still needs at least 4its,maybe u are not using your ram in dual channel?

[native halo]

u could try to downgrade driver to test if it works , but its an annoying process,i first had to uninstall current driver with DDU then restart then install new one,restart again then delete venv folder and let it regenerate again then good 2 go

u could try with driver version 536.99 or with 531.61

when i downgraded to both versions i had to do the same process of DDU and several restarts so it takes a while

@supple raptor whats your it/s on a 1024x1024 img on sd 1.5 with euler a sampler 20 steps?

[supple raptor]

Let me start up sd real quick and check. I know I gen at batch size 2 1920x1088 at ~2s/it

3.42 it/s

That's in A1111 with a 3080 12gb so Comfy should be a little faster. ^

[native halo]

yea idk why he gets 1 it on a 3070 with 1.5

[supple raptor]

Could be running out of vram if HW acceleration isn't turned off. that gen used 7.8GB on my system.

Using the wrong Commandline Args could also cause slowdowns.

[native halo]

yea or maybe hes using wrong python version, needs more info 😔

[silk narwhal]

i am on a 3060 with python 3.11.7 and
onnxruntime_gpu-1.16.3-cp311-cp311

let me do a test in comfy real quick

got prompt
model_type EPS
adm 2816
Using pytorch attention in VAE
Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
Using pytorch attention in VAE
missing {'cond_stage_model.clip_l.logit_scale', 'cond_stage_model.clip_l.text_projection'}
left over keys: dict_keys(['denoiser.sigmas'])
Requested to load SDXLClipModel
Loading 1 new model
Requested to load SDXL
Loading 1 new model
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:17<00:00, 1.12it/s]
Requested to load AutoencoderKL
Loading 1 new model
Prompt executed in 25.05 seconds

[supple raptor]

Since he said he just got a new card, my recommendations would be

1. Run DDU in safemode to completely remove all traces of graphics drivers, then reboot and reinstall the latest version.
2. Check python version. 3.10.6+ is recommended, though 3.11.x should also work now.
3. Delete Venv and rebuild. Latest torch is 2.1.2+cu121 and the compatible xformers version is 0.0.23.post1

[native halo]

so 1 it with sdxl 😔

[silk narwhal]

sdxl turbo yea

but think my hdd or RAM is bottlenecking

[native halo]

yea for sdxl 32gb ram its better

[silk narwhal]

4 minutes to load a SDXL model into RAM 😄

think genning at 512 and then multipass upscale is faster than genning at 1024

[native halo]

yea its faster to 512 then highres fix to 1024

[nova remnant]

Harley Quinn, a villain, anti-hero, unique brand of chaos, unpredictable nature, fascinating character.

[native halo]

its where u have webui installed

[supple raptor]

It's a folder inside either your WebUI or Comfy folder.

[silk narwhal]

ComfyUI Windows Portable does not generate a Venv, in case he is using that

at least that is the one from the official github: https://github.com/comfyanonymous/ComfyUI/releases

https://github.com/phineas-pta/comfy-trt-test

very very limited as i can tell

[supple raptor]

Good to know, I avoid all of the portable versions. I'd rather have control over what versions of everything it's using.

Just clone the repo to the the folder you want it to be in. If you already have git installed then the command is just git clone https://github.com/comfyanonymous/ComfyUI.git from the terminal. It will put the comfy folder in the current path that the terminal is in so make sure to cd to wherever you want it to be downloaded to first.

here's the manual install instructions https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file#manual-install-windows-linux

[native halo]

what resolution?

that looks so slow 😬

[supple raptor]

What commandline args do you use?

[nova remnant]

ohmygawsh im dying....

so much funny stuff at #1019361238234443776

[native halo]

are those comfyui arguments?

what about auto1111 what arguments u have there?

[supple raptor]

Yeah I'd just do a manual install so you can make sure the correct versions of everythign are installed.

[native halo]

u could also try to download a1111 version 1.6 and install it to see if u get the same speed there

just download the zip from here https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases

after that just open webui-user.bat and it will install everything

[native halo]

now try with 512x512

yea that looks better

is that on auto111 or comfy?

what arguments are u using on auto111 ?

yea that looks good,so what did u do? driver downgrade or just reinstall comfy and auto?

yea u prob installed something wrong thats why it was slow

sorry idk about comfy workflows in a1111

it doesnt have workflows so u dont need to do the same thing after u reopen it

oh yea just clikc the blue arrow under the orange generate button

when u open the webui again it uses the prompt u were using before u closed the webui

if u use it while u have it open it will just reset the prompt again,it only works when u just opened the webui and havent typed anything

[supple raptor]

Better yet, go through and set everything to the default values that you want then go to settings and in the defaults section you can set the current values to your defaults so that everything will be as you normally want it every time you load the WebUI.

[supple raptor]

you just have to clear out the positive prompt to make it work if webui is already running. You can also drag an image from the outputs folder into the prompt field and then click the blue arrow to apply the full workflow.

[native halo]

oh i didnt know about the drag image part 🙇‍♂️

[shadow palm]

In this moment, the girl became a living work of art, her nails an expression of her personality and a testament to the beauty that could be found in the details. The stable diffusion of admiration for her nails echoed in the room, creating an atmosphere of awe and appreciation for the artistry on display.

[nova remnant]

you can set some specific settings in ui-config.json under your webui folder so that you don't have to repeat setting them over and over every time, you can also set prompt-related workflows in your style.csv file under webui folder

[near elbow]

model loading mostly depends on SSD speed (hoping you have no spinning disk drive anymore 😉 ) - e.g. a 2GB SD 1.5 model loads in 4-5 seconds with a good m.2 SSD

and 6GB SDXL in much less than 10s

and 30 steps dpm++2mk with 1024x1024 SDXL takes around 5 seconds on a 4090 (no turbo) - forget about the it/s, it's the actual generation time that counts

[supple raptor]

The ui-config.json is no longer required. Just set them in the UI then change the current parameters to defaults in settings.

[nova remnant]

oh thats sweet, thanks for the update.

[brisk valley]

keep in mind that the defaults function in the settings updates ui-config.json

[nova remnant]

yes thats where it stores the values w/o needing to manually open that file in editor

[brisk valley]

nubby said it in a way that meant ui-config.json didn't need to exist anymore

[supple raptor]

I meant editing the ui-config.json is not necessary.

[brisk valley]

yesh

[nova remnant]

damn.. the color came out epic

[rich tartan]

Double sampling, heun for environment, uni_pc for the character
And it really fixes the anatomy of heun while keeping the background/environment

you loose a little effects patching heun, but it's still better as no/bad background

at 0.5 denoise btw

[steady grail]

interesting

[rich tartan]

Maybe training another lora will fix the background's anatomy of the char that's still having some issues. Any methods you guys use to get the prefect anatomy?

I dunno, but i may try out

[steady grail]

some model are simply full of mutations

cursed by anything mix

[native halo]

yes depends on model

[steady grail]

that model was simply too cursed lol

[rich tartan]

hmm, yeah.. Did you guys manage somehow to fix/help those models?

[steady grail]

nope, got rid of them

i simply deleted mostof the oldest mix

[native halo]

you could try to fix them if u know how to play with supermerger and weights

[steady grail]

especially because, with time, i noticed, the models are improving

[rich tartan]

supermerger? i'll look into that, thank you

[native halo]

you would also need the recipe of the model u are tryin to fix

its easier just to download a new model,unless the one u tryin to fix has certain style that hasnt been replicated yet

[rich tartan]

that's my issue

[steady grail]

oh, so much grave of the fireflies vibe

that one still hurts inside

[brittle rampart]

hurts inside? why?

[steady grail]

never saw that anime?

it's sent on the end of ww2. a true master piece

[brittle rampart]

I think no. unless it's Gosick

what anime is this?

[brittle rampart]

I want a box like this with the same gift!

[vital raptor]

same

[steady grail]

https://en.wikipedia.org/wiki/Grave_of_the_Fireflies

[brittle rampart]

ah this is!

[steady grail]

yep, that one

my heart is still bleeding

i life changing story

[silk narwhal]

sadly my ssd's are too smol so got all my Ai gen stuff on HDD :/

[rich tartan]

Thats.. slow..

[unreal ridge]

Dang, this one is new for me

[silk narwhal]

2x12 resolution at 104 steps 😛

[rich tartan]

yh xD

[rich tartan]

alright.. something is def wrong

it's really wrong as it has such less mutations

[unreal ridge]

You're probably just improving :D

but wtf happened to the image?

[rich tartan]

im currently researching of new ways and dumping many methods to find the best settings for better anatomy, it worked in my other workflow, but suddenly in this workflow it's messed up

But im really suprised as it keept a good anatomy

[unreal ridge]

I wish I could send the stuff that I'm generating here

[rich tartan]

ah.. scheduler exponential makes it pixelated

it my other workflow much less, but pixelated compared to others

but this one has the less mutations

tried with karras and so on, all made mutations

Gonna try to come up with a fix

the left one is the double sampled one

[steady grail]

looks like a defective controlnet model

[silk narwhal]

throw a 1-2 blur and one more ksampler 8 steps with a low denoise 0.32 works to fix edge artifacts or pixelations, before doing a ydetailer hand and face fix

[rich tartan]

yeah trying to do, but more mutations come up with a 3rd sampler.. Putting the denoise of the 2nd to 0.7 fixes the pixelation, but mutations come with it

[silk narwhal]

hmm..can you throw me the prompt .. lets see what my wf does with it

[rich tartan]

i need it at 0.4, my results shown me that it fixes mutations very well compared to other schedulers

[rich tartan]

yh sure, but i need to make one without the loras

wait a sec

[rich tartan]

here, just move the pic into your workspace and it should load in a few secs

The thing is, the upscale method makes a big diff, nearest-exact is more pixelated and the result is also

[rich tartan]

But in the other hand, nearest exact gives the 2nd sampler more to work with

[silk narwhal]

let me save my own WF 😄

[unreal ridge]

have you tried 4x Ultra sharp?

[rich tartan]

i didn't but can it fix the sampler?

currently the one im using

[silk narwhal]

never got clean results with RESRGAN

[unreal ridge]

https://openmodeldb.info/models/4x-UltraSharp
I think it's this one. Just put it on esrgan folder.

[rich tartan]

thanks i'll try out

but this one also fixed a lot of pixelations

[cerulean brook]

never tried using canny and depth controlnets together but from what I see depth can improve quite a lot details, before the trigger on the guns was impossible to render, and grenades ended up as part of the clothes. Anyway merry Christmas to everyone

[rich tartan]

works, but youre not upscaling and improving mid

exponential dks up when using an "broken" imagine

interesting, the way you input the latent, the way it also comes out

But this breaks the cause why i wanted to use it, it fixes mutations better as the other schedulers

[brittle rampart]

Both refused to sit in the box!

[rich tartan]

That's the cause it did better.. i gave a not broken latent in

alr at least something i researched

i was dying to use this

[silk narwhal]

experimenting with a few thing

[rich tartan]

Normal >karras > exponential > sgm_uniform > simple > ddim_uniform

yeah, i think lora will need to take hand here for better anatomy

One thing that i could improve from my dumping now is setting my denoise 0.1 down on the second sampler fixes more anatomy
(left one 0.4 right one 0.5)

[silk narwhal]

lol ..cursed gen

🤡

[rich tartan]

hm.. 0.7 denoise gave even better

[rich tartan]

u controlnet/mask the face?

[silk narwhal]

ydetailer messed up 😄

[rich tartan]

oh, never used any detailer

left one ddim_uniform, right one exponential both at 0.7 denoise, quite some good differences, and at 0.7 it's not much pixelated anymore

[nova remnant]

do you like the background @native halo

[rich tartan]

you like pvc style?

i mean, i love it so much, don't even bother using something diff currently

[silk narwhal]

it looks good .. decent mix between realistic and anime

[nova remnant]

pvc style lmao

[rich tartan]

you can get good results tho

[rich tartan]

But this one is the most "PVCish" style i got to

[nova remnant]

[rich tartan]

haha, the mirror broke, guess why

Fixed it(now shes happy?)

[silk narwhal]

alright .. this wf keeps throwing nice gens with this prompt 😄

[silk narwhal]

lcm?

[native halo]

i love that massive hat

i rate 40/5

aye

[nova remnant]

deepbonk

[nova remnant]

@nova remnant you must be hungry for a lil love

that thing you did was a unique bonk stuff

lmao

[native halo]

my fav girl from my fav anime

[silk narwhal]

Close up of Suu from Monster Musume 😄

[nova remnant]

@supple raptor I think implementing buzz system is inspiring user engagement and activities on civitai

needing it is not a bad thing btw, maybe a lil discretion for the public eye :p

and if that need ever stops human race will go extinct no matter how you put the argument

unless otherwise age reversal is a thing and death becomes avoidable

are those with low sampling steps or lcm?

or are you using any LoRA style to add grain?

i see, what weight are you using?

i can picture that third image to look fantastic if it were bit crispier

thats bit too much btw, .. try with maybe 0.3

can you please share that metadata?

ahh ok

ok saved, booting up sd now

👀 1970s \(style\), 1girl, above clouds, aerial fireworks, american flag, astronaut, aurora, balcony, blue hair, blue legwear, campfire, cape, card \(medium\), christmas, christmas lights, christmas tree, city, city lights, cityscape, cloud, constellation, constellation print, crescent moon, desert, dress, dusk, dust, earth \(planet\), embers, fireflies, fireworks, full moon, galaxy, garter straps, globe, gradient sky, halftone, halftone background, hat, hirschgeweih antennas, horizon, lamppost, light, light particles, lighthouse, looking at viewer, milky way, moon, moonlight, night, night sky, outdoors, pine tree, planet, print headwear, purple sky, reindeer, rocket, shooting star, shore, short hair, sky, skyline, skyscraper, smile, snow, snowing, solo, space, space craft, space helmet, spacesuit, sparkler, star \(sky\), star \(symbol\), starry background, starry sky, starry sky print, summer festival, sunrise, tanabata, tanzaku, telescope, thighhighs, twilight, ufo, ultimate madoka, window, winter, witch hat

thats a mini dictionary btw

its hard to imagine what that would generate exactly

booru tag guesses words randomly

often gets carried away from the main focus lol

[native halo]

but if u use a 1.5 anime model u want to use the booru tags

[nova remnant]

sure thing, but adding that diverse words don't show an actual image concept

yep, thats cause it picks up every random thing

[native halo]

try with a 1600 words prompt 🤠

[nova remnant]

booru tag style is cool if you can give it a focus with words but the booru tag extension we have kinda sucks

steven i love using booru style lol, it makes things lot easier to convey ideas but i keep them focused on the idea

one of the prompt ideas i play with is ask google bard something along the line of ...
write a creative prompt based on the following words: girl, flowers, garden, stunning visual, sunset
that's just one example of the things you can play with your ideas

can't do that, cause the booru tag system on a1111 isn't too meaningful other than throwing every possible words of an image, but here is my logical interrogation of that image... 1girl, blue hair, witch hat, short dress, cape, stockings, scepter in hand, starry sky

[native halo]

garter belt

[vital raptor]

A lot of prompting is also needlessly repetitive and convoluted. Most models don't need similar concepts to be repeated, whether it's 1.5 or XL. For example: "christmas, christmas lights, christmas tree" in the above example, something like "christmas decorations" covers all of that and will do just as well. It also covers a number of other concepts, like winter, pine tree, etc. Sometimes it's needed to add an extra concept to specifically push the model in a certain direction, but generally it's best to keep them as simple as possible. By default the 1.5 model only accepts 75 tokens to begin with and the UIs are just using tricks to break prompts into chunks to process them different which just waters down the terms used making most of them ineffective

[nova remnant]

not a fault lol, its not perfected

[native halo]

u also need negatives like this

[nova remnant]

faints

[vital raptor]

yeah, 90% of that won't even do anything, you're better off getting a good negative embedding or 2

[nova remnant]

thats just adding in words from template w/o having any meaningful ideas

its funny tho 🙂

but guess what, studying those and questioning those can help learn prompting lol

[vital raptor]

I saw a post on reddit that was something like "the perfect negative prompt" and it was 12 pages of text

[silk narwhal]

lowres, polar lowres, bad anatomy, bad face, bad hands, bad body, bad feet, bad proportions, bad leg, more legs, worst quality, low quality, normal quality, gross proportions, blurry, poorly drawn, text,error, missing fingers, embedding:bad-hands-5, embedding:ng_deepnegative_v1_75t, all i need 😄

[nova remnant]

oh yeah those are more focused on what you are trying to fix with anatomy and general image quality

[silk narwhal]

just do 5-12 passes and all the ugliness gets fixed with it 😄

[nova remnant]

one thing i also learned is that by using a lot of textual inversion can affect the model you are using

[vital raptor]

It's better to start with a blank negative prompt and only add things you actually need it to remove from the generations it making, if you need things like 'blurry' or 'lowres' (especially when already using negative embeddings) it's a garbage model and you should just get a new one

[nova remnant]

i think @supple raptor mentioned that in discussion and i saw that in my example over time

[vital raptor]

yeah, textual inversions are fairly similar to loras as far as what they can affect, though it affects a different part of the process, but they can be just as destructive (or helpful)

[nova remnant]

you know that's one of the thing SDXL tried to reduce

@nova remnant if I visualize to generate that image and tried to use that long list of prompt to achieve that would be like attempting a lottery

what was the point tho, randomness?

cause if you were trying to generate random images i think that's fair to add lots of keywords

you would have a certain pattern to it tho 🙂

but honestly the more i think on it, the more i feel like creating a reality of a new kind w/o any of the current preconceived patterns lol

we have reached the end of 5th dimension

i mean 4th

[silk narwhal]

i start generating random.. and then i lock the seed and tweak the prompt or add a 2nd or 3rd pass

[nova remnant]

its cool 🙂

randomness is just vastly something beyond our logical inference but also come from patterns of possibilities

some of the contradictory words in that prompt were....
above clouds, astronaut, aurora, balcony, campfire, city, city lights, cityscape, cloud, constellation, constellation print,desert .. etc

i feel like a god to be prompting the entire universe ya know?

trimmed 1girl, aerial fireworks, blue hair, blue legwear, campfire, cape, crescent moon, desert, garter straps, looking at viewer, shooting star, short hair, smile, snow, thighhighs, twilight, witch hat

[nova remnant]

@nova remnant btw i didnt mean to discourage you, was sharing some ideas with you, and that prompt of yours gave me some fun ideas too that im generating now

[nova remnant]

some technical bits, for latent upscaling the denoise value is best used 0.5 and up, for esrgan based upscaling 0.5 and below

when using hires fix .. its ideal to use 1/3rd the sampling steps, so if your sampling steps are say 30 then you'd use hires steps to around 10 steps

[supple raptor]

deepbooru is pretty cursed. Use WD-Tagger instead. If you're using A1111, here's the extension https://github.com/picobyte/stable-diffusion-webui-wd14-tagger.git

I think Swin V2 tends to be the most accurate

Yeah, it should try to tag everything in the image since all of it is important for training which was the intended usecase.

[native halo]

no cause that thing is a mess

[nova remnant]

https://github.com/NVIDIA/Stable-Diffusion-WebUI-TensorRT
https://www.nvidia.com/en-us/geforce/news/game-ready-driver-dlss-3-naraka-vermintide-rtx-vsr/
in case you run into a common bug

https://github.com/NVIDIA/Stable-Diffusion-WebUI-TensorRT/issues/27

getting a strange warning when launching webui

i wonder

this is the entire line of warning message im getting

could that be related to recent updates on tag autocomplete?

[arctic nova]

I get this one too. Tag Autocomplete: Could not locate model-keyword extension, Lora trigger word completion will be limited to those added through the extra networks menu.

Auto complete still works tho

[nova remnant]

ahh ok, works here too but why would that warning show up

[arctic nova]

no idea

as long as it works

have been getting it for a while now

probably more than a month alrd

[nova remnant]

there are some new files released i saw in git pull

[native halo]

because u need the model-keyword extension

[nova remnant]

whats that?

something i need to install or can skip?

[native halo]

u can skip it but if u want it here it is https://github.com/mix1009/model-keyword

[nova remnant]

i'll install it

nice that solved one issue with the model keyword