🍥｜anime

[wispy canopy]

0 help, 0 loras, pure txt2img, on my generic model, y ou did give me an idea how to handle that camera angle!

[wispy canopy]

mantisblades -> not known

[cyan burrow]

oh

ok then, armblades

[wispy canopy]

<form>bending doesn't seem to have strong correlations either

[cyan burrow]

I mean they were used thousands of years back

then use halftone instead of databending

[gloomy scroll]

[wispy canopy]

hand gun barrel pov

😄

@cyan burrow sorry, they don't turn out very interesting

i mean, they're pretty for sure... just... bland

sorry, my model can't do mantisblades or databending

[cyan burrow]

yeah I mean I'm trying to create an elaborate idea through short prompt and that's not gonna work it has to be very very broad

[wispy canopy]

unless the model knows all the concepts by heart, it's going to be a tough uphill battle, dall3 could probably do it

[cyan burrow]

the end of the world!

in the matrix

a digital world being sundered

thats not the prompt btw but still it might work

[stable crane]

neat

[cyan burrow]

ah more stuff I can't show @wispy canopy - wanna see it?

I mean I can't show it here

[wispy canopy]

casure ^^

[delicate vessel]

what model is this?

[cyan burrow]

dark gemini 3 (my own sd 1.5)

https://civitai.com/models/6209/dark-gemini-v3

I'm working on evolving it though

so this might be a merge ahead

but mostly it bases itself off DG3 so youll find similar results

[dawn jasper]

This is powerful stuff.

[cyan burrow]

I released it like 6 months ago, then I went into back to back cancers and surgery and recovery and now I'm trying to see if I can make a better version

[gentle musk]

Playing around with what I can get by mixing SDXL Turbo and a regular SDXL Anime model

[primal wharf]

Katana becomes tail

[whole portal]

this is exactly what i want

[nova remnant]

anyone using SDXL turbo?

[gentle musk]

I'm playing with it at the moment

[nova remnant]

is it good with render, i saw their page it takes 1 step?

[gentle musk]

I've not managed to make it do anything within 1 step

Needs at least 4 usually

[nova remnant]

can you post an image ?

just want to see how good the output comes out

[gentle musk]

Of pure SDXL Turbo, or the stuff I'm messing around with mixing it with an anime model?

[nova remnant]

yeah the turbo rendered image

im fiddling with juggernut sdxl

that model is extremely realistic

[gentle musk]

That's 3 steps

[nova remnant]

hmmm

[gentle musk]

4 steps

[nova remnant]

2nd pic is much better to reflect on chacters face

and 3rd too

thats with 4 steps using turbo?

[gentle musk]

Different schedulers make a fair difference as well, so you have to mess about with it a bit

I've tried the Turbo scheduler in the custom ksampler nodes, but it doesn't seem like it's any better.

Also with these I'm using Kohya Deep Shrink

[nova remnant]

ahh ok ... i can get some decent output with lcm sdxl lora

[gentle musk]

Otherwise you are stuck with 512x512

[nova remnant]

with 8 steps

oh yeah about that limitation with 512x512.... you can't use any model to render above that res?

[gentle musk]

Turbo is the model

And you get weirdness and duplicates over that res, like you do with SD 1.5

[nova remnant]

oh man

[gentle musk]

But Deep Shrink sorts it mostly

[nova remnant]

did you try with deep shrink?

[gentle musk]

Yeah all those images I'm posting are using it

Otherwise it just looks like a mess at those resolutions

[nova remnant]

awesome

how long did it take to render one of those image you posted above?

[gentle musk]

a second

[nova remnant]

damn

what gpu are you using?

[gentle musk]

3080

[nova remnant]

sweet

i have 3060

so all it requires is download the turbo model and render?

[gentle musk]

it seems slightly better than the LCM Lora, but not by a huge amount.

[nova remnant]

even if its slightly better than lcm that's a good difference

[gentle musk]

Yeah, just make sure you use a CFG below about 1.5.

And if it looks like a mess try a different scheduler.

[nova remnant]

scheduler?

sampler?

[gentle musk]

both

[nova remnant]

im using a1111 lol

[gentle musk]

Oh

Dunno if it works in that

It probably will, as in it will render, but I don't know if it's setup that it will look normal.

I think it should, but you might have to change the sampler settings

[nova remnant]

just to give you an idea what i have on my end with lcm atm

[gentle musk]

Comfy has a feature where you can merge models on the fly as well, so I've been testing merging Turbo with counterfeitXL and it works pretty well

[nova remnant]

i see

not sure if a1111 can merge on the fly

[gentle musk]

usually needs a couple more steps to make it not as blurry

[nova remnant]

so ...when you use turbo model you'd have to merge with other models to get a desired output it seems?

[whole portal]

i am gonna get crazy with lcm and some loras 😦

[gentle musk]

Well, for anime yes, because the default SDXL is crap at anime

And turbo is distilled SDXL

[nova remnant]

ahh ok

yeah well im using blue pencil for anime and juggernaut for realism

and a1111 won't let me merge on the fly

[stable crane]

turbo+ DS?

[gentle musk]

Yeah, for more than 512x512 images

[nova remnant]

i should check out deep shrink btw

its an extension right?

[gentle musk]

Dunno how it works in A1111, Comfyui has a node for it built in

[nova remnant]

think that's the one

[gentle musk]

Without Deep Shrink

With

Ignore the crap image, it's just default SDXL Turbo

[nova remnant]

yeah i used to get a lot of those on sd1.5 if i broke the resolution definiton

https://www.reddit.com/r/StableDiffusion/comments/185z05u/fix_doubleheaded_glitches_in_stable_diffusion/

[stable crane]

seems deep shrink mostly works on SDXL

without deformations

[nova remnant]

im using sdxl atm

[stable crane]

I wish it worked perfectly for 1.5

[nova remnant]

gonna give it a run

just checking that video first

[gentle musk]

LCM still works pretty well though. This is 6 steps with DeepBlueSDXL with LCM

[nova remnant]

lcm is pretty good so far

[stable crane]

so with 1.5, I can already generate at 1024x1024 with rarely a problem

[whole portal]

lcm works nice, until you decide to use it with some loras

[stable crane]

and if deepshrink only let's you get to 1024x1024, it's not amazing

[gentle musk]

You can go as high as VRAM allows

[whole portal]

those which requires lower cfg like 5.5 or 4 with "normal" semplars

[stable crane]

but suffer mutations in 1.5

[gentle musk]

Ah ok, it's fine on SDXL

[stable crane]

I wouldn't say "I can go as high as VRAM allows" if the image is unusable

yeah SDXL is great

but style is not so great

[gentle musk]

You might be able to get it to work by changing the settings for it. But I don't really use SD 1.5 anymore

[stable crane]

though, sdxl+ds+maybe turbo as base -> upscale with 1.5 could maybe get the stlye back

[nova remnant]

i kinda got into the habit of using set resolution for my workflow so i rarely get deformities from free canvas sizes

[stable crane]

can you share a pic of the mixing nodes?

[gentle musk]

Left is LCM, Right is Turbo mixed with DeepBlue

It's model merge simple

[nova remnant]

turbo looks a bit washed out

[gentle musk]

It might be a custom node

[stable crane]

kinda nice to be able to do temp merges

saves on disk space for experimentation

[gentle musk]

yeah, it's because I'm having to mix more of SDXL Turbo in, which takes away the animeness of the other model, so it does that a bit.

[nova remnant]

ahh ok

[gentle musk]

I think LCM is probably better

[nova remnant]

juggernaut is freaky realistic with human like blemishes

[gentle musk]

But you are limited to the LCM Sampler, with Turbo you can use others.

[nova remnant]

i love lcm a lot ... have been rendering lots of images with it lately

quality isn't too compromised .. they can look pretty sharp and good given the right settings

[gentle musk]

I've noticed that LCM doesn't play nice with Deep Shrink. But you don't really need it as much anyway.

[nova remnant]

yeah given you stay within these resolutions you should be fine to avoid split ups
SDXL
1024x1024 default

Variants:
640x1536
768x1344
832x1216
896x1152

also 768x1024

[stable crane]

wait, are you supposed to load LCM lora, or use LCM sampler?

[whole portal]

lcm works best with LCM, Euler A and UniPC (at least based on my tests)

[whole portal]

both actually 🙂

[stable crane]

😮

ok let me try that again lol

[whole portal]

but you can use lcm lora with those samplers i just mentioned

[gentle musk]

Yeah, I think LCM is better with other models if you want to keep that models style. You can mix in Turbo, but there's a balance over it being blurry or just turning into too much like base SDXL.

Like this is LCM and still looks anime

[stable crane]

before or after deepshrink?

[gentle musk]

This is mixing and it's taken away some of the style

[gentle musk]

In comfyui I don't think the order matters for the nodes

[nova remnant]

this looks pretty good for sdxl anime

[supple raptor]

SuperMerger can do that in A1111 as well. As long as you have the save model option unchecked, it only merges in vram and you can gen as normal with with the merge, even from the txt2img tab until you change models.

[nova remnant]

ahh sweet, need to check that

think i need more ram to not have my pc stutter when rendering with sdxl and have youtube playing

[gentle musk]

How much do you have

[nova remnant]

16gb .. i know i should get 32gb

[gentle musk]

Yeah

I had 16GB and it kept trying to lock up my PC

I've got 32 now and no issues

[nova remnant]

yeah

definitely need to upgrade

[stable crane]

my comfy keeps dying in vae decode

maybe I gotta restart PC

Killed
thx

[stable crane]

I made a horror

[nova remnant]

.

[stable crane]

I really gotta get a webm node

for comfy

12mb png is ridic

[nova remnant]

png quality is better i think

[nova remnant]

i switched to jpg for few days but jpeg images looked like they were missing some depth

[stable crane]

jpg is bad

also surprised comfy doesn't come with filetype nodes

[nova remnant]

amazes me the quality lcm can render

[stable crane]

oh LCM works as a sampler even without the lora

[gentle musk]

Yeah, it won't let you get low steps though with random models, it will still look rubbish

[bitter latch]

and the lora works without the sampler.. gasp

😄

tho wouldn't recommend setting the lora at 1

[nova remnant]

yeah lcm lora works with ther samplers too, i was using dpm++ 2m sde karras but had to keep lora weight to 0.3 or the images came out badly overcooked

[stable crane]

oh can you increase lora str with LCM sampler?

[nova remnant]

you can use weight on lora

lora:lcm-lora-sdv15:0.7 im using that with lcm sampler atm

[stable crane]

interesting

[stable crane]

still seems to need at least 8 iterations

[nova remnant]

8 sampling steps, 4 hires steps, 2 cfg scale... the quality is totally fair

[stable crane]

looks nice!

[steady grail]

wow

[bitter latch]

hmm nope cant' post what I just got.. nope

[steady grail]

i know what do you mean, even with a joke prompt like the knight maid, i often get unexpected results

some models have a dirty mind XD

[gusty palm]

is this the extension form of the thing u guys were talkin about the other day

[steady grail]

omg, how did you do this? i love it

[bitter latch]

yeah and i've learned to double check if it's actually swimware and not bedware

[native halo]

yea cuz its not ai 🌚

[steady grail]

it's exactily what i was looking for

ohhh

this image is mind blowing, i would love to replicate such layered effects

till now, i could do something like this, only by mistake

[native halo]

you could probably do it maybe with cd tuner and set contrast to dark

[steady grail]

the last time, i obtained something close to a layered detail by adding a noise image on the lineart controlnet

it was weird, but the effevt was very cool

[gentle musk]

time for nightmares

[bitter latch]

haz some idea's.. hmmmm

[steady grail]

plz, at least, spoier the image

it's very unsetling lol

[gentle musk]

it has been removed

You can have this instead

[steady grail]

i don't think is too much gore, but the feeling is very weird

wow, charming

not my type, but it's objectively cute

i'm more for the energetic type

[nova remnant]

i wasnt talking about it till today, its an extension that can prevent rendering cloned images when the canvas is off proporitons

[bitter latch]

dark forest, bright light coming through tree's from above, high contrast foreground, deep shadows, girl blowing magic bubbles

[native halo]

try with (bubbles,soap bubbles:1.5)

[steady grail]

Fly me to the moon
Let me play among the stars
And let me see what spring is like
On a-Jupiter and Mars

[native halo]

take this one instead

[gusty palm]

i wish i could share the gens i been doing

[steady grail]

year ago, i knew that type of girl

oh god, she totally destroyed me

[gusty palm]

tgirls rock

[steady grail]

such little devils know how to play with man's hearts lol

[steady grail]

charming

[bitter latch]

well that took a odd turn 😄

we'll find out when the bot times you out

😄

[gusty palm]

i tweaked the prompt quite a bit to make clothes appear lol

[bitter latch]

still on the adult-sexualized side so ya need to be careful

[steady grail]

even if the image is pretty tame, i think undervear shouldn't be exposed like this

but i don't know, sometime, my perception is a little off

[gusty palm]

ill try for swimsuit

[bitter latch]

yeah got me timed out the other day tho might have been the tattoo

[native halo]

its not underwear if you are in the beach or in the wrestling ring

[steady grail]

is even the type of of wear, that was without a doubt underwear lol

[gusty palm]

i merged a yiffy model at low weight into mistoon and something else and using animal wildcards and like most batches produce something good

[native halo]

try to make a model that creates mechas that dont look like gundams

[white jacinth]

Can I get a boy character inspiration????

[steady grail]

how about a prompt with this song?

https://genius.com/Frank-sinatra-fly-me-to-the-moon-lyrics

usually gives me crazy results

[gusty palm]

hah

theres a lora called 'steal ai' or something and the author of that makes good looking loras that might be useful for that

[bitter latch]

had to try

[nova remnant]

the right image looks like michael jackson

and the left one has a touch of david copperfield

[bitter latch]

Frank Sinatra by Ghibli

[nova remnant]

capn kirk

[terse ore]

I'm trying to inpaint a bandana from camo to plain green using Photoshop with a Stable Diffusion plug-in but not successful. Original:

After inpaint:

The ears, nose, etc is not aligned properly

[terse ore]

And I seem to be able to inpaint only in a rectangular area basically

[near root]

After months of plugging and chugging I'm finally starting to figure it all out 🥳

[bitter latch]

ruhh roh

[bitter latch]

haha John McClean Movie by Ghibli

Die Hard Movie by Ghibli

[bitter latch]

Furry Version of Die Hard The Movie by Ghibli,

hmm let me try that with the fuzzyanime merge I made 😄

DE HAAD

[bitter latch]

haha

[bitter latch]

Furry anthro Die Hard Zootopia move poster by Ghibli

[bitter latch]

eh that tail looks odd

[supple raptor]

The art style makes this a pretty easy one since the colors and shading are flat. Just paint over the camo bits to get rid of the colors you don't want. Then inpaint at a fairly low denoising to blend it all together better if you aren't completely happy with the result.

[terse ore]

I tried that and it's much better. It still shows that box thing where the area containing the inpaint occured though. Is that something I need to figure out in the settings, ie inpaint only the masked area?

[supple raptor]

I'm not sure which PS extension you're using. In the Krita extension that I use you can define the inpaint mask by creating a new layer and painting over the area you want masked with a solid color, using a soft edged brush such as the "airbrush" to blend the masked area into the rest.

[terse ore]

I'm using the Abdullah Al Faraj one. The problem is that I can mask but I have to make a rectangular drag selection for the mask, which changes everything in the box. I guess I'll have to look it up or experiment to see what works. But thank you, your idea did make it look much better

[supple raptor]

From the video, it looks like the rectangular selection box is meant to just be the input image and you should be using a clipping mask to define the actual inpainting mask in a similar way to how it works in the Krita extension.

[terse ore]

I see, I'll try to figure it out

[supple raptor]

The video I was watching is accessible by clicking the image under the Demo section on the github page. Unfortunately it's just a sped up video of someone creating an image with a mix of txt2img, outpainting, and inpainting so it isn't very easy to follow along with what exactly they're doing. I would install it and try myself, but I still use the CS6 Adobe Suite which isn't compatible.

[terse ore]

I'm using the "unofficial" PS which is why I use it lol, otherwise I wouldn't

[supple raptor]

I don't use PS much personally. I mainly have CS6 for Illustrator which I use for work and that version is more than capable for my needs. I don't see Adobe ever convincing me to switch to their subscription BS.

[keen hollow]

evenin'

i miss a few days and there's some new turbo thing

[bitter latch]

and oh boy it's 13 almost 14GB really? wow

hmm ok 12.9GB

50 steps in 1 step? this I gotta see

[native halo]

tried to get it without a lora but no,at least i got a close one

[keen hollow]

yeah I dunno, I havent tried it yet

[steady grail]

cute

[keen hollow]

haven't had time to figure out what it's all about

[bitter latch]

https://huggingface.co/stabilityai/sdxl-turbo

[keen hollow]

until I can use it with web ui, i'll spectate ;P

[bitter latch]

https://stability.ai/news/stability-ai-sdxl-turbo

[native halo]

thats a model for sure

[bitter latch]

sure taking a long time to load

so far not working in a1111

or at least I havne't figured it out

[keen hollow]

yeah it'll require some updates

i think they posted some instructions for using it

[native halo]

with a lora is easier

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1179282575286599750/00031-1030362120.png?ex=6579377b&is=6566c27b&hm=7fa0b0f4f5cb1ce518ec9e9b0ddddfdffd5b2a8995a9ca5bc043222bc9bf36bc&=&format=webp&quality=lossless&width=441&height=662

[nova remnant]

this video is a crazy reminder of this evolving ai tech https://stability.ai/news/stability-ai-sdxl-turbo

but i dont suppose i could do all that real time locally on my rig

https://civitai.com/models/215418?modelVersionId=242733

[gusty palm]

i found a cool new asthetic word

cyberdelic

[gusty palm]

i knew real time was coming

didnt expect it to even come close for another year or so at least

[nova remnant]

yeah its amazing , btw you can use the turbo model with 768 and 1024 res

[gusty palm]

on a gtx 1060?

i have a feeling movies and video games will be rendered in realtime ai at some point

and probably vr but i dont have vr so i dont care about that

and then volumetric displays will be holodecks

[nova remnant]

definitely. remember nvidia showcased a game rendered in real time? thats gonna be main stream

actually not real time, but the npc was creating stories based on player interactions

those are all coming together

get ready for a big convergence

pretty soon the tech will come at a pace society won't have time to think but to go with the flow

ofc past the linear thoughts

[gusty palm]

i didnt see that

the real time thing. i got to see how much faster path traced rendering got in the past 5-10 years though

[nova remnant]

you couldn't have missed it but here is the link https://youtu.be/5R8xZb6J3r0

[gusty palm]

wow

that video is so wide

the lighting looks really good

im guessing the dialog was like, generated with a llm given the prompt of the story lore or something

[nova remnant]

the conversation takes place through generative ai based on player specifics what he did in the game

btw have a look at this for fun how cool this turbo xl is https://clipdrop.co/stable-diffusion-turbo

im gonna grab the model to run locally

[gusty palm]

wat kinda reqs do it need

is there 1.5 turbo

[nova remnant]

just the sdxl turbo at present

no special req othar than being able to run sdxl models

basically enough vram

[gusty palm]

im too invested in 1.5 to switch

[nova remnant]

i hear they are also working on sd 1.6

so 1.6 could be more gpu efficient

probably with all the newer parks you get with sdxl

if you have 8bb vram or higher you should be ok

i'd recommend using this model to play with since its fine tuned based on turbo sdxl https://civitai.com/models/215418?modelVersionId=242733

and its in alpha now so lot of improvements to come weeks from now

[gusty palm]

6gb

[nova remnant]

yeah standard size for sdxl models

[gusty palm]

lots of shared mem tho

[nova remnant]

with 6gb it could be bit slow but you can try the sdxl medvram arg on webuser.bat file

--medvram-sdxl

[gusty palm]

oh hmm

[nova remnant]

it wont affect your regular sd 1.5 models

that one is specific for sdxl models

you may also need to use --no-half-vae

[wispy canopy]

sdxl turbo with realtime prompting is actually kinda fun 😮

here, bigger pew

[nova remnant]

you can generate them at higher res

1024 too

im gonna fire up a1111 to test this turbo model in a bit

waiting for my food delivery

[wispy canopy]

they become pretty inconsistent tho

haven't found a way to make kohya's work nicely with them

[nova remnant]

i see but according to emad what he wrote 512 gives weird results

[wispy canopy]

sdxl turbo is for 512 iirc

[nova remnant]

thats that i read too on hugging page

but according to emads post you can use 768 and 1024 both for decent output

[wispy canopy]

if you accept cloning

[nova remnant]

`Oh erm for higher quality you can also just generate at 1024x1024 but 768x768 will be fine.

512x512 is bit awkward at times.`

[wispy canopy]

euler default is pretty bad

got it with lcm and freeu -> pretty good

[nova remnant]

lcm and dpm++ 2m sde karras is evidently working good

the fine tuned model at civitai also work well with those two samplers

[wispy canopy]

lcm + freeu is my combo to go

can even crank the cfg to 1.5, making negative prompts working

i mean... just doing random shit and have it render realtime is just too much fun 😮 (kitty has no recollection of warcrimes)

[nova remnant]

is that on clipdrop?

[bitter latch]

ok I have lcm and freeU, hows this work?

150 sampling steps 😄

[wispy canopy]

using comfyui?

that's the one for the realtime renders

https://comfyanonymous.github.io/ComfyUI_examples/sdturbo/

[bitter latch]

nope just regular a1111

[wispy canopy]

dunno if it'ssupported by a1111 yet

[bitter latch]

well it loaded the ext

[wispy canopy]

ah, someone already made ext? 😮

[bitter latch]

loaded from the git link

lora:LCM_LoRA_Weights_SD15:0.8, Rabbit girl Miku in summer time on beach
Steps: 150, Sampler: LCM, CFG scale: 3, Seed: 428834036, Size: 512x512, Model: CounterfeitV30_25, VAE: clearvae_v23.safetensors, FreeU Stages: "[{"backbone_factor": 1.2, "skip_factor": 0.9, "backbone_offset": 0.5}, {"backbone_factor": 1.4, "skip_factor": 0.2, "backbone_offset": 0.5}, {"backbone_offset": 0.5}]", FreeU Schedule: "0.0, 1.0, 0.5", FreeU Version: 2, Lora hashes: "LCM_LoRA_Weights_SD15: aaebf6360f7d", Version: v1.6.0-2-g4afaaf8a

[wispy canopy]

but that isn't the sdxl turbo model, right?

[bitter latch]

no just regular model

Rabbit girl Miku in summer time on beach lora:LCM_LoRA_Weights_SD15:1
Steps: 15, Sampler: Euler a, CFG scale: 3, Seed: 2820788942, Size: 512x760, Model: cybertoonrealMerge.fp16, Clip skip: 2, FreeU Stages: "[{"backbone_factor": 1.2, "skip_factor": 0.9, "backbone_offset": 0.5}, {"backbone_factor": 1.4, "skip_factor": 0.2, "backbone_offset": 0.5}, {"backbone_offset": 0.5}]", FreeU Schedule: "0.0, 1.0, 0.5", FreeU Version: 2, Lora hashes: "LCM_LoRA_Weights_SD15: aaebf6360f7d", Refiner: aZovyaRPGArtistTools [c7751e6108], Refiner switch at: 0.8, Version: v1.6.0-2-g4afaaf8a

hmm interesting

[nova remnant]

im getting whacky output

using sampling steps 2, no hires fix, cfg 2

with 768px

[bitter latch]

SDXL_turbo

[nova remnant]

can you use adetailer with it?

something is nutty on my settings, i cant use adetailer and i also cant use positive prompt like (highres, highly detailed:1.1)

gives me crap load of error in console

and i get totally distorted image

also what about using hires fix with turbo?

[bitter latch]

this isn't very fast either

[nova remnant]

should take 1 or 2 sec to generate

but im having issues of some sort

weird fkin issue

this prompt throws me error for w/e the f* reason
(highres, highly detailed:1.1), aesthetic, masterpiece

UnboundLocalError: local variable 'h' referenced before assignment
Time taken: 0.9 sec.

[bitter latch]

in other words "let me load... nope nm "crash

[nova remnant]

this is with 0 sec .. instant gen but also rubbish lol

thats with LCM

but if i used 2m sde karras that will throw that error above

im testing with steps 1, cfg 1

[bitter latch]

hhe

testing u detailer (fork of adeatiler) and kohya hires fix

tho not sure what i'm doing

[nova remnant]

i think i found the mis match

adetailer will work if i have enough steps, like above 1

[bitter latch]

ooh

back to regular

😄

Frank Sinatra by Gihbli, very complex, extra smoothing of details, elegant, concept, 3.5D, 4K, unreal render, illustration

[nova remnant]

ohmygwarsh

i found out what was going on

this turbo model im using has limiting token for prompt

[bitter latch]

yeah it's sorta like a demo.. not their best work

[nova remnant]

im using this fine tuned model altho its in alpha https://civitai.com/models/215418?modelVersionId=242733

holy crap

so i can't use prompt freely with it as of now

which means its pretty much useless past the point of demo

6gig of excitement dumped

[bitter latch]

yeepppp

[nova remnant]

lol

[bitter latch]

12.9gig on the version 1

deletes it

[nova remnant]

oh yeah i saw that one too

[bitter latch]

oh 13.5

[nova remnant]

man i was starting to have fun

this is the best i can do for now with limited token

took 1 sec

768x1024

[bitter latch]

ok off to work, sigh

[wispy canopy]

oh fuck

cheat found

merge the sdxl turbo model!

add in lcm lora too

3 steps, 1 second

776x640

[ornate echo]

What if you try to drop deep shrink in right after the FreeU_V2 node before the sampler...does that allow you to bump to 1024x1024 and get a cohesive image? (You might have to bump the steps up a bit when doing that.)

[wispy canopy]

deepshrink doesn't want to work on my comfy

i dunno why

[ornate echo]

I've been using it on mine just fine...even with Turbo.

[wispy canopy]

and i think i already can 😮

just doesn't do anything 😮

[supple raptor]

1.6 is just a different finetune of the 1.x architecture. It's available to try on dreamstudio. Seems better than 1.4 or 1.5 were as base models, but I'm not convinced that it's any better as a base model than the good old NAI leak as far as anime goes so I'm not sure it will have much impact on anything for us if released publicly.

[nova remnant]

that's unfortunate, i was hoping community would develop better fine tuned models with sd 1.6

[wispy canopy]

1 second images too stronk

[steady grail]

still lcm gens?

i was thinking of using lcm as a base for controlnet

[supple raptor]

We might. The big advantage that it has is that all the stuff we already have for 1.5 models should be at least mostly compatible with it. There could be some weirdness when it comes to LoRA compatibility simply because there is usually some weirdness there between different base models. But they all should at least technically work.

[wispy canopy]

this is brightprotonuke turbo lcm 😄

i mean.. for a 2 step generation, you can't even complain

prompt was nakama power lol

[steady grail]

this name lol

[wispy canopy]

oh wait... was 6 steps

[steady grail]

what's this lcm? a model? or a lora?

[nova remnant]

and its not just compatibility thing but also better engine altogether with sd1.6?

[wispy canopy]

6 steps -> 1 second 768x768

[wispy canopy]

the lcm lora, lcm sampler

freeu

[fringe drum]

anyone knows where can i find a product showcase models ? - like a place holder where i can put a picture of real life product ( printer and scanner ) and its description

[wispy canopy]

and a modelmerge using turboxl to speed up generation even more

[steady grail]

turboxl still need the vram of a normal sdxl model?

[wispy canopy]

it's worse -> using the comfy dynamic merge it needs both models loaded XD

but yeah, in essence, it's just a sdxl model

[steady grail]

brightprotonuke is an xl model

[supple raptor]

No, the "engine" is the architecture. The architecture is denoted by the first number in the version, in this case "1".

[steady grail]

lol

i suppose lcm is better with xl models

the gens are more tidy

[nova remnant]

ok

btw excuse me but the issue i was having with trubo sdxl and my prompt tokens wasn't actually about prompt token limits at all, it was the file name convention that glitched out with the model which i have fixed now after talking to the model author, and here is what i get

[00:29<00:00, 3.29s/it]

[gentle musk]

This seems to be the quickest I can get at the moment
1344x768
SDXL Turbo Mixed with CounterfeitXL + LCM Lora + Kohya Deep Shrink
4 steps CFG 2

[wispy canopy]

good shit, amiright?

brings generation time down a lot

[gentle musk]

Yeah if you bring the steps up by 1 or 2 it makes it look a decent bit better as well

FreeU actually seems to make it worse, but only sometimes

[wispy canopy]

default freeu not good

[nova remnant]

how was your merge process, did you merge turbo base with counterfeit and what % did you give counterfeit to reflect on the new version?

[gentle musk]

70% counterfeit 30% Turbo

[wispy canopy]

got a lot more turbo in there

[nova remnant]

ok thanks i think i need to download the turbo base too or do you suppose i could use fine tuned turbo instead of base?

[gentle musk]

I'm using the simple node, because I don't understand what the blocks stuff is

[wispy canopy]

@gentle musk, take not of my freeu values, i changed them a bit, added lcm as sampler

[wispy canopy]

turbo base

[wispy canopy]

low blocks -> composition, midle blocks, summary of composition, ready to become output, output -> actual image output

[nova remnant]

ok

nubby needs to get in this game

1 sec render time and high quality images

[gentle musk]

Still makes it look worse at my current settings

[wispy canopy]

😦

[wispy canopy]

my workflow in image

[gentle musk]

it can change image to image which is the problem

Sometimes FreeU makes it look better, sometimes it doesn't

[wispy canopy]

that's a merge issue

[gentle musk]

Sometimes having the LCM Lora makes it look better, sometimes it doesn't

[whole portal]

lcm is a tricky beast

[nova remnant]

are these necessary with sdxl bad anatomy, bad hands, bad proportions, distorted, cloned

[gentle musk]

I barely ever use negatives with SDXL

[nova remnant]

sweet

[wispy canopy]

no

sdxl uses little to no negatives

join the LCM turbo hyper gang

[gentle musk]

I can get this with 2 steps after messing around with the model merg settings

[wispy canopy]

imagine if you gave it 6 😮

[gentle musk]

It's not fully done though, you can still see the weirdness around the edges

[wispy canopy]

still, we went from unrecognizable bs in 2 steps to nearly image complete in 2 steps

[gentle musk]

6 steps is

With your FreeU settings

And with mine

[wispy canopy]

i like mine just a bit better, seems a bit cleaner

but that's just preference

freeu does clean up a lot of the "blotchiness"

[gentle musk]

I don't think I even need the LCM Lora, this is without

4 steps, using dpmpp_2m_sde

3 steps just lacks that last bit of detail

[wispy canopy]

yeah, i like 4 steps better too

nothing vs freeu vs freeu+lora -> need to tweak some settings some more i think, lcm lora is a tough cookie to balance

[gentle musk]

2 steps with Uni_PC

[nova remnant]

you guys might wanna incorporate this to your merges maybe? https://civitai.com/models/119032?modelVersionId=175819

[wispy canopy]

bit blurry

i mean, it's coherent

but, quality could use a step or 2

[gentle musk]

It actually get worse for some reason at higher steps with Uni_PC

[wispy canopy]

ah, yeah, i think the sde and lcm sampler are the only sampler that can handle the noise multiplication that turbo and/or lcm lora create

[gentle musk]

5 seconds to generate with an IPAdapter fed into it

10 seconds if I use 2 IPAdapters, but that's because I start running out of VRAM

[last vale]

Dumb question, but how do you achive so low steps without the LCM Lora?

[gentle musk]

you need to merge SDXL Turbo and another model

[wispy canopy]

STRAIGHT HIT!

[last vale]

What weights did you use? (I am not sure if it is called weights xD)

[gentle musk]

I'm still testing different weights

I'm using the modelmergeblocks, but I don't really understand what it's doing so I'm just changing stuff until it looks good

[last vale]

You are using comfyui to merge them?

[gentle musk]

yeah

[last vale]

Oh nevermind I was stupid

It only works with XL models right? If yes, what model did you use there? I haven't really found an usable anime XL model...

[gentle musk]

CounterfeitXL

[last vale]

For me they always come out blurry as if they didn't pick up the "turbo style"...

[gentle musk]

3 steps never seems to get the eyes right

not bad though

[nova remnant]

oh damn, even her nails came out right

[wispy canopy]

love my model ❤️

it gets "so much just right"

[nova remnant]

im new to merging, i just got all the necessary files now to look for a basic to do guides

[wispy canopy]

she's marveling her own hands 😮 wooow i never thought we could have such pretty hands

[gentle musk]

I've broken something

There we go, 10 images 15 seconds

[nova remnant]

these anime eyes are great, i dont like downward slope a lot of the models have

i wonder how you achieved that by merging counterfeit

[gentle musk]

That one was a different model

[nova remnant]

lol i see

[gentle musk]

This is counterfeitXL

[nova remnant]

this eye also looks good

although a bit disempowered

the images i see at civitai counterfeit page, those eyes are too mellow

[gentle musk]

If I put more of Counterfeit into the merge it changes them a bit

[nova remnant]

this new stuff with lcm lora and turbo given a new life to sdxl

the fast gen really removes some of the hesitance to use it

i hope this will lead to putting more focus on quality now

the training

[nova remnant]

trying to give this anime look ...
a portrait of a woman, blonde hair, blue eyes, coral pink lips, hoop earrings, pearl necklace, tank top, navel cutout, miniskirt, <lora:lcm-lora-sdxl:0.7>

not working out too well with anime style

[gentle musk]

I wouldn't use "portrait" as that tends to try and to the top half only

[nova remnant]

ah yes good point

[gentle musk]

And full body as well, is what I needed to stop it doing close-ups

[nova remnant]

(anime style:1.3), a blonde girl with blue eyes, coral pink lips, wearing hoop earrings, pearl necklace, tank top with navel cutout, and a miniskirt, <lora:lcm-lora-sdxl:0.7>

[nova remnant]

mine's not anime :/

[gentle musk]

I'm using CounterfeitXL which is heavily loaded towards anime

It makes it by default usually

[nova remnant]

yeah i know but im trying to use a generic xl model to bring out both anime and hyper realism

the authors page shows all those images in the style of anime, seme real, and realism, im trying to do that by prompt

but heck, im so glad xl model now doesnt feel like a slow struggle to generate images with

[nova remnant]

is that sd1.5?

[gentle musk]

no

[nova remnant]

lol ok it looked like it

[gentle musk]

lmao

[nova remnant]

hah

[gentle musk]

The extra speed is nice though when I'm just messing around seeing what I can get.

[nova remnant]

its awesome

i dont notice much difference now from sd1.5

except i still need to get 32gb ram

[wispy canopy]

@nova remnant something like this ?

can't get the hoop earrings to stick tho

[nova remnant]

kind of but not that flat

[nova remnant]

something along this style

[wispy canopy]

you mean the cell shading?

[nova remnant]

or this

[nova remnant]

close

[nova remnant]

yeah close but not blending in the whole image naturally

[wispy canopy]

well, i'm not specifying anyhtning for the environment, so it just does whatever

[nova remnant]

i would love it if its possible to bring out that sd1.5 .. 2.5 D look

[nova remnant]

that

blending in more naturally

[wispy canopy]

?

[nova remnant]

you have that touch to it but needs to blend in smoother

[wispy canopy]

softer colors?

[nova remnant]

contrast

[wispy canopy]

she do be sexy tho

[nova remnant]

lot better, a bit softer contrast would be nice

[wispy canopy]

yeah, trying 🙂

[nova remnant]

cause with high contrast the subject kinda sticks out the image .. that hurts the eyes lol

[wispy canopy]

i can't get the "roughness" from 1.5 models tho, my model kinda... sucks at that lol

making some fine damsels tho

[nova remnant]

try some comparison with detailed background and see how the subject blends in with the scene

[nova remnant]

the color is ok but contrast is too high

symmetry are good too

except few of the renders have artificially thin neck

[wispy canopy]

working at a resolution the model doesnt'quite like

[gentle musk]

I've noticed this turbo/LCM setup starts to struggle with things like eyes from far away

[wispy canopy]

again the neck tho

aight, imma go play some star ocean ^^

[wispy shard]

[nova remnant]

wooo much better with my anime theme, merged blue pencil with turbo base

head body proportion is bit off

a girl with brown hair, hazel eyes, in jogger's suit, standing in the garden. (anime theme:1.1), i think i got it...

where is @untold glacier these days?

[untold glacier]

A day in the life...

[nova remnant]

oh hey

cant read those, they are blurry

im trying a few merges to get turbo xl work with anime theme sdxl

[untold glacier]

https://tenor.com/view/feel-me-think-about-it-meme-gif-7715402

[nova remnant]

lol well i could zoom in

btw how do you like my anime merge with turbo xl

[untold glacier]

Looks neat. Very clean.

[nova remnant]

really fast gen too with lcm lora + turbo xl

but no hires ... dont need it

[supple raptor]

I think it could really benefit from a couple more steps or something. It has like a highly compressed jpg effect. Overall composition is quite good though.

[nova remnant]

currently using 5 steps.. and cfg 1.5, i think i could add a few more steps and see how it comes out

steps 8, cfg 2

this is a custom merge btw, first time doing it so might not be fully optimum

also not using hires fix

[supple raptor]

That seems quite a bit better, it's still there if you really zoom in, but at native res it's not noticeable at all.

[nova remnant]

yes, wonder if hires can fix those detail

actually i can't do hires while using sdxl with my 16gig ram

[supple raptor]

Not even with tiled vae at like 1.5x upscaling?

[nova remnant]

not sure, i havent tried tiled vae, when i was using sd1.5 his res fix by 2x worked very nicely so didnt bother to look into tiled vae

[nova remnant]

you dont see the metalic or plastic feel

the artificial metalic gloss was one of my biggest issue with sdxl, and im glad i dont have that in my custom merge

[terse ore]

Yay it works! I deleted my automatic1111 (I use comfy to generate) and the PS plug-in, and reinstalled to the newest versions (both were months out of date). Now I can be a pro

[terse ore]

Has anyone tried using the toyxyz blender rig for controlnet? It generates openpose, canny, depth, mediapipeface, and a bunch of other controllers all at once. It can also be combined with a 3d character rig, but also comes with a rig customizable by body type with sliders.

[nova remnant]

@wispy canopy these pics are all rendered with merged model using turbo sdxl and lcm lora

[native halo]

its one of the models out there

[supple raptor]

Wait... why using both? SD isn't vodka, it doesn't need double distilled.

[nova remnant]

i think i lcm lora has a level of coherence

didnt want to miss that

also hands are much better with it

let me render that last image by its seed and not use lcm lora

missing a lot of detail

i wonder if the turbo aspect of it is working on this custom merge

im not using lcm lora on it btw

[supple raptor]

tbh, based on that image I'm not sure that merging the SDXL turbo model is really helping things.

[nova remnant]

well i dunno but im not getting that harsh metalic glow

same seed and same settings but with lcm lora ...
settings are steps 8, cfg 2

different seed

going to render that one with bluepencil model w/o turbo

i had 0.7 multiplier from blue pencil

[stable crane]

interesting

model looks pretty good overall

[nova remnant]

thanks but im not fully aware of what i did lol

first time

[stable crane]

so that model has sdxl turbo + lcm lora?

• bluepencil it sounds like

[nova remnant]

yeah well i merged bluepencil and reproduction then took that new model and merged with base turbo model

[native halo]

just a few more extensions like lcm + another sdxl turbo re-release and we will finally manage to replicate the sd 1.5 effect

[nova remnant]

something i've been missing with sdxl lol... but this model merge surprised me

[stable crane]

I was trying to get an LLM node in my comfyui then got hit with cuda bullshit

[stable crane]

nice

[nova remnant]

😬

sd 1.5 feel

and by removing (anime theme:1.2) with same seed

[stable crane]

nice

[nova remnant]

i deleted a whole bunch of sd 1.5 models from c drive

[stable crane]

I see reproduction is based on https://huggingface.co/Linaqruf/animagine-xl-2.0

might almost look better than reproduction

[nova remnant]

ooooh yeah it looks better

grabbing that one

not the same seed

but those 2 images have similar look

im rendeirng these at 768x1024 w/o hires fix, and with adetailer and deepshrink enabled takes me about 15s

[stable crane]

some you can't really tell are sdxl

[nova remnant]

well there were two things i was unappy with sdxl ... the metalic gloss and slow render time, both of those are sorted now

and i can actually appreciate the coherence of sdxl along with lcm lora

[stable crane]

Idk what you mean by metallic gloss

but it did feel like, too hyperreal for anime

or something

[stable crane]

like this one still looks SDXL to me

[stable crane]

but not this one

[nova remnant]

thats what i mean .. the shine felt artificial

[nova remnant]

merged 3 models this time then took the new model to merge with turbo xl

[keen hollow]

evenin' - what's new?

[native halo]

Kissinger died

[keen hollow]

hm?

[native halo]

👍

[keen hollow]

huh.. how 'bout that

100 years old

[nova remnant]

3rd merge

[keen hollow]

gold jewels

have some dairy queen...

[nova remnant]

lol they made a mess

[keen hollow]

yeah it's a more realistic dq

[nova remnant]

cool gen 🙂

[keen hollow]

Amazon Warrior

[nova remnant]

3 different merges

cool images you got there

[keen hollow]

another amazon warrior

[nova remnant]

nice but why the spoiler?

[keen hollow]

eh I dunno

belly button? 🙂

[nova remnant]

@bitter latch the token thing i was having issue with, it was a slight bug with the file name that was conflicting with my prompt, there is no ilmit to turbo xl models, i even merged models to make a custom turbo xl with lcm lora

[bitter latch]

kewl

hmm , "Epic Ghibli sci-fi art style" makes some interesting art

sigh I wish my check would drop early for once. sigh

[bitter latch]

ha so tired been putting miki instead of miku (facepalm)

[wispy canopy]

high quality, and good speed!

@nova remnant, did you try the turboxl and lcm combo on some other models too?

[nova remnant]

i merged three models for that combo

using my custom merge for images now

[wispy canopy]

reason i'm asking, models perform aobut the same when you just swap in other base model, for this merge strategy?

[nova remnant]

well i also made three separate merges, and as it turns out one of them stood out that im keeping for use

[wispy canopy]

^^

[nova remnant]

btw where we left off, these two images here looks much better to blend in with the whole scene

[wispy canopy]

^^ yeah i kinda got to work again

am doing a model merge again, this time even merging down to elemental level

i know my version 1.2 has had good hands, so i want those elements back, so very carefully now going through multiple prompts at the same time, checking output is still ok across all level, improving hands and details, and retaining it's newfound playfulness

[wispy canopy]

8 steps no turbo, this model is becoming quite good

[nova remnant]

im fully leveraging lcm lora

with 8 steps and 2 cfg

[wispy canopy]

i think a pure anime model will do better anime at 8 steps than mine will

[nova remnant]

my merge is not entirely animne tho but good at it

i can generate some realistic images by taking off (anime theme)

[wispy canopy]

then again, on the same seed i have also this and this , making sure the composition is generally model wide good, not just a single specific thing

comfy autoqueue really coming in clutch

[nova remnant]

how do i make tiled vae work? do i just enable the settings as shown in this image and enable hires fix ?

i was thinking tiled vae was a seperate process and didnt need hires fix to be enabled

[wispy canopy]

it checks if the total image size is larger or smaller than set tile size -> and then splits up vae if needed

so i always try to get it to NOT split on lower generation, and then fire on highres fix

[supple raptor]

I check every box and set 1344 encoder tile size and 128 decoder tile size on 12GB VRAM. Lower VRAM will want smaller tile sizes. Below 8GB you don't want to move VAE to GPU either. Encoder is used when converting pixels to latents, and decoder is latents to pixels. You'll see when it activates in console.

[nova remnant]

so i'd have to use hires fix and depending how large that image will get that will trigger the tiled vae script given its checked?

[supple raptor]

Usually, yes or some other method that allows genning at higher than typical native resolutions.

[nova remnant]

ahh ok that makes sense now ty both

trying one now with 2x

[supple raptor]

Ideally you want both tile sizes as large as possible before running out of VRAM at the resolution and batch size you're generating.

[nova remnant]

the thing about hires fix with sdxl is that it eats up my ram not vram

[supple raptor]

If it's eating up RAM it's because torch is falling back to system RAM because there isn't enough VRAM. That's what Tiled VAE is going to try to avoid.

[nova remnant]

it took me a while to render one image 02:15<00:00, 10.25s/it]

usually w/o hires it takes me 15s

[supple raptor]

Yeah that's definitely falling back to system RAM.

[nova remnant]

how do i prevent that, i had the tiled vae boxes checked

[supple raptor]

It's possible that the actual inference is exceeding your VRAM, in which case your only real option is reducing the render resolution. Comfy is also considerably more VRAM efficient so that is something to try as well.

[wispy canopy]

alice in wonderland hare put on his business suit, no more mr nice hare

[nova remnant]

ooooh

ok in that case i should just render at high res and not use hires fix

i can render ok at 1500px

what gets me about hires fix is that when it tries to add the detail

[supple raptor]

Tiled Diffusion or SD Upscale script can be used to upscale instead of Hires fix.

At least I think those work with XL. I use comfy if I want to run XL personally.

[nova remnant]

this one you mean

i didnt see much option to upscale the resolution within tiled vae

[wispy canopy]

comfy is more vram friendly for sure

also a lot less lingering memory issues, which happens at times with a1111 still, imho

[supple raptor]

Yeah, I dislike the noodles, but if I really want to run XL then it's worth it.

[wispy canopy]

don't have vram to spare?

[nova remnant]

im doing fine with a1111 its that upscaling thing cause i miss hires that was in my worksflow on sd1.5

[supple raptor]

I have 12GB. I can run it on A1111, but it's considerably faster on Comfy.

[nova remnant]

me too

but comfyui is not an option for me

if i were to mess with all the wire i wouldn't have the incentive to generate images

[wispy canopy]

it's faster in general, imho

i mean, a1111 just runs heavy

[supple raptor]

Stable Swarm might be worth looking into. It uses comfy as the backend so you get all the efficiency benefits but with an actual UI.

[nova remnant]

why is there a move to kill a1111?

[wispy canopy]

nah, i mean, a1111 for sure isn't bad

[nova remnant]

is the a1111 dev team not keeping up i thought a1111 was the most popular

[wispy canopy]

it's VERY easy and comfortable in use

i'd say both A1111 and comfy have their own uses

but for clean, mean running -> comfy beats out a1111 for sure

[supple raptor]

It does, but it's still faster than using comfy if you change stuff after you factor in the time spent fucking around with nodes that would have taken 2 clicks in A1111

[gloomy lodge]

leggingssss

[nova remnant]

exactly how i feel

[supple raptor]

A1111 is just a clusterfuck of code held together by bubblegum and scotch tape. Nothing beats it for ease of use and available features, but it sure would be nice to get something that does.

[nova remnant]

so obviously a1111 dev are slacking

[supple raptor]

He's been doing a pretty good job at maintaining it lately, but it's kind of to a point where fixing the problems with it would be nearly impossible without destroying compatibility with many of the extensions which would then require countless other devs to update their code.

[wispy canopy]

true.

that ease of -> just click and go

[supple raptor]

Controlnet in A1111: add image, check box, click radio button, move a couple sliders.
Controlnet in Comfy: add these 15 nodes if you want this specific controlnet, but you need those 12 instead for this other one, make sure to connect all the noodles just right or nothing is going to work.

[nova remnant]

i remember when sdxl came out and how a1111 had no support for it when everyone was excited to try new sd model, then even after a few hacks a1111 didn't have the speed to keep up with comfy, and few weeks later a1111 pulled through and had good benchmark ... the ease of use with a1111 is an important factor for me and its also more intuitive than comfy ui

[supple raptor]

I would probably bite the bullet and get used to Comfy if inpainting wasn't so cursed and the mobile experience was usable. I gen from my phone quite a bit and that's so bad on Comfy that it might as well not even be possible.

[nova remnant]

im not giving up on a1111

@nova remnant hewo

been a while

idk

at least some hours since i've seen this chat

[nova remnant]

https://tenor.com/view/thicc-slap-cat-bouncy-gif-20058397

i've been afkish for few days

didnt have the time to play with ai