🍥｜anime

[finite herald]

Wise choice

[bitter latch]

but hey I do have DM's 😄

[supple raptor]

Most models are either aesthetic finetunes, or merges of several aesthetic finetunes. Partly because most people don't have the hardware to train on millions of images.

[bitter latch]

grinz innocently

[vital raptor]

the more I try to weight for anthro cat and 'uwu' the smaller and cuter she gets 😄

[bitter latch]

yeah most models , unless trained for it, doesn't do furry well

and the one's that are. .ooooh boy

[vital raptor]

yeah, my own merges are trying for "as flexible as I can, but within reason" which basically comes down to a specific style/look being more overtrained than the others. If I tell it to do a different style, it will, but otherwise it just makes a reasonably good image

[bitter latch]

yeah thats why I merged the one's I did in my NOMD merge model

[vital raptor]

I generally don't set out to make anything specific though, I'll just put random keywords into the prompt and see what happens, then tweak from there if I want to adjust it

[bitter latch]

random scene, random girl, random boy, random time of day, random weather and such is fun

[vital raptor]

something like the pony model would be much more useful if I wanted to generate specific characters in specific styles I guess

although, merging it with my own model I've been working on has been working fairly well so far, the extra knowledge it brought to the table is good

[bitter latch]

why she look like she's about to cry? what?

[nova remnant]

[untold glacier]

I wish there was one of these for SDXL

[nova remnant]

Put (dark:1.8)

Recommendation by @native halo

👍

[native halo]

masterpiece

[wooden coral]

[brittle rampart]

very cute eyes!

[gloomy scroll]

Did I do it right?

[bitter latch]

well then

[acoustic locust]

Model?

[bitter latch]

NuOrangeMixerDream (NOMD)

it's on civitAI

and yep it's mine 😄

well ok it's a merge 3

[acoustic locust]

Thanks

[bitter latch]

welcome

hope you get some good results

I'm watching Full Time magister, wish it was english dubbed instead of chinese and sub

oh well

[supple raptor]

Stability released one alongside SDXL. I haven't used it yet. https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_offset_example-lora_1.0.safetensors

[brittle rampart]

she is very serious

[lyric zealot]

She's a tsundere

[brittle rampart]

ah this is...

[bitter latch]

tough love?

[lyric zealot]

something along those lines

[bitter latch]

sigh I should be in bed but dammit I want to finish season one of this anime!

only 2 more ep to go

[brittle rampart]

guys who know this is project?
https://github.com/LykosAI/StabilityMatrix

[brittle rampart]

or tell me what the problem might be. I can't download packages from GitHub... I get this message

try to get some node package for Comfy

[bitter latch]

and done, wow that is a good series, and chinese at that. ok off to bed. nite

thuds

five more seasons to go.. must not coninue.. must noooot

[gloomy scroll]

[bitter latch]

haven't watched many chinese anime. didn't mean any offense

[bitter latch]

yeah.... let me get this out of my brain

hmm the face. dangit

There we go!

[nova remnant]

[bitter latch]

ooh toasty

[nova remnant]

Small medium

[bitter latch]

and with this. bed.. for real

[dapper ore]

question:
from 1 image of the character anime to make in different poses and emotions
keep the consistency of clothes, face, hairstyle, general kallait
2) if you create a lora from 1 image, what script to use.

[native halo]

https://tenor.com/view/emoji-meme-jar-gif-4658690738464352739

[bitter latch]

... sneaks one last one in

[gloomy scroll]

https://tenor.com/view/closing-the-lid-fruits-juice-kiwi-blueberry-gif-14588286

[native halo]

https://tenor.com/view/emoji-meme-jar-gif-4658690738464352739

[gloomy scroll]

You should take a another look at her

[lyric zealot]

yup, just another closer look

[gloomy scroll]

Just bluntly put you should remove this

[lyric zealot]

yup lmao

[native halo]

u right i didnt saw the thing ill remove it and post it again edited

[native halo]

shes clean now

[gloomy scroll]

[gloomy scroll]

I couldn't bother searching for bottled waifu so here have Miku inside bubble

[nova remnant]

what did you post lmao

[lyric zealot]

not me lmao

[nova remnant]

[lyric zealot]

not allowed please

even if you censor it like that

[violet abyss]

I'm sorry

[lyric zealot]

no problem

[silent kraken]

What did they post

[lyric zealot]

nsfw stuff

[brittle rampart]

cute art style. model or LoRa?

[lyric zealot]

there should be metadata

[brittle rampart]

then must start server....

[brittle rampart]

GitHub experts i am need help...

[gloomy scroll]

Unsafe spaghetti hand holding

[broken osprey]

Does anyone know how I can get this to look as if they are looking at each other, I am using matrix regional prompting but idk how to define openpose images for regions :)

[nova remnant]

pretty

[keen hollow]

Picking up my Blender hobby again... plus, might make for some fun img2img options. 🙂

Feel free to use this Tifa. 🙂 3840x1600, rendered w/ Blender Cycles, 5000 samples.

[obtuse trellis]

[nova remnant]

AnimaPencilXL a merge of BluePencilXL and AnimagineXLv3

[nova remnant]

Snow

[nova remnant]

Boba

[untold glacier]

Oh fantastic! Thanks for the link.

[broken osprey]

Hey guys can I get some tips on improving image quality out of the box in a111

8 feel like every image I do there comes out bleak even with same settings as comfy. Comfy brings out more color, positioning and general composition

Is there a way to chsnge the backend, AFAIK comfy uses a different backend for noise and the like so, are these two just going to produce different results all the time or is there a way to change a111 and make it better

[steady grail]

extensons

[bitter latch]

what do these look odd/dirty/ bad scetched or such?

interesting style to be sure but.. shrugs

It's a combo of sketched, holographic neon color effect, ikabana

experimental style i'm working on.

tho prob gotta remove that lora cause it's effecting it to much

can't sleep so might as well put my brain energy to work

[grave snow]

intentional, experimenting. most of the stuff I made in that style has strictly been weird stuff like this so I took a detour to the more normal world to see how it fared

[bitter latch]

hmm or could be the heun sampler

[bitter latch]

hmm that didnt' go well

thats better

[vital raptor]

chibi godzilla

and other random chibi things.. i'm bored 😄

[sour shadow]

small winter miku is the best

[acoustic locust]

what does "score_9, score_8_up" and so on mean?

when I don't use it i get something like this

when i do

like this

which is better, but why is it so messy?

any way I can get rid of the messiness?

[lyric zealot]

are you using that ponyxl model?

[lyric zealot]

the tag score_9, score_8_up and so on were trained on the model

[acoustic locust]

Yeah, but what does it mean?

[acoustic locust]

It sucks.

[lyric zealot]

basically they are trained on those tags as quality tags?

so in theory, it's kinda like putting "masterpiece best quality" and such

[acoustic locust]

ok

also

WHY

IS

CIVIT

AI

SO

FUCKING

SLOW

it keeps timing out

and it takes forever to load

it's driving me insane

[nova remnant]

Kek

[lyric zealot]

like you trying to download stuff? It's slow maybe cos there's a lot of users at this time

not sure tho

[viscid sapphire]

How can I fix these hands? I have copied some real hands and I am trying to use inpaint but it doesn't work

also i tried using controlnet

openpose hand

[nova remnant]

Can't you hires?

Image seems to be 768x

Hires helps already

[acoustic locust]

Model?

[lyric zealot]

same as before, SoushikiV1

[gloomy scroll]

Waiting for that SoushikiXL

[lyric zealot]

sadly, i will probably not be using XL model, not till when i upgrade my GPU

[gloomy scroll]

Same

[lyric zealot]

[gloomy scroll]

[lyric zealot]

https://tenor.com/view/gpu-gif-27003482

one day

[viscid sapphire]

i will try, thanks

[pure oyster]

Doesn't have red eyes, 0/10

[lyric zealot]

why is there a need for red eyes?

[pure oyster]

8/10

[brittle rampart]

and here is a question about building a prompt. what should be the sequence of constructing a prompt for two characters. (meaning a description-controlled appearance prompt).

[lyric zealot]

For two characters, you will definitely want to use regional prompter for that

[brittle rampart]

ah... only use RP.. oki-doki

[lyric zealot]

That will ensure that whatever you wanna prompt for character one will stay for character one, and the same for another character

[brittle rampart]

Otherwise, it will be either very difficult or impossible to achieve this using the usual construction of a prompt?

[gloomy scroll]

I think this may work also

Char1
AND
Char2

[lyric zealot]

It will be hard to dictate what prompts goes to what characters, you can partially do this with using BREAK

[brittle rampart]

AND - this is not for RP only?

[lyric zealot]

But if you have access to regional prompter, it’s still the best

[gloomy scroll]

It's webui default

[brittle rampart]

oki!

[brittle rampart]

of course i have.. but interesting try without it.

[lyric zealot]

Well good luck!

[brittle rampart]

[dawn barn]

Make sure you selected a VAE

[dawn barn]

Otherwise the difference is that ComfyUI does it from CPU

The seeding

While A41 does it from GPU

[bitter latch]

merged my NOMD merge with the Aroyal RPG tools one for more deail. worked well. tho now less anime

[brittle rampart]

you was right use RP correctly and easy way.. many thank.

[lyric zealot]

No problem! It’s probably the best way for multiple characters generation

[broken osprey]

The thing is I select the same VAE's keeping everything down to the size and all equal, yet it doesn't come out as goood, maybe its the way it creates noise idk

[native halo]

nothing to do with that,its just a matter of skill

[vital raptor]

a1111 has options, you can use cpu or gpu for seed generation

honestly i'd be surprised if there wasn't a similar option for comfy

@broken osprey the 'backend' is the one thing that is the same between comfy asnd a1111. it's stable diffusion, comfy and a1111 are just web based interfaces for the same program

the noise is mostly generated by the sampler, although it is affected by a lot of things

a good place to start is to pick a random seed and run the exact same model/prompt/seed/etc on comfy and a1111 and then post both images with their meta data, it's hard to figure out a problem when we don't even know what the difference looks like

[lyric zealot]

If you are using different weights, then there will be a differences as well. Since Comfyui handles weights differently compared to auto1111

[broken osprey]

what did you prompt to get this image? or did you use regional prompting for the bg and then character?

[broken osprey]

yeah but I try to find guides and they all say the same thing, not even touching most of the time how to get certain styles, I ask here but to no avail so yeah can't do much .-/

[gloomy scroll]

This image has pnginfo

[dawn barn]

There is

[dawn barn]

Then that was added recently

[vital raptor]

pretty sure it's always been there, can't find any reference to it in the patch notes, in august they added a 3rd "NV" mode that attempts to emulate nvidia seed generation for non-nvidia users to replicate nvidia seeds

so at the very least it was before the 1.6.0 update, but i know it was there before sdxl support, which was added in july

a lot has changed for both UIs in the last ~6 months

[broken osprey]

I got this result, the model doesn't know about the symbol I guess cus I am prompting for it

[nova remnant]

a lot of factors play into getting an image to look in a certain way, such as using LoRA and different models can have different results how they are trained, but if you are focused on text prompts you would prefer more to use XL models

[broken osprey]

Yeah but im not sure though, like look the other one is using regional prompting as is this one, but this one is using subsetsuna model and it looks like shit lmao. I tried a ton of different VAE's with this one and only latent regional prompting seemed to produce something good, while most models in comfy come out good out of the box :c

Imma try this same prompt in comfy and see the results with just prompting and no regional, or maybe some reigonal just gonna take some time

[nova remnant]

about VAE you wont need anything more than these 3 for most tasks, the first 2 are for sd1.5 and the last link for sdxl
commonly used for realism https://huggingface.co/stabilityai/sd-vae-ft-mse-original/tree/main
commonly used for anime https://huggingface.co/hakurei/waifu-diffusion-v1-4/tree/main/vae
sdxl https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main

[broken osprey]

which one of the anime vae?

I have the pruned 84000 and 560000 both seem to work for realism, for anime I have anime2.ckpt

[nova remnant]

if you are already using those vae by those authors show in those links, you might need to look at the models more closely

i say that under the assumption you are expecting better results

[nova remnant]

this image looks blurred out a bit, faces look smudged, lot of rooms for improvements

im also guessing you are using comfyui, but with a1111 there is a very good extension called After Detailer that does amazing work on faces

for comfyui similar extension might be called face detailer

off to gaming ... big event

[steady grail]

@nova remnant what are you playing?

[nova remnant]

guildwars2

[broken osprey]

Nop that one was using regional prompting on A1111, using i think 960 * 576 or smth like that, I haven't been able tor eplicate this behaviour closely in a111 it is wayyyyy harder to do so I use this thing. Tbh what weirds me out is that the image looks bleak even with a suggested vae.

I think Ill try it again with kf anime2 vae and see if it comes out better

[vital raptor]

are you using the same sampler, number of steps, cfg, etc? those all make more difference than regional prompting

[broken osprey]

I havent tested this image in comfy I am going to run first in png info then to txt to image, then adjust regional prompting matrix regions and see how it goes

Im not sure if this means something wrong like no ema what is that ;c

I wasn't aware the VAE was capable of introducing detail or correcting the image in any way, Im not sure what changed apart from the ADDCOM to ADDBASE in my new test

[gloomy scroll]

I use RP little bit differently but ADDBASE is base prompt that contains default description for image and I see that you have ADDCOL in you prompt which just splits image half to 2 regions and prompt before ADDCOL affects left region and prompt after affects right region

[broken osprey]

Oh yeah I noticed that indexing starts from zero so its like using .at(0) for the first break, then at 1 and so on so you go n-1 in the regions ^^

I also noticed that I can define controlnets in each region so now I am happy ^^

[gloomy scroll]

How did you define multiple cnets for RP I have used only one so was just wondering?

[broken osprey]

Basically tailored made for the size of the image and the divisions, This one is I think 960 * 600, using two divisions I mapped out each division in the controlnet image by giving a single skeleton for that region.

Something like this, with this you can manage both poses and that it ^^

[gloomy scroll]

[broken osprey]

I wouldn't know how to do ip adapter to a single face or depth for a single part of the image though, that probably needs more control like comfyUI conditioning and the like

[supple raptor]

What you're thinking of is that Comfy generates seeds on CPU while A1111 generates them on GPU by default. You can change to CPU for seeds in A1111 if you want, but it has no effect on quality. The only reason CPU is sometimes considered better for generating seeds is because it allows seeds to be the same regardless of GPU manufacturer. From what you described, it definitely sounds like you haven't manually selected a VAE and are using a model that has a broken VAE baked in.

[supple raptor]

ema weights are used for continuing training. You don't want them for image generation so they are typically pruned out of the model before uploading to make the filesize smaller.

[broken osprey]

ah I see, well the model came out better when I changed form anything vae to kf anime so I think it was hat ^^

[supple raptor]

Yeah, anything VAE is garbage. It's just NAI vae renamed. Bad textures and dull colors on most models.

[supple raptor]

Here's a better version of kl-f8-anime2 btw. I adjusted the contrast down by 10% which which helps a lot with the blown out colors that it can have at times. https://huggingface.co/nubby/kl-f8-anime2-blessed/blob/main/WD1-4-kl-f8-anime2-bless09.safetensors

[broken osprey]

oop well comic babes produces good outputs haha, well they recommend that vae so yeah

[broken osprey]

but whcih one of those files should I download?

[supple raptor]

Just the one I linked directly to. WD1-4-kl-f8-anime2-bless09.safetensors

[broken osprey]

ah oke ^^

[supple raptor]

If you find a model that needs more or less contrast you can use the other safetensors versions. The number at the end corresponds to the contrast adjustment. "09" is 90% of original, "08" is 80%, "1-05" is 105%, "1-1" is 110%.

I use 09 for basically every 1.4/1.5 based anime model.

in hindsight the naming I used was pretty bad

[broken osprey]

nah not bad, jujst needs that in the model card ^^ in any way I will try it a while, thanks for the info

[viscid sapphire]

what model?

[broken osprey]

comic babes

I prompted for the ring and all but nothing came out

Even increased the region size but then it became a rock solid circle with some corners

[supple raptor]

I think it came around the same time as SDXL support, maybe the update before it. It's relatively recent by version numbers, but it's still been quite a while in actual time that it's been there.

[broken osprey]

became this

Also just as a curiosity, could a model not behave well when used in regional prompting, cus I was trying to get a model with some normal clothing and it refused to come out correctly, literally told it every piece of clothing and it made her naked and the like, even with the same VAE I used for older images it produces weird contrasting edges and cuts in the image :c

[nova remnant]

What would you guys say is the sweet spot of steps for restart ?

[nova remnant]

in general use case with most samplers 30

ofc you could bring that down to 6 steps even by using turbo models

[broken osprey]

How did you manage to create such a detailed pose?

[nova remnant]

thats all prompt based and depends a lot on the model

[broken osprey]

ah damn I tried a dynamic pose with a model and it never comes out good, fullbody prompts come out horrible and adetailer can't fix most

look :c

I tried using controlnet and adetailer and nothing :c

[nova remnant]

controlnet is for character attributes not the image quality btw

[broken osprey]

I know, I tried posing

.-. I used adetailer too and nope

[nova remnant]

also adetailer default settings should be good for any images for face correction, can you show your adetailer settings?

[broken osprey]

Literally the basis

no added prompts no tweaks

[nova remnant]

are you using face_yolov8n.pt

[broken osprey]

yeah

look at this image This is in stable diffusion

same model, same fullbody image, same vae and it looks normal

without adetailer btw

[nova remnant]

what was the prompt for this?

[broken osprey]

I think its embedded

Indoors, photo studio, studio, simple background, white background, studio lightning, clear lights, photo shoot, model, 1girl, clothed (full_body_shot:1.3), (high quality, highres, , absurdres, 8K, image, magazine cover BREAK
Simple brick wall, wall, studio lights, ambient lightning, (studio lights:1.3), simple wall, BREAK
((1girl) , full body, looking at viewer, posing, dynamic pose, smile :D, perfect skin, perfect eyes, perfect hair, (short hair:1.3), (pink hair:1.3), wearing a (baseball cap:1.3), long (oversized hoodie:1.3), (oversizzed_hoodie_with_floral_print:1.3), miniskirt, extremely , (green eyes:1.5), long eyelashes, thick eyebrows, pink_eyebrows, (thighhighs:1.3)) BREAK
Simple brick wall, wall, studio lights, ambient lightning, (studio lights:1.3), simple wall,

I didn't turn on regional prompting if that is the issue

but even then it shouldn't kill the image that much right?

[nova remnant]

i wouldn't recommend playing with regional prompting before you have good control over basic stuff

[broken osprey]

BRO how can I learn the basic stuff when no one talks about it? :c I tried looking around for guides and the like and always the same 75 tokens max 77 if you are on comfy. Test models out with different vaes and prompts, dont use too long ones (except the model allows for them) use controlnet use ipadapter, what is the basic stuff then

[nova remnant]

also regional prompting is mostly about placements in different areas of the canvas, not a lot to do with image quality

[nova remnant]

ok are you using a1111 atm?

[broken osprey]

yeah I am using this thing

im running the same prompt with nothing on see if it creates something good

[nova remnant]

ok, dont worry about the tokens, 75 token limit is outdated, as for actual prompting either try using combination of keywords to convey your concept but not repetitiously, or you could try narrating your concept in language pattern

[broken osprey]

also I can't run fast test on anything cus my gpu is slow affffff and 8gb on amd run out in a second since no optimizations

[gloomy scroll]

With Webui you don't rly need to care about token count

[broken osprey]

I tried using natural language and it seemed to work better in comfy.

[nova remnant]

it wouldn't make difference between a1111 or comfyui they are both just webui

[broken osprey]

This is comfy, natural language and keywords

[gloomy scroll]

Tags vs natural language depends more on model you use

[nova remnant]

they are the interface, the underlying engine is stable diffusion

[broken osprey]

this is last prompt with nothing but keywords and 840000 vae

[nova remnant]

also look for a decent model to play with

are you mainly focusing on anime ?

[gloomy scroll]

Can you share image with pnginfo?

[broken osprey]

Not exactly more like this, I use this with a chinese model and a lora

[nova remnant]

so hyper realistic?

[broken osprey]

But the chinese model is meant to produce anime ^^ and it produces good results in comfy :c nothing I do comes out good in a111

[broken osprey]

Adetailer couldn't fix it :C,

[nova remnant]

cute clothes

that's a webp file?

[broken osprey]

And then there is the issue that colors and prompting blend into each other, I wanted to fix that with regional prompting but yeah

[nova remnant]

on nvm

Indoors, photo studio, studio, simple background, white background, studio lightning, clear lights, photo shoot, model, 1girl, clothed (full_body_shot:1.3), (high quality, highres, , absurdres, 8K, image, magazine cover Simple brick wall, wall, studio lights, ambient lightning, (studio lights:1.3), simple wall, ((1girl) , full body, looking at viewer, posing, dynamic pose, smile :D, perfect skin, perfect eyes, perfect hair, (short hair:1.3), (pink hair:1.3), wearing a (baseball cap:1.3), long (oversized hoodie:1.3), (oversizzed_hoodie_with_floral_print:1.3), miniskirt, extremely , (green eyes:1.5), long eyelashes, thick eyebrows, pink_eyebrows, (thighhighs:1.3))

that's a huge clutter buddy

i'll suggest narrowing that down to a meaningful prompt

[gloomy scroll]

What are these safe_pos BadDreamAnimeAndRealisticFitCorder?

[acoustic locust]

What gpu do you have?

[broken osprey]

negative embeddings, renamed for me

[broken osprey]

Dude I literally try simple prompts, but then again, I can't run 30 iterations on a simple prompt and hope one lands on what I want :c thats just a waste of power and time.

For example there idk I could change that, the start describes the general image and the basckground, the other describes the person to detail

[broken osprey]

Rx 5500 2019 8gb

[acoustic locust]

it works with an rx gpu now?

[nova remnant]

always worked ? >.>

[broken osprey]

works like shit but works

now this is new

[nova remnant]

5500 is one of the tough ones to run SD

[broken osprey]

removed quality prompts

[nova remnant]

You got unlucky haha

[broken osprey]

cat knows my pain

[supple raptor]

https://cdn.discordapp.com/emojis/859306742868869153.gif?size=64&name=neko_pat&quality=lossless

[nova remnant]

you won't get to culture your own style if you just copy long list of prompts, you coudld definitely play with prompt but first thing you should work on is your own prompts no matter how simple they are and thats how you get to build up along with following examples

[supple raptor]

poor 4gb cat

[nova remnant]

[acoustic locust]

what gpu do you have with only 4gb vram????

[nova remnant]

I have gen'd like 10k images on my pc

It was enough for me

[broken osprey]

but thats my prompt, I didn';t copy it :c I wanted a girl with pink hair and green eyes in a sports outfit :C

[nova remnant]

rx 580

Its preeetty old

[broken osprey]

thats even more pain

[nova remnant]

simplify the prompt .. but i think you also have to pick a decent model, but others are asking you quesitons too lol answer them

[broken osprey]

but who?

[nova remnant]

beeg ears

[gentle musk]

beeg ears best ears

[acoustic locust]

they made a 4g model of that?

🫡

my condolences

[broken osprey]

I know I should do that but like okayu

1. removing the quality tags suddenly relieved the model and it can think!!!

2. What else shbould I remove, supposedly in sd you cant concatenate keywords like simple_white_background cus it just breaks it down when passing through the clip so what else :c

[acoustic locust]

how long does a gen take???

[nova remnant]

[broken osprey]

amd creating weak ass cards :c, no thats a lie, its actually really good for gaming, but like thats it

[nova remnant]

try this model to play with
https://civitai.com/models/47067/pikas-new-generation?modelVersionId=71733

[broken osprey]

*until the graphics card runs out of memory and overclogs the main ram pool and stutters like an idiot

[nova remnant]

2s per it without hires, I can't hires for some reason but I can up the resolution lmao

[acoustic locust]

that's actually not bad

I used to get 8s/it on my 1070 ti

[broken osprey]

I swear that loooks like most of the anime models I have ( i have 165 models)

Could use aniverse
animesh
realtoon
toonyou
toonyou jp
fiamix nsfw
divine elegance
meina pastel
meina normal
astrea pixie

[broken osprey]

WHAT

YOU TAKE SHORTER THAN MY CARD?

7.76s/it

I HAVE double the vram and it takes longer?

[nova remnant]

which commandline args are you using

[broken osprey]

the ones you gave me before, no half and lowvram, even medvram dies

[nova remnant]

what about 3

[broken osprey]

I literally don't understand what my gpu is doing

[nova remnant]

you have to use no half ?

what are these negatives? Negative prompt: EasyNegativeV2 safe_pos BadDreamAnimeAndRealisticFitCorder

no-half kills my gpu

[obtuse trellis]

I don't use no-half, it's very slow with

[nova remnant]

i know BadDream is an embeddings but you got that all wrapped up with something else

[nova remnant]

some cards need it to work

like mine needs it if I want to inpaint or do something similar

[obtuse trellis]

my card makes 1 image in like 5 min, I think I can't use some models? Or some features, I notice the error before which is fix by no-half

[broken osprey]

supposedly, even cs1o said I should

[broken osprey]

Okay

Easy Negative is easy negative the negativeembedding that is recommended for anime
safe_pos is literally safe pos its callede like that, look it up on civitai

Bad dream is bad dream and the author is fit corder

This is the kind of stuff I wanted to stop with regional prompting, notice how the simple_white_wall became a flower embedded wall which becomes fully translucent at the top and beneath has bricks lmao

[nova remnant]

[broken osprey]

also those are quite a lot of fingers in one hand :=

[broken osprey]

Arent these your ui args?

[supple raptor]

honestly, for an image like this you really don't need regional prompting at all. might want to delete because of pantyshot though.

[broken osprey]

mb

Its so low quality I didn't even notice

[nova remnant]

pruned the prompt for your by removing contradictory words, you can work on the weight but try without weights first:
indoors, white background, studio lightning, photo shoot, 1girl, full body shot, looking at viewer, dynamic pose, smiling, intricate detail on skin, eyes, hair, short hair, pink hair, wearing a baseball cap, long oversized hoodie, miniskirt, green eyes, long eyelashes, thick eyebrows, thighhighs

[nova remnant]

I use --medvram --opt-sub-quad-attention --opt-split-attention-v1 but I hardly use my pc for genning, only if I really want something tested

[broken osprey]

so you use colab for generation?

[nova remnant]

my favorite service is @lyric zealot pc

[broken osprey]

But then this loses the effect on the wall and the detail on her clothing ;c

[nova remnant]

add them

[lyric zealot]

[nova remnant]

you had like simple, wall, brick wall, simple wall, etc

[nova remnant]

those are not good prompt practice

[broken osprey]

and where was I supposed to know this :C

[nova remnant]

thats common sense

you dont need to get too technical with prompting

[nova remnant]

[nova remnant]

sorry if that sounded rude but i reallyh mean its common sense dont make it too techy

[broken osprey]

I don't see them as contradictory, like what do you mean? If I have wall, brick wall, aren't those two literally referencing the same thing, a wall made out of bricks :c

[nova remnant]

simple brick wall is enough to describe what you are asking, if that doesn't, apply weight

[broken osprey]

Your prompt produced two ppl in them :c

its going through finetuning in adetailer though

[broken osprey]

how can u use medvram on 4gb?

isn't that like definition of lowvram?

[broken osprey]

But for example keeping the oprmpt too vague wouldn't do this at all

[nova remnant]

no clue, either my hires doesn't work so Im taking medvram lmao its faster

[supple raptor]

medvram should be plenty of offloading even with 4gb. lowvram is pain. if you actually need it for an sd1.5 model then you should really just pay $10/m to generate on an online service.

[broken osprey]

I mean I dont need it rather if I dont want to run out of memory after the 10 or 15 gen on a set of 30 I need it

this is the image iwht your prompt @nova remnant

[nova remnant]

thats because of the modle you are using and im done explaining

[gloomy scroll]

No idea what's the problem so I just tried running your prompt with my setup

[broken osprey]

okay then ^^

[nova remnant]

i wasn't looking for arguing over irrelevant stuff thought you needed suggestions

[broken osprey]

but I do need them :c sorry if I seemed angry or smth :c

imma go back to my testing see if changing the model works or smth

[gloomy scroll]

Nubby strikes once again

[acoustic locust]

What model?

[supple raptor]

Soushiki

[supple raptor]

https://cdn.discordapp.com/emojis/930664540423745658.webp?size=64&name=theSaddest&quality=lossless

[acoustic locust]

Did they ever update that?

[supple raptor]

It's still a relatively new model and performs well. I don't see a reason it would need updated.

[acoustic locust]

Did you merge or finetune it because those are way better than what I get using it.

[supple raptor]

nope, likely just a matter of learning how the model wants to be prompted and using a simple and effective negative prompt.

[acoustic locust]

what neg do you use?

[supple raptor]

(Worst Quality, Low Quality:1.2), border, 2girls, nsfw, skimpy, grayscale, (watermark:1.2)

[acoustic locust]

any loras?

that looks like an add detail lora

[supple raptor]

not in those images

[acoustic locust]

wow

soushiki is really good

[steady grail]

my favorite model

[supple raptor]

yep, it's pretty easily my favorite 1.5 based model. I like it better than my own model for most images.

[gloomy scroll]

I use Soushiki also but I have 3 loras on top

[nova remnant]

not wanting 2girls

[acoustic locust]

It's better than most xl models

then again most xl models suck

[supple raptor]

I usually forget to remove that even when I do want 2girls

[gloomy scroll]

I never remember what I have in negative

[supple raptor]

One note about Soushiki is that it looks a lot better after Hires fix. My images usually look pretty bad during the low res pass.

[acoustic locust]

2x?

[supple raptor]

even 1.5x is enough to make it look a lot better. I run 2x though.

starting res is usually 544x960 or 960x544 for me.

[acoustic locust]

same

odl

idk

my genspeed is so weird

like i get 10it/s on the first pass

and then 2 on the next

[supple raptor]

I don't like it at Clip Skip 1 either. I always use Clip Skip 2

[acoustic locust]

what does it do?

[supple raptor]

Clip Skip 2 means that it skips the final layer of CLIP (text encoder) which generally helps to make the image more coherent at the expense of potentially not listening to the prompt quite as well. Soushiki doesn't have a problem with listening to the prompt at Clip Skip 2.

[acoustic locust]

What does clip skip do entirely?

[nova remnant]

explaining ai things more than "it just works"

[vital raptor]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/5674

explains clip skip

[keen hollow]

Some dalle roller derby based on the ones posted the other day

[acoustic locust]

honestly, it just works

ok

that makes more sense

thanks varghoss

[supple raptor]

NAI came up with using clip skip 2, they explained it here under the Using Hidden States of CLIP’s Penultimate Layer heading: https://blog.novelai.net/novelai-improvements-on-stable-diffusion-e10d38db82ac

[acoustic locust]

how do I change it?

[supple raptor]

The TLDR is that 1.5 based anime models are basically all trained with Clip Skip 2 which makes them also perform significantly better when you use that while generating.

[supple raptor]

I'd just put it in your quick settings list. Find the quicksettings list in settings and add CLIP_stop_at_last_layers, apply settings then restart webui. Then you'll have this slider at the top and you just set it to 2

If you don't already have the VAE selection there, you can add sd_vae while you're at it.

[acoustic locust]

which one?

[supple raptor]

The top one.

[acoustic locust]

Prompt?

[gloomy scroll]

Those have pnginfo

[supple raptor]

Negative prompt: (Worst Quality, Low Quality:1.2), border, 2girls, nsfw, skimpy, grayscale, (watermark:1.2)
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7.5, Seed: 357179407, Size: 544x960, Model hash: 51e4b0d5e1, Model: SoushikiV1.0, VAE hash: 235745af8d, VAE: WD1-4-kl-f8-anime2-bless09.ckpt, Denoising strength: 0.51, Clip skip: 2, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate erode: -28, ADetailer mask blur: 8, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 12, ADetailer version: 24.1.1, Hires upscale: 2, Hires steps: 15, Hires upscaler: 4x_NMKD-YandereNeoXL_200k, Discard penultimate sigma: True, Version: v1.7.0```

[gloomy scroll]

[acoustic locust]

jesus

I don't even need the seed i'm going to get the same gen lmao

i'm regarded

[supple raptor]

Easier for me to just copy/paste it so you can get whatever parameters you want

[acoustic locust]

that's the metadata

[gloomy scroll]

Here have a jump scare

art by aha, watercolor \(medium\), masterpiece, best quality, (detailed background),
BREAK
(arctic tundra, snowy landscapes, frozen lakes, day, frosty, harsh, ethereal:1.3),
BREAK
(magic circle, detailed magic circle behind her head:1.3),
BREAK
1girl, solo, (open hair, long hair, straight hair, white hair, blunt bangs, azure eyes:1.3), pointy ears, long eyelashes, choker, blush, fangs,
thin legs, (white oversized hoodie, sneakers:1.3), from above,
happy, :D,
<lora:LoraMerge_dim128:1>
<lora:ahafinetagging128dim-epoch-000008:0.7>
<lora:loconshanzhagao128-epoch-000008:0.5>
Negative prompt: ((simple background, white background:1.5), (skirt, cape, sleeves, fence:1.2), see-through, skin tight, thick thighs, navel, dress, Coat tail, cape:1.3), verybadimagenegative_v1.3, (low quality, worst quality, bad anatomy, bad hands, normal quality:1.5), artist name, white border, censored,
(holding stick, off shoulder, chain, claws, holding, holding bag, black hair, animal ears, ahoge, twin tails, clothes writing, (hair intakes, hair ornament), hood, hat, ribbon, hair bow, bow, headwear:1.3), (leotard:1.3), holding skirt, lifting skirt, clothing pull, animal ears, depth of field, navel,
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1602681394, Size: 448x768, Model hash: 51e4b0d5e1, Model: soushiki_v10, VAE hash: df3c506e51, VAE: kl-f8-anime2.vae.pt, Denoising strength: 0.5, Clip skip: 2, RNG: CPU, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 0, ADetailer use separate steps: True, ADetailer steps: 10, ADetailer model 2nd: hand_yolov8n.pt, ADetailer prompt 2nd: hand, ADetailer negative prompt 2nd: negative_hand-neg, ADetailer confidence 2nd: 0.3, ADetailer dilate erode 2nd: 4, ADetailer mask blur 2nd: 4, ADetailer denoising strength 2nd: 0.4, ADetailer inpaint only masked 2nd: True, ADetailer inpaint padding 2nd: 0, ADetailer use separate steps 2nd: True, ADetailer steps 2nd: 10, ADetailer ControlNet model 2nd: control_v11f1p_sd15_depth [cfd03158], ADetailer ControlNet module 2nd: depth_hand_refiner, ADetailer version: 24.1.1, ControlNet 0: "Module: none, Model: control_v11p_sd15_openpose [cab727d4], Weight: 0.4, Resize Mode: Just Resize, Low Vram: False, Guidance Start: 0, Guidance End: 0.4, Pixel Perfect: True, Control Mode: Balanced, Hr Option: HiResFixOption.BOTH, Save Detected Map: True", RP Active: True, RP Divide mode: Mask, RP Matrix submode: Horizontal, RP Mask submode: Mask, RP Prompt submode: Prompt, RP Calc Mode: Attention, RP Ratios: "1,1", RP Base Ratios: 0.2, RP Use Base: True, RP Use Common: False, RP Use Ncommon: False, RP Options: ["[", "F", "a", "l", "s", "e", "]"], RP LoRA Neg Te Ratios: 0, RP LoRA Neg U Ratios: 0, RP threshold: 0.4, RP LoRA Stop Step: 0, RP LoRA Hires Stop Step: 0, RP Flip: False, Hires upscale: 2, Hires steps: 8, Hires upscaler: 4x_foolhardy_Remacri, Lora hashes: "LoraMerge_dim128: 45063301a4db, loconshanzhagao128-epoch-000008: 63f809656656", TI hashes: "negative_hand-neg: 73b524a2da12", Version: v1.7.0

[acoustic locust]

prompt engeneering scares me

[gloomy scroll]

I just realized that I didn't even hit 2k limit with that yet seems like I need more extensions

[supple raptor]

Prompt engineering is easy, just add and remove shit until you get something you like

[vital raptor]

I try to keep it more simple, but i'm using XL models

Steps: 30, Sampler: DPM++ 3M SDE Karras, CFG scale: 4, Seed: 712785966, Size: 1024x1024, Model hash: 83b7e5434a, Model: ProtoLarkXL_v6-pony-experimental-vae_fp16, VAE hash: b94a1cb9dc, VAE: sdxl_vae-fp16-fix.safetensors, Denoising strength: 0.25, Clip skip: 2, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer model 2nd: hand_yolov8n.pt, ADetailer confidence 2nd: 0.7, ADetailer dilate erode 2nd: 4, ADetailer mask blur 2nd: 4, ADetailer denoising strength 2nd: 0.4, ADetailer inpaint only masked 2nd: True, ADetailer inpaint padding 2nd: 32, ADetailer version: 24.1.1, Hires upscale: 2, Hires steps: 7, Hires upscaler: 4x-UltraSharp, Version: v1.7.0```

[supple raptor]

@gloomy scroll my face when I realise I've gone 2 days without genning Marisa toes

[acoustic locust]

mf

does medvram not mean anything to you

[gloomy scroll]

Good job on limiting yourself

[gloomy scroll]

Do you use Tiled VAE?

[acoustic locust]

I can imagine you shaking like a chihuahua behind the screen

[supple raptor]

medvram just offloads parts of the model to cpu, it doesn't change the amount of vram used by that much either.

[acoustic locust]

idk

[gloomy scroll]

https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111

[acoustic locust]

does this affect performance?

[vital raptor]

low and med vram are supposed to break the process into "fits in vram" sized chunks so it OOM less (very basic description) but I don't think a lot of extensions and stuff use it well

[gloomy scroll]

It's little bit slower but decreases vram usage

[acoustic locust]

ok

[supple raptor]

it's a lot faster than the alternative which is overflowing into system RAM.

[acoustic locust]

where do I find that tab

[vital raptor]

I haven't really looked into it much though since i don't need it 😏

[gloomy scroll]

It's dropdown under prompts if you installed extension I linked

[supple raptor]

yeah, they don't really work that well if you use more advanced extensions

[acoustic locust]

ok

[supple raptor]

problem solved

[gloomy scroll]

Smh

[supple raptor]

https://cdn.discordapp.com/emojis/978879333248667718.webp?size=64&name=AYAYAYA&quality=lossless

[gloomy scroll]

Since you added 3 feet emojis to AnimeAI server I guess we need to use those

[supple raptor]

https://cdn.discordapp.com/emojis/1025170604636242011.webp?size=64&name=owo&quality=lossless

[keen hollow]

green

[vital raptor]

heh, 'green eyes' tends to bleed with this model

[keen hollow]

it's a good green

[acoustic locust]

Y'all ever crack your neck so hard csf leaks from your nose?

[vital raptor]

every day

even when i try to prompt clothing color to get rid of all teh green, it just shifts it somewhere else

[acoustic locust]

should i be concerned?

[vital raptor]

lol no, i don't actually have a csf leak, but i do have hypermobility so I crack all my joints a lot

[acoustic locust]

I like cracking my joints.

I don't think it understands sweat pants

is there another name for those pants that isn't a verb?

lmao here's jogging pants

[stable crane]

anyone have a good llm node for comfy?

[vital raptor]

yeah, jogger pants or 'joggers' sometimes works

[acoustic locust]

it just makes her jog

lmao

i just did baggy pants

even before I had the tiled vae it was doing this, why?

soushiki is a master at hands

[vital raptor]

(black jogger pants:1.1) works on my model pretty well

[acoustic locust]

are you also using soushiki?

[vital raptor]

sometimes the weight isn't needed, but at least in a1111 'grouping' terms with a weight makes it think harder

no, i'm using a custom XL model

[acoustic locust]

custom?

[vital raptor]

my own merge + training

[acoustic locust]

ok

[vital raptor]

yoga pants works too

[vital raptor]

sometimes generations are so randomly not what it's been generating for 20 minutes, but sometimes that's ok

[nova remnant]

skinny legs @primal wharf

[primal wharf]

Not skinny enough

[gloomy scroll]

https://tenor.com/view/comfort-comfy-cute-cat-cute-animals-happy-cat-gif-17947113

[gloomy scroll]

I will save the day! 🥢

[primal wharf]

🥢

[nova remnant]

[vital raptor]

i tried to make her legs skinny and it took off her shorts..

[acoustic locust]

tiled diffusion is really good

prompting regions are really cool

[gloomy scroll]

Regional prompter can do this too although it does it little bit differently

[nova remnant]

Kek

Focus on something, other things go away

[vital raptor]

yeah

[acoustic locust]

that's cool, prompt?

using the tiled prompter, why am I getting artifacts like this?

[acoustic locust]

what model?

[nova remnant]

bluepencil

[gloomy scroll]

Battle ready

[nova remnant]

Magic circle

[gloomy scroll]

I have few of these

[nova remnant]

I know you love those

I'll try magic square when I'm on PC

[vital raptor]

i think the older version of my model does this better

keeps to the prompting better at least

[nova remnant]

There's a character that is like a chimera in a story I read, she looked like your gens

And I always think about it lmao

Like she's going to transform any time

[vital raptor]

maybe she will 😛

[nova remnant]

What is your prompt

[vital raptor]

BREAK
standing in a intricately detailed colorful magic circle of glowing runes,```

[nova remnant]

Default break or extension

[vital raptor]

default

[nova remnant]

Does it look a lot different without it

[vital raptor]

mostly it just tends to make the magic circles all one color, usually matching the dress

[nova remnant]

I like throwing flavor words at the AI when it's like magical stuff

[vital raptor]

added random streaks of glowing particles flowing through the scene background, to the prompt

the more you put into it, the more it fills out an image

[gloomy scroll]

@vital raptor I went through all this trouble and you just did it with prompt

art by aha, watercolor \(medium\), masterpiece, best quality, (detailed background),
BREAK
(desert, sand dunes, cactus, sunrise, hot, isolated, otherworldly:1.3),
BREAK
(magic circle, detailed magic circle behind her head:1.3),
BREAK
1girl, solo, (open hair, long hair, straight hair, white hair, blunt bangs, azure eyes:1.3), pointy ears, long eyelashes, choker, blush, fangs,
thin legs, (white oversized hoodie, sneakers:1.3), from above,
happy, :D,
<lora:LoraMerge_dim128:1>
<lora:ahafinetagging128dim-epoch-000008:0.7>
<lora:loconshanzhagao128-epoch-000008:0.5>
Negative prompt: ((simple background, white background:1.5), (skirt, cape, sleeves, fence:1.2), see-through, skin tight, thick thighs, navel, dress, Coat tail, cape:1.3), verybadimagenegative_v1.3, (low quality, worst quality, bad anatomy, bad hands, normal quality:1.5), artist name, white border, censored,
(holding stick, off shoulder, chain, claws, holding, holding bag, black hair, animal ears, ahoge, twin tails, clothes writing, (hair intakes, hair ornament), hood, hat, ribbon, hair bow, bow, headwear:1.3), (leotard:1.3), holding skirt, lifting skirt, clothing pull, animal ears, depth of field, navel,
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1578653998, Size: 448x768, Model hash: 51e4b0d5e1, Model: soushiki_v10, VAE hash: df3c506e51, VAE: kl-f8-anime2.vae.pt, Denoising strength: 0.5, Clip skip: 2, RNG: CPU, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 0, ADetailer use separate steps: True, ADetailer steps: 10, ADetailer model 2nd: hand_yolov8n.pt, ADetailer prompt 2nd: hand, ADetailer negative prompt 2nd: negative_hand-neg, ADetailer confidence 2nd: 0.3, ADetailer dilate erode 2nd: 4, ADetailer mask blur 2nd: 4, ADetailer denoising strength 2nd: 0.4, ADetailer inpaint only masked 2nd: True, ADetailer inpaint padding 2nd: 0, ADetailer use separate steps 2nd: True, ADetailer steps 2nd: 10, ADetailer ControlNet model 2nd: control_v11f1p_sd15_depth [cfd03158], ADetailer ControlNet module 2nd: depth_hand_refiner, ADetailer version: 24.1.1, ControlNet 0: "Module: none, Model: control_v11p_sd15_openpose [cab727d4], Weight: 0.4, Resize Mode: Just Resize, Low Vram: False, Guidance Start: 0, Guidance End: 0.4, Pixel Perfect: True, Control Mode: Balanced, Hr Option: Low res only, Save Detected Map: True", RP Active: True, RP Divide mode: Mask, RP Matrix submode: Horizontal, RP Mask submode: Mask, RP Prompt submode: Prompt, RP Calc Mode: Attention, RP Ratios: "1,1", RP Base Ratios: 0.2, RP Use Base: True, RP Use Common: False, RP Use Ncommon: False, RP Options: ["[", "F", "a", "l", "s", "e", "]"], RP LoRA Neg Te Ratios: 0, RP LoRA Neg U Ratios: 0, RP threshold: 0.4, RP LoRA Stop Step: 0, RP LoRA Hires Stop Step: 0, RP Flip: False, Hires upscale: 2, Hires steps: 8, Hires upscaler: 4x_foolhardy_Remacri, Lora hashes: "LoraMerge_dim128: 45063301a4db, loconshanzhagao128-epoch-000008: 63f809656656", TI hashes: "negative_hand-neg: 73b524a2da12", Version: v1.7.0```

[vital raptor]

😂

[vital raptor]

that was actually my whole goal moving to sdxl, my sd1.5 model required so many embeddings, negatives, loras, etc etc, I got tired of it

only negs i'm using is fox, tail for this prompt, no loras, no embeds, just a well curated model

[gloomy scroll]

I'm still waiting for that SoushikiXL update

But oh well time to hide under comfy blanket

[native halo]

when a100 drops to 300$

[vital raptor]

all nvidia needs to do is release a consumer level AI card with just whats needed for SD and LLMs, no video output, no gaming features, and around $1000 or less and they'd be constantly sold out

the problem is all the datacenters would buy them all

[keen hollow]

need something different in here

someone do a character made out of bacon or covered in fried eggs or something

maybe dueling with king kong

[keen hollow]

nice!

those are good haha

[native halo]

they sure love bacon

[keen hollow]

How about hanging a framed photo on the wall?

Or exploring the pyramids?

Or doing make-up to be a clown.

[nova remnant]

xl failed ...
a boy made of bacon and covered in fried eggs dueling with king kong

[keen hollow]

or being a plane "wing walker"

[nova remnant]

did you do them with bing?

[native halo]

yea sd struggles with the bacon suit,maybe with ipadapter it can

[nova remnant]

they came out great and pretty precise

i was thinking at first they were done on sd with some of your magic

[native halo]

cant do it in sd 😔

[nova remnant]

oh well i tried with xl too, but didnt work out

[native halo]

early xmas celebration

[nova remnant]

lol isnt that too early

a boy made of bacon and covered in fried eggs

how did that algorithm get a pass for release

a guy made out of eggs and bacon

[native halo]

now im hungry

[nova remnant]

bacon man

sausage man

[nova remnant]

octopus man

a girl drowning in a milk pond

[keen hollow]

Today's Tifa

[bitter latch]

still anime or....?

[nova remnant]

witch, voodoo magic, witchcraft symbols, old wooden cabin

[bitter latch]

borrows

[nova remnant]

witch, voodoo magic, witchcraft symbols, old wooden cabin, magic potion, smokes, candles, dim light

[bitter latch]

I made some more styles, 😄 , gonna try those on this new merge

reminds me need to upload the new one

[keen hollow]

world's smallest car behind her, but kinda liked this one..

[nova remnant]

water world, underwater colony, ocean depth, atlantis, advanced technology, epic view, stunning visual.

[vital raptor]

hanging a picture is a difficult one this is as close as i've gotten so far

[vital raptor]

reflections are hard

[brittle rampart]

guys. Tell me how to make sure that a random seed is taken during "Batch count" generation?

As far as I understand, the seed is incremented(+1) from started value seed

[acoustic locust]

xl model?

[gloomy scroll]

NAIv3

[acoustic locust]

?

Link?

\

[gloomy scroll]

https://novelai.net/image

[acoustic locust]

Is NAI based on SD?

[gloomy scroll]

Yes but apparently they modified it to fit their needs

[acoustic locust]

[acoustic locust]

gendump time

[acoustic locust]

Model?

[zinc flax]

Personal merge

[zinc flax]

Squish

[gloomy scroll]

https://cdn.discordapp.com/emojis/1170247494161137714.gif?size=64&name=headpat_UwU&quality=lossless

[acoustic locust]

Of?

[zinc flax]

Aurora + 3A + several loras i'm testing lol

[acoustic locust]

3A?

[zinc flax]

https://huggingface.co/sleepotimer/Model_A

[acoustic locust]

Ok thanks

[acoustic locust]

Model? If it’s not dalle.

[nova remnant]

Dalle

[zinc flax]

Yeah that's dalle

[acoustic locust]

Model?

[acoustic locust]

Ok

[nova remnant]

Bodysuit

[acoustic locust]

What model?

[native halo]

custom mix

[native halo]

💪

[nova remnant]

👍

Very tight

[nova remnant]

Huh that's a character

[native halo]

yes i only prompt chars

[nova remnant]

No anime allowed

[native halo]

[nova remnant]

I want to gen tight shorts elves

Here we go

@native halo (small shorts, tight shorts:1.3)

👍

[native halo]

(high waist shorts,skintight,latex:2.5)

[olive musk]

First time trying to make a short Manga using SDXL

[nova remnant]

@native halo This library came out as a mess

[native halo]

i would like to open that book

[nova remnant]

[bitter latch]

actually looks like a realistic library

[nova remnant]

[bitter latch]

like this

anyhow back to bed.. thud

[native halo]

like this library? 📚

[nova remnant]

That hair could harm people

[native halo]

dw the hair is soft like a cat hair,she also purrs when u pet her

[nova remnant]

[native halo]

take this bigger hair

[nova remnant]

take a guess, that's from infamous nubby .... ;p

[nova remnant]

@acoustic locust Kakigori v3 https://civitai.com/models/100505/kakigori

[keen hollow]

It was. But I've had similar in the various sdxl models 🙂

[nova remnant]

crisp and clean

[steady pumice]

ty friend

[nova remnant]

np 🙂 you should archive them on civitai and for viewing pleasure

[steady pumice]

I thought about it but I have so much stuff I would need to upload

[nova remnant]

yeah its good to put the good ones there for keeps

[steady pumice]

I think my fav image I've done is this one

[nova remnant]

looks very good

[steady pumice]

took a lot of inpainting to get it the way I wanted

[gloomy scroll]

Are those bear ears I see?

[steady pumice]

YES

[nova remnant]

I love books @native halo

[steady pumice]

I too like these books

[supple raptor]

That's correct. With -1, the first one is randomly generated then the rest in the job are incremented +1 from there. A1111 doesn't have a way to randomize each one in the batch because it doesn't really matter. Seed 2000 is not inherently more similar to seeds 1999 and 2001 then it is to seed 1 or 99999999

[silent kraken]

Damn what's the prompt

[silent kraken]

Where is this from?

[olive musk]

Different for each panel, but something like this:
Monochromatic Manga Ink drawing supercloseup of glass of water on small nightstand, eerie atmosphere, professional artwork, nighttime, dark
And then I just used for my negative prompt
vibrant, saturated, happy, childish
..but I also used a Junji Ito lora with a very light weight, as well as IpAdapter with about 3-10 panels I found visually appealing for reference.

if you know ComfyUI, here's my workflow:

[silent kraken]

Yeah this seems about as intensive as I'd expect for that masterpiece, thanks for the workflow. Looking forward to more mangas from you 😂

[olive musk]

Thanks, I will post more when I'm done! :)

[silent kraken]

Intentional onesie or?

[finite herald]

no

[nova remnant]

[wheat falcon]

3 anime chars as pixelart

[untold glacier]

My eyes have been opened to PonyDiffusion now that I've found a major issue with my workflow and fixed it. After following the advice of someone used to the model, what one can make with it is incredible. It's like NAIDv3 but better, and free.

[steady grail]

ponydiffusion is an xl model, right?

[slate narwhal]

Any tips? (Any examples of a good prompt with some style selection?) I haven't had luck with the model, I've only got some good results with the booru tags but the lack of artists and styles makes it difficult (I've tried that 100 artist list and even found some of my own, but no good results)

[supple raptor]

They have a 1.5 model too, not sure if it's worth using. The XL model is the one people have been talking about.

[steady grail]

oh, ok

[nova remnant]

ysemshiro

pony is really good

I can't use it yet

[untold glacier]

https://tenor.com/view/minor-spelling-mistake-gif-21179057

[untold glacier]

Let me put together a short guide list.

It may take a second, as one of my hands is occupied with a grilled cheese

[slate narwhal]

👍No problem, no hurry

[native halo]

Guilty Crown

[untold glacier]

Shows how little I remember of that anime. Loved the music though.

[native halo]

yes all the budget went to Sawano and the animation

[gloomy scroll]

Bunny!

[wraith raptor]

gonna try to make adult version :D

As in just mature

[gloomy scroll]

[keen hollow]

tell me more about this ponydiffusion thing. Have not heard of it.

[native halo]

of all the XL models out there,ponydiff is one of them

[acoustic locust]

Teaching friend how to use SD

he made cirno porn

[steady grail]

😓

the little fairy....

i don't whant to know who he is

oh god, how much i love the touhou saga

[gloomy scroll]

https://tenor.com/view/cirno-unintelligible-dies-touhou-animation-gif-26529131

[steady grail]

my hands are still trembling thinking about the games

[nova remnant]

Kek

Gotta test the capabilities

[acoustic locust]

He's never using it again lmao

it's not like he prompted it or anything

[nova remnant]

[nova remnant]

I wouldn't know it existed if not for this discord chat

[steady grail]

ehehe

[acoustic locust]

You really, really don't

[steady grail]

you should play it. it's avery stimulating experience

[steady grail]

sweet lol

[acoustic locust]

Is the reason my gen speed is slow pcie 3.0?

[nova remnant]

Blurry intensifies

[acoustic locust]

that's really weird, what model?

[nova remnant]

What is your gpu

[acoustic locust]

3070

I have a b450 lmao

[nova remnant]

[slate narwhal]

It's from the bot, a mix of anime and photorealism in the prompt, and some upscale tests; from top to bottom it goes from anime to a real life human body

[acoustic locust]

Yeah, It's really weird

[slate narwhal]

Yeah, it got the effect right

[slate narwhal]

The good old "bokeh in everything" from the base SDXL

[gloomy scroll]

If you are using uncensored model you need to prepare for everything

[acoustic locust]

it was soushiki

He specifically prompted for it

[nova remnant]

Shushiki is gud

[acoustic locust]

It's really good

[nova remnant]

Vampire alert