🍥｜anime

[stable crane]

I'm now thinking how awkward it will be when our cars achieve sentience

[native halo]

https://tenor.com/view/girlmode-girlmoded-girlmoding-transform-transformer-gif-26013054

[supple raptor]

rotary engine, nice. she's probably a mazda

[worthy iron]

damn I love this gif

is that from an actual anime?

[stable crane]

transformers

[native halo]

Henkei Shoujo

[nova remnant]

[nova remnant]

Very suspicious that you know that

[native halo]

i know better ones

[raw rose]

What model is this? I really like this style!

[bitter latch]

Miku soldier in power armor with mini-gun, mdjrny-v4 style, nousr robot, dreamlikeart

forgot the styles. oops

whoa

And thats with the anime model AOM3

switches to an XL model

ok this is gonna take a while

[bitter latch]

taps foot

[bitter latch]

Epic sci-fantasy, Miku soldier in power armor, dreamlikeart, on battlefield, far view, wide angle, fine art illustration, 3.5D, elegant shading, elegant detail, <lora:Ink scenery:0.5>magical, Vivid multi-colored Ink's, detailed outline's, lightly shaded
Negative prompt: kkw-extreme-neg, blurry, deformed hands, deformed face, deformed fingers, multiple limbs, ugly, bad anatomy, bad hands, ugly hands, deformed, disfigured, poorly drawn face, mutation, mutated, extra limbs, poorly drawn hands, (grayscale), weird colors, censored, deformed glasses, lowres, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark
Steps: 70, Sampler: DPM++ 2M SDE Exponential, CFG scale: 8, Seed: 3580538802, Size: 1024x512, Model: astreapixieXLAnime_v16, VAE: sdxl_vae.safetensors, Denoising strength: 0.5, Clip skip: 2, Hires upscale: 1, Hires steps: 20, Hires upscaler: Latent, Lora hashes: "Ink scenery: 26343a3a36a0", Pad conds: True, Refiner: campursariXL_v1 [c8014a292e], Refiner switch at: 0.9, Version: v1.6.0

came out pretty good. but i'm done with xl takes to long

[wispy canopy]

output from my unreleased model mix, basically a generic model where i asked "gimme something that looks a bit more crude"

[nova remnant]

cool coloring but i dont think you need all that negative prompts with more recently trained 1.5 mods

[bitter latch]

da who what?

[nova remnant]

let me try that prompt w/o those negs

i see you have 3.5D prompt in there, whats the for?

[bitter latch]

gives it more of a 3D look

I thnk...

[nova remnant]

im gonna use some textual inversions tho instead of negative prompts

[wispy canopy]

good morning ^^

[nova remnant]

i see, but you can only push dimensions that far on a 2d image canvas

[bitter latch]

takes out neg stye

[nova remnant]

hi eface

this is what i got tora

not really very colorful but that's different mod i guess

gonna try landscape with more color detail

[wispy canopy]

dis is what i got XD

[nova remnant]

nice

i like his original prompt .. i edited that slightly ..

Epic sci-fantasy, Miku soldier in power armor, dreamlikeart, on battlefield, far view, wide angle, fine art illustration, 3.5D, elegant shading, elegant detail, magical, Vivid multi-colored Ink's, detailed outline's, lightly shaded

dream like art btw

[bitter latch]

Epic sci-fantasy, Miku soldier in power armor, dreamlikeart, on battlefield, far view, wide angle <---- typed this, rest is my soft anime style

[nova remnant]

neat

[bitter latch]

dreamlikeart is leftover activation keyword from a mdjrny4 lors

lora

[nova remnant]

ahh i edited that line lol... with spaces in words

[bitter latch]

worked good

[bitter latch]

very nice

[nova remnant]

for negative this is what i have ...
FastNegativeV2, BadDream, negative_hand-neg, (low quality, worst quality:1.2), (grayscale, grainy, border), (watermark:1.1)

[bitter latch]

hmmm ok...

[bitter latch]

nice

hmm putting 3.5D in neg and 2.5D in pos.. yeah didn't work well

[nova remnant]

there are anime mods that are trained to generate either 2D or 2.5D w/o prompting

[bitter latch]

hmm tho it could be art by mark brooks that could have done it

[nova remnant]

damn this is a huge power armor

[nova remnant]

i like these two prompts ... far view, wide angle

those two are doing the perspective magic

[bitter latch]

cyberman? prompt: Epic sci-fantasy, Miku soldier in power armor, dream like art, on battlefield in cyberpunk2077, fine art, soft colors, black outline style, elegant, very detailed Cel-Shaded With shaded detail's, soft shadows

[nova remnant]

hmm with miku soldier in power armor focues more on soldier in power armor

i removed soldier

[bitter latch]

hmm point

this started from a screen shot from my character from starfield

[nova remnant]

🙂

[bitter latch]

2nd character, taking a totally diff route

[nova remnant]

Epic sci-fantasy, Miku in power armor, dream like art, on battlefield, far view, wide angle, fine art illustration, elegant shading, elegant detail, magical, Vivid multi-colored Ink, detailed outline, lightly shaded.

could add dynamic pose

[bitter latch]

ponders bed

[bitter latch]

deploy ultimate!

[nova remnant]

just ultimate or with pose?

think i need a mecha mod for armored anime

[bitter latch]

no no I mean as their ultimate weapon, as in they deployed it

[nova remnant]

oh

[bitter latch]

and someone just crapped themselves

[nova remnant]

:p

[bitter latch]

ok gonna play some more starfield then bed. nite

thuds

[nova remnant]

anyone getting TOS violation crap when trying to upload images to civitai... i've encoutered that a couple of times with two different images

and those didnt have any nudity at all

or is that have anything to do with upload cap? cause one image gets in and 2nd image gets tos violation w/ no desciption

[nova remnant]

you making new merges eface?

[wispy canopy]

yeah, been at it for a while, and i think i'm finally getting a bit closer to a good mix

[nova remnant]

cool :p

i shortened the prompt to Miku in power armor, sci-fi, fantasy, battlefield, far view, wide angle, magical, multi-colored. for contextual reasons

[wispy canopy]

yeah, this upscale went REALLY well

that inspires confidence ^^

[nova remnant]

i'll go check your page real quick to see what you have up

didnt know you worked on hentai mods ;p

[wispy canopy]

that was my first accident into model merging actually

[nova remnant]

heh

[wispy canopy]

i did a random thing and all of a sudden it seemd to outperform the parent models

[nova remnant]

so darkalfa is sd 1.5 ?

[wispy canopy]

yeah, is 1.5

[nova remnant]

you had sdxl up tho right?

[wispy canopy]

yeah, curent murge is sdxl

[nova remnant]

ahh ok

i think you had one up few weeks ago

[wispy canopy]

yup, brightprotonuke 1.1

trying to merge for a 1.2 for over a couple of weeks now 😛

[nova remnant]

i wish i could play with sdxl more lol.. i have the rig but render time is so slow

[wispy canopy]

highres upscales take long either way

[nova remnant]

to be fair ... i have problem with 6gb files for mods, i dont have that much space in ssd and load time on 16gb ram is kinda slow too

[wispy canopy]

you might wawnt to give that distilled sd a go?

[nova remnant]

which one is that?

[wispy canopy]

not anime, but just model testing ^^

[nova remnant]

hmm tell me about it

is that from using base and refiner ?

i still have those 2

[wispy canopy]

distilled sdxl is a reduced parameter size i think

think it works on sd.next

[nova remnant]

a1111 should work too, since it has good support for sdxl now?

[wispy canopy]

different kind of model type

[nova remnant]

oh

[wispy canopy]

https://civitai.com/models/188863/segmind-ssd-1b?modelVersionId=212123

you need diffusers -> that means sd.next or comfy

[nova remnant]

i think i'll configure the sdxl again just for the sake of fun

[wispy canopy]

https://huggingface.co/segmind/SSD-1B

that one for a1111?

might wanna give it a go ^^

[nova remnant]

so i could use this for a1111 https://huggingface.co/segmind/SSD-1B/blob/main/SSD-1B-A1111.safetensors ?

[wispy canopy]

i think so, yeah

[nova remnant]

ahh sweet

[wispy canopy]

if he says it's compatible

[nova remnant]

and i load that up just like regular models?

[sick idol]

wait eface u using a tensor rn?

[nova remnant]

4gb .. nice

[wispy canopy]

no, not using tensor, as it stops freeu from working

and freeu is a huge quality upgrade imho

[sick idol]

thought so

[nova remnant]

what vae does SSD-1B require btw?

[wispy canopy]

i have no fckn clue xD it's own vae i suppose

[nova remnant]

i dont see any mention of that

guess i'll try the sdxl default vae

[bleak cape]

guys does google colab not work anymore?

[supple raptor]

WebUIs on Colab are dead. People moved to Paperspace after google banned WebUIs. It isn't free, but you can get faster GPUs on it.

[supple raptor]

@gritty spruce we got a spammer

[gritty spruce]

Thank you!

[nova remnant]

can't seem to match the SDXL aesthetics similar to SD 1.5 :/

btw i realized i can render sdxl images pretty fast when not using hires fix

but meh, the sdxl texture feels so artificial

the texture of hair in that image i just posted is totally bs

[bleak cape]

is there a free alternative?

i was fine with colab anything i could use anything similar to it

[nova remnant]

i already forgot why im making this effort to get sdxl working

but to be fair, i think i got one good sdxl mod for realism

[zinc flax]

Another free alternative would be those free ai generation site like pixai, yodayo, etc

[vestal crypt]

Prussia design

[native halo]

i wanna live in prussia now

[wispy canopy]

well, that was a nasty system whatever the fuck that was

[raw rose]

Really loving the results. Do you plan to ever release it? I’d love to use it to experiment with a comic I’m trying to create. Most anime models look way too polished and painted, i like that this looks more like a drawing

[wispy canopy]

of course ^^ right now i'm doing a lot of testing to make sure it actually makes good images in all the other situations too

(this is from the same model)

[raw rose]

Neat!! Can’t wait to try it. Do you find it difficult to train models? I tried about 8 months ago but the results were awful. Maybe stable diffusion gives better results now when training from scratch?

[wispy canopy]

i didn't actually train the model

but selected models wwhich i felt had good parts, and got to merging

[raw rose]

Ohhhhh i see! ok interesting

[wispy canopy]

so, not just going for all thepretty models, but also looking for the creative models, so you can get interesting images

or smart models

[raw rose]

Gotcha! Thanks I’ll have to play around with that :)

[wispy canopy]

personally, i like smart models the best -> the rest can be added in with lora's and stuff, but you can't fake model intelligence (how to hold a gun, for example)

[raw rose]

I see, so models that look to have the best understanding of anatomy and how the world is actually structured

Rather than too much nonsense in the outputs

[wispy canopy]

exactly

they're rare though

damn, this one came out great ❤️

[raw rose]

Wow, that looks awesome. Person is perfectly proportioned, buildings are logical, plants are rendered right. And its colors and style are so pretty

[wispy canopy]

it's using regional prompting though, so not pure txt2img

in this case it was a 50/50

next up, going for this kind of setup

can already see in the preview, i should regional prompt more often xD

0: skyscrapers
1: girl
2: grass, weeds and sky
3: grassy pavement

let the model figure out the rest ^^

[whole portal]

i am pretty new to this stuff, so i need your help people

can someone recommend a nice realistic anime model that behaves well with this hypernetwork?
https://civitai.com/models/140117/latent-layer-cameras
I want better control of a view, without too much play with prompt

[wispy canopy]

i think most models should be able to handle it, just get one that you like the style of most

i'm not the guy to do model recomendations: i don't use anime models lol

[whole portal]

it looks like anything v5 does not want to play nice with this hypernetwork 🤷🏻‍♂️

[wispy canopy]

you could always give my old model a try, darkalfa

https://civitai.com/models/64843/dark-alfa

[whole portal]

can you please tell me what I should use when the model asks for clip skip: 1 (like yours) but lora wants clip skip: 2? should i use 1 or 2? or i just need to experiment? or should I give up and not try? 🙂

i understand the difference (at least i hope i do 🤣) but i am not sure what actually happens when you mix resources that recommend different clip skip. i am very new to all this, so try to understand me 🙂

[wispy canopy]

experiment 🙂

best way to find out

[whole portal]

and yes i am aware this is not a place to ask questions like this one, i will ask in support, but i hopped you know the answer if you exprimented

[wispy canopy]

support is more for technical, i suppose, if it comes to image creation, i think this chat should be fine

you want to run clip skip as low as you can. some models become incoherent when they do, so then increase

[whole portal]

at least you have much more experience than me, so i guessed that you might already experienced that situation before

[wispy canopy]

well, not when it comes to hypernetworks/lora's and the stuff -> i only prompt stuff 😛

[whole portal]

have you used epi_noiseoffset2 for those?

or did you use some other method to get "dark"

[cyan burrow]

I didn't use any noise offset

The gallows in silhouette, an like a cut out, swung by an executioners shadow, onto the victim, blood gushing out, surrealistic, parallax, 5d, Storybook Illustration, Polychromatic-Colors, ProPhoto RGB, Artificial Lighting, back lit, Contre-Jour, Posterization, Post Processing, Tone Mapping, 8k, hdr

Negative prompt: EasyNegativeV2 Negative_Hand title, watermark, trademark, copyright, game like, cgi

Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Size: 768x768, Model hash: ffb0f06c40, Model: XCell_Sindrome_v0.1, VAE hash: df3c506e51, VAE: kl-f8-anime2.ckpt, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer version: 23.11.0

I even mispelled part of the prompt

[whole portal]

nice

[tidal phoenix]

If you know, you know

[stable crane]

nice dragonmaid

[wispy canopy]

lyricsprompt testing my model

[wispy canopy]

this model is gonna be the one, i'm 95% sure at this point! i mean, i can pretty much throw anything i want at it, and it does it

[stable crane]

interesting, good mix of anime + realistic

[stable crane]

I love pic 2

[wispy canopy]

it's a really well following generic model 😄

[wispy canopy]

well, now the other hard part comes

making all kinds of preview images

[nova remnant]

lmao

"This is fine selfie"

while a flood is happening

[keen hollow]

'afternoon

[autumn violet]

gotta have a good picture for your funeral I guess

[stable crane]

@ me when you launch it 🙂

[wispy canopy]

gib prompts for image previews!

[stable crane]

(masterpiece:1.3), (best quality:1.1), (tall:1.2), (1girl:1.2), oni horns, (highleg panties under ninja shorts:1.2), spoken heart, shiny skin, (ninja mask:1.1), white eyes, hair over one eye, (extremely long white hair:1.3), red shoulder pads, black gloves, black fishnet ninja garb, wide hips, narrow waist, order, outline

[wispy canopy]

oh boi. this.. uh...

@stable crane i suppose i'll have to dm the results xD

no way in hell that i can post those here without getting a temp ban

it's... juicy

[supple raptor]

[keen hollow]

lol, i dunno... random 😛

those ones came out well

[bitter latch]

added text and border in GIMP

[keen hollow]

ok, that's enough of that series for now 😛

[terse jacinth]

[nova remnant]

She is not centered on the image

Unplayable

Ungennable

[untold glacier]

[nova remnant]

You can run Fooocus-MRE in Colab? Are there any tutorials on how to do that? I tried setting up both ComfyUI and InvokeAI in Colab, and couldn't get it to work. Well, I did manage to get them set up, but for HD space reasons I wanted to host my models in Google Drive and point Colab to them, that's the part I couldn't get to work. Spent a whole day on it, to no avail. 😦

[slate narwhal]

Fooocus-MRE is not maintained anymore; the dev just stopped because he had to repeat code to get equal to the main Fooocus, so most people are using the main Fooocus or the other forks like Fooocus-Control (for lots of control nets) and RuinedFooocus. MRE worked fine in Colab, it has a link to the colab in the main page.

There are no tutorials I know of on how to run Fooocus.

Nowadays I use the main Fooocus myself; the latest updates have improved it a lot, and there are very specific things that lllyasviel (the genius behind ControlNet) created for it.

As for Colab, I pay the $10. The thing is that after your credits go out (which usually do in a day), it simply keeps working fine without any limits on hours. In the past, or if you used Auto, it would only allow 4 or 5 hours without credits and then ban you, but because of Fooocus having Comfy's optimizations, Colab is not banning anymore. And I can change the runtime to a higher memory, so I can change SDXL models/refine very quickly.

Comfy won't work as it should in Colab, not even in the paid Colab. The version that was online a few months ago didn't have the Manager, so it was basically useless.

Someone released a version with the Manager already baked in, but it would give errors with custom nodes that you don't get on a local computer, so devs don't care about Comfy on Colab, it's too complicated. And as you've said, downloading the models is a nightmare.

The solution Fooocus gives to downloading the models is much simpler, quick, and easy; it's even better than Auto1111. I don't use Google Drive anymore.

I just paste a simple text list of models—Loras, TI's, the new Face IPA, etc.—and hit the play button. It then downloads everything from CivitAI/HF and puts everything in the right directories. If you want to try a new model or Lora, you just change one line of code (the download link) and hit play again. It downloads in seconds as Colab downloads at around 148 to 200 MB/s, so you can load lots of

[slate narwhal]

SDXL models in a single run, which would be really difficult in Google Drive as each model is usually 6gbs. After you try that for the first time, you don't want to go back to Auto1111 ever again. In Auto1111 for colab, as you probably know, you had to upload everything manually to Google Drive, and then the interface should load your models from there. Not easy to do for lots of 6gbs models.

I've never tried Invoke on Colab. As my main setup, I only use SDXL, and I only use Fooocus. If I need SD 1.5 models, I just refine them as a VAE using Fooocus (it's basically SD 1.5 running on Fooocus). Sometimes, very rarely, I go to auto to merge SDXL models, or if there is some plugin that only works with auto.

lllyasviel already said he is on to create an original solution to the main problem in SDXL, the blurriness at distances, distorted faces at distances, and upscaling in general, as he thinks Adetailer and upscalers for Comfy do not solve the problem as they should, and I agree. If he solves it, Fooocus, at least for me, will cover every SDXL need.

[bitter latch]

wtf is with the wall of gd spam HFH

[nova remnant]

I think I just learned this. 😦

"Fooocus-MRE is not maintained anymore; "

When you say most people are using the main Fooocus, do you mean in Colab? Is there a link for that? I really liked the extra features in MRE, more control, at least at my early stage in this adventure.

I'm not really very code savvy, but I can follow a good tutorial or instructions and usually get it done. When you have a minute, can you explain exactly how to add models in colab?

I would definitely pay for Colab if I knew I could get it to work. On the weekend, thinking I could run the models from GD, I upgraded my storage there. If I can get Colab to work I'll downgrade and put that money into Colab.

Thanks for you reply/ info!

[stable crane]

chat needs more pics 🙂

[slate narwhal]

Answer to your questions @: #✨｜sdxl message

1. How do you copy and paste models into Fooocus MRE (Colab)

A: I could send you, it's a very simple txt file. You copy it and paste it under the "%cd /content/Fooocus" line

"Where do Outputs save to?"

A: It saves your images to Fooocus/Outputs, very similar to Comfy

"If you want to save Outputs, then you have to save them before closing Colab, otherwise they are gone, right?"

A: Yes, you could click the "History log" to save all your prompts and images or download all images from the directory, but you need to do it before closing your session. I do it like this: After generation I save each image I like to my own computer, or I put it in Fooocus upscale directly and then save. It's really simple.

Will answer your other msg soon.

[wispy canopy]

i recognize this one!

[stable crane]

hehe

I fell in love with this ninja

so what if she only has 4 fingers, love will find a way

I'm really looking forward to trying your new model

going to make so many ninjas

[untold glacier]

If there's someone I don't want to hear that from, it's you, lol. You're the one with the stream of consciousness with your image posting.

[untold glacier]

Here you go.

Oops, wrong version.

[untold glacier]

So much... plot.

[stable crane]

I never saw this show but was requested a comm of it, is it good? lol

[nova remnant]

Thank you!

"I could send you, it's a very simple txt file. You copy it and paste it under the "%cd /content/Fooocus" line"

[bitter latch]

at least it's on topic most of the time. not a wall of spam.

well , spam to me at least. alot of non-anime technobabble

[untold glacier]

An hour and a half went by with nothing going on and he was responding to an earlier discussion. There's no issue there. If you want more waifus, post them.

I see this channel as an anime community in and of itself and there's a lot that goes on here betwixt the members that isn't explicitly anime posting all the time.

https://tenor.com/view/chillin-anime-wallpaper-live-gif-25167647

[bitter latch]

ok.. now thats out of our system.. moving on 😄

[wispy canopy]

you yoinked my prompt that fast? 😮

[untold glacier]

Haha, nope. That's actually an earlier gen with the previous edition of the model.

[wispy canopy]

ah, the other test 😄

[untold glacier]

Yeah, my fav was this one

And try going low to like 3 with the CFG. You'll be surprised at the texture that it brings to the gens.

[wispy canopy]

such hires 😮

yeah, this model does great with very low cfg values

[untold glacier]

Oh, I gotta share this one. It was haunting and yet profound. This is one that I'll bring into the latest and hope that it gens it again but just better.

[wispy canopy]

it should, afaik

[bitter latch]

hmm 3 cfg you say....

I really should be in bed but a few more then bed

I hope

[bitter latch]

tweaks the prompt and changes to a cyber model

[wispy canopy]

😮 randomtora lost the 4g 😮

[bitter latch]

there

[wispy canopy]

😄

😛

lol

4Gface? hmm... sounds similar in my mind

[bitter latch]

there

[wispy canopy]

😄

[bitter latch]

hmm doesn't really look anime to realistic

[wispy canopy]

oh in that case i should be 24Gface

[bitter latch]

haha

[wispy canopy]

pretty cool 😄

[bitter latch]

this cyberrealistic model isn't doing so good.. hmm

[bitter latch]

last one of that

[wispy canopy]

highres. very nice ❤️

[bitter latch]

nice

[wispy canopy]

such ghibli vibe ❤️

[bitter latch]

well yeah I used a lora 😄

tho I had dogs waiting for owners in prompt. seemed to be ignore. odd

gives up and finally goes to bed

[wispy canopy]

gnight tora ❤️

[slate narwhal]

"When you say most people are using the main Fooocus, do you mean in Colab? Is there a link for that?"

I mean in general. Most people use Fooocus locally. The colab link is: https://colab.research.google.com/github/lllyasviel/Fooocus/blob/main/fooocus_colab.ipynb

"I really liked the extra features in MRE—more control, at least at my early stage in this adventure."

Moonride did excellent work, but after Illy put in the Ipadapter image prompt, you can do everything you could in MRE, at least in my use cases. Pyracanny, ImgPrompt, and CPDS do a lot more than Control Loras on MRE did, and if you need more, there is always the Fooocus-Control fork. Moonride merged some of the features of MRE into the main Fooocus but we've lost some, like the prompts saved in the pngs. You can still use his fork if you want.

"When you have a minute, can you explain exactly how to add models in Colab?"

I will do it in my next message.

[modest anvil]

yoo been a while

[untold glacier]

Anyone need a clean car?

[wispy canopy]

dis model gonna be lit!

[untold glacier]

[slate narwhal]

It's pretty simple:

1.Open the colab at: https://colab.research.google.com/github/lllyasviel/Fooocus/blob/main/fooocus_colab.ipynb

2.Go to the line %cd /content/Fooocus[put the cursor here] and put the cursor after the word Fooocus[here]. Press enter.

3.Paste the contents of the txt file.

4.Press the Play button and wait.

5.Click the link to open Fooocus, and generate all you want

This is my personal list at the moment and it downloads almost 160gb of models in one go. If you don't want to wait (around 3 minutes) just remove the lines and put only 1 or 2 models, so it will load pretty fast. Remove the models, Loras, Embeddings etc, as you would like by editing the text, or change to your own. If you would like to add a new model, copy and paste a line from one of the existing models and paste the download link from CivitAi or huggingface in the link part of the line and change the name.safetensors to the model you are adding. Colab has around 160gbs of space so you can load a lot of models. I recommend that you "change the runtime type" on Colab to High-Ram if you want to refine models.

[zinc flax]

Is this another alternative to a111?

[slate narwhal]

Thanks for backing me up. Yumeshiro, you're right. I've been in this channel since the beginning, and it has always been normal practice to explain technical things.
Nubby, for example, is our master in this area, and he explains a lot of non-anime technical stuff in here and helps a lot of people. I learned a lot from him and read all of his explanations with attention.

The other day, Eface helped me with some technical prompt stuff. And I remember the majestic explanations by SirVeggie on prompting and on SD in general. In the old days, this channel was like a hub for everything SD; it was the second most visited channel, only losing in popularity to the main #1072236442463518882 chat.

I know the message was kind of verbose on my part; I write too much, but it was only 2 messages, and I thought it could help others understand Fooocus and running SDXL easily as most in here use Auto, so I'm giving another perspective from another tool.

And as you've said, the channel was empty at the time. So that's why I've decided to stay public instead of going to DM. I don't know why it bothered RandomTora so much. I will try to write shorter messages in the future.

[untold glacier]

💗

Don't worry about it. Like you said, I think convos like this are great. We all get into detailed discussions from time to time, and we still love anime as well. This is a great community we have here, and you're a part of what makes it that way.

[slate narwhal]

Yes, the difference is that it was created by lllyasviel, the guy who created ControlNet, so as you can expect, he does things his own way, "out of the box" and with a lot of specific things (like models and configs) he and his team create just for the tool. He created Fooocus as a way to make SDXL compete with MJ, and the tool is just great.

[untold glacier]

Now that I've generated so much, I've made my room so hot as to make it difficult to sleep in. I'm off to try and get some rest.

G'night folks. May your dreams have you soaring above the clouds.

[slate narwhal]

Good night, and thanks for the kind words

[wispy canopy]

dafaq was that autocorrect lol

[untold glacier]

Lol

[wispy canopy]

gnight @untold glacier

[cyan burrow]

testing my latest version of my model, did a character concept merge with blocks, so far so good

[iron star]

cool cg style 🙃

[wispy canopy]

steampunk denim fashion. love it

[nova remnant]

Sweet! Thanks! I'll do this after work. I appreciate your help on this. 🙂 I'm also still pretty new to IG, so I'm learning my way around. What does it mean to "refine models"?

[untold glacier]

Afternoon, peeps.

[wispy canopy]

@stable crane @raw rose https://civitai.com/models/148871?modelVersionId=219180

new model merge is out!

[terse jacinth]

@wispy canopy

[wispy canopy]

hi ^^ get on the boat, newest, most badass sdxl model is out!

shameless self promo 😄

[terse jacinth]

im not even sur if i can run SDXL currently, lol

im freshly getting back to things, brand new webui install

what's been going on this alst year i should be aware of ?

[wispy canopy]

hmm, sdxl obviously the big thing (for stability.ai)

then dall3 decided to go slap around everyone

and then uhh... yeah, the channel went pretty silent

[terse jacinth]

lol

[wispy canopy]

with the loss of pow and cow, i suppose people left a bit

[terse jacinth]

pow is gone ?

lol

[wispy canopy]

want some mashed zombie brains? 😄

[terse jacinth]

im already mashed out enough

got this cold going

[wispy canopy]

the same cold that's ravaging nl? if so, good luck

you'll be stuck with it for a while

[stable crane]

yay

[wispy canopy]

@stable crane now you gotta show me your best gens with it too!

(or worst, like mashed zombie brains in a jar)

[stable crane]

whoa cool

[wispy canopy]

this is what my model does so well imho, i just actually mashed two prompts as alternating prompts

"here you go, go figure it out and make it work"

[stable crane]

any style tips for bringing it closer to anime? default looks to be photoreal

[wispy canopy]

yeah, default is photoreal, just slap anime on

that'll do the trick usually ^^

it's a generic model, after all

[stable crane]

some other aspect of my current prompt must be bringing it back to real

[wispy canopy]

what's the whole prompt?

[wispy canopy]

"when the mask comes off"

[raw rose]

Thanks for the link! Going to play around with it. Any certain prompt words i should be using to achieve those rouhgly drawn looking anime image you created a day or so ago? :)

[wispy canopy]

NEG: 3d, photo, realistic, movie, details, high quality, gradient, shading, soft light, shadows, sullen```

change where wanted, but that's a way to do it

so you intentionally tell it to make "sucky" stuff

i think amateurish drawing did a lot of the heavy lifting

[raw rose]

Awesome! going to try that out, thank you :)

[untold glacier]

The model is the best.

[slender barn]

let me know what you think or how I can improve the style!

[brittle jay]

wiat hwat

since when no cow/pow?

[wispy canopy]

haven't seen them in ages

@brittle jay this your look right now? 😮

[brittle jay]

ofc

[stable crane]

how long does this take to gen? you must have a good card?

[brittle jay]

@wispy canopy do you still have that server wghere you said I was not allowed in because I am not 18?

[wispy canopy]

discord? 😮

[wispy canopy]

well, rendertime is purely steps x sampler

1024 50 steps rstart-> 2x 20 steps restart takes about 3 minutes iirc

[brittle jay]

reply to wrong message

[brittle jay]

what no

steps x sampler x (gpu multiplier)

something like taht

[wispy canopy]

of course gpu

but gpu is a static for everyone

[wispy canopy]

yes, a 3090 🙂

[brittle jay]

wdym

it is different for each gpu

[wispy canopy]

it is. but given YOUR gpu, it's a static variable for YOUR environment

and you cannot influence it besides better software, or replacing gpu

what you can do, is change sampler and steps

[brittle jay]

so what was the point of what you wrote?

they said you must have a good card?

[stable crane]

oh, me?

yeah I have a 3080 but it takes me forever to highresfix a 768x768 in sdxl

11.7secs to generate the 768x768 though

I think I have to re-enable the TRT optimizations for this

[wispy canopy]

make sure to build one for both resolutions

they're a great timesaver tho

very nice prompt @stable crane ❤️

[stable crane]

can use adetailer to avoid upscaling being absolutely necessary

it actually kinda understands tan camo

though I'm not sure where it's getting the blue from

blue eyes maybe

[supple raptor]

maybe try desert camo with arctic camo in the neg

[stable crane]

that made it go full blue

I should really just build the TRTs

and stop complaining about highresfix time

[cyan burrow]

what a lovely shower

[wispy canopy]

you bringing out the thiccness tho

[cyan burrow]

efacer

[native halo]

camo works,now my eyes are focused on the skin

[stable crane]

not thicc enough

[nova remnant]

When do they cut you off when you're on the Free plan? I made 12 or 15 images last night on the T4 GPU. I was expecting to get shut down at some point, but didn't.

[brittle jay]

why are people ok with the fact that all of their beliefs or things they think they know are based on just unjustified feeling?

[nova remnant]

[brittle jay]

the basis for logic cannot be logic/logical by itself

if we go down the chain of basis's for everything, we have to get to an end eventually

and at that point, that thing is just baseless

so it is just an assumption

an intuition

how are people ok with this?

absurdism hitting hard

[nova remnant]

#1008652627204132904 @brittle jay

[brittle jay]

sorry

[native halo]

https://tenor.com/view/who-asked-gif-21634393

[brittle jay]

I asked

[native halo]

https://tenor.com/view/the-end-of-evangelion-gif-25422997

[brittle jay]

https://tenor.com/view/moai-muiei-gif-22873582

[brittle jay]

where are thje alians? this is vfake pictuteers

[native halo]

where are the Nadeko images

[nova remnant]

I have no clue what that is

[native halo]

https://tenor.com/view/skull-issues-skill-issues-bruh-hell-naw-how-aaaaaaa-gif-23728802

[nova remnant]

When I pasted in the models and ran it, I got an error saying I was out of space. I take it that's part of where being a paid subscriber comes in? More space?

[brittle jay]

ofc nadeko from the monogatari series

[native halo]

💦

[brittle jay]

bruh

[nova remnant]

:water:

[brittle jay]

https://tenor.com/view/guh-gif-25116077

[native halo]

(slime:1.5)

[brittle jay]

my step brother returns from japan today/tomorrow (now is 2:40am he return 6am)

I am finally gonna get quest 3 headset I asked him to buy

very cool

[native halo]

https://tenor.com/view/anime-milk-cat-anime-milk-gif-23177860

[nova remnant]

And then you're gonna roleplay as a cat in vrchat ?

[brittle jay]

no

[nova remnant]

bruh

Why then

[native halo]

im going to,i have my sexy nadeko 3d char ready

[brittle jay]

because I hackk into the matrix

color a0

cd ..

cd ..

tree

neofetch

[native halo]

https://tenor.com/view/watch-tv-binge-couch-potato-bored-in-the-house-kick-back-gif-20078369

[brittle jay]

ok now to be real

vr is very cool

but my quest 1 pretty broken

almost unusable

quest 3 also has good ar

[brittle jay]

which nadeko

[native halo]

the one that likes to play twister

[brittle jay]

what is twister

oh I see

that game

[nova remnant]

@native halo I think you need to visit the hospital

[brittle jay]

so monogatari nadeko

that is pretty sus

[native halo]

💉

[brittle jay]

isn't she like 13-14

https://tenor.com/view/happycupcake-gif-22997102

[native halo]

idk let me call her mother and ask her

sorry i think i lost her number

https://tenor.com/view/shoko-asahara-asahara-shoko-aum-shinrikyo-aum-gif-25819344

[brittle jay]

https://tenor.com/view/edp445-pogfish-gif-21237217

[brittle jay]

yo isn't this form that one cult's anime

[native halo]

https://tenor.com/view/jesus-is-my-shepherd-gif-5499790

[brittle jay]

https://giphy.com/gifs/ZorabianFoods-dum-biryani-chicken-recipe-TPKTeKMPQuDe3QNiuf

[steady grail]

how 90s felth https://www.youtube.com/watch?v=1NGDSpCeq4A

[native halo]

idk what happened here

[gritty spruce]

😂 you’re going to deflate them with the needle!

[native halo]

anime generation contest when 😔

[gritty spruce]

If all of this is a sim, you are the matrix.

[gritty spruce]

I strongly second this move

We can actually do one here, just for the anime channel

There needs to be some rules

AND sfw hahahaa

[native halo]

yes been savin some extra wide images for that moment

[gritty spruce]

Extra wide or extra thicc

[native halo]

[gritty spruce]

We can totally run a private contest here, for the anime freaks, I think a lot of people would participate, this channel is active

[native halo]

perfect contest for@nova remnant he says he knows lots of anime

[steady grail]

i have some ideas in mind

i have other things to do, but i can try to postpone them. this looks funny

[gritty spruce]

Let’s hear some ideas guys, most of the team comes by from time to time and I can also run then with staff, let’s do it!

The most basic rule probably should be that it has to be SD based. Ie: no dalle or midjourney etc

[steady grail]

i'm for a 90s vibe

[native halo]

what about free 4090 for the 3 winners 🙏

[steady grail]

uahahah

[gritty spruce]

Did you see they are almost gone? No stock anywhere lately

[native halo]

[gritty spruce]

Or $2400 where there’s stock

[steady grail]

they are burning too many microchips

[cyan burrow]

gotta fix faces

well that's enough block tweaking and model merging

all from sd 1.5 though

[native halo]

small gun

[keen hollow]

evenin'

[keen hollow]

Trying the jungle spider queen

[nova remnant]

[native halo]

now do Inori with her original flower outfit and post her here

[keen hollow]

i dunno who that is

[native halo]

🌚

[keen hollow]

still spider queens...

never know who is afraid of spiders 😛

switchin' to bumblebees

[steady grail]

i'm afraid of spoilers 😱
🤣 🤣 🤣

[keen hollow]

hmm...

ok let me s ee what I can do

[native halo]

this one is better

[steady grail]

🤣

[steady grail]

are you trying to heal me? XD

[keen hollow]

trying...

[steady grail]

uahahaah

[keen hollow]

quiet tonight...

[stable crane]

all my gens are realistic tonight

or things like this 😄

[nova remnant]

[native halo]

my leg hurts i need a personal nurse 🤠

[nova remnant]

@stable crane those look like sdxl

hello lewds

[stable crane]

it is

hiya

I've been playing with it today

trying a dedicated anime checkpoint as soon as it finishes downloading

[nova remnant]

oh nyoo, thats a big problem
@stable crane big nurse

[stable crane]

I have to change back to 1.5 to gen a BIG nurse

[nova remnant]

The colors are like solid on sdxl

[nova remnant]

its pretty weird

[stable crane]

it's amazing at absurd art

my fav of today

[nova remnant]

1.5 is also amazing at absurd||ly big|| art

[nova remnant]

3 billion dollars painting

[keen hollow]

for whatever reason, dalle glitched and it has not been taking my tokens at all tonight

just unlimited gens

[stable crane]

bonus

[stable crane]

sucks to go back to sd1.5 hands

[keen hollow]

gotta love hands

[zinc flax]

Flex them hands

[stable crane]

a lot of layers of hands in that pic

[keen hollow]

that uniform is not practical

random: every time someone pronounces costume as "coshtume" a puppy dies.

[stable crane]

people do this?

[keen hollow]

good grief yes

it's growing out of control

[native halo]

since u didnt know Inori 🤠

[keen hollow]

shrugs

[native halo]

sensei 🙏

[native halo]

idk what happened but looks like this is steaven seagull

[keen hollow]

yep that's him

I had Discord on the NES back in the day.

Potatoes was the best game, though.

[stable crane]

dang it takes so long

[stable crane]

awesome

[native halo]

discord accordin to dalle

[stable crane]

discord mod

(plz no ban)

[native halo]

[slate narwhal]

Refining, in its most basic definition, is when an image with a lot of noise is passed through a model that takes out the noise and makes it sharper and of better quality. This "refining" is meant to happen in the last 80% of the final step count. In Fooocus, things are a little different, and there are 3 refining modes, and in one of these, you can refine SD 1.5 models. I won't go into the technical details to avoid too much text in the chat, but the basics of the 3 modes in Fooocus are:

As a VAE: SDXL -> SDXL (Refiner) and SDXL -> SD 1.5 (Refiner). In this mode, the SDXL model keeps intact its main qualities even after the refining happens, but depending on the strength of the refiner, it will absorb the style of the refiner, keeping the prompt. The SD 1.5 part works by using a special "interposing" model and negative embed tricks to do the magic. You need to heavily adjust the SDXL prompt for style and artists to appear in the SD 1.5 refiner if the SD 1.5 model is stronger in strength. This mode is also excellent for boosting the quality of a model and using the great SD 1.5 models in or as SDXL.

As a Joint (Junction): SDXL -> SDXL (Refiner) . In this mode, the models get mixed in the refining process; similar to a merge, the characteristics of the main model will mix with those from the refiner. You can, in a rude manner, predict a merge of two models and do other types of junction tests.

As a Separate (Separation) SDXL -> SDXL (Refiner) . In this mode, several elements from both models get inverted or separated in the refining process. You need to do tests to understand it better.

This can get even more complex and crazier when it is used with the special IPAdapter models Fooocus has, creating complex and fun combinations. I hope this makes sense.

[slate narwhal]

Yes, this is probably what happened; it works fine here. Here is a simpler version of the file, only with three of the best SDXL models; it should load really quickly.

About your other question: "When do they cut you off when you're on the free plan? I made 12 or 15 images last night on the T4 GPU. I was expecting to get shut down at some point, but I didn't."

They never cut me off after my credits expired. But I'm on the monthly plan. I don't know about free. The cut-off only happened when using Auto months ago after credits expired. Using Fooocus, I've tested it with more than 14 hours of non-stop generations, and there were no bans or cut-offs even after credits expired.

[nova remnant]

I bit the bullet and got Colab Pro about a half hour ago. Now I'm trying to figure out which GPU to use. What is your "go-to' GPU? Does it matter in terms of Compute Units. i.e. Do the different GPU's eat up different amounts of Compute Units? What is the best GPU to use for image generation?

[nova remnant]

AT this point in my journey this is WAY over my head. LOL

[slate narwhal]

My go-to GPU is the good old T4. Yes, it does matter a lot, and yes, different GPUs consume different amounts of compute units. The A100 will consume approximately 13 units per hour, while the T4 will consume only 2.35 (using high ram). There is no need to use the A100 or others for image generation, because if you use Fooocus, it's like you are using Comfy, with all the memory optimizations and fast speed.

The setting I do alter is the "Shape" in the "Change Runtime" to High ram. This is important if you want to load 12gb models, helps when changing models (makes it quicker), and eliminates any "out of ram errors".

[nova remnant]

I loaded the first txt file you sent, and there were about a half dozen that aborted the download. I fired it up on T4 (not High Ram though), and selected the Paradox1 model, a refiner and a Lora and it Errored out twice, is that because I didn't have it set to High Ram?

[slate narwhal]

Probably yes. I don't get this kind of error using high ram. High Ram helps a lot if you are using 3 to 5 Loras at the same time, plus refiners and others.

The abortion errors are normal, and they are not to be totally trusted. It says aborted, but then it tries again a few seconds later and downloads it so quickly that it won't give any warnings. You can just check later in Fooocus if the model is there. Sometimes there can be a problem with the Civitai link, and then you just go there and copy a new link, but this is rare.

[nova remnant]

Ok, thanks man. I appreciate all your help. 🙂

[slate narwhal]

I always avoid to use 5 loras, even Comfy can get memory errors with that many Loras + all the other stuff

[nova remnant]

I really don't even know what I'm doing yet. As I mentioned earlier, I don't understand what everything is/ does, like refiners, lora's, embeddings etc. I've only been playing with this for a couple of weeks. Super fun though. I'll learn, in time.

[slate narwhal]

💖 Np. I hope you have a lot of fun with all this stuff

[untold glacier]

Just building my next Waifu.

[slate narwhal]

Yeah, no hurry. In a few weeks, you will understand a lot more. 👍

[untold glacier]

Eface's new model! https://civitai.com/models/148871

SDXL.

[native halo]

sdxl kinda sucks for anime so its normal

🤠

[stable crane]

highres fix helps a ton

but it's slow af

[native halo]

even with that they still look like sd 2.1

[stable crane]

my upscale failed, but have a lowres version

I must sleep

[untold glacier]

Three things: time/experience with it, a good model, and natural language prompting. A good place to start with a model when prompting is to tell a story, mention the subject, talk about what's going on in the scene, and give general quality/style modifiers. Try not to prompt with single words, but rather sentences that help it understand what's going on.

The rest is understanding how the model works, general settings adjustments with samplers, CFG, and steps, and how to upscale, etc. Most of that comes with trial and error, and learning the system.

[untold glacier]

Yeah, I gotta sleep too. G'night Lewds.

[stable crane]

it looks bad until you upscale

[native halo]

now try to upscale it to 16k

[stable crane]

it took a long time just to do this 😢

but I should probably try to mega upscale it before bed

leave it running

[untold glacier]

It's necromancy day in homeschooling.

Ok, I really am going to bed now.

[stable crane]

damn that's cool

must take a long time to run

[untold glacier]

Generally 2-3 minutes I believe. That's the base gen, the upscale, and the face detailer.

[stable crane]

my ice cream took 3min

oh interesting

that's quite a lot faster

I gotta sleep too

[untold glacier]

ComfyUI is the only way I can make this work at all on my system. If I were to try this with current Auto1111, it'd be like 15 minutes or so.

[stable crane]

oh shit

I forgot comfy was faster

ok maybe I have to try that out

[untold glacier]

Yeah, it's a super pain to set up and to change anything, but the template I used and customized helped a ton.

[untold glacier]

Taking this and altering it to my purposes was the best decision, and helps so much with genning with ComfyUI: https://civitai.com/models/119257?modelVersionId=179750

Anyway, we should rest, lol. I am for sure, absolutely going to bed now before I come up with anything else to say. Goodnight world.

[stable crane]

yes

[wispy canopy]

anyone need a heal?

[glacial bane]

Finally a genshin fan

[wispy canopy]

didn't post my salad gundam yet?! this is an anime channel, gundams are important!

[magic wave]

Daisy

[brittle jay]

got my quest 3

the ar is so good

wow

I mean there is some flickering and distortion with these cameras

but it is just so good

[wooden coral]

Dont make me finally buy a VR headset

[brittle jay]

VR?

this is XR

MR

[wooden coral]

Doesnt it also do VR? Im not super interested into the AR stuff cause ill probably vomit within minutes

[untold glacier]

It's a VR headset. That's its primary purpose.

[zinc flax]

Smug

[wispy canopy]

@brittle jay how is the image quality?

[brittle jay]

very much better than quest 1 and 2

[wispy canopy]

cries in rift s

[brittle jay]

I have a serious complaint

the battery is utter shit

[wispy canopy]

how long is battery life?

[brittle jay]

probably the worst battery I have ever had in my life

[brittle jay]

idk I got it on 30 something percent and it lasted like 40 minutes with almost no use except one gaming session of like 15 minutes

you can look at the reviews

they all say the saem

[wispy canopy]

way too puny battery?

[brittle jay]

noww I am trying to charge it

not even workking

it's plugged in for some time now

still 1%

when I get back my 67watt charger I will try with that

[brittle jay]

there is almost no noticable screendoor effect and I would go as far as to say that it is pretty much unnoticable unless you really try to

it is not blurry

actually good screen for once

maybe I am just too used to quest 1

[wispy canopy]

nah, rift s was even worse

[brittle jay]

now I can see my vrchat kittens in higher quality

[brittle jay]

wdym bad face results

show

examples

[cyan burrow]

its the chin

[untold glacier]

Well, a face detailer (or Adetailer if you have an Auto1111 build) can be helpful. Those don't look too bad--just a particular style.
Try adding this to the head of the prompt:
Style: "Anime". (flat colors. 2d art. Lineart. Clean lines. Stylized:1.2)
And start the rest of your prompt with "simple anime artwork of" or "anime artwork of"

[cyan burrow]

does the type of prompt new in sdxl?

where you can write "style:"

[untold glacier]

It's a lot more aware of such things, yeah, and works quite well with it.

[cyan burrow]

instead of "anime art style" and then maybe "splash art style infusion" in 1.5

[untold glacier]

In my admittedly early experience with it--only about twenty hours of constant use and experimentation or so--it's very much sentence reactive.
#🍥｜anime message

It doesn't work the best when you're trying to describe a scene or subject if you're using single words or a more tag-like booru system.

[cyan burrow]

This is what my latest model version (still private testing) does, it's 1.5

[untold glacier]

Though you can use quality or qualifier modifiers that way, like "masterpiece, moody, sketch, lineart" etc.

[cyan burrow]

It works best with a mix of natural language and booru/clip descriptors

[untold glacier]

Your model or SDXL models?

[cyan burrow]

my model, those images

sd 1.5

can do it dark too

[untold glacier]

Yeah. SDXL really brings it to another level though, where you can say:

"A hauntingly atmospheric piece by Hajime Sorayama, depicting an evil dark sorcerer meticulously piecing together a homunculus of a beautiful girl. The scene is bathed in dim crimson light, casting eerie shadows across his workbench littered with arcane objects and ingredients. In the center stands the unconscious form of the young woman, her limbs and features gradually taking shape beneath the sorcerer's deft fingers, as if she were a macabre puzzle brought to life. Her body is taut and pale, accentuating her femininity even in this unnatural state. The air is thick with magical energy and malice."

And then get exactly what you ask for:

[cyan burrow]

the woman isn't standing though, except the one on the left inside the table. She's not pale. I don't sense the malice 😛

but it's good

[untold glacier]

It's the best. When you have a huge paragraph like that that gets the gist and all the important factors right, that's super important in a model, let alone with such high quality.

Oh, I forgot to give a neg prompt. Gimme a sec

(photo. photorealistic. 3D. CG art. disney:1.2). realism. disfigured

That's what I use in my anime gens with Eface's new model.

[cyan burrow]

(beautiful young woman:1.20), (Elven ears with detailed identical horns above them:1.17), medium_breast, Soft Skin, (Beauty filter:1.12), alone, (realistic viridian eyes:1.15), (realistic neck-long messy cobalt_blue hair:1.15), (Lamia:1.15), evil, (solo:1.15), (wearing primitive fur and leather dressings:1.14), (digital illustration:1.13), (digital art:1.13), (anime_art style:1.15), (line art design:1.12), (Acrylic Style Infusion:0.90), cowboy shot, (soft lighting:1.08), 8k, ProPhoto RGB, Crepuscular Rays, Ray Tracing Reflections, Tone Mapping <lora:epi_noiseoffset2:1.2>

Negative prompt: big breast, trademark, watermark, title, trademark, copyright, brand name, company name, eroge, hentai, nude Negative_Hand Bad_Negative_Anatomy

Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1455061902, Size: 768x768, Model hash: 44e9fea24a, Model: XCell_0.3, VAE hash: df3c506e51, VAE: kl-f8-anime2.ckpt "epi_noiseoffset2: d1131f7207d6", Version: v1.6.0-2-g4afaaf8a

it works pretty well, it used to respond almost entirely to natural language but after some merges, it requires a combination

so I segment them

[untold glacier]

That's good practice, both in SDXL and 1.x models.

[cyan burrow]

and if there's a scene, that's what I end with

It unfortunately made it snow inside for some reason, but still.

(adorable girl, moe, solo:1.12) (fox ears, soft fur:1.15), (deep ice green eyes:1.1), (arctic white long hair, arctic fox tail:1.12), (Log cabin in background, one window lit:1.11), dim lighting, soft lighting, natural light, medium shot, full body shot, (ProPhoto RGB, Post Processing, Post-Production:1.11)

Negative prompt: big breast, trademark, watermark, title, trademark, copyright, brand name, company name, eroge, hentai, nude Negative_Hand Bad_Negative_Anatomy

reason I include big breast in negative is that so many models that I merge with has this...unmistakable bias towards massive tits

this was made in an earlier merge, I haven't tested how it works in the latest

[brittle jay]

as they should

jk

[cyan burrow]

heh

[brittle jay]

all sizes are good

[cyan burrow]

bayonetta boobs on everything kind of distracts

if my character needs big tits, I'll prompt for it

@brittle jay latest merge is working good so far

[untold glacier]

What CFG are you running that at?

[cyan burrow]

the distortion thing can be fixed with patches and software tweaking

[short coral]

lookin a lil blurry/smudged

[brittle jay]

wwhat is

[untold glacier]

Yeah, it looks like the CFG is a bit too high to my eyes.

Slightly burnt, a little overbaked and smeared details.

[short coral]

yep

[untold glacier]

None of this is upscaled though, and some of that would be fixed then.

[brittle jay]

I said that like a couple months ago about some older versions

nobody understood me

[untold glacier]

CFG can be a hard thing to explain and spot to someone not familiar. Just try and be me a year ago trying to explain and show how Xformers very slightly deteriorates the visual quality.

I don't think even Nubby saw what I was talking about, lol.

[untold glacier]

#🍥｜anime message

[brittle jay]

.

[cyan burrow]

huh? What is smudged or blurry? the cfg is low at 7

[brittle jay]

yes

[untold glacier]

You can't... see that?

[cyan burrow]

be more specific cause I posted a lot of images

[brittle jay]

look at the thread/conversation following this #🍥｜anime message

[wispy canopy]

songmaster going loco!

[short coral]

quick dump cos we announced NAIDv3 today

[brittle jay]

wwaht is that

[cyan burrow]

some I can see have issues with the eyes, the woman in black is smudged

the woman in red with red eyes has a blurry face

but no adetailer on that

or fix

[brittle jay]

models?

[short coral]

all just one model

[brittle jay]

model?

[short coral]

NAID 3.0, unreleased

[untold glacier]

When the entire image is smudged and blurry, it can be hard to distinguish between a non-smudged one.

[short coral]

ye it's super clear, although ig it ultimately depends on your quality standards

[cyan burrow]

ok so you mean those are areas with blur or smudge?

[brittle jay]

high cfg effect

[untold glacier]

They're particularly egregious areas, but the entire images are.

[brittle jay]

artifacts

burning

[cyan burrow]

the woman in red I can at least blame on it being a matrix prompt 😛

[untold glacier]

Turn your CFG down to 4 and then gen the same images.

[brittle jay]

there are several ways to try to explain it

[cyan burrow]

dude, my CFG is at 7

stop saying it's high

[untold glacier]

Turn it down to four, and you'll see a difference.

[brittle jay]

who said it's high?

[wispy canopy]

BrightProtoNuke 1.2. Love it ❤️

[brittle jay]

it just looks like the effect that high cfg does

[wispy canopy]

finally found some prompt that takes songmaster for a GO

[cyan burrow]

8:48 PM]crazykazoo: high cfg effect

[untold glacier]

High is relative. The effect I see can be caused by many things including "high" CFG.

[brittle jay]

can you give me png info of fourth one?

[cyan burrow]

It can also be a prompt from an older build not working on a newer one

[short coral]

sure sec

[cyan burrow]

since I'm copy pasting older prompts onto a new build

[brittle jay]

it is the high cfg effect. the effect that is associated with high cfg / that high cfg does

[stable crane]

so 7 CFG is not "high" generally, the model might just be burnt?

[cyan burrow]

ok but it can be related to anything?

[stable crane]

or maybe VAE?

[brittle jay]

ye

[cyan burrow]

the vae might no longer be compatible

[stable crane]

I'd think of it as more of a "burn" effect (that you often see on very high CFG)

[cyan burrow]

or maybe I need to put it on clip 2

[wispy canopy]

Edge and Pearl - Machinae Supremacy

[untold glacier]

This will help.

[brittle jay]

exactly

man I talked about this like 7 months ago

like this entire topic

[stable crane]

you can see it on overtrained lora too, or even just lora with str too high

[cyan burrow]

though sometimes it's not so much burn as it is color smearing, pixel noise etc

[untold glacier]

I think we all have, lol.

[untold glacier]

So you can see it? What we're talking about?

[cyan burrow]

yes but you never explained it well @brittle jay , I mean no offense...But you provided no deep explanation. You just pointed to the areas

[brittle jay]

#🍥｜anime message

[cyan burrow]

yes I can see it most of the time

[untold glacier]

[stable crane]

neat

[brittle jay]

cause I do not know what the underlying reason is

how would I explain it

[untold glacier]

Don't worry. We've been explaining it plenty.

[short coral]

I can't find the original with metadata saved god damnit

[brittle jay]

L

[cyan burrow]

I'm not coping, I wanted to learn how to fix it from more than just seeing where

[brittle jay]

😭

[cyan burrow]

copium is when you don't want to learn, you just need an excuse

[untold glacier]

Lower CFG. But only after you try Clip Skip 2.

[untold glacier]

I'm well aware, lol.

[cyan burrow]

Is there any sampler that has tendencies towards burning, blurring etc?

that I should avoid? Generally

[untold glacier]

No. That's usually all up to the model, and if you're using network extensions.

[brittle jay]

I mean on lower steps

some samplers do that

[wispy canopy]

restart best sampler

slowest sampler too

[untold glacier]

I will say that DPM++ 3M SDE has some problems though with certain anime models.

[wispy canopy]

if restart can't handle it, none can

[cyan burrow]

restart? I don't have that

or do I?

[wispy canopy]

if you have a1111 1.6, you should

[stable crane]

@wispy canopy I'm totally not sold on FreeU

[short coral]

Euler my beloved

[untold glacier]

Depends on the model.

[wispy canopy]

I can't live without it

[cyan burrow]

yeah DPM++ 3M is a bit...miffed imo

[stable crane]

I do love Restart now

[brittle jay]

euler🤮

[wispy canopy]

to each their own, i suppose

[stable crane]

EulerA is a champ on 1.5

[wispy canopy]

I suppose freeu helps with my bullshit prompts 😮

[untold glacier]

Oh hey Nubby.

[brittle jay]

nubby the goat

[untold glacier]

[supple raptor]

like many things in this field, it's different, but not inherently better IMO

[untold glacier]

Really does depend on the model.

[cyan burrow]

EularA's sample number can drastically alter the looks too

[untold glacier]

And your CFG. Lower is generally better.

[cyan burrow]

20 can be one generative, 40 could be 100% different

[brittle jay]

no

dynamic thresholding

[untold glacier]

With FreeU, yes.

[brittle jay]

oh idk about freeu

[cyan burrow]

ok so karras or regular?

[brittle jay]

karras forever

[untold glacier]

Depends on the sampler.

[brittle jay]

it is still preference

[untold glacier]

2M, absolutely. SDE, not so much if you can help it according to my studies.

[brittle jay]

2S a I prefer without karras

[cyan burrow]

and what's the recommended minimum steps on restart, or most proficient and efficient?

[brittle jay]

2M with

[cyan burrow]

I've dealt with the others and I know the steps on them

[brittle jay]

@wispy canopy

[cyan burrow]

but on restart I don't know

[stable crane]

no FreeU

[untold glacier]

Generally you want to go higher with Restart, since it restarts itself. 50 I'd say.

[stable crane]

should I be using this setting @wispy canopy on your model?

[untold glacier]

Otherwise you might get cohesion issues.

[cyan burrow]

thank you

[stable crane]

I just click ✅

[brittle jay]

wtf is that

[untold glacier]

And yeah, I'm not a big fan of Restart like some are, lol.

[cyan burrow]

and I never knew what Prompt word wrap length limit to have

[wispy canopy]

if that's the SDXL template, yeah

[untold glacier]

I'd bring stage two backbone 2 scale to 1.1, but otherwise I'd say it looks good.

[cyan burrow]

and damn, where can I find unet models?

[brittle jay]

is there explanation for this somewhere?

[untold glacier]

They aren't worth it.

[wispy canopy]

lcm lora is amazing!

[short coral]

we're so back

amazing quality, highres, aesthetic, faux traditional media, photoshop (medium), oekaki, jaggy lines, [sketch, lineart, hatching (texture)], 1boy, toned male, sitting, solo, holding cup, mug, stubble, facial hair, bags under eyes, green eyes, brown hair, messy hair, short hair, @_@, round eyewear, green sweater vest, collared shirt, loose necktie, looking afar, tired, upturned eyes, spoken squiggle, open mouth d:, single earring, indoors, office, detailed background, depth of field

---

neg:
{very displeasing, worst quality, bad quality, monochrome}, bad image, lowres, bad anatomy, cropped, jpeg artifacts, 3d, multiple views, {{chibi}, realistic, younger, loli}, [gigantic breasts], bad hands, {close-up}, card (medium), dakimakura (medium), [touhou, touhou (pc-98)]

Steps: 30 | cfg: 7.4 | Sampler: DPM++ 2M | Seed: 2545747747 | neg strength: 130% | Rescale: 0.78 | NS: Exponential

[wispy canopy]

that's free-u

[brittle jay]

your syntax is wrong

[short coral]

no it isnt

you're just not using NAI

[brittle jay]

I mena unless comfy or something?

[cyan burrow]

ok and word wrap? should I put it on like 42 and follow the faith of the universe?

[brittle jay]

do yyou mean nai as a service?

[untold glacier]

This saddens me, lol.

[brittle jay]

oh shit

[untold glacier]

Poor NAI people

[brittle jay]

this is a rare moment

[short coral]

we're still goin strong

[brittle jay]

actually using nai as a service

inconceivable

[untold glacier]

I've seen their latest model and tried it out. I love the artist name recognition that it has--that's amazing. But other than that, I'm sad that it's pretty far behind everything else.

[supple raptor]

that should be far more than necessary with restart. add ~10-20% to the number of steps you use on DPM++ samplers and you should be good. You don't need as many steps as the older samplers like Euler, DDIM, etc.

[short coral]

you haven't seen v3 examples yet?

for a general model it's pretty wild

[supple raptor]

NAID v3 is looking very impressive

[untold glacier]

If it's restarting halfway through the process, that means that generally the numbers will need double the amount right? I really haven't had any good gens with lesser numbers, but granted I haven't used it a ton.