🍥｜anime

[keen hollow]

ah well, maybe ill stay with what I have then... lol

Swarm never worked for me right. Tried it on multiple occasions.

I have a comfy install but I avoid it if at all possible. 😄

[supple raptor]

What sort of problems did you have with Swarm?

that's what i use for basically all of my gens now. i only load up Forge when i want to be able to gen from my phone since the Swarm mobile layout is unusable

[keen hollow]

I've had issues getting it to run newer releases. (SD3 and Flux both had issues)

But I have a somewhat custom environment.

[supple raptor]

oh... i bet you were still on the old StableSwarm version, not the current SwarmUI one. mcmonkey, the guy who mad StableSwarm, left stability and had to create a new fork so that he could continue maintaining it since the StableSwarm repo belongs to stability

this is the repo that is actually still being developed https://github.com/mcmonkeyprojects/SwarmUI

[keen hollow]

I did try the SwarmUI version, one of the early ones after it first shifted.

[modest fog]

@supple raptor do you know if swarm supports zluda

[supple raptor]

it just uses comfy so it should. you'll want to do the custom install option and add your existing comfy install that has zluda already functioning though. also, i know zluda development is in a weird place for now since it was an AMD sponsored thing that they decided to stop. the dev is continuing work on the project but he had to start over from where he was before AMD sponsored him

but if you can get comfy working with zluda in a python venv (put the venv in the root of your comfy directory) then all you have to do in swarm is add a custom backend with the StartScript pointed at the main.py inside your comfy directory

[modest fog]

[brittle rampart]

Thanks a lot. let's try.

[modest fog]

You'll become noob funya now

[brittle rampart]

nooooo... i want try use SwarmUI

[modest fog]

[vital raptor]

SwarmUI is good, that's what I use too

[brittle rampart]

if I made these settings. Will the upscaler start? Am I right?

[vital raptor]

yes

also, if you use that but don't want it on all the time, there's a 'Refine Image' button in the More menu that will use those settings just for the current image so you don't have to turn it on/off all the time

[brittle rampart]

this is?

[vital raptor]

no, the Refine Image one, not Upscale

I added the Refine Image option because the Upscale 2x button doesn't use the refiner settings

[brittle rampart]

how and where i can do this?

[brittle rampart]

stole the idea, right?🤣

[vital raptor]

do what?

[brittle rampart]

added the Refine Image option

[vital raptor]

oh, yeah I meant I added it to SwarmUI.. I contributed the code for it 😄

there is a setting under the User -> User Settings tab where you can change which buttons show though, so you don't have to click on the 'More' menu each time

those are the ones I have showing, the rest are still in the More menu

[near elbow]

"imitation is the sincerest form of flattery" 🙂

[brittle rampart]

and where in SwarmUI for SD1.5 models can I see and set the CLIP_skip parameter?

[vital raptor]

it's under advanced sampling

'CLIP Stop'

[modest fog]

You're smart cat

[vital raptor]

I mean, if you want something done right sometimes you just have to do it yourself 😛

Now that swarmui has proper extension support, i'm going to make a couple extensions for features that won't go in the main release

[brittle rampart]

@vital raptor @supple raptor SwarmUI has something similar to the Tagger extension in Forge?

[bitter latch]

tho i'm not sure it's her's.....

[modest fog]

Ffgk!

Sounds like a messed up swear

[bitter latch]

ok it's creeping me out

[modest fog]

[bitter latch]

goes back to bed

[zinc flax]

https://tenor.com/view/anime-frieren-fire-flames-hot-gif-7599056546983191152

[keen hollow]

@steady pumice This one is fun, how'd you do it? The axe looks good.

[steady pumice]

lots of inpainting

[twilit portal]

❤️‍🔥

[zinc flax]

Looks super clean, almost like an actual game cg lol

[steady pumice]

thanks that's what I was going for when i trained the style lora it's using, that's the first gen using it and I'm happy with the way it turned out

[zinc flax]

Style looks familiar pretty familiar but don't know the exact name

[steady pumice]

it's using a mix of things

ZzZ, genshin, honkai, star rail, Wuthering Waves, and other gacha games cgs that have close enough style with each other that they wouldn't conflic to much

[zinc flax]

o it does look like ZZZ style

Nice

[steady pumice]

I recently started playing it and paid up to get burnice lol

[zinc flax]

Based

https://tenor.com/view/zenless-zone-zero-zzz-dance-dancing-groovin-gif-13484640386046654412

She's a fun character

[vital raptor]

SD3.5 is... inconsistent so far..

[keen hollow]

need more examples

[vital raptor]

same prompt

extremely high resolution anime screenshot of a small female child with short white hair and purple eyes. She has cat ears and a tail and is wearing a frilly dress. Standing in a forest over a vast valley full of trees and flowers with mountains in the background and fluffy clouds in the sky. She is facing the viewer.
Not a great prompt either, but I was hoping for more

[keen hollow]

I guess I need to try it out

ugh... means I have to use comfy

[vital raptor]

I'm using swarm

[keen hollow]

I suppose I could try swarm again.

[vital raptor]

the model is very repetitive as well, normally don't see so many repeated patterns without going outside the normal resolution

7.8s for a large size model is nice though, just gotta figure out why it generally looks like garbage

most images are blurry or pixelated

[keen hollow]

eh maybe ill wait a while before setting up all this crap

[vital raptor]

i'm hoping i'm just doing something wrong, otherwise it's another flopped model

[zinc flax]

Back to sdxl we go lol

[umbral depot]

Back to 1.5 still

[zinc flax]

1.5 also looks good

[umbral depot]

All of my loras and ti and models and prompts are 1.5 so going to sdxl is a pain

And then 3.x models don't work on a1111 do even more reason to not go that route

[zinc flax]

literally me moving from pony to illust

had to delete old loras cuz i won't be using that again lol

[glacial coral]

So we still gonna be using Flux and Pony models?

[vital raptor]

well it's only been out for a few hours, i'm sure general usage will improve

[glacial coral]

True I’m still excited for community stuff that will come out for it

[vital raptor]

personally, I'll probably be sticking with pony models for a while since the benefits of flux and sd3.5 don't really make up for what it can't do

[glacial coral]

What’s ur favourite pony model u like using

[vital raptor]

i use autismmix_confetti

[glacial coral]

Yeah that’s a pretty good one I used to use that quite often

[vital raptor]

i found a few styles I like from it and havent had a reason to find anything new

[glacial coral]

Fair, u got some fav loras u like using

[vital raptor]

i don't use any

[glacial coral]

I may start creating anime stuff again

https://civitai.com/models/404154/wai-ani-nsfw-ponyxl I used to use this quite often also

Well a variant of that

[glacial coral]

quite nice

[green temple]

thx

[keen hollow]

Yeah you still using confetti?

[vital raptor]

yeah

[keen hollow]

Seems interesting... Seems like it can do backgrounds with the characters

[keen hollow]

Haha right on. I messed around with confetti a little while yesterday

[keen hollow]

I need a good workflow to create multiple characters and a good background in widescreen. Probably need to combine multiple generations and outpainting or something.

Would like a final fantasy battle scene

I feel like I've mastered just solo Tifas 😎

[near elbow]

your prompt, flux fusion, anime style, first result

[vital raptor]

yeah, flux tends to follow the prompt better in my experience

it still seems to have odd quality issues sometimes, and it's pretty bad at anime in general unless you want a fairly generic style

[glacial coral]

flux fusion? what model is that

[vital raptor]

plus it's slow, ~15s for a base gen with no upscale

[glacial coral]

what GPU u got?

[vital raptor]

but, i've only tried the base dev model, haven't tried any finetunes or merges yet

[vital raptor]

4090

[glacial coral]

ah ok i was considering grabbing a refurb or used 3090 for sub 600 GBP

[vital raptor]

if you can get it cheap, it's probably worth it, my 3080TI does pretty good

really wasnt worth the cost to get a 4090 tbh.. the speed is nice, but so expensive

[glacial coral]

I have a budgte build

so 4090 maybe a issue price and the performance

[vital raptor]

yeah, without a system and powersupply that can handle a 4090 it wouldnt be worth it anyway

i spent around $6K on my system including my 65" 4K display

[glacial coral]

yh i have a 5600 as my cpu so either AMD GPUs or used 3090

[brittle rampart]

but I got carried away with Scaling. 🤣

[vital raptor]

the prompt i used in SD3.5 in autismmix_confetti (with the pony tags i use added)..

[keen hollow]

I'm gonna consider a 5090 😝

[vital raptor]

i'll be considering it, but it's likely not in my budget for a while

as much as I kind of regret the 4090 purchase, I'm also doing a lot of AI stuff and need another gpu anyway, 5090 might be it, might be a used 4060TI or something.. who knows

i'll be waiting until the 5090 is released at the very least

[rough oxide]

anything known about 5090 yet?

he's way too much into this vn

[near elbow]

https://civitai.com/models/630820?modelVersionId=936976

[near elbow]

got the 4090 when it came out, best purchase ever

[near elbow]

she's not much better herself 🙂

[rough oxide]

supr into her music. probably listening to her own tracks xD

also dang, "super saiyan fuchsia" just works -todd howard

[keen hollow]

Yes quite a bit is available now. Release at CES in January.

[rough oxide]

o it actually gonna have 32gb vram? i need that

[brittle rampart]

Guess in 3 tries who came for the sweets! 🤣

[supple raptor]

not flan, remi didn't let her go trick-or-treating

[near elbow]

miko wants one too

8GB more than now with the 4090, hmmmm... but 600W instead of 450W

[keen hollow]

600W is pretty crazy

gonna cost $9.00 to generate an image

[rough oxide]

600w? ughhhh

i'd have to up my psu

still on 850w

if i didn't brick my 1200w psu from my prev pc....

xD

[glacial coral]

U lot looking at 5090 and stuff while im still considering a 3090 😭

[rough oxide]

3090 in my curr pc rn

[glacial coral]

I have a 6600 xt rn in mine

[rough oxide]

3090 works for sd3.5 large so it's gud.

[glacial coral]

I mainly wanna get one because it’s sub 600 GBP and I can come off my MacBook

I’m starting to figure out until then I may have to GPU rent

Quick question also what’s some good style words I can use when doing a prompt for anime?

[keen hollow]

I'll strongly consider a full new PC in 2025.

9800x3d, 5090, 64GB

[rough oxide]

depends on your model of choice. you gotta state this first.

also... the kind of style you're trying to achieve

like if you're using noobai you can just directly state artist names

[glacial coral]

Yh that’s true

[glacial coral]

I’ll prolly use this

[rough oxide]

_<

pony is scourge of serious anime gens

lol

[keen hollow]

why?

[rough oxide]

the way they did their training & such

it completely forgot basic concepts

https://civitai.com/models/833294/noobai-xl-nai-xl

this/illustrious v0.1 is my recommendation

also

pony is soon going to do the thing where they scrub artist names

so just overall pony is just a lulz model rn

lot of loras made for illustrious already too

[steady grail]

but noobai is noob, dah!

[rough oxide]

iz noob friendly

dun need 2 paragraph long prompts

[keen hollow]

I gave illustrious a good try and found it yield a lot of meh results

[steady grail]

that's not because it's friendly, but just because it's too noob to manage more

[glacial coral]

This looks real nice I’ll give it a go

[rough oxide]

it is understandably underbaked. but that's why things like noobai came along + the creator illustrious might still release the future versions (but it is a company model)

the stuff it can do now far surpasses NAI but that's like pretty far away if that ever comes

[keen hollow]

I haven't seen much from noobai.

[rough oxide]

noobai though is an additional 12m finetune for 5 epochs and it smooths it out

[keen hollow]

Need to see Tifas to make a call. 😄

[supple raptor]

pretty much everything i've posted in the last few days has been from noobai

[glacial coral]

Looks like a really good pony model

[rough oxide]

no it's an illustrious model

lol

well

[glacial coral]

I need to grab those glasses real quick💀

[rough oxide]

pony has its own prompting structure & stuff like that

noobai is a finetune off illustrious which by itself is an uber powerful base model

just needs more training is all

[keen hollow]

ah right on

if nubby uses it then it must be legit 😄

[glacial coral]

Also just noticed it’s SDXL lmao

Yh ima need to grab those glasses quick

[steady grail]

i would love so much more light weighted base models 😦

[modest fog]

"serious anime gens"

now ive seen everything

[glacial coral]

(kinda just took the prompt) its pretty nice for sure

[modest fog]

amd just announced something 9000x3d

they literally announced nothing more, just gave us a date to tell more lmao

[keen hollow]

yeah there's already leaked cinebench scores for it

[modest fog]

can you link me that o:

[keen hollow]

how about this? https://youtu.be/tkJJHLP_me8?si=WLEr0jF5LEkFQ6-_

haven't watched it yet

But I'm confident thatll work. 🙂

[glacial coral]

how much more gaming performance is it vs 7800x3d

[keen hollow]

no one knows yet

[modest fog]

also I hate this guy

hes so annoying sounding

[keen hollow]

LOL

ok one moment

[glacial coral]

I love him

[keen hollow]

Here watch Riley 😛 https://youtu.be/USEX-ho8ZZI?si=P-R3rp7NCe6YLbH8

[glacial coral]

Riley is also great

[modest fog]

its fine

I just watched muted

his hair is stupid too

(literally like my hair)

[modest fog]

amd either launches something great

or extremely not worth the prior one

lmao

[glacial coral]

Better that than Intel CPUs

[modest fog]

nubbyai when

[modest fog]

surely next intel one is not a 5% upgrade

[glacial coral]

💀

[supple raptor]

he's one of the few tech youtubers that actually knows wtf he's talking about

[modest fog]

I know

Lmao

[supple raptor]

Kakigori-XL might become a real thing now that we have a base XL model that isn't garbage for me to start from

[modest fog]

I like watching hardware unboxed

[supple raptor]

hw unboxed is also good

[modest fog]

It's not like I need much of that info, just really like the videos

When I'm eating and stuff cx

I thought noobai was a base model

Didn't realize it was illustrious

[supple raptor]

level1techs is very good for info, but doesn't cover a lot of stuff that average people/gamers would care about

[supple raptor]

it might as well be thought of as a base model. it's not an aesthetic finetune, basically just a continuation of illustrious training

[modest fog]

Everything confusing

Let me gen something

[supple raptor]

noobai is gud, use noobai

1girlyesyes

[modest fog]

No, that requires too much work (not genning inside discord)

Booba

Less confusing just Booba

[modest fog]

Big

[rough oxide]

how'd i end up in a fantasy world? this is still very confusing.

[modest fog]

Where are the tentacles

[rough oxide]

off screen to the right

[modest fog]

They are coming

Hmm

Would tentacles be safe in this server

[rough oxide]

probly not

prob not in the main anime server either.

the inner machinations of my mind are an enigma.

i have no clue what i'm cooking up anymore

[modest fog]

The hell is a main anime server

The main anime server for me it's quite different

[rough oxide]

i'd consider touhou project AI the main one

lotta devs there too from various anime projects

[modest fog]

Look in dm

[green temple]

guys what does 4koma & 2koma mean in prompt ?

or negative prompt

[supple raptor]

number of panels in a comic strip

[green temple]

oh ok

[modest fog]

nubby is wrong

put that in positive

and make many smols at the same time

@green temple

[supple raptor]

FlanPeeved

[green temple]

smol forever

[steady grail]

cute

really like the style

[modest fog]

cursed controller

[rough oxide]

probly some third party controller

[modest fog]

from e-wisbay

[supple raptor]

Pokemon boy got the lil' bro controller

[vital raptor]

it's just a left handed controller

[modest fog]

Left handed people

[vital raptor]

got a problem with left handed people?

[modest fog]

goblinite

[modest fog]

its not right

[bitter latch]

good thing we left it behind

[rough oxide]

dayum that's good. which model?

[near elbow]

flux fusion (remix from schnell and dev plus some magic)

not even a specialized anime model 🙂

[terse ore]

There was an option where, whatever typed in the prompt would give a list of options for auto fill of booru tags. I reinstalled ComfyUI and can't remember how to enable it again.

[rough oxide]

https://github.com/pythongosssss/ComfyUI-Custom-Scripts

maybe you're referring to this?

[terse ore]

Ah, that's what it was, yes. Thank you! 🙏

[rough oxide]

ookami mio?

<3

[brittle rampart]

yes

[modest fog]

No

It's 1girl

[rough oxide]

fusion almost seems too close to schnell to me... >_< i still probably like the outputs of dev better overall mayhaps. although not sure what you meant by some magic 👀

[rough oxide]

elder scrolls vi will absolutely have a mod for this day 1.

if not then the modding community has failed.

[modest fog]

The devs already made a mod for Thomas for fun

(probably)

[rough oxide]

most likely.

kept hidden in vaults so copyright demons can't strike.

[near elbow]

well the de-distilling and merging and whatever else they did... but i totally like the results. 8 steps are fast enough for my taste

[near elbow]

too late

[rough oxide]

onyoooos

[modest fog]

Grabby ears

[green temple]

left handed people what

[modest fog]

Funya never let me grabby

[modest fog]

Hamburgers for mouse girls?!

A whole lot of cheese

[brittle rampart]

nooooo! just a "double cheese"!

[lyric zealot]

https://tenor.com/view/steal-racoon-cute-funny-animals-cat-food-gif-15431336

[supple raptor]

smol stole my shion

marisad

be sure to gib her lots of headpats at least

[lyric zealot]

https://media.discordapp.net/attachments/1291041773266796575/1291337505160953981/nuero_.gif

[modest fog]

Ugly

[supple raptor]

@lyric zealot here's that lora btw https://civitai.green/models/834822/nyalia-illustrious

[lyric zealot]

https://tenor.com/view/neuro-sama-vtuber-shoot-gun-pov-gif-7305956933862925905

[modest fog]

[lyric zealot]

that's a style right?

[near elbow]

bad trigger finger though

[supple raptor]

yeah, i wasn't sure at first either

[supple raptor]

"no u"

[modest fog]

kok

https://tenor.com/view/kok-cat-gif-25438805

Those are my last words

dies

[near elbow]

"angry anime waifu firing an assault rifle, evading bullets, superior firepower"

[brittle rampart]

now we are waiting for Marisa 😉 🤣

[lyric zealot]

she's busy eating

[supple raptor]

she's finished eating and is on her way now

btw... heres a merge of the noobai vpred model that aims to fix some of the issues it had due to being trained on a much smaller dataset. seems good from the first 2 images i've made. if anyone needs it i can add the keys necessary to run it in forge/a1111 dev branch and re-upload to hf https://civitai.green/models/873533?modelVersionId=977843

[lyric zealot]

... is on her way to another burger joint for her burger

[supple raptor]

https://tenor.com/view/touhou-burger-mogu-gif-25778324

[lyric zealot]

does that mean if you add that whatever, i can use it on my super outdated forge?

[supple raptor]

nyo, it'll need to be updated for the keys to work. depending on when your version of forge is from, you might be able to just make a copy of SDXL's config yaml and add parameterization: "v" in the params section though

then name that file the same as the model and place it in the same folder as the model.

[lyric zealot]

depending on version. surely my version will work

[supple raptor]

probably tbh. i'm pretty sure that worked basically from the start of forge until he abandoned the last bits of A1111 inference code and switched to obfuscated stolen ComfyUI code instead.

[lyric zealot]

did he really

[supple raptor]

yesh

[lyric zealot]

the i made it myself, totally not copy and paste

[supple raptor]

this PR is kinda memey in the sense that parsee doesn't actually expect it to be merged, but it is all real info https://github.com/lllyasviel/stable-diffusion-webui-forge/pull/2151

[lyric zealot]

oh, this just happened

a developing story i see

[supple raptor]

yeah

[lyric zealot]

https://github.com/Stability-AI/generative-models/blob/main/configs/inference/sd_xl_base.yaml

do i just copy it from the link above, and then add what you mentioned at line 6?

[supple raptor]

i think that should be right. that worked on 1.x models anyways, but i see a reference to EpsScaling there under denoiser _config. if it doesn't work, your output will be just noise

[lyric zealot]

wait, what do you mean by 1.x models?

[supple raptor]

SD1.4 or 1.5

[lyric zealot]

i thought that's sdxl based model?

[supple raptor]

yesh, this one is an SDXL model and the yaml you linked is for SDXL. the 1.x yaml didn't have any reference to eps and those are the only models that i've loaded this way personally

so i'm just not completely sure that it works the same way on sdxl models

[lyric zealot]

i guess it didn't work with my version of forge

[modest fog]

[lyric zealot]

https://tenor.com/view/мип-mip-mahouka-koukou-no-rettousei-raihousha-hen-shiba-miyuki-gif-21596781

[modest fog]

hentai gif

[lyric zealot]

fine, it's time to switch to swarmUI

[supple raptor]

time to switch to SwarmUI i guess

[lyric zealot]

https://tenor.com/view/jodiewalton-vacation-trip-packing-bags-gif-26033180

[modest fog]

but swarm ugly

[lyric zealot]

people should totally make a "transfer my stuff from a1111/forge to other ui" stuff

[supple raptor]

In Swarm you just go to the model browser, Edit Metadata, and select

[lyric zealot]

you ugly, and still here as well so

[modest fog]

What do you mean im still here

I am always watching

[supple raptor]

i just left all my shit in the A1111 directories and set Swarm (and all half a dozen other UI's i have installed) to access it from there

[lyric zealot]

https://tenor.com/view/youre-already-dead-omae-wa-mou-shindeiru-fist-of-the-north-star-gif-14326210

[lyric zealot]

been doing that for my forge as well, but might be time to finally start cleaning up my stuff

[supple raptor]

even my outputs folders are all actually a subfolder inside the A1111 outputs directory that i just symlinked out to wherever the UI's expect them to be since i was too lazy to tell each of the UIs where their outputs folders are

[lyric zealot]

time to learn another webUI

[umbral depot]

We going back to 1.5 models

Can we go back to wd2?

[supple raptor]

image shows how to do hires fix since it's not intuitive for someone coming from A1111 or similar.
for face detailer, use <segment: in the prompt. it will give autofill with an explanation, you can either use clipseg to just type what you want detailered, or you can load a yolo model to make it just like adetailer

[lyric zealot]

adetailer is built-in right?

[supple raptor]

here's a nice face yolo model that i've been using in case you don't have it. it uses segmentation to mask only the face instead of doing the whole bbox around the face. https://civitai.green/models/293448/anzhcs-face-segmentation-prototype-or-yolov8-or-adetailer-model

[supple raptor]

yep, booru tag autocomplete is also built-in

[lyric zealot]

too many info

is there still stuff like xformer?

[supple raptor]

yep, i think by default it will just use sdp, but if you install xformers it will use that instead

good time to make sure you're updated to pytorch 2.4.1+cu124 and xformers 0.0.28.post1 as well since those were a decent speed increase for me

oh, xformers 0.0.28.post2 wheels just came out yesterday for pytorch 2.5.0 support. maybe i should see if i can break my venv again

[steady grail]

why not using 2.6.0.dev20241002+cu124?

[supple raptor]

i'm lazy and don't like updating my pytorch versions often. i just finally got everything switched to 2.4.1 the other day

theSaddest

[steady grail]

lol

i should install xformer now that i think about it

not installed when i reinstalled comfy

anyway that update busted stuff a lot

[supple raptor]

you'll have to try to build from source for pytorch 2.6

[steady grail]

much more faster now

i'll see

later

[supple raptor]

last time i tried building xformers from source i got skill issued and went back to a working pre-built whl

[steady grail]

lol

i do stuff for fun

[lyric zealot]

i totally know how to update my pytorch

[supple raptor]

with venv activated pip3 install -U torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124 then to install/update xformers pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu124

if you ever need a specific version of torch to go with xformers just add version after torch and it will get the right torchvision and torchaudio for it like pip3 install -U torch==2.5.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

required torch version for the latest xformers whl is listed in the xformers repo https://github.com/facebookresearch/xformers

and if you need a different cuda version for your gpu you can just replace the cu124 on the end of that url with either cu121 or cu118

oh waow torch 2.5.0 brought my gen time down to ~50 seconds compared to ~60 seconds with torch 2.4.1

[lyric zealot]

I don't even know where is my virtualenv for swarmui

is it the newly installed comfy backend?

[supple raptor]

yeah it'll be in there

it just gives you the comfy portable .zip which includes it's own version of python. i don't rember how to update that since there isn't actually a python venv to activate i think

[lyric zealot]

https://tenor.com/view/dog-doggo-pup-puppie-dig-gif-17364765

let me find it real quick

how do you link your ESRGAN stuff

[supple raptor]

my setup for some of that stuff is a little funky since i use a standalone comfy (non-portable) install but i think you should be able to use the extra_model_paths.yaml inside the comfy backend to do the same as me

mine is like this

[lyric zealot]

i will just copy and paste that since i already linked the others

[supple raptor]

That also works since those models are small anyways

[lyric zealot]

do i have to activate the autocomplete thingy?

[supple raptor]

yeah, User > User Settings > AutoComplete. may also want to change your image format to PNG while you're in there.

here's my auto complete config

[lyric zealot]

i totally missed that page

[supple raptor]

it's easy to miss

[lyric zealot]

got to download the danbooru csv

[supple raptor]

i think it should be included iirc

if not, here it is

[lyric zealot]

[lyric zealot]

breasts

[supple raptor]

idk where these came from if they aren't included by default. maybe i stole them from the A1111 extension

[lyric zealot]

[supple raptor]

i think i used the e621 once when i was messing around with pony, but i don't think i've ever used any of the other ones

[lyric zealot]

what does this even mean

[supple raptor]

comfy didn't find the model. looks like you'll need to do the extra_model_paths.yaml thing anyways

[lyric zealot]

[gloomy scroll]

Yue has dress like that also

[steady grail]

i hope i can compile xformer without installing another venv and ptorch

i don't have much space left on the pc at the moment

[modest fog]

steals dress

[modest fog]

Just delete those millions of images and models

[steady grail]

still need to experiment on them and i have installed a game that weights 180gb

i have something like 80gb of lora

maybe i should eliminate something of them

[modest fog]

[steady grail]

i must collect cute things

powershell is doing something

fingers crossed

so many errors

[modest fog]

She will help

@green temple smol help

[steady grail]

stable-diffusion-webui if i'm not wrong is the name of auto1111 directory

how dear

lol

yup, gotya

[lyric zealot]

[steady grail]

i think i did it

xformer loaded

let's try genning something

xformer loaded

model loaded, clip loaded, lora loaded

loading complete, genning

error...

xFormers wasn't build with CUDA support...

this package of xformer is probably not compatible with this version of pytorch...

😭

[supple raptor]

theSaddest

[lyric zealot]

my swamui installation is so cursed, im thinking of giving up

[steady grail]

what an unlucky day

we should blame the developers of stable diffusion

🤣

[supple raptor]

maybe just delete the comfy backend stuff and git clone comfy separately then add it as a backend to swarm. that's how mine is set up and It Just Works™

[lyric zealot]

where are you supposed to store your yolov8 models?

[green temple]

smol help = big help

[supple raptor]

.\StableSwarmUI\models\yolov8

[lyric zealot]

my swarmui is definitely fucked. It's there, it appear in autocomplete, but there's some error saying comfy cannot find it

[supple raptor]

let me check if i have them somewhere in my comfy stuff too

[lyric zealot]

might work now

new error

[supple raptor]

wait, do you mean the same error you had before? or is it a similar but different error

[lyric zealot]

[supple raptor]

oh what error now

[lyric zealot]

it was same error, but now there's some other error, but i guess it's because i didn't specify what to detect

[supple raptor]

are you using <segment:thing> or <segment:yolo-yolomodelfilename>?

[lyric zealot]

the first one works

the second one didnt

[supple raptor]

ok, which yolo model are you trying?

[lyric zealot]

the same as you

[supple raptor]

thinkeroo that shouldn't need any special inputs

[lyric zealot]

it works, but that's using <segment:face,0.1,0.1>

[supple raptor]

I use this to specify 0.4 denoising and 0.8 confidence
<segment:yolo-Anzhc_Face_seg_1024_v2_y8n.pt,0.4,0.8>

[lyric zealot]

same error

[supple raptor]

go to the server tab, then logs and pick ComfyUI in the dropdown. might be like ComfyUI-0 or something like that. that will show you the full console from comfy

[lyric zealot]

the error is so cursed

[supple raptor]

if all else fails, the official SwarmUI server is helpful. i'll send you an invite since invite links can't be posted here

[lyric zealot]

[steady grail]

ok, i reinstall the proper torch, how do reinstall torchvision, looks like torchaaudio don't get disinstalled and gives conflicts

[supple raptor]

both commands should work on both windows and linux
latest stable with cu124: pip3 install -U torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
lastest nightly with cu124: pip3 install --pre -U torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu124

[steady grail]

looks like nightly is too updated to work with xformers

no, it works

i was wrong

let's try it

if i didn't mess up all the requirements, now i should have xform...

error

[supple raptor]

sometimes the nightly wheels will work for xformers. just depends on whether the xformers wheel has been built for it or not. i'd suspect there aren't xformers wheels for nightly pytorch yet since they only came out for the latest stable pytorch yesterday

[steady grail]

yep

[steady grail]

i think my vram became too obsolete for the new xformers

[rough oxide]

always wondered but do you need to manually update the nightly build like every time it updates if you wanna keep it up to date? if so is there a way to do that without reinstall

[supple raptor]

yeah, you have to update it, same as if you were on the stable release. you could add an update to your startup script, but that will slow down launching a bit since checking for the update takes some time.

[rough oxide]

geh need to do force reinstall to do that i think. well... can probly do it once a month or something rather than a regular thing xD

[supple raptor]

the real problem is if you want to use xformers, you also need to know which version of pytorch the latest xformers wheels are compatible with since there is often a significant amount of time between torch being updated and xformers catching up.

[rough oxide]

i've just been using python's own cross attention thing.

xformers very slightly affects quality of image iirc

[supple raptor]

xformers is more memory efficient than sdp and is also deterministic. if you want sdp to be deterministic you have to disable the memory efficient attention which makes the gap in memory efficiency even larger.

that being said if you aren't running xformers and want torch updated all the time, then you can just add one of the following to your startup script right after the venv gets activated
latest stable:
pip3 install -U torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
latest nightly:
pip3 install --pre -U torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu124
replace the cu124 with whichever cuda version you use if your card doesn't support cuda 12.4

[bitter latch]

art!

[modest fog]

This is @green temple

[supple raptor]

ono! she fell down. help her up!

marisad

[modest fog]

No

[bitter latch]

hmmm interesting

[bitter latch]

haha

what?

[wooden sail]

Is that SD3.5?

[bitter latch]

no

[rough oxide]

thanks for that, i'll save it and use it when i need to update. (for some reason couldn't find any command for easily updating thru google when I checked)

[modest fog]

make it a dark magician girl

can that model do it

👀

[near elbow]

"anime dark magician elf girl eating a donut"

[zinc flax]

You're asking too much with flux lol

[modest fog]

sus

haha

Its the extra episode one

its just like her!!

[zinc flax]

Huge

[brittle rampart]

I feel sorry for the bird... she's hot!

[supple raptor]

Here's a site you can just bookmark. It covers other system configs. Don't forget to activate your venv first otherwise you'll be installing it to your system python and it won't be used by webui https://pytorch.org/get-started/locally/

[rough oxide]

hmmm. i just use the one system python & comfyui, no venv ;_;

spent a while on the image model voting thing, pretty sure i like 3.5l the most aesthetically right now, but it feels like some others might follow the prompt a bit better sometimes

[supple raptor]

If you have everything just in system python anyways then no need to activate venv

[rough oxide]

aye, usually just have a couple things setup at a time that don't really conflict with eachother so haven't needed venv yet

[supple raptor]

I like venv because if something gets screwed up, I can just shift+del the venv and create a new one.

[vital raptor]

i just put everything in it's own VM 😄

i think pretty much everything i've tried except comfy comes in a venv or conda env anyway, not sure why comfy doesn't

[supple raptor]

Comfy expects that you know what you're doing enough to decide what sort of environment you want to use if you opt to not use the portable version which uses its own included python anyways (just like a1111's portable version)

[vital raptor]

maybe, but unfortunately the vast majority of users don't know what they're doing and just run into trouble because of the way it's done, and they don't even have install instructions available for if you did want a venv/conda setup. I imagine it probably causes more support issues than it's worth

doesn't really matter to me personally, I can figure it out just fine, but I'm not a typical user

[gloomy scroll]

https://tenor.com/view/special-ralph-the-simpsons-football-gif-5491308

[vital raptor]

[supple raptor]

Yeah, I don't particularly agree with the choice to not include instructions for it, but that's how it is. They really just want people to use the portable .zip instead of cloning and managing their own venv, but I find that creates even more problems because dependencies are more difficult to fix when a problem comes up.

[rough oxide]

pretty soon there'll just be a standalone program it looks like, will make it a lot easier to recommend to less technical friends who have good hardware

[supple raptor]

yeah just saw that too. wonder how that will effect SwarmUI's install process since that just uses the portable .zip version of comfy by default which they said will eventually be replaced by the standalone electron app

[vital raptor]

I'd imagine they'd still have some kind of headless API mode since so many things use that.. but you never know

unless the electron app is only replacing the 'portable' install, then swarm could just pull the normal version, complicates the install on swarms side a little, but shouldn't be too bad

[supple raptor]

yeah, the app will only be replacing portable. here's the blog post https://blog.comfy.org/comfyui-v1-release/

[near elbow]

she's sceptical

[modest fog]

renoodlefied

[supple raptor]

for the way i use my system, a standalone app would be less convenient. i like to have my ui in a browser tab because it allows me to more easily switch between other related things, such as looking through danbooru to get new prompt inspiration, or just looking up tags or checking civ/hf/github for new things

[modest fog]

eepy

[vital raptor]

same, i don't even run them on the computer i'm generally sitting at

[near elbow]

yeah, it's just so inconsistent and hard to combine and all... i'm happy with forge (i.e. enhanced automatic1111). tried comfy when flux came out and it was the only way to run it, but once forge supported it i was super happy to escape the spaghetti 🙂

[vital raptor]

i just use swarmui for the UI, still has a comfy backend if i ever need it

[near elbow]

yeah, rendering rig with forge here and browser based access (just in my intranet of course)

[supple raptor]

i go back and forth between Swarm and Comfy mostly. if i know that i just want to connect noodles when i launch it, i'll just launch comfy standalone, but if i might want swarm features i launch that instead. Forge is only used when I'll be generating remotely on my phone

[modest fog]

swarm 🐝

[near elbow]

text with flux fusion is fun 😉

[modest fog]

@green temple bee

[modest fog]

swats

[near elbow]

i don't advise to try that - she has quite a sting 🙂

[modest fog]

make it smug and :3

[vital raptor]

technically that is smug, but this outfit is trying to lewd everything

[vital raptor]

better if i force it to be a dress

[modest fog]

how dare this AI thing tries to make lewd stuff

so horrible!

big bees 👍

[brittle rampart]

We urgently need to find a recipe for repelling bees!

[vital raptor]

What do you call a ghost in a bee costume?

||A boo bee 🐝||

[green temple]

ngl this is me

[modest fog]

Lazy

[supple raptor]

you called?

[chrome moss]

hi everyone, I have a problem when upscaling the image with tiled diffusion and tiled vae with sdxl model is that the lines appear prominently on the image according to the tiles, thanks.

Is there any way to solve this problem?

[supple raptor]

lower denoising is the main solution

[bitter latch]

same prompt clip skip to 12 😄

[bitter latch]

ooh

[keen hollow]

I'm with you on this one.

So comfy is changing to an app? Did i read that right?

Oh. Closed beta.

[steady grail]

[chrome moss]

I use denoise in range 0.1->0.2, but the result is still the same. Could you show me your tiled diffusion and VAE settings? I think the main problem lies there.

[supple raptor]

i only have an old sd1.5 image handy at the moment since pretty much all of my sdxl stuff is done in SwarmUI/Comfy and i don't often find it necessary to upscale past my 1.5x highres fix pass.

[chrome moss]

ok, thank you, btw, the insancely detailed image.

[steady grail]

gotta go, see ya

[modest fog]

Worst elf in existence

Too much forehead, eyebrows and hair look stupid

[supple raptor]

https://tenor.com/view/sinner-cat-kitty-cute-kitten-gif-17576170

[modest fog]

i speek truf

[supple raptor]

https://tenor.com/view/frieren-sousou-no-frieren-anime-warm-campfire-gif-16192143661019154751

[modest fog]

7th image

@rough oxide are you still using imagen

[rough oxide]

stable diffusion 3.5 large

which is why if hands ever appear they're usually bad

these hands r fine tho

and by these hands

i mean excluding the first 2

[brittle rampart]

OH YES! We can definitely handle this... 😁

[near elbow]

mecha demon fun times

[unique mortar]

it is crazy.........only make a pic for 2134x1440, it used up the 8GB of Vram on card and also 16GB share Vram on system ram......Almost all the available video ram....

[bitter latch]

haha

ok time to go listen to peple bitch

[near elbow]

doesn't go higher than 2048 here, so this is 2048x1440 just for fun 🙂

Steps: 8, Sampler: Euler, Schedule type: Simple, CFG scale: 1, Seed: 234841387, Size: 2048x1440, Model hash: 126324b1fc, Model: fluxFusionV24StepsGGUFNF4_V2GGUFQ80, Version: f2.0.1v1.10.1-previous-583-g0c97a534, Module 1: clip_l, Module 2: ae, Module 3: t5xxl_fp8_e4m3fn
Time taken: 44.7 sec.
A: 18.71 GB, R: 19.97 GB, Sys: 22.1/23.9883 GB (92.1%)```

same prompt in 2048x896 takes only 16s, no swapping to real RAM needed i guess

[supple raptor]

something strange is going on here. you have the resolution set to 960x720 in the screenshot, but hires fix is recognizing it as 1067x720. also, that generation should probably take more than 8GB of vram, but certainly not 14.5GB more. i suspect there are multiple bugs happening at the same time here that all contribute to much higher than expected vram usage.

[supple raptor]

for reference these screenshots are using a a slightly larger size since 1067 isn't divisible by 8 so isn't a valid size and using an SDXL model which takes more vram to run than sd1.5 models. it was only about 8.7GB vram during the highres pass at 2144x1440 and peaked at 10.5GB during the VAE decoding at the end

[bitter latch]

hmmms

[rough oxide]

whai are sword textures so fun to play around with now xD

[green temple]

jean ?

[bitter latch]

just girl

😄

[green temple]

ok

[keen hollow]

wow I'm having a heck of a time re-creating images

I think Forge has changed a lot of things in how it processes over time

[supple raptor]

yeah, it's basically completely different. switched from A1111 inference code back when it was forked to what is most likely (hopefully) just stolen Comfy code that's been obfuscated

[keen hollow]

Why "Hopefully"?

[supple raptor]

because trying to hide that he stole the code from Comfy without giving credit is one of the few explanations for why it would be obfuscated that isn't "because it's hiding something malicious"

[keen hollow]

Yikes - that's in Auto now?

[supple raptor]

auto is still auto. it's still kind of cursed code, but it's not obfuscated and afaik is generally considered safe (extensions are another matter)

[keen hollow]

Oh... so Forge is the one with obfuscated code?

Well crap.

[supple raptor]

yeah

[keen hollow]

That's my primary guy.

This makes me sad.

[supple raptor]

I've heard mostly good things about ReForge, which i'm pretty sure was forked prior to the obfuscation

[keen hollow]

😐

I was thinking ReForge was behind again.

Maybe the time has come to fully just switch to comfy.... ugh

eh nvm.... no one cares... 😛

[bitter latch]

turns off dtg

[supple raptor]

Invoke has become somewhat interesting. I played around with it a little bit. For in/outpainting it's quite nice. Documentation is seriously lacking though and from what I remember of their old documentation they used unique prompting syntax. Prompts also seem to maybe be limited to 75 tokens, though not entirely sure on that one.

[keen hollow]

comfy has a pretty big overhaul coming.. might be the right time once that is out of beta

building workflows for even basic features is such a royal pain, though. 😐

but even worse is trying to FIND an existing one that doesn't suck

the workflow repositories are all filled with garbage

[supple raptor]

i wouldn't really call it much of an overhaul. it's basically just an electron wrapper for the existing browser based version. it sounds like it might get a few unique features such as multiple tabs to accommodate having multiple workflows loaded, but you can already do that by simply opening another browser tab for comfy. the main benefit of it will be ease of installation.

[keen hollow]

I thought they were baking in the manager functions and such.

[keen hollow]

Plus some additional security/vetting procedures.

[supple raptor]

the manager functions amount to just having the Comfy Manager extension pre-installed if i understood correctly and i think the security/vetting is already in place in Comfy Manager.

[keen hollow]

What are you using mostly these days?

[supple raptor]

mostly Swarm. occasionally i'll load standalone comfy if i know i don't plan to use any of the swarm features for a while

[keen hollow]

which swarm features do you like?

[bitter latch]

hmm?

[rough oxide]

swarm's kinda just comfy upgraded

[keen hollow]

Yeah it puts a UI on top that's decent. It's tricky for me as it requires some funky workarounds in the sandboxed environment I use.

[supple raptor]

the image history and model/lora/embeddings browsers are miles ahead of the ones in A1111/Forge, autocomplete for special prompt features is great (example: adetailer is done by simply adding <segment:yolo-Anzhc_Face_seg_1024_v2_y8n.pt,0.4,0.8> to the end of your prompt. it gives an gives suggestions for the feature as you start typing and then gives an explanation of how to use it within the tooltip like in the screenshots), inpainting is a much better experience than base comfy, same with ip-adapter and controlnets

[rough oxide]

oh right, swarmUI is my absolute favorite way to load loras

[keen hollow]

ok... you've convinced me to try it again

Does it still have the dotnet dependencies?

[rough oxide]

yes

dotnet 8

[keen hollow]

those were a little tricky for the sandbox.. will have to figure out how I did it before 😛

[supple raptor]

yep, literally just pick your lora from the browser which even has a subfolder tree on the left and search filter. and it appears next to the model, ready to change the strength if needed. or, if you like A1111's method you can even do <lora:... in the prompt instead

[keen hollow]

I made a mistake of downloading Loras and not organizing them well at all

Does Swarm have any sort of help for that?

[bitter latch]

😄

[keen hollow]

I know that auto/forge had that Civit Browser thing. it was a little janky.

[supple raptor]

the only feature that swarm is really missing for me is a good mobile layout

[supple raptor]

not for automating sorting them. as you can see in my screenshot you can organize them yourself though and it doesn't become a nightmare if you have more than like 5 total subfolders like Forge/A1111 do

[supple raptor]

it does have a model downloader tool built in though so you can download stuff from a Civitai or HF link (mostly useful if you're genning from a different machine than it's actually running on)

[rough oxide]

lol i just realized is anime the last dreamer community remaining? haven't checked channels & roles list in a while

[supple raptor]

yeah, i think it's been the only dreamer community for quite a while

Agony

[bitter latch]

thats what I wanted... kinda

[supple raptor]

the other ones each had like 5 people who would maybe post an image per day for the most part. some of them were completely dead for days or weeks at a time

[rough oxide]

makes sense. there is anime and there is non-anime.

[supple raptor]

there is anime and there is nothing else

[rough oxide]

i think i only ever used anime, the current SD one like sdxl when it was new

[supple raptor]

Patchy_smug

[rough oxide]

and the bot channels when they were testing new models

[supple raptor]

i posted a couple of images over the years in other channels, but even most of those were anime that just happened to cross-over into other categories

[rough oxide]

almost fax. but i do like how good sd3.5 can do realism, made me want to do more realism stuff even though i totally avoided it before

see? great realism

XD

(....i still probably inject anime into 90% of the realism prompts)

[bitter latch]

hmm ok... I guess princess peach is on the train/tram?

[rough oxide]

i often wonder, if an image almost always has to try and conform itself thru steps to match what the prompt is... then... how do you end up with results like that unless it identified "something" as princess peach

[rough oxide]

cutee

reminds me of

rilliona from yugioh

whai is it so hard to find sfw artwork examples of rilliona ;_;

[keen hollow]

I kinda miss when the image sharing was more active.

Not kinda... I do. I do miss it.

[bitter latch]

what am i? chopped mutton?

well don't everyone say no at once 😛

[keen hollow]

chopped mutton... nice

Who wants Domino's tonight?

[rough oxide]

dominos makes good pineapple pizza

[rough oxide]

you mean here or elsewhere? it's still pretty active here which is part of why i keep coming back

[keen hollow]

This full server.

It's decently active comparatively, but not like it used to be.

[rough oxide]

i think a lot of ppl splintered off into the more specialized servers

[bitter latch]

it got confused

groans

https://tenor.com/2zW4.gif

[rough oxide]

i want fruit cake now.

adds to cart

[bitter latch]

now quadruple the fruit

I just had a whole log of that.. ooh I dink I fuxed up

[bitter latch]

interesting way to sit.. couch

ooh I don't think my blood suger likes me right now

https://tenor.com/bNVoY.gif

[green temple]

indeed

[keen hollow]

Like for the individual UIs?

[rough oxide]

nay

like there's sort of a main anime AI server, and there's lots of other servers for specific services like novelai & such

[keen hollow]

oh I see

are they any good?

Is speed (it/s) viewable in Swarm?

[rough oxide]

ye, the main anime server is a great place.

did swarm not show anything when you generate on the commandline?

[keen hollow]

show's me time when it's done...

ignore the speed, settings were off... but that's what I see.

nothing for current generation speed

[rough oxide]

ah yea that's usually wrong too iirc. guess maybe it doesn't then

since that'd be the place to output it normally

well it was wrong when i was doing batching stuff anyways

been back to using just regular comfyui for a while myself

[modest fog]

Big

[green temple]

noooo

[modest fog]

Small

@green temple have small agent ready to arrest you for liking small too much

runs away

@green temple Minecraft

[bitter latch]

sigh, time to work.. blah

[green temple]

herobrine 😱

[bitter latch]

gotta love my split shift tho 😄

[modest fog]

no scary stuff allowed

[green temple]

do minecraft alex shifting

[modest fog]

@green temple . o ( )

[modest fog]

I don't know what that is D:

[modest fog]

[modest fog]

I tried it on prompt but I got something so big I can't even post here

I don't think nai knows who it is

[supple raptor]

idk if it/s is visible somewhere else, but you can see it in Server tab > Logs tab > ComfyUI in the dropdown. this will show the full comfy console so if you have any errors that don't give a clear message in swarm, you can check there too

[vital raptor]

yeah, in the Server -> Logs -> ComfyUI-0 it shows the it/s

the swarmui console shows the total time

[finite herald]

how has no one done this with miku yet....

[vital raptor]

at the bottom of swarm by the tabs it shows estimated gen time, but it's using previous times and averages, its usually wrong

[vital raptor]

because i'm not good enough to be seen that close next to miku

[finite herald]

awareness is the first step

[vital raptor]

[rough oxide]

i keep getting miku on a screen xD

that was only one that came out like that

i probly need to redo my prompt

why yeet message? owo

[bitter latch]

realized I was being rude

[rough oxide]

it's fine lmao

[bitter latch]

https://tenor.com/UlKv.gif

[supple raptor]

@finite herald miku ate too much candy

[keen hollow]

Nice, I'll check that out.

[bitter latch]

mmm candy

[keen hollow]

evenin'

Workin' on acclimating to Swarm. I will say the setup went smoother this time. Seems to be working without much issue. 🙂

[rough oxide]

aye, should just be dotnet and git. i think it uses a venv for python right?

then everything else installs as it needs

[keen hollow]

I use some sandboxing which makes it a little weird, and I use a bunch of custom folders for model storage and such

but yeah it did ok

Now I need to figure out some quality-of-life stuff, like where to store my favorite negatives. 😛

[rough oxide]

hmmm. i only find sandboxing useful if the program is malicious/can be exploited/etc... safetensors & common sense is pretty good as far as SD goes in general

[keen hollow]

comfy with the plugin malware issue and forge with the obfuscated code begs to differ 😄

[rough oxide]

....guess i dodged a bullet not using comfyui manager til now (and still not yet...) xD

fair then. didn't know there were malware problems with plugins

any specific names to avoid?

[keen hollow]

there was one in particular but it's long since been eradicated 😛

it's a lot of self-updated stuff, git pulls i'm sure you aren't reviewing for updates and such... I dunno. I just feel better sandboxing it off.

[rough oxide]

that's good

i review all my updates ;D

swarmui auto updates tho

[keen hollow]

you do code review of every commit?

[rough oxide]

yup

there's not a lot of code to review

[keen hollow]

right on

[rough oxide]

i check every update and only pull when i see something relevant to what i need

but up to that point i've reviewed all the code

some of it goes over my head though, especially the mathematical stuff to do with samplers & what not

i figure if something happened there then there'd be a loud outcry though xD

[keen hollow]

eventually, yep

Is there a built-in "png info / metadata viewer" in this?

I have my program I can use, but nice to have a built-in also.

seems like you can just drag an image in.. interesting

[rough oxide]

u can on the comfyui workflow side

then send the workflow to the gen side

[keen hollow]

I gotta figure out this segment thing

so I should be able to add a segment token to a generation to hone-in on a particular part

but adding it to the prompt is going to modify the full result. hmm...

I guess you tag it on to the end and it'll be ok

yeah that does it 😄

gonna try their CLIP Seg thing over the adetailer yolo models for a while, see how it goes

wow you can re-render JUST the segment

that's pretty nice

[supple raptor]

it just uses whatever comfy is using for python and all the deps, so if you let it install the comfy backend it uses the bundled python and that comes with the .zip