🍥｜anime

[wispy canopy]

(feel the pain we sdxl users have)

[gusty palm]

1.5 can already do that does it reduce the doubling effect or somen

[wispy canopy]

bingo

[gentle musk]

Yes

[wispy canopy]

i mean, i can generate at 4096x2048 just fine with any model

it's just not coherent

and messy

[gusty palm]

hmmm

im doing like 2:1 aspect ratio i might try it

[gentle musk]

There's some Gradent Latent thing that Kohya has made as well recently, but I'm not sure the point of it.

[gusty palm]

384*768 i get some doubling but its not super bad

[wispy canopy]

yeah, tha'ts just solved with khoya's highres, when set up right

image composition then is just clean

does require some tinkering to get right, you need a feel for it

[gusty palm]

might try if i get fustated with wide pics

[wispy canopy]

did try those, ezpz

[gentle musk]

It takes stuff like this mess

And you get this instead

That's 1080p with SDXL

[gusty palm]

ah

[gentle musk]

It takes a bit longer and you need a decent amount of VRAM, for SDXL at least, but you get decent results.

[wispy canopy]

tried for sd1.5 models (darkalfa), works just fine

[gusty palm]

is the one u linked up there for 1.5

[gentle musk]

It's for both

[gusty palm]

oh ok

[wispy canopy]

kohya's is just a calculation method for sampling

fuuu

[gentle musk]

It resizes the u-net I believe

[wispy canopy]

can't extract 4096x2048 image from gpu with 24gb

the vae needs a whopping 36gb to calculate the image

[gentle musk]

So intead of stretching out the u-net to match the image resolution, which causes all the doubling, it resizes it and does "things" to it

[gusty palm]

thats lewd

ill pray for u
net

[wispy canopy]

not to mention, the fckn 10 minutes of calculation time

that hurts the most -> 10 minutes ot end up in "cuda out of memory"

@gusty palm 2048x2048 sdxl, same as 1024x1024 sd1.5

[gusty palm]

figure out the upperlimit and render at like just below that

[wispy canopy]

2048x2048 is over the limit already 😄

it's just to make the point -> you can do a single pass image, and have the image be consistent

at over double the resolution

@gusty palm same image, darkalfa (sd1.5) 1536x1024, single shot, first image that came out

different model

rendertime, 60 seconds on a 3090

not sure if sd1.5 people like that tho

[gusty palm]

thats cool

[wispy canopy]

no super highres, 50 seconds 640x640x8 (sd1.5)

1:06, 8x768x768 (sdxl)

from my testing, sd1.5 starts becoming slower when using kohya's at really high res, so might be best to just use regular upscaling in that case

channel be dead like always lately, lets burn bridges. uhhh...

[wispy canopy]

risky gamble, i suppose

[gentle musk]

cheater

[old dirge]

a cute cat

[whole portal]

I am using klF8Anime2 for 1.5 models; any recommendations for Vae for SDXL?

[gentle musk]

I just use the base vae, I've not seen very many other vaes for it really

[wispy canopy]

the 0.9 one

[wispy canopy]

base 1.0 vae is bugged -> you need the 0.9 one

[gentle musk]

Yeah sorry not the base the 0.9

[gentle musk]

Yeah I never used the 1.0 one, because I saw the awful things it did to images

[wispy canopy]

god, don't remind me

it WAS awful

people be like "dude, the vae is bugged"

stability be like "should be fine"

nope

[gusty palm]

what pixel resolution is considered 4k

2048-4096?

[whole portal]

has anyone tried this one? https://civitai.com/models/152040/xlvaec?modelVersionId=202582

[gentle musk]

the example they provide shows very little difference, just slight colour changes, which you could do very quickly after generation anyway

[wispy canopy]

any multiplication of dimensions where the total would be over 4.000.000(si) pixels

[stable crane]

looks good

[wispy canopy]

hotheaded

[wispy canopy]

yeah, that's gonna be hard pass for me

the colors are pretty good, but the raster effect is fckn atrocious

colors can be corrected afterwards, image artifacts (Raster/screendoor effect) cannot

[whole portal]

https://civitai.com/models/118483/sdxl-vae?modelVersionId=128527

this one?

i am to lazy to download manually from hf 🙂

[wispy canopy]

@cyan burrow this one loves your efforts!

[whole portal]

f... no file

[wispy canopy]

@whole portal https://huggingface.co/nyanko7/sdxl-vae-0.9/resolve/main/sdxl-vae-0.9.safetensors?download=true

😘

[whole portal]

thanks 🙂

[nova remnant]

https://huggingface.co/stabilityai/sdxl-vae/blob/main/sdxl_vae.safetensors

[wispy canopy]

one of my fav prompts!

[wispy canopy]

isn't that the 1.0 broken one?

[nova remnant]

dont think its broken

[wispy canopy]

1.0 is broken, for sure

on low angle constrasting slopes you get artifacting

[nova remnant]

not sure about that

hugging face is still hosting it

[wispy canopy]

between hosting it, and being a broken file, are two seperate things

the 1.0 vae is broken, and no self-respecting model uses it

[whole portal]

this one doesnt work on my mac

Loading VAE weights specified in settings: /Users/viking/stable-diffusion-webui/models/VAE/sdxl-vae-0.9.safetensors
changing setting sd_vae to sdxl-vae-0.9.safetensors: RuntimeError

[wispy canopy]

model? vae? ui?

what isn't working?

[whole portal]

this vae

gives me runtime error

[wispy canopy]

what model are you running?

and what ui?

[whole portal]

a1111

https://civitai.com/models/200061/hephaistosnextgenxl-lcmbf16hybrid?modelVersionId=225116

[wispy canopy]

it "should" work

but then again, any model has a vae embedded

so you can try without

and iirc, hephaistos uses the 0.9, like eveyrone

[whole portal]

https://civitai.com/models/152040/xlvaec?modelVersionId=202582 - this one works fine

[whole portal]

i didn't see that info

if it has it, then i will not use it

[wispy canopy]

people just embed it in the model, like humans breathe 🙂

it's just there 🙂

[whole portal]

this is image without any vae

and this one is that vae you didnt like

i do not thing it has vae in it

[wispy canopy]

^^ in that case, keep using it

obvious improvement!

[whole portal]

i will try 0.9 with some other model, to see if it is working at all here

[wispy canopy]

damn, very nice common prompt 😮

common prompt: anime, creative fantastical reality, symbolic, lineart

[whole portal]

Loading VAE weights specified in settings: /Users/viking/stable-diffusion-webui/models/VAE/sdxl-vae-1.0.safetensors
Applying attention optimization: sub-quadratic... done.
VAE weights loaded.

1.0 loads without a probel

0.9 produce errors 😦

[wispy canopy]

ugh. probably need the fixed 0.9 vae

can i be lazy and not look for it? (am cooking atm)

[whole portal]

0.9 will never run here

with 1.0

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1178044637605281892/00095-178219412.png?ex=6574b690&is=65624190&hm=478272014ac79d83f6aca574e21b6e8e880a3d70f1f3b406175cc8dc62e47c5f&=&format=webp&width=530&height=662

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1178053354367234149/00136-3291350613.png?ex=6574beae&is=656249ae&hm=1b435fd9cd97a94de9147f661ba3f5f32f97a50cd13c5efd3de596396d07e12a&=&format=webp&width=441&height=662

[wispy canopy]

this chibi aspires to becomes as cool as nee-chan

[gusty palm]

RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 2 but got size 4 for tensor number 1 in the list.

anyone know what this means

[nova remnant]

That means you got an error

[gusty palm]

i grabbed a random hugging face model with no readme info or anything so i dk what to expect

[wispy canopy]

restart webui

[steady grail]

weight of the model?

[wispy canopy]

unless it's an actual error

[wispy canopy]

@gusty palm i get that error with kohya's a lot -> usually means somethign shortcircuited

[nova remnant]

torq

[steady grail]

may be incompatibility between model and lora or embeddings

[gusty palm]

ok ill try no lora

it gave a big error message in the log ttoo but i didnt wanna spam the chat with gibberish

RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 1 but got size 2 for tensor number 1 in the list.

something is different but it still not doin stuff, it seems to be generating it doesnt do this til after the gen

[autumn violet]

Compelling storyline there

[autumn violet]

The symbolism of the cup disappearing between the bottom two images is very deep

[gusty palm]

thats so good

i switched back to my merge and its working so maybe that mystery ckpt is just incompatible or something

[nova remnant]

thanks, was fiddling with controlnet

[supple raptor]

Most likely it's an sd2.x based model

[gusty palm]

was thinking that might be it

[whole portal]

what is the name of the model?

[nova remnant]

He ate the cup

[cyan burrow]

@wispy canopy here's a prompt for you Quantic demon with razor-hair dreads, sharp facial features, reptile eyes, onyx robe, obsidian sword, digital art, dark fantasy illustration, cinematic shot, wide angle, 70mm Lens, 8k, HDR

I haven't tried it

[wispy canopy]

is kinda non anime, as i'm on a non-anime model, hope you like it ^^

[cyan burrow]

ah looks pretty good, here's my results. First SDXL

SD 1.6

SD 1.6 got it best

[wispy canopy]

conceptually yes. blurry, yes

[cyan burrow]

anime art style

I'm still weirdly good at prompting

but I suck at using extensions

getting my new computer mouse tomorrow, Logitech PRO X SUPERLIGHT Wireless

[hallow thorn]

Can someone pretty pretty please help me with outpainting one anime artwork?

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1178083760357523517/00122-3291350599.png?ex=6574db00&is=65626600&hm=eefec0e72baf4f4857b69564a0aed8136e54c31405c5e25ea4dfcbf7e3dc314a&=&format=webp&width=441&height=662

[crystal prism]

ran it through a few different models

[cyan burrow]

cool, none really got close to what I was envisioning but still cool

[nova remnant]

very mortal kombat

@native halo more body suit maid stuff

[steady grail]

lcm is not good for upscaling

[supple raptor]

I mean... lcm is just worse in general than not using it for everything except speed.

[steady grail]

agree

i'm trying to mix it, followed by a ksampler with karras , or sandwiched between to karras ksampler's

the result looks always worse

the low cfg impacts terribly on the prompt fedelty

[whole portal]

lcm should be used with lcm sampler, unipc or euler a

cfg MUST be 1-2 not more than that

[steady grail]

yeah, and i think that's the bigger problem. a low cfg means low details

[whole portal]

low details?

[steady grail]

yeah, the image get influenced less from the prompt, so you get a more bland image

[steady grail]

look at the land in the image

[whole portal]

this is lcm

[steady grail]

that's because the model

[whole portal]

Steps: 10, Sampler: Euler a, CFG scale: 1.7

[steady grail]

with low cfg, the reult is less influenced by the prompt, the more the cfg, the more the result follow the prompt

[whole portal]

same model, just with wrong settings

[steady grail]

generating many images, you will notice the similarities, based on the bias of the model

and no metter the prompt, they will always look similar between each others

[stable crane]

nice

[supple raptor]

Even the arxiv paper for LCM admits that DPM++ samplers produce outputs with higher CLIP Score as low as 8 steps. LCM is a method to increase speed, by allowing passable outputs at very low step counts which makes things like animation and realtime generation more practical. If you don't mind waiting longer for the image, then the average quality will be higher by using proper settings without LCM.

LCM is neat, especially for people with low compute who would otherwise be waiting several minutes per image, but you're only fooling yourself if you think it's providing better results than you could get at higher step counts without it.

[whole portal]

i need LCM

or i will wait forever

Steps: 8, Sampler: LCM, CFG scale: 1.5, Seed: 1729141576, Size: 512x768, Model hash: ec41bd2a82, Model: photon_v1, VAE hash: 63aeecb90f, VAE: vae-ft-mse-840000-ema-pruned.ckpt, Denoising strength: 0.3, Hires upscale: 2, Hires steps: 4, Hires upscaler: R-ESRGAN 4x+, Lora hashes: "LCM_LoRA_Weights_SD15: aaebf6360f7d", Version: v1.6.0-2-g4afaaf8a

Time taken: 11 min. 27.0 sec.

[steady grail]

too much uncreative for me

[whole portal]

i would wait an hour for this without LCM

[supple raptor]

What GPU do you have?

[steady grail]

i have only 4 g and i just need 2 min to gen an image with 35 steps and 25 with a full upscale

[whole portal]

M1 mac

a1111 is far for optimized for macs

but at least it works

[supple raptor]

Ah... yeah you're missing out on a lot of optimizations that nvidia GPUs get. LCM is a very good option for M1 silicon.

[steady grail]

comfy don't works on mac?

i suppose

[whole portal]

it does

[steady grail]

if you get a good workflow, you will generate way faster with a low spek pc

too bad you will spend more time having fun finetuning the workflow, than generating images lol

[whole portal]

my point was that R-ESRGAN 4x+ is great for lcm upscale

[steady grail]

at least that's what happened to me

[whole portal]

not to show creativity with that picture

[supple raptor]

I'd probably recommend looking into InvokeAI for M1 support. https://github.com/invoke-ai/InvokeAI

Not sure if they have LCM yet, so that would be something to look into.

[steady grail]

i like these very much

hat, fatal and animesharp, are very sharp

but real hat is not very good for anime upscales

[whole portal]

i usually use R-ESRGAN 4x+ Anime6B

[steady grail]

i feel esrgan lacks in details :/

looks like a little crude as an upscaler

[supple raptor]

I haven't bothered to use another upscaler since switching to YandereNeoXL-200k https://openmodeldb.info/models/4x-NMKD-YandereNeo-XL

[steady grail]

this model is for sdxl?

[supple raptor]

ultimately, unless your denoising is too low or they're oversmoothing, they all make very little difference.

[supple raptor]

No, it's just an ESRGAN upscaler like all of the ones in your list.

[steady grail]

ah, got confused by xl at the end

[supple raptor]

Understandable. It's just a coincidence though. It was released in 2021.

[steady grail]

oh

look at this fools, making an upscaler for sdxl even before sdxl lol

seriously, i'm not serious XD

[supple raptor]

Maybe stability made SDXL for the YandereNeoXL-200k upscaler

[wispy canopy]

not with lcm sampler/lora

[steady grail]

mh... touché.

[wispy canopy]

think of lcm as GIGANTIC multiplier

both in step and cfg sensitivity

[steady grail]

so, why the results are so bland? is the sampler itself?

[wispy canopy]

what is bland?

[supple raptor]

It's essentially distilling the model.

[steady grail]

oh....

[wispy canopy]

this is on lcm... not bland

imho

but, some stuff is required to set it up properly -> lcm lora (if not a native lcm type model merge) + lcm sampler + correct steps/cfg

preferably freeu

[steady grail]

maybe my problem are the not compatible models

[whole portal]

what setting did you use with lcm?

[wispy canopy]

regular models will work just fine with the respective lcm loras

[steady grail]

the suggested once, lora at 1, lcm at 1 and 5 steps, with lcm sampler and sgm_uniform scheduler

1.5 model with 1.5 lora

[wispy canopy]

my lcm settings (both sd1.5 and sdxl) a1111
LCM lora strength: 0.8
sampler: LCM
steps: 12-20 (because i like quality)
Freeu

hope i didn't miss stuff

[steady grail]

why lcm lora at 0.8?

[wispy canopy]

1 starts overcooking images imho, but that's just personal preference

[steady grail]

even with cfg at 2?

[whole portal]

i am using 0.7

[supple raptor]

31.1s for 20/15 steps on DPM++ 2M Karras Vs 26.5s for 8/8 steps on LCM. I'll spend the extra 5 seconds every time.

[whole portal]

or 0.3

lora:LCM_LoRA_Weights_SD15:0.7

cfg 1.5

[supple raptor]

In addition to just looking worse, the LCM result tried to make the eyes greenish and ignored the hairbow.

[whole portal]

10 steps

[wispy canopy]

why is that quality so bad? 😮 and so slow

[supple raptor]

because LCM is not great. It's faster, but it will almost always look worse than comparing it to sane settings with a better sampler.

[whole portal]

i use lora:LCM_LoRA_Weights_SD15:0.7 if i do not use any other lora, if i do, i lower it to lora:LCM_LoRA_Weights_SD15:0.3

[wispy canopy]

i do not agree, to be honest

[whole portal]

and i usually set other lora to 0.8 or 0.9

[wispy canopy]

people don't believe the power of LCM

[whole portal]

i agree

https://civitai.com/models/200061/hephaistosnextgenxl-lcmbf16hybrid is my favorite model in this moment

[wispy canopy]

lcm has the prompt comprehension ddim has, and no quality drawbacks

[supple raptor]

lora at 0.7 and at 0.3

[wispy canopy]

"when i don't use it properly, it will fail"

yeah. it does

use lcm sampler and use high lora strength, or use dpm++ 2sa karras with low lora strenght with penultimate sigma off

[supple raptor]

viking just said they use it lower than 1.0 so I tried it

[wispy canopy]

that's the only 2 settings LCM lora will function.

i've tested all samplers

[whole portal]

your problem is sampler

[supple raptor]

I'm using LCM sampler

[wispy canopy]

a1111?

or comfy?

[supple raptor]

a1111

[wispy canopy]

freeu?

[supple raptor]

no, and I didn't use freeu for the dpm++ image either

[whole portal]

can you paste your generative data?

cinematic film still, young geisha, cyberpunk, armor, tattoo, smiling, undefined lora:LCM_LoRA_Weights_SD15:0.7, A style of drawing that emulates the distinctive visual style of Japanese anime and manga, often featuring colorful characters with large eyes and exaggerated expressions, smooth soft skin, beautiful eyes, intricate colored hair, symmetrical, anime wide eyes, soft lighting, detailed face, by makoto shinkai, stanley artgerm lau, wlop, rossdraws, concept art, digital painting
Negative prompt: Watermark, Text, censored, (((internal-organs-outside-the-body, internal-organs-visible, anatomy-description, unprompted-nsfw, worst-human-external-anatomy, worst-human-hands-anatomy, worst-human-fingers-anatomy, worst-detailed-eyes, worst-detailed-fingers, worst-human-feet-anatomy, worst-human-toes-anatomy, worst-detailed-feet, worst-detailed-toes, camera, smartphone, worst-facial-details, ugly-detailed-fingers, ugly-detailed-toes, fingers-in-worst-possible-shape, worst-detailed-eyes, undetailed-eyes, undetailed-fingers, undetailed-toes))), (((bad anatomy, disfigured, mutated, extra limb, missing limb, floating limbs, disconnected limbs, disalocated limbs, abnormal limb position, disconnected head, malformed hands, long neck, mutated hands and fingers, bad hands, missing fingers, mutation,bad hands, fused hand, missing hand, disappearing arms, disappearing thigh, disappearing calf, disappearing legs, missing fingers, fused fingers, abnormal eye proportion, abnormal hands, abnormal legs, abnormal feet, abnormal fingers))),
Steps: 10, Sampler: LCM, CFG scale: 1.5, Seed: 1040800543, Size: 512x768, Model hash: ec41bd2a82, Model: photon_v1, VAE hash: 63aeecb90f, VAE: klF8Anime2.safetensors, Lora hashes: "LCM_LoRA_Weights_SD15: aaebf6360f7d", Version: v1.6.0-2-g4afaaf8a

[supple raptor]

masterpiece, best quality, BREAK c.c. green hair, code geass, (tsurime:1.3), crossed bangs, yellow eyes, long hair, hair bow, white sundress, looking down, shy, bashful, standing, crossed ankles, from behind, looking back, adult, medium breasts, outdoors, nature, flowers, wisteria <lora:LCM_LoRA_Weights_SD15:1.0> Negative prompt: (Worst Quality, Low Quality:1.2), border, nsfw, skimpy, grayscale, 2girls, (watermark:1.2) Steps: 8, Sampler: LCM, CFG scale: 1.5, Seed: 1594577000, Size: 544x960, Model hash: fea68357f5, Model: 9-AAAAAAA.fp16, VAE hash: f88e1cfb7c, 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: 23.11.1, FreeU Stages: "[{\"backbone_factor\": 1.2, \"skip_factor\": 0.9}, {\"backbone_factor\": 1.4, \"skip_factor\": 0.2}]", FreeU Schedule: "0.0, 1.0, 0.0", FreeU Version: 2, Hires upscale: 2, Hires steps: 8, Hires upscaler: 4x_NMKD-YandereNeoXL_200k, Lora hashes: "LCM_LoRA_Weights_SD15: aaebf6360f7d", Pad conds: True, Version: v1.6.0-2-g4afaaf8a

[whole portal]

try with klF8Anime2.safetensors

[supple raptor]

That's what I am using, I just lowered the contrast on mine by 10% to resolve the color bleeding that the wd VAE normally has.

[whole portal]

and try without upscaler first

[wispy canopy]

sd1.4 no lcm euler vs full lcm setup, same render time (2 seconds, 15 steps)

seriously, get your lcm setup correct.

[supple raptor]

tell me where it's set up wrong then.

the last one even used FreeU.

[wispy canopy]

8 steps is iffy, but should be doable

is that the old wd1.4?

ah no, just the vae

cfg can be raised to 2

[whole portal]

can you try without upscale?

[wispy canopy]

i dunno what clip skip 2 has as effect, i did not test that

but i do know freeu doesn't like it

[whole portal]

clipskip depends on the model and loras

[wispy canopy]

544x960 -> probably use kohya's highres if you want to do that

[whole portal]

it is not related to lcm

[supple raptor]

10/10 steps. a little better, but still very plasticy

[steady grail]

kohya works again?

[wispy canopy]

it works all the time?

i even figured out how it works 😮

[steady grail]

yesterday or 2 days ago, it didn't change the output at all, no matter the settings

[supple raptor]

15/15 steps

[steady grail]

still bland

[wispy canopy]

cfg at 2, lora strength at 1?

[supple raptor]

those were 1.5 and 1. I'll try that next.

[wispy canopy]

single shot 1792x1792, had to modify the script a bit tho

[supple raptor]

30/30 steps

[whole portal]

change vae

[nova remnant]

btw nubby im using 0.4 weght for the LCM LoRA while using LCM sampler

[wispy canopy]

unless you know what you're doing, LCM is not going to give a lot of extra over 16 steps

[supple raptor]

these problems are not related to VAE

[whole portal]

it can be

[wispy canopy]

yeah, i think that vae should be ok

[whole portal]

i had the same problem with sdxl

until i changed vae

[supple raptor]

The same VAE was used for the DPM++ image

SDXL only has a single usable VAE...

[wispy canopy]

@steady grail getting all the settings right puttings this in single shot generation is really cool tho

[whole portal]

see?

[supple raptor]

10/10 steps cfg 2. better, still considerably worse than 20/15 on DPM++

[whole portal]

https://civitai.com/models/152040/xlvaec?modelVersionId=202582

this one solved the problem with XL for me

[wispy canopy]

nubby, i do know that LCM listens a lot better to prompts

is any of those tags associated with grayish images?

[supple raptor]

it definitely does not

[nova remnant]

<lora:lcm-lora-sdv1-5:0.8>

[wispy canopy]

it does. especially across breaks

i mean, i'm using songmaster abstract shit

[supple raptor]

vanilla kl-f8-anime2 just to prove the point that the problem is not the vae

[wispy canopy]

anyway

[steady grail]

almost a 2000x2000

[wispy canopy]

sd 1.4 tried nubbies prompt xD

[steady grail]

how detailed it is at low res?

[wispy canopy]

512?

[steady grail]

y

[wispy canopy]

this is on hentai model 512x512, no loras, no negatives, 15 steps

i have no lora, so the only thing i prompted was c.c. green hair, code geass, (tsurime:1.3), crossed bangs, yellow eyes, long hair, hair bow, white sundress, looking down, shy, bashful, standing, crossed ankles, from behind, looking back, adult, medium breasts, outdoors, nature, flowers, wisteria

no negatives

i'm not sure where nubby is hitting a roadblock

[nova remnant]

with data

[whole portal]

Steps: 10, Sampler: LCM, CFG scale: 1.5, Seed: 1574975432, Size: 512x768, Model hash: a1535d0a42, Model: AnythingV5Ink_ink, VAE hash: 63aeecb90f, VAE: klF8Anime2.safetensors, Lora hashes: "LCM_LoRA_Weights_SD15: aaebf6360f7d", Version: v1.6.0-2-g4afaaf8a

lora:LCM_LoRA_Weights_SD15:0.7

[wispy canopy]

1024x1024 single shot with kohya's (not dialed in properly)

i think the color issue is really something on your end nubby

[supple raptor]

lcm low res 10 steps cfg2 with freeU vs DPM++ 2M Karras 20 steps cfg 7.5 no freeU. The comparison is much closer here with LCM winning at sharpness but the wisteria looks very artificial.

[supple raptor]

The colors aren't even the biggest complaint, it's the plastic like textures that it applies to everything.

[wispy canopy]

do the stuff the other are posting have it too?

[whole portal]

i see no problem with your prompt on AnythingV5Ink_ink with lcm

[wispy canopy]

i'll say it again with full confidence -> LCM is actually really good at getting your prompt and model in the actual result, closer than the other samplers do

[supple raptor]

Idk, nobody else is showing a comparison of their model and prompt with sane settings on DPM++2m Karras or some other decent sampler.

[wispy canopy]

it responds better to fucking weak tokens, over multiple breaks where only ddim ever did that

and with better contrast

[wispy canopy]

i have comparisons on my model, but that's sdxl

not sure if interested?

prompt has anime lineart, you tell me which listened better

[supple raptor]

It looks like these are running at the same CFG though right?

[wispy canopy]

might be right about that one, don't know if i can adjust that one in jinja

fuck

my disk crashed again

i can't do gens for the next 20 minutes

and should go to bed

sigh

[wispy canopy]

oh, it responded to my threat of immediate termination

seriously, i gotta RMA that disk

[whole portal]

lora weight 1 - do not use 1

0.9

[supple raptor]

DPM++2m Karras 7.5 CFG 20/15 steps no FreeU vs LCM 2 CFG 10/10 steps FreeU. Even on this model that has a more simplified art style, LCM is still a simpler gen.

[whole portal]

0.8

0.7

which one looks the best for you?

i like 0.8

[supple raptor]

for good measure here's the same 2m Karras config as above, but with FreeU enabled.

[supple raptor]

colors look better in 0.7 to me but 0.8 has details that make more sense. They all show some signs of the same texture problems I was talking about though.

[steady grail]

what settings do you use on free u?

[supple raptor]

I rarely use it. That was just the default SD1.4 settings.

[steady grail]

oh

i still need to study it properly

[supple raptor]

Like most things, it seems to sometimes help, and sometimes make things worse so it's largely just more buttons and switches to move around while you play the SD Gacha.

[whole portal]

0.75

cfg 2 made even bigger mess than lora weight of 1

[wispy canopy]

one got borked, the other got borked too, i'm missing a very important setting, probably

[supple raptor]

which looks better?

[wispy canopy]

2?

[ornate echo]

Depends on what you're going for. The 2nd image has cleaner lines, but you might like the look of the first image better.

[whole portal]

cfg 1.7 + weight 0.7 gives similar result as cfg 1.5 + weight 0.75

[supple raptor]

agreed, the issue I have with LCM is that it makes every model look basically like the second image.

[ornate echo]

Well yeah...essentially it's boiling down the data which in turn reduces variability.

[wispy canopy]

prompt understanding -> it's all the same prompt same seed lol (just 6 steps, so whatever pops up first is the actual image xD)

[supple raptor]

This is just proof that the model has no clue what to do with the prompt.

[wispy canopy]

which one?

[supple raptor]

all of them

[wispy canopy]

at 6 steps. amazing there's even coherent images

ah fuck no

that other issue again

lemme check for certainty.

somehow i can't lock seeds

[supple raptor]

Yeah, I mean the quality of them is pretty good for 6 steps, but the model clearly doesn't have any idea what the prompt should look like.

[wispy canopy]

yeah, was unlocked seeds

[nova remnant]

now prompt for maid bodysuits

[supple raptor]

this makes more sense.

[wispy canopy]

i agree with you on this one nubby, this one hits much better on karras

then again, dpm++2sa karras does like anime

correct comparison would be 12 steps dps a karras cfg 7 vs 24 steps lcm cfg 2 (will take the same time)

although i think with 24 steps i'd be able to raise to cfg to 3 perhaps?

[whole portal]

i finally understand why i had a problem with some images with LCM

[wispy canopy]

?

[whole portal]

i used UniPC until i installed animatediff to enable LCM sampler

[wispy canopy]

unipc won't work xD

tested it, really bad image output

[whole portal]

thats why i used 0.3

[wispy canopy]

again, the proper comparison is 12 steps cfg 7 vs 24 steps cfg 2 (time taken)

[whole portal]

and 0.3 is not enough for LCM

[wispy canopy]

it is, if you're using dpm 2sa karras

[whole portal]

UniPC

with 0.3

[supple raptor]

LCM just completely nuked the style that was prompted for.

[wispy canopy]

watercolor?

[supple raptor]

Hires fix denoising? Yeah that should normally be between 0.41-0.51 for ESRGAN upscalers and 0.55-0.7 for Latent upscaling

[supple raptor]

(sketch:1.2), (watercolor \(medium\):1.1)

[whole portal]

i am not using any upscaler now, just testing normal image generation

i am talking about lcm lora strength

[supple raptor]

oh ok

[whole portal]

i am figuring out what will give too pale images and what will give a s..t like this #🍥｜anime message

[wispy canopy]

i like the part where lcm is just more coherent in the images

[whole portal]

lora:LCM_LoRA_Weights_SD15:0.4 should be max for Uni PC

on the other hand lora:LCM_LoRA_Weights_SD15:0.7 should be min for LCM sampler

[wispy canopy]

absolutely

you'd almost think that the LCM sampler is made to work with the LCM weights 😄

[whole portal]

https://civitai.com/models/200061/hephaistosnextgenxl-lcmbf16hybrid has LCM lora embeded, so it is faster and easier 🙂

https://civitai.com/models/202108/bluepencil-xl-lcm

nice

lcm embeded models started to popup

[wispy canopy]

of course

[whole portal]

https://civitai.com/models/202308/reproductionlcm?modelVersionId=227755

[wispy canopy]

holy hell, that's fucking overcooked tho

[whole portal]

cfg 2

[wispy canopy]

fucking. overcooked

[whole portal]

it would probably look better with 1.5

[wispy canopy]

the blotchiness is just really bad

[whole portal]

or even 1.2

since you cant controlo lcm strengt (because it is embeded), just lower the cfg

it should be better

[wispy canopy]

this be my final comparison. does watercolor sketch painting just fine

[wispy canopy]

or just have lora ^^

anyway, 3:30 am

gotta go to sleep

lol

cya laters ❤️

[whole portal]

nope

3:39

ahahaha

i guess i should sleep too

[steady grail]

oh, we are close, guys

i complete the last gen and i go to sleep

[terse ore]

I added SAMLoader to the face detailer and use the sam_vit_h model. It makes the face problems go away for most part. Much more consistent and better now

[steady grail]

good

good, happy to know

not what i expected at all, but i dig this style so much

[steady grail]

going to sleep

see ya

[zinc flax]

Lewd

[native halo]

dont forget to talk about the anime img generation event 🙇‍♂️ 🙏

[slate remnant]

hey mods, this is more than over the top don't ya think?

[stable crane]

😢

[slate remnant]

I've seen people get timed out for less

[native halo]

https://tenor.com/view/ned-flanders-gif-24570713

[stable crane]

well, if you wanna see that pic, it'll be on my twitter

[slate remnant]

actually wanna see mods be fair and time ya out since they seem to be time outing people lately

[native halo]

i just want the anime img contest maybe a 4090 as prize ( its ok we can settle for a 3090) 👍

[stable crane]

sounds like crab bucket mentality

[steady grail]

wishing other's misfortune is pretty dark

[native halo]

tbh we need more bodysuits

[nova remnant]

I got you

@native halo

[native halo]

(latex bodysuit:2.5)

[gritty spruce]

Hey, let's please follow our guidance, we've gone thru this many times, please use the spoiler/blur filter when in doubt and If you want to share spicy stuff, do via DMs between yourselves and not on the public channel.

[nova remnant]

[native halo]

needs more peace signs ✌️

[gritty spruce]

😜

[nova remnant]

lmao I still had maid outfit in the prompt

[nova remnant]

@native halo

[gritty spruce]

these are from a glitchy/graphic finetune i'm working on now

[nova remnant]

I like this one

[nova remnant]

is it a lora

[gritty spruce]

Still on it but that's the idea

[gloomy lodge]

gross

[zinc flax]

@mods

[gloomy lodge]

gents let is clown this man

[nova remnant]

Why do I always miss the weird stuff

I'm curious

Anyways I think this is the night for me

sleep time

[bitter latch]

mumra the ever living DUCK

[wispy canopy]

awww

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1178354610629517343/00142-2359669363.png?ex=6575d73f&is=6563623f&hm=21d7eb0be903e5349782aac11db64fc462d1b3cb3c18b8e1a18a419cc4d792fd&=&format=webp&width=530&height=662

[primal wharf]

Smol sus is back for more images!

[lyric zealot]

And now back to sleep

[primal wharf]

[magic wave]

Thats really good

[gritty spruce]

kinda disturbing that they don't blink but it's a wip

[bitter latch]

hava bunny miku

[bitter latch]

say that again? (safety off click)

ok back to regular miku

[steady grail]

it's weird what this model does with this prompt XD

[neon linden]

is colab still banned for free use?

[unreal ridge]

I'm pretty sure they managed to make it unusable. There's probably other free options, but you'll probably have to look it up.

[neon linden]

that's sad, been trying to find a way to run it online with a custom model and I can't manage to find a way

I don't mind that it is slow as long as I can use it

[gritty spruce]

The free tier is a joke with SD right now, they push you to the higher ones to make it usable, it's a drag.

[gritty spruce]

this looks really good, is it with LCM? I can't get LCM to look that good

at least over sdxl, on 1.5 a bit better

[neon linden]

do you guys know an alternative?

[gritty spruce]

i don't sorry, runpod is really great but not free

[neon linden]

oh rip, thanks though!

[wispy canopy]

yeah, that's sdxl lcm

even on my generic model too

was trying to find a prompt prefix and negative so i could do that cell-shaded 2d art

can immediatly see if i go over that 75 token limit -> the effect becomes lost a bit

the cons of a generic model, it's hard to keep it in line

ah, found a way to fix it ^^

well, time for me to hit the bed ^^ cya guys later

[wispy canopy]

my sdxl and sd1.5 settings for lcm: (lcm lora with own model)
lora strength 0.8
cfg 2.5
20 steps LCM sampler, can go lower, but anatomy generaly is fixed up better on higher steps, also when prompt editing having more steps can help a lot
freeu

[gritty spruce]

mine are kinda similar... cfg 2 steps 12-20, i'm missing freeU tho

going to test

[dawn jasper]

@nova remnant Awesome images, also how's it going, guys?

[gloomy lodge]

parents catching you smacking the crap out of your sibling old image

[terse ore]

Anyone use toyxyz's blender openpose+rig? I'm trying to import a rig but not sure how

[astral jasper]

@supple raptor

@primal wharf

[supple raptor]

hai imi

[astral jasper]

sup

[primal wharf]

[nova remnant]

big emperor

[primal wharf]

Yes

May the emperor live 10000 years

[gusty palm]

dooooood

[wispy canopy]

@gritty spruce i forgot, i also enabled "Always discard next-to-last sigma" and Negative Guidance minimum sigma at 3.0

did a lot of testing on those, what got me the best image output

[primal wharf]

No need to stand on high ground when wings can put you high up in the air.

[wispy canopy]

found a combo of tags that got me reliable 2d anime ffrom my model ^^

[primal wharf]

Ancient warrior meets modern equipment

[wispy canopy]

using a lora for those guns?

[primal wharf]

Nah. Bing image generator

It seems to only know AR-15 unless it's prompted futuristic gun

[wispy canopy]

no wonder the images were that crisp 😮

although, the dithering on bing/dall3 images are pretty visible

[primal wharf]

I guess that could explaon the coherency, only knowing AR-15, save for futuristic guns

Those are examples of futuristic guns

[wispy canopy]

my model, not even half bad 😄

[primal wharf]

Waiting til the day I can generate some futuristic AK-12 2016 inspired rifles.

[wispy canopy]

checking if sag can improve things somewhat

[primal wharf]

What is sag?

[wispy canopy]

yet anonther method/extension in attempt to improve image quality

it feels a bit cleaner?

[primal wharf]

Police with 4 katanas lul

Probably used for throwing

[wispy canopy]

in anime, one of the best moves, in real life... uuhh...

[primal wharf]

XD

[wispy canopy]

your police has nothing on this!

[primal wharf]

Buster sword is heavy AF

That exoskeleton sure provides lots of extra strength

[wispy canopy]

i can still hear people shout "sdxl can't generate anime"

[wispy canopy]

meow

[primal wharf]

Go ahead, I'm going to sleep too.

[nova remnant]

@wispy canopy are these cute ?

[wispy canopy]

i suppose they are ^^

[nova remnant]

damn this model is good

100% on point

[wispy canopy]

damn, i have to work hard in prompting to get something similar

then again -> i can switch up the style at any given moment XD

[native halo]

depends what type of anime u mean,if u mean the generic one then yes it can

[wispy canopy]

i mean, any extreme style is going to hard to pull off...

[native halo]

yes unless u have a lora

maybe in the future we will get a model that can do any style with just a prompt

[wispy canopy]

i mean, can get general anime in various gradations from my model, but more stylized stuff like Panty & Stocking with Garterbelt is going to be really hard

without directly referencing that style

[native halo]

try to make Yasuomi Umetsu style without a lora

[nova remnant]

make a what

[native halo]

one of the many famous japanese artists

[nova remnant]

okay google (Yasuomi Umetsu black leather bodysuit tight wet:2.3)

[nova remnant]

I'm clueless about anime, google'd it but only looks like old anime to me

[native halo]

he worked on a lot of r18 ova's so make sure u ready for the nudity in google results

[wispy canopy]

@native halo yeah, my model knows the name is associated to anime, but doesn't know the style

[nova remnant]

https://tenor.com/view/no-way-wayment-holup-hold-up-wait-a-minute-gif-15124840

[nova remnant]

time to start merging

[native halo]

or try with Hirohiko Araki the creator of Jojo to make something like this

[wispy canopy]

pretty sure i can't pull that off with my model (yet)

[native halo]

yea and the loras for it suck so cant be done yet

try with Satoshi Urushihara,in his style the most important part are the eyes and the hair,also pretty sure there is a lora for it for XL but last time i tried it sucked

[nova remnant]

@native halo

rate these

mechanical stuff

[native halo]

i like the big explosions

[wispy canopy]

explosions? XD

[native halo]

(latex bodysuit,skin tight:2.5)

[wispy canopy]

ah yeah. no, i didn't use those yet

lets see what happens if i do

@mods -> @native halo made me do it!

[nova remnant]

is that with the 2.5 weight ?

[wispy canopy]

yeah

[native halo]

its for science im makin a thesis about effects of latex on skin i swear

[wispy canopy]

my model can take insane weights

sd1.5 models would get burned on those so hard

look ma, no controlnet

[gritty spruce]

😂

[wispy canopy]

@gritty spruce if all else fails, just get brightprotonuke too ^^ -> certified quality, no controlnet needed

[gritty spruce]

For stills it’s fine, I’m having trouble with animateDiff and LCM Anime

On sdxl can’t get good sharpness, lines, res

Maybe has to do with the sdxl motion module, since 1.5 ones do fine

[wispy canopy]

ah, specific animatediff?

[gritty spruce]

This one is 1.5 lcm at 12 steps (with nubby’s ckpt)

On sdxl can’t get near that even at 25 steps

[wispy canopy]

and i think you mean the image sharpness?

pretty wicked animation ❤️

[gritty spruce]

Yeah it’s terrible

[wispy canopy]

damn, really having to up the steps to get LCM to correct all the small errorrs (getting the lines very crisp again)

i like crisp images tho

12 steps... credit where credit is due -> LCM does amazing on low steps

[gritty spruce]

Those are sharp, and she can even breathe! 😂

[wispy canopy]

are you referring to chest sizes that are good for 20 litres of air?

[gritty spruce]

Lol the latex one seemed a bit in need of a mechanical respirator

[wispy canopy]

latex is a material that gives -> she'll be just fine

[gritty spruce]

She cant even move her arms forward of the tension

[wispy canopy]

can you here the latex squeeking with every step she takes?

[gritty spruce]

Hahaha

One thing that models never get right is the navel over latex/rubber, feels weird to have it visible

[wispy canopy]

i really like higher step LCM tho -> it fixes most, if not all weird anatomy

[gritty spruce]

I guess you can negative prompt it

[wispy canopy]

yeah, navel in negs usually does it

same as i did with the nai based models, where skirts were glued to the crotch

negative prompt leotard fixed that

but SOOO annoying

[nova remnant]

just remove it then

[wispy canopy]

fun fact: your navel is the first scar you have

[gritty spruce]

Interesting fact

[wispy canopy]

my face when 10 steps got the best image

[wispy canopy]

i'm not saying i found a common prompt that i really like, but goddamn, do i like this common prompt 😄

it's great having not to worry about anatomy

yeah, gotta find a way to bookmark this style, it's awesome

[gusty palm]

LMS karras is kinda fast

[wispy canopy]

cat

[nova remnant]

Bot

Now do a cat themed spaceship

(first thing I thought of)

[zinc flax]

fbi open up

[vestal crypt]

https://media.discordapp.net/attachments/1027429244478955530/1178673578108014622/00078-3714751713.png?ex=6577004f&is=65648b4f&hm=577a69af9bcdd68667207df934805780060acfaaf9bb1c0d93d245c00585aa5a&=&format=webp&width=441&height=662

[nova remnant]

would bind her with a spell

[gusty palm]

https://i.imgur.com/5jpqxo4.gif

which extension is doing this pop up when i mouse over stuff?

is there a way to like, make it not pop up over the thing im trying to click on

[gritty spruce]

model: kakigoriv2 https://youtu.be/ZgUomyIz9wE

[gusty palm]

i love it

did u make the music too

id love to know how u did that

[gritty spruce]

that's a track called "Dropper" by Phrase

[gritty spruce]

comfyUI, animateDiff, promptTravelling

[gusty palm]

uwu ty

[terse ore]

In comfyUI, should I put the controlnet conditioning in the first Ksampler or second or both?

When I want to use high res fix upscale for example

[stable crane]

looks like you sync'd it to the music too @gritty spruce

it'd be interesting to add to the prompt with the current lyric at timestamp

seems like @wispy canopy would be all over this kind of thing 😄

[wispy canopy]

yeah, good shit ^^

pretty sure myprmopting methods would work that well tho

[dawn jasper]

@stable crane Good for private messages and the like?

[gritty spruce]

yeah, I have a workflow that triggers the prompt travel with sound, it's not 100% "reactive" because of the context window in aDiff but since it triggers changes, they fit the beats

[gritty spruce]

cnets before the sampler

[terse ore]

Only before the first one, and skip the second one?

[gritty spruce]

i don't know your workflow, what do the samplers do? first one gens, second upscales? details?

i'd do it before the first sampler in the chain

[terse ore]

Yes, the second one is the upscale

[gloomy scroll]

Gib prompt

[stable crane]

wym? if you mean will I accept DMs, yes 🙂

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1178754268757495828/00019-3107161075.png?ex=65774b75&is=6564d675&hm=7d7b071c484725fb04d9d6af535cd3a77a20f71159f881bbc4e6da69b0917b46&=&format=webp&width=441&height=662

[gloomy scroll]

Wall

[gloomy scroll]

The continuation of wall

[gentle musk]

Different outfits thanks to IPAdapter

[dull jackal]

D r o p p i n g this survey HERE: https://forms.gle/ZefWVxhXwiHBZ4y37 Please feel free to fill it out at your convenience! Byyyyyyyyyyyyyyyyyyyye!

[nova remnant]

@native halo do you like jackets ?

[native halo]

i like big jackets

[stable crane]

me too

more cyberpunk!

[whole portal]

any recommendation for a realistic anime model, something more like digital painting, something like this

i made this with photon, which is primarily made for photorealistic images

[native halo]

id say this one maybe? https://civitai.com/models/182723/kakarot-28d

[whole portal]

it's "too realistic"

i just found one nice

https://civitai.com/models/204997

oh, this is lora

is there a model similar to this?

[native halo]

this one more "flat" but idk if u like the color https://civitai.com/models/193275/kakarot-25d-cozy

[nova remnant]

[native halo]

but since that is a lora u could try to combine a good realistic model like dreamshaper/absolute-reality with that lora to get the desired effect

[stable crane]

does kohya deep shrink work on 1.5?

[native halo]

yes

[stable crane]

looks like it mutilates hands though?

all my hands are mega mutant

[native halo]

are u using it in comfy?

[stable crane]

yes

haven't pulled lately

[native halo]

in a1111 it conflicts with adetailer and deforms faces so i just removed it

[stable crane]

seem legit?

[native halo]

i think eface found the perfect settings to make it work 😬

i dont use it cause it renders img at a high scale so get oom

[stable crane]

I think he was using it in SDXL too

[stable crane]

tried tiled vae?

[native halo]

yea but rendering at high res causes deformities so high res fix works better for me

maybe thats why u gettin deformities

[stable crane]

in SD1.5

it's fine up until the hands basically

[native halo]

try to use Freeu and SAG

[stable crane]

it works great with SDXL though

[native halo]

i get slightly better hands on 1.5 when i use freeu and SAG so maybe that will help

[stable crane]

IDK how to use those things

[native halo]

for auto1111 is easy just install and boom idk about comfy spaghetti tho 🍝

ask eface he probably has a workflow with all lcm+deep shrink+freeu+sag

[zinc flax]

sleep

[dawn jasper]

I like big jackets and bigger chests, yes.

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1178955520988692501/00087-2956283206.png?ex=657806e4&is=656591e4&hm=554f0915eaa5aaadacafe97c797d0d6763febce9a707a9ec0607c2048f878ce9&=&format=webp&quality=lossless&width=590&height=662

[wispy canopy]

new model merge is coming along nicely

some stuff still to be figured out, but not in a bad spot

[gritty spruce]

ThiccaLewds 😜 👁️

[stable crane]

nice gun

[gritty spruce]

She needs exosuit legs to carry it

[stable crane]

4 spider legs

lower half tank

still a bit messed

time for zzz

[bitter latch]

MumRA The ever Living! tho I think he lost his kool

[gloomy lodge]

anyone here have any idea how to use hugging face i wanna download dalcefo_painting2 from it but im a bit tarded and the bottons confuse me im so lost on what to download

[bitter latch]

ok that took a turn

[supple raptor]

There's a small down arrow button next to each file name. You can just click the one next to the file you want to download it.

[gusty palm]

did they fix the weird noise issue with upscaling img2img cd tuner images

i seemed to not have that issue last few times i tried it

it popped up when i tried the color enhancer post process extension though, tiled diffusion renoise seemed to help though

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1179023728663871569/00222-2596829799.png?ex=6578466a&is=6565d16a&hm=7cc2e5f707746feeee40d08925ecff6c6a8dfbf1b126b3f81b5e2b6442bec0e2&=&format=webp&quality=lossless&width=441&height=662

[zinc flax]

hug

[wispy canopy]

new merge attempt

[nova remnant]

angyy

[lyric zealot]

sorry for the super late reply, but here

[lyric zealot]

[gloomy scroll]

Oh yeah I forgot already

[lyric zealot]

so sorry, was literally out for the whole day

[gloomy scroll]

Np

[swift lava]

Found something interesting. I have always wondered what the point of the "Latent" upscaler is because by the time you set denoising high enough to get rid of artifacts the picture has changed completely. But recently I have been playing around with watercolor styles, and those have an issue where many upscalers destroy the watercolor aspects. If I use Latent upscale and set the denoising around .5, it keeps the picture mostly unchanged and the artifacts actually add to the style.

[bitter latch]

FINALLY

French Maid Sailor Moon

[bitter latch]

better

and people say eular A isn't good...

gets more consistant results than most i think

[wraith raptor]

Most people use euler a, in most situations you'll have better use for 2M Karrass, but it's not a game changer either way

[bitter latch]

(Far wide perspective view:2) ---- oops

[bitter latch]

true

[wispy canopy]

the shrine maiden offers you a weapon, do you accept?

[dawn jasper]

Tiny girl there.

[primal wharf]

Must be an important item to finish a quest.

Accept that, to scare the enemy's general.

[wispy canopy]

It was a ruse. now hand over your waifus!

[primal wharf]

Royal guard will handle this

[wispy canopy]

with the power off dall3, the evil empire?!

[primal wharf]

His mask says it all about his trigger discipline.

xD

[primal wharf]

I would say neither evil nor good.

The prince play pretend as royal guard

A time traveller illustrated in an old Chinese poem.

[wispy canopy]

no denying how good dall3 is

[primal wharf]

Imagine training SD with some DALLE images.

[wispy canopy]

that's what paradox, paradox2, are

by extension, that's what brightprotonuke is

A time traveller illustrated in an old Chinese poem.

[primal wharf]

Here's my prompt for that soldier in a Chinese poem: A blend of anime and ancient Chinese painting style. It depicts a time traveller who is wearing a modern military camo uniform, with ballistic plates and chest rig, a tactical helmet and goggles, and assault rifle, standing next on a mountain with waterfall and lake. Vertical Chinese text as poetry describing him.

[wispy canopy]

brightprotonuke1.3 unfinished (0 negatives, prompt copy paste)

[wispy canopy]

i mean, any sdxl model trained on dall3 stuff is in essence just a low level distillate emulation 😦

[gloomy scroll]

Miku accepts this gift

[wispy canopy]

how can i convert my model

[wispy canopy]

the rules say you can't challenge eface with lora's, controlnet, or dall3 😮 did you? (very fine image you made)

[cyan burrow]

just got my Logitech Superlight PRO X Lightspeed (long names every mouse has now...I remember when it was just M500)

[gloomy scroll]

You think there was model that could have done that image without help 6 months ago?

[wispy canopy]

0 🙂

[wispy canopy]

if that's wihtout help, i applaud you and my chibi lost the battle

[cyan burrow]

superlight, glides nicely, fits nice in my hands and has friction pads to place on the mouse...

[gloomy scroll]

It was some lora

[wispy canopy]

lora's ARE strong, i can't deny that

quick prompt to try to get a similar ffect

[cyan burrow]

@wispy canopy quick prompt? young woman, cybernetic augmented, posing with mantis blades, Kirlian, Luminogram, Databending, anime, HDR

I dunno what that would do