🍥｜anime

[trim current]

should be fine

one of my ssds arent mounted cos im lazy

my lil kingston

boot drive

[vital raptor]

yeah, loose ssd is fine as long as it's not hitting anything else

[summer night]

i like to store important data on my hard drive because hard drives are easier to recover and less prone to sudden catastrophic failure
i need to set aside some money to get a large SSD to backup my data

[vital raptor]

easier to recover until the read head takes a dive into the platter

also more likely to fail to begin with

and slow af

[summer night]

i've already replaced my hard drive once and the heads up the clicking sound gave me provided enough time to get a replacement and transfer over everything

if an SSD dies, it dies suddenly

[vital raptor]

there's a few HDD in my server, but they're painfully slow when data gets pulled from them so I used 2 1TB mirrored ssd for a cache, and my AI stuff is on a 2TB nvme cache

having mirrors and/or raid setups is nice, I can lose a few drives and no data will be lost, just swap them out and let the array rebuild

expensive to get setup at first though

[summer night]

right now the only drive i have a backup of is my boot disk

[vital raptor]

ah yes, the least important one 😄

[summer night]

and that was because i had to do MBR2GPT

[summer night]

i don't have enough space anywhere for backups of the other disks

can't afford anything right now after buying a 4070 super

plus i'm saving for a new pc

[vital raptor]

you have over 3TB of space.. it shouldn't be too hard to free up a chunk of that if needed

unless it's all ripped movies or something

I'm only using maybe 2-3TB of my 12TB NAS, and most of that is AI and backups of other systems, also the ~12 VMs I'm running

eventually i'll get around to ripping my DVD/bluray collection

[summer night]

most of it is games
my download speed is 20mbps so its a hassle to download a game if i want to play it on a whim, so i just leave games installed and uninstall them if i need to make space for something else

[vital raptor]

yeah, games take so much space these days, i just delete them when i'm not playing them

but, I have fast internet, so I get keeping them around

having slow internet sucks

[summer night]

treesize is a useful program for finding out what is taking up space

my downloads folder is so large because that's basically where i store misc files like most programs and stuff

music folder is 30 GB because i rip all my CDs to itunes

[modest fog]

On my PC and phone I have a 100~ songs playlist from 10 years ago and that's enough x3

[summer night]

i don't have any apple products other than an old ipod touch, itunes is just a fairly cromulent music manager

it keeps pestering me to sign in to apple music though

[supple raptor]

in that case, noob is the superior model as it was trained on a much larger and less aesthetically biased dataset than pony and it also knows a far wider selection of characters, concepts, artists, and tags in general than pony

[dusky linden]

Noob and other illustrious spin offs are goated

[modest fog]

Can it make girl drinking water

From a cup

[vital raptor]

except i think characters are needless data that could have gone into more meaningful concepts or details elsewhere

sure, if you want X character, it maybe superior, but I don't gen existing characters, so its meaningless to me

not saying pony is better, but I haven't switched over because I've already got a workflow that works and illustrious/noobai don't really have anything I want

[dusky linden]

staw or actual gass

glass

[modest fog]

Glass

Same thing

[dusky linden]

dang

[modest fog]

[dusky linden]

sipp

[modest fog]

That's a fairly long straw

I want a long straw too

[dusky linden]

suction too strong

[vital raptor]

straws are easy

[dusky linden]

YOU GOTTA LIFT THE STAW by a LITTLE

not on the BOTTOm

you make a

VACUUM

[modest fog]

Have you ever had one of these

[vital raptor]

yes

[modest fog]

They look so weird

[dusky linden]

i add straw to negatives

man

the more i try to make it drink from the cup it self the more NSFW it gets

man

[vital raptor]

most models won't do drinking from the glass itself because it obscures the face and no one bothers to train stuff like that, pretty much every model should be able to do a straw though

[dusky linden]

i mean theres one that does it

but its forced beer drinking

closest your gon get

[modest fog]

I can't remember the last time I drank water from a cup

[dusky linden]

yeah but its tea

[modest fog]

I always just have a bottle around

Lmao

[vital raptor]

yeah i drink everything from a 40oz water bottle 😄

[dusky linden]

i just carry those water jugs around

[modest fog]

I'm afraid to prompt for water jugs

Ratsplosion

[dusky linden]

pretty mild

[summer night]

eh maybe noobai has always been messy @wispy canopy
this is with hires fix and rescale cfg

[wispy canopy]

yeah, i've been suspecting that too

[dusky linden]

NoobAI needs style loras no

[summer night]

i might try generating a bunch of images on wai nsfw illustrious with my preferred style mix and then training a lora on it to use on noobai

[supple raptor]

That's not how this works at all. Training extra characters doesn't mean that it didn't learn some other concepts because illustrious and noob scale loss based on tag frequency. This is what allows for tags with as few as 40 uses in the dataset to work even within the same prompt as a character like Miku. Pony didn't even use tag dropout yet alone tag frequency based loss scaling.

Noob knows an absurd number of characters explicitly because of its massive dataset which also translates into it knowing just about every other concept in that massive dataset.

[summer night]

is there a specific reason why noobai has muddy details?

maybe it was overtrained

because you also have to use some kind of dynamic cfg in order to get the image from being way too oversaturated

[dusky linden]

isnt it the same case with base ponyV6?

[supple raptor]

It's definitely not overtrained. The TE training they did is questionable though

Also, I've never used any form of dynamic cfg or anything similar with any of the noob models.

[wispy canopy]

still, i have the same issue -> can't get output stable 😦

[summer night]

using a significantly lower base CFG seems to help
left image is euler a, right image is euler

[supple raptor]

Just takes some experimenting with artist combinations to figure out what you like with the model. Also, if you're not using almost entirely tags for your prompt then you're simply not likely to get good results because that's what it was trained on and it wasn't overfitted on an aesthetic dataset to force it to output something pleasing even when it's wrong.

[supple raptor]

Since 0.6 for eps and 0.5 for vpred versions you want 4 CFG with Euler or Euler_A. Before those versions, 5 CFG was usually best. If using CFG++ samplers then you want closer to like 1.25-1.5cfg but I rarely use those.

[summer night]

huh, they should update the model page with that information

i've been trying to get it looking good with a base cfg of 5 this whole time

[supple raptor]

Never trust model pages for generation parameters. Almost none of the people who are interested in large scale training do much actual generation.

If they aren't just settings they copy/pasted from someone else, then they're usually just something that seemed good enough from some test generations.

Always best to run some grids of your own and figure out what actually looks best.

[summer night]

i'm still going to try generating a bunch of images with my preferred mix of style tags on wai-nsfw-illustrious and then train a style lora on them and see if i get a better result on noobai than just using the style tags the same way

[supple raptor]

LoRAs can certainly help with that sort of thing as you're basically doing a small aesthetic finetune

[summer night]

this is without rescale cfg, it looks oversaturated to me

Steps: 30, Sampler: Euler, Schedule type: Automatic, CFG scale: 4, Seed: 2923251954, Size: 832x1248, Model hash: ea349eeae8, Model: noobaiXLNAIXL_vPred10Version, Clip skip: 2

[supple raptor]

cattower is another noob variant you might want to check out. it's very popular and I've seen at least 2 blind model rankings that it's won in

[summer night]

this is the official art of the character for comparison

[summer night]

i'll check it out

[supple raptor]

1. Don't use clip skip with SDXL models. CLIP's penultimate layer is skipped by default in XL which is what CLIP skip 2 was used for. Ideally setting it to 2 does nothing if the UI is well made, worst case you're skipping an extra layer that you shouldn't be.
2. Euler A is consistently better in my own tests on noob vpred models. For scheduler, I prefer AYS/Align Your Steps, which needs specifically 28 steps to work right.

[summer night]

scheduler is new to me since i'm just now upgrading from the old version of forge to reforge

[summer night]

so should this be enabled or disabled then?

or does clip skip 2 do the same thing

[supple raptor]

I think the wording on that is bad. I'm pretty sure unchecked means it ignores the clip skip setting for both clips (default SDXL behavior, meaning only penultimate layer is skipped regardless of what you set). Checked would mean that the clip skip setting is actually used. You can test by locking all parameters and seed, then try to re-generate an image with clip skip set to 3 or 4. If it's the same as the image with it at 2 then it's being ignored

[wispy canopy]

@supple raptor thx for the idea of rescale -> that stabilized it for me

[summer night]

i tried switching between clip skip 1 and 2 earlier and the outputs were the exact same

[supple raptor]

I think that was @summer night, I don't rescale my cfg

[wispy canopy]

well, rescaling was the trick anyway, i wouldn't have thought of it otherwise

[summer night]

what setting are you using

[wispy canopy]

0.7

but with 3 pag, and 3 cfg

[summer night]

what is pag

[supple raptor]

Cool then yeah it's being ignored as it should be and you can ignore that comment

[wispy canopy]

perturbed attention guidance

pretty sure i can optimize the schedulers tho

[summer night]

the default is 0.7, the model page says 0.2, but i found that 0.5 produced the least messiest and oversaturated result for me

to me it's hard to optimize settings because you can spend a long time trying to get the "best" result on a single seed, and then as soon as you generate with a different seed, the new image looks bad

[summer night]

there's also cases where it can be hard to tell if one option is objectively better than the other

[wispy canopy]

lots of testing

[summer night]

the most in-depth testing i've done was comparing almost every sampler and arriving at euler a which i had already been using
i did a tournament bracket style comparison where i compared 2 against each other and the better one moved on

[supple raptor]

Yeah, don't test a single seed or prompt. When you're trying to dial in settings you want to test using grids each with a different seed and multiple prompts for each adjustment.

[wispy canopy]

and you should also put in schedulers and a few other things in that comparison... xD

[summer night]

i didn't even have the option of changing schedulers independently of samplers because i was still on the old version of forge

[wispy canopy]

one of the advantages of comfy

[summer night]

are outfits like bunny suits allowed here?

[wispy canopy]

like, there's quite a few things going here xD

[summer night]

[supple raptor]

Depends on context. If they expose too much or the character appears young it's an easy bonk

[summer night]

i guess not then because of the nature of a bunnysuit, which are basically designed to show cleavage
small breasts -> character looks too young
large breasts -> too much cleavage

[wispy canopy]

except for the weird staff being missing at some point, decent image 🙂

[summer night]

i find this happens a lot with poles

if it becomes obscured by anything, it can easily misalign, or skip a section if it's obscured twice

[summer night]

also why does the filename have upscaled_lewd in it

[wispy canopy]

it's the lewd setup 😄 i started making the flow for lewds, but just works just as well for any other anime

[summer night]

when you make workflows in comfy, can you have it set up to where you can generate an image, and press a button to confirm if you want it to move on to the next step like upscaling?

the thing i like about a1111 based UIs is that you can generate a batch of images at base resolution and decide which ones you want to upscale

this handy button

and then after upscaling i can send it to img2img to use adetailer *if it needs it

[wispy canopy]

i do it like that -> i'd have to drag the image i want to upscale back in -> and then enable the upscalers

[summer night]

also can you make the lines follow straight axes and make 90 degree turns?

[wispy canopy]

yes, that's possible

and there's also get-set node

so you can make the connections "bridged"

i should probably do that too

still, getting a lot crisper output from wai tho

if not a bit more boring tho

[summer night]

i would be bothered by the inconsistently sized houses

[summer night]

basically making comfyui look like these nicely managed perfboards and breadboards

[trim current]

use reroute nodes

[desert island]

yes

[summer night]

idk if i'm really going to use comfy
i have it installed, but i might use swarmui instead

iirc swarmui has both a simple ui and the node ui

i'm waiting for it to become more developed though

[dusky linden]

Thats why we use reverse bunnysuits :^)

No cleaveage issues then

[desert island]

that's what I've been told... I haven't moved over to swarm cause I haven't needed to yet but I think I might jump over soon.

[dusky linden]

Swarm is great, i wfh today so i can show you the ui n all if you want

First walking the dog tho

[desert island]

nope, not ready yet... I'm in the middle of too many different things and I can't lose focus 😄 , otherwise I'll.... ohhhh what is that... mmmmm shiny....

[supple raptor]

Not really much to wait for with swarm. Only thing it's missing is a mobile layout.

[dusky linden]

God i wish

Tablet works better tho

[summer night]

cat tower seems to produce a bit of a better result than base noobai but it still probably needs a style lora

Steps: 28, Sampler: Euler a, Schedule type: Align Your Steps, CFG scale: 5, Seed: 2215415859, Size: 832x1248, Model hash: 107552c070, Model: catTowerNoobaiXL_v15Vpred, Denoising strength: 0.3, Clip skip: 2, ENSD: 31337, Hires CFG Scale: 5, Hires schedule type: Align Your Steps, Hires upscale: 1.5, Hires steps: 20, Hires upscaler: 4x-UniScaleV2_Sharp

oh damn i forgot to turn the cfg back down to 4

alright it didn't make that much of a difference regenerating the image at 4 cfg

[dusky linden]

What about dpm ++ 2m?

[supple raptor]

You might try other samplers/schedulers with that one. I haven't used it personally so idk if it does better with some other combo

[summer night]

huh i'm looking at images on civitai that are using it at a cfg of 7
i'm gonna try that

[dusky linden]

Sometimes i get better with 2m karras, sometimes not

Less tags mean less cfg

[green temple]

try noobvpencilxl is good too

[summer night]

having a hard time deciding which of these is better

[desert island]

v80 looks like it might be more consistent strangely... Do another batch. 🤔

[summer night]

i'm going to try generating images of sinon and i'll go with whichever model consistently generates her side hair clips correctly (with 3 holes)

why do ai models love giving her hair clips 4 holes so much
her hair clips should be one of the easiest concepts in her design

https://tenor.com/view/sword-art-online-slap-anime-kirito-gif-23549745

[desert island]

well I guess you're going with 90 then cause one of them has three

Going to bed I've got the ground work setup to accept LLM generated strings for specific portions of the prompt for the manga generator project. You can toggle it in the settings to either use the User Generated strings for characters, actions, setting/scene, camera, and lighting, or have the AI fill that out once I implement it tomorrow.

Tomorrow I think I'll need to enter in an easy way for users to alter the rules for specific scenarios... like if the camera is this use these rules, or if the setting is sci-fi, for example, use these rules etc. Then determining the amount of panels needed for a particular scene might also be tricky for it... lots to consider but 🤷‍♂️ that's a tomorrow problem.

Night all 😄

[green temple]

brain not braining

[vital raptor]

I wasn't talking about how well or not the models know lesser known characters or how much better it was trained than pony, I'm also not arguing that either one is better than the other.

My point was that a model can only contain X number of "things" it knows, I'd rather those things be concepts, poses, backgrounds (locations, etc), types of clothing, actions, expressions, or any number of other things and not characters and the details of their clothing and appearance

Again, not saying pony (or anything else) is done that way, or better at it, just saying that I think training hard on characters and their details is not what I want from a model. I'd rather have a model that can generate every angle, pose, action, expression and location with a generic figure and needing a lora trained for a specific character than the other way around

As an example, I'd rather the model be able to produce a subject drinking a glass of water. I don't care at all if it knows who Miku is, that's useless information and better relegated to a lora

[steady grail]

yeah, i agree, characters and styles can be added with lora

hello guys

but training different characters, gives more flexibility to the model

the point is to find that fine line where characters and concepts are genned fine enough

[summer night]

omfg
i'm so used to models having baked in vaes now that i didn't even consider that noobai might not have a baked in vae

manually setting the vae to the SDXL vae didn't really seem to have any benefit on quality

[ashen cloak]

Hi everyone! I'm currently using Illustrious models and I'm having the issue that all men I generate have these really overpronounced iliac furrows. Anybody got any idea how to get rid of those? (ideally without segmenting/adetailer)
I've already tried putting any of these tags in the negative prompt: iliac furros, adonis belt, apollo's belt, muscle, bara, hip bones - and it had zero effect

my latest attempt was downloading a muscle lora and running it with a negative value which did somewhat work but it also creates whacky anatomy for more complex prompts sadly

[vital raptor]

maybe try abs in the negative

[dusky linden]

or give the male character some more weight (as in kilograms)

[ashen cloak]

right, forgot to mention, abs and muscular are in the negatives too

[ashen cloak]

when I try that they either go full overweight or they just get a big belly while retaining the iliac furrows

I'm starting to feel like I'll have to train a lora for it

[desert island]

That's a crazy workflow. 😱

[wispy canopy]

how much proper upscaling matters

[wispy canopy]

it took a while to get there tho

i really like how non-strict tile controlnet upscale can do -> it can make up it's own new details

[desert island]

well it looks great... 😄

[wispy canopy]

og -> upscale stage 1 -> upscale stage 2, i love seeing the extra detail being added

[desert island]

Yeah it's really nice. The difference between all of them is REALLY noticeable. Great stuff! 👍

[steady grail]

wow, what's that? segmented attention?

[wispy canopy]

PAG, perturbed attention guidance + controlnet

combined, extremely potent when used as upscaler

[steady grail]

i should try

ty

[wispy canopy]

oh right, take notice -> my upscale steps have a denoise ratio of 1 (!!!!!)

controlnet the only thing keeping the image from exploding lol

brother cat and his spirit sister? uhh... he got converted into killing machine? sounds very similar to the story of aphelios lol xD like the idea of mecha murder cat more tho

i like how it actually improves upon the fur

[wispy canopy]

artist combo makes keeping that cat appear lol

[wispy canopy]

the kong 😄

[modest fog]

@green temple

🗡️

[dusky linden]

Thats a man

I just know it

[modest fog]

?

[dusky linden]

Man in maid outfit

You wont fool me chaos

[modest fog]

Stop being weird

[dusky linden]

Thats it, im making mikus just for you

[modest fog]

Should I block you too

[dusky linden]

Damn i thought we were buddies

[green temple]

Why look at me like that

Last one

[summer night]

anyone know what resolution i should use for a 4:5 aspect ratio?

these are the only ones i have shortcuts for

[modest fog]

Need to go to sleep earlier

else shes gonna use her sword

[modest fog]

916*1145 ? sounds weird

[summer night]

it's a common aspect ratio, but it has to be at an uncommon resolution to work good with SDXL

[modest fog]

sdxl isn't super well trained on those

mostly 3:2, 16:9 and 1:1 I believe

the others can get buggy like stretch bodies or do repeated patterns in some way

[summer night]

i thought it was trained on only 1:1 images

[supple raptor]

No, SDXL was trained with bucketed resolutions. 4:5 is not one of the buckets used though.

[modest fog]

would that apply to finetunes too ?

I only have weird things happening if I go over 1024*1024

and tall/wide ones like 16:9

[supple raptor]

Depends on the finetune, but unless the person training had the big dumb, or an incredibly limited dataset, yes.

[keen hollow]

happy weekend

[supple raptor]

also, this resolution won't work at all since both values need to be evenly divisible by 8. you could do like 920*1144

[modest fog]

I think on a1111 when I used it I had something to automatically fix those into divisible ones

[supple raptor]

It's been a long time since I last used it, but I seem to remember it being possible to type invalid values and then you just get an error. If you just use the sliders to set the values to whatever is closest to the desired value then they should be valid unless you've screwed with the config file for the sliders.

it's also entirely possible that i'm misremembering or that it's since been changed

it's been 7 months since i've used a1111

[dusky linden]

basic ahh style

much prefer this

[wispy canopy]

artist tag or lora?

0.5sec before annihilation

[dusky linden]

Vanripper, natively in illustrious

Lora though is more consistent

[desert island]

Manga Panel And Page Creator with Griptape support for ai prompting of certain elements. Still a work in progress but I think it's a pretty good start.

[modest fog]

Ugly

[lyric zealot]

https://tenor.com/view/miku-hatsune-miku-meme-looking-at-camera-looking-gif-14156409521403913189

[wispy canopy]

nevermind the bad hand, lovely stile (already mixed with another artist xD)

[green temple]

raphtalia ?

[modest fog]

No, it's fox

[green temple]

focs

[modest fog]

It jumps

[modest fog]

Cheddar

[green temple]

fat

[modest fog]

Fat not allowed only smol

[green temple]

fat smol

[modest fog]

[dusky linden]

Holo

[green temple]

ok

[trim current]

anneelooze its dtetct

[summer night]

how do you get inpainting to not make a mess

i've never been able to figure it out

[vital raptor]

usually a higher denoise makes a bigger mess, so i just lower it to clean up after fixing bigger issues

like, i'll use 0.9-1 to replace a hand, then use 0.6-0.75 to fix finger details, then 0.45-0.5 to clean up the weird smudges if needed

[summer night]

i'm just trying to remove this watermark

so i guess i'll try 1 just to get it gone, and then repeat with a lower value to fix the smudging

[vital raptor]

yeah, i'd use like 0.95 to remove the watermark, then a second pass with 0.6 or less to clean up

[summer night]

i'm also not really sure what masked content option to use, i've been using latent noise

[summer night]

i feel like fill and original wouldn't be good for removing the watermark because it would fill the masked areas with the colors from the watermark too

they might be good for fixing the smudging though

is there an inpainting controlnet for SDXL? i remember looking for one a few months back and couldn't find anything

idk if it would even be any better than the built-in inpainting

[vital raptor]

i would probably try original, i only used fill for things like removing extra fingers not sure how it would handle a watermark

i never got good results with latent noise for anything

[summer night]

just from the descriptions, latent noise sounds like it would be the best at completely removing something

since fill, original, and latent nothing all take colors from the masked area

[vital raptor]

yeah, but you want it to blend it, not erase it completely

[desert island]

Bow before god empress Miku! 😆

[summer night]

i used latent noise to get rid of the watermark at 1 denoising strength and then used fill at 0.6

it looks okay if you don't look to closely

i did expand the masked area a little on the 2nd pass

[vital raptor]

you can always do it again, as long as you're not using the same seed you can keep generating that area until you get a better result

[summer night]

i usually generate 3 at a time
and pick the best result
which is what i also do with adetailer

[vital raptor]

or even just adjust the mask to the current problem area and fix a little at a time

sometimes i have to mask and remask a few times at different levels to get a clean result

hell, i was doing an inpaint on an image yesterday where the characters tail was connected in the wrong spot, had to remove it from one spot and make it attach somewhere else.. i ended up with over 200 gens of the same image with minor changes 😄

[modest fog]

I just draw watermarks out if I need

I need to learn how to copy and paste stuff in krita

And move things

I go back in mspaint to do that lmao

[vital raptor]

that's what i do in swarm, either erase them from the image directly or just paint a color over them before the inpaint mask, but i don't think a1111/forge can do that

[modest fog]

Too many buttons

[vital raptor]

and i'm too lazy to use a different app 😄

[modest fog]

Kinda of relaxing to use krita

Just draw things

Anime stuff is generally easy to do a bad job and be good enough

[summer night]

i'm going to make an inpaint mask in paint.net, maybe square edges will look less obvious

[modest fog]

((also no one will pixel peep that exact area of the image to see the slightly different pixels))

[summer night]

i will

[vital raptor]

i would just generate over it with a new mask 🤷‍♂️ it's easy to smudge things around until it blends in

then again, doing part of it in a paint app is faster on slow hardware

[summer night]

maybe i should inpaint before upscaling

but i feel like hires fix gives better quality than the upscaler in the extras tab, and you can only do hires fix on the txt2img tab

[vital raptor]

you can do hires fix with img2img but you have to setup the upscaler in the settings somewhere, it's not intuitive

one of the many reasons i don't use those UIs anymore

not that swarmui is always better, but it makes more sense overall (to me)

[summer night]

i wish that the inpainting masking tools were a little more intuitive
things like a square brush, drag box select, eraser would be nice

just a round brush with an undo button isn't that great

[trim current]

invokeai has good inpainting

krita comfy is pretty good if you know how to use it

[vital raptor]

i imagine the round brush is specifically because it's easier to hide rounded edges, squares stand out

[dusky linden]

What ui do you use? Swarm b
Has all that and more (no square brush) but you can even add other images into the image and turn that into masks

[summer night]

i've been using reforge

i just moved on from the old version of forge

[dusky linden]

You can even "paint" on a image with the background color, mask it and then impaint and it looks great 99,9% of the time

Within the ui

I never used reforge I'm afraid but swarm is a total recommendation

[summer night]

this thing right here is literally a make or break for me
i stayed way from reforge for a bit because extension sorting was weird and i couldn't have it right under the generation parameters
can i have something like this in swarm?

[dusky linden]

Yes

Its natively build in

For portraits and width

And it's a nice dropdown menu

I can give you a tour but I'm expecting some handymen in a bit

[summer night]

i already have swarm installed but i haven't messed around with it much, i need to figure out how to point it to my automatic1111 models folder, because that's where my old forge and reforge installations point to for their models

[dusky linden]

Oh thats easy

Also screenshot of the ui

Under resolution you can see that setting

Also for the models, its somewhere under server configuration

Iirc you can just copy as path on the model/lora etc folder and just paste it there

[summer night]

are you able to do things like hires fix an image you just generated, and easily send it to img2img or inpainting like these buttons on reforge?
💾 is the save image button
🖼️ is the send to img2img button
🎨 is the send to inpaint button
✨ is the hires fix button (regenerates the current selected image with your hires fix settings, but skips the image generation since it already has the final image to work off of)

i get a crap ton of use out of these

when you generate images they go to the "generated images" directory, but when you click the save button, the images go to their own "saved images" directory

[dusky linden]

Its automatically saved but if an image is generated you can star a image to get their own folder. Theres a edit image, (impaint) and use as init

If you click reuse parameters and its your last image it will skip the image generation and you can use the refine/upscale(hiresfix)

[vital raptor]

there's also a refine image button that will 'hires' whatever the current image is

[dusky linden]

Use as init =img2img

Theres also a quick and easy upscale 2x button

[vital raptor]

that doesn't use the refine settings though

[summer night]

this is how i have my image directories set on reforge

[dusky linden]

Yeah

[vital raptor]

the refine image button uses the current refine settings

[dusky linden]

Possible in swarm

[summer night]

notice that generated images go to the output folders but images saved using the save button just go to the root images folder

i guess the star button works the same?

[vital raptor]

basically

[dusky linden]

There's a output folder thats sorted by date, and "starred"

Also sorted by date iirc but you can edit the output folder

And image names, i changed mine so its just model+time

[summer night]

really the only reason i haven't moved on from something automatic1111 based is because i'm so used to it and don't want to give up any comforts, but if swarmui has everything i need i guess i can move to it

[vital raptor]

mine is set to put everthing in Output/new-images and the star button copies them to Output/Starred

[summer night]

i use adetailer a lot too

[vital raptor]

swarm is different, and not everything works 100% exactly the same, but it's overall better and much more upto date when newer things come out

[summer night]

does swarmui have something similar that uses the yolov8 models?

[vital raptor]

yes

it's called segment in swarm

[dusky linden]

Theres <segment:thing you want edited> (prompt for edit)

[summer night]

as long as i have quick resolution buttons, hires fix or something similar, inpainting, and adetailer or something similar, i'm good

[dusky linden]

But i recommend enabling mask grow in the "init image" menu

Benefit of segment, its in the prompt box so even if you want to reuse the whole parameters you can in the future

Built in image library of what you generated and you can just "reuse parameters" ez

[summer night]

not totally related but the reason i've kept using forge/reforge is because automatic1111 hangs for a bit at the end of image generation for some reason

forge doesn't do that which saves a lot of time

[dusky linden]

Swarm has comfy ui backend should be faster

I went from 22s on forge per sdxl to 13s on swarm

(3070ti)

[vital raptor]

the hanging at the end is a vae decoding thing because a1111 has like 0 memory management

[summer night]

i think when i was using a regular 3070 i was getting about 15 seconds per image (30 steps, 1024x1024) on forge

[dusky linden]

The only thing i do miss is to be able to edit the open pose output (drag pose lines) but with the newer models i hardly use it anyways

[summer night]

this is what i was getting on forge with a regular 3070, i don't remember what hires fix settings i was using

[gleaming flame]

not sure if it still counts but at least it was made and refined with anime models ;p

[summer night]

the 4070 super is just a little faster

[dusky linden]

Im waiting for the 5090

Already got my alarm set for jan30

[summer night]

i bought a 4070 super a few weeks ago
i went through a bit of a conundrum after the 5070 announcement but decided that i'd just keep the 4070 super instead of returning it

i did not want to go back to a 1070, even for a couple months (the 3070 i have is fried)

i'm passed the return window now so there's no point in stressing about it anymore

there were basically 2 lines of thought running through my head:
if i return the 4070 super and wait for the 5070, i might end up not being able to get the 5070 at msrp and i'll feel like crap
if i keep the 4070 super and then the 5070 goes up for sale at msrp with no stocking issues, i'll feel like crap

so i just went with the safer option

if i returned the 4070 super and couldn't get the 5070 at msrp, i'd be sol, because 4070 supers are no longer in stock

anyways, i'm done with that image i was desperately trying to remove the watermark from

[dusky linden]

Once i got time i can take you through the swarm settings if you'd like to get a quick start

Will take a few hours tho

[summer night]

i have no idea where to even put in my models path

i have it open

[modest fog]

Booba

[dusky linden]

Server

Server configuration iirc

[summer night]

alright well i figured out some of it

[dusky linden]

Yeah thats the spot

If some results look off, be sure to enable a sampler

[summer night]

idk if i need to fill out those 3 and the wildcards folder

[dusky linden]

Last 3? Nah

Did you use that cursed autocompletion extsension from wildcards?

[modest fog]

Every time I see other uis I remember how nice and simple a1111 was

[summer night]

i have no idea so probably not

[dusky linden]

Thank the lord

Yeah wildcards are pretty easy/convenient in some cases

[summer night]

it's going to be confusing not having a dedicated img2img tab

[dusky linden]

Generate a image and click edit image

That shrimple

Ah handyman is later, walking the dog and i got time in like 30m i think

[summer night]

i was wondering why my models weren't showing and then i realized that for some reason a backslash was omitted from the models paths

can i do some kind of tag autocomplete from a csv file?

on reforge i have a few different tag autocomplete files i can switch between

[dusky linden]

Yes

User settings

[summer night]

alright i found it
i clicked the documentation link and found out where to put the files too

does swarmui automatically save my settings?

i kinda know how to use swarm now, but i think i'll keep using reforge for a bit
but now i can move over to swarm at any time if i need to

the interface in automatic1111/forge/reforge is just more comfortable to me

[dusky linden]

Fair, there's a lot of work being done on swarm so new features are being added constantly

[supple raptor]

tbh swarm is simpler. it's well documented and basically everything in the UI has tooltips that actually explain what they do and how to use them along with links to the full documentation if you still can't figure it out.

[gleaming flame]

had to blurry the message a bit xD

[summer night]

i can't get flux to work on swarm

i installed the gguf support extension but whenever i click on generate it just says this

ok nvm, even though it prompted me to install the extension and i clicked the install button, i went to the extensions tab and it was not in the installed section
it isn't even in the available extensions section

great webui

[dusky linden]

did you have it in the right folder?
w

what folder did you put it in

works fine for me

[summer night]

it's in the folder with all my other models

[dusky linden]

thats the problem

Models > diffusion_models

thats because of comfyUI

not swarm

mcmonkey is also annoyed by that lol

[summer night]

so where do i put it then if it's in my automatic1111 directory

[dusky linden]

oh yeah you link it manually

i would move it over tbh

or a copy. its like <20gb anyways

[summer night]

hmm, automatic1111 has a diffusion_models folder

or maybe swarmui created it since that's the directory i have it set to

[dusky linden]

swarm would not create a folder within a1111

[modest fog]

surely that says rock on

[summer night]

i moved it to the diffusion_models folder and now i get a different error

i assumed my 12 GB card could load the Q5_1 model

[dusky linden]

hmmm it should just work, any reason your using the dev over scnhell tho?

[gleaming flame]

surely

[summer night]

2025-01-18 02:47:31.637 [Error] [BackendHandler] backend #0 failed to load model with error: Model loader for flux1-dev-Q5_1.gguf didn't work - architecture ID is missing. Please click Edit Metadata on the model and apply a valid architecture ID.

that's probably the issue

[dusky linden]

try to restart swarm

it should get the missing nodes etc

[summer night]

i already did

[summer night]

i heard it had higher quality

[dusky linden]

i mean its the base model no

[summer night]

hmm

[dusky linden]

wait are you by any chance using stability matrix?

[summer night]

i have no idea

[dusky linden]

did you update swarm since you have downloaded it?

[summer night]

this is where i got my flux model from though
https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main

[summer night]

yes

[trim current]

thats gguf

[summer night]

indeed

[dusky linden]

https://civitai.green/models/645943/flux-unchained-by-scg using this

[summer night]

what is civitai green

[dusky linden]

the SFW version of civitai.com lol

oops

[summer night]

oh

[trim current]

dont use gguf

[dusky linden]

but yeah GGUF is also supported iirc

[trim current]

oh

[dusky linden]

but its machine heavy

[trim current]

idk i use comdy with its node

gguf is suppose to help with lowend vram machines

[summer night]

that's gonna take a few hours to download and i'd really like to figure out how to get what i have working

i need to figure out what metadata to fill out on the model i have

[dusky linden]

found this in the swarm disc

[summer night]

what is the prediction type flux uses?

[dusky linden]

did it work on forge from the getgo?

[summer night]

i haven't tried it on regular forge

my forge install is the outdated version before they upgraded gradio

reforge doesn't have flux support i don't think

[modest fog]

flux

[dusky linden]

https://github.com/mcmonkeyprojects/SwarmUI/blob/master/docs/Model Support.md#gguf-quantized-models

try

Flux.1

as architecture

[summer night]

maybe i don't need to fill out the prediction type in the metadata, i just entered the architecture id and it's downloading something now

[dusky linden]

or Flux.1 Dev

[dusky linden]

yes thats intended. its getting the nodes required to run it

[summer night]

does swarmui support cogvideox?

i couldn't get it working on comfyui with sebastian kamph's workflow and guide

[summer night]

[dusky linden]

currently unsupported but when hunyuan img2img comes out its gonna be irrelevant iirc

[summer night]

i'll try hunyuan or mochi for fun even though image2video is what i was trying to do

[dusky linden]

oh yeah the impainting thing i wanted to show

watermark

[summer night]

holo IS farming

[dusky linden]

just paint over lmao

masked it

HIGH init creativity

ah forgot mask grow and blur sec

[summer night]

maybe through playing around with flux on swarm i'll get used to it and start using sdxl models on it

[dusky linden]

oh yeah funny thing if you put init to max and use the grow function

[summer night]

i feel like reforge is probably just going to be maintained the way it is and not get any new features

[dusky linden]

but dont edit your pompts

[summer night]

like when you select "only masked area" on inpainting options in auto1111 based uis

[dusky linden]

yeah lol

mask blur is good for it tho

[summer night]

do you see why i didn't want to have to download another flux model

[dusky linden]

oh yeah definitely. last week i used to run on 20mbs up/down

now im on 1000/1000

[summer night]

20mbps download
5mbps upload

[dusky linden]

after a quick upscale, wham its gone

adding a depth map blur because i want to blur the background

[summer night]

cox is putting in fiber in our area which is honestly really surprising

we live out in the middle of nowhere, my neighbors literally raise cows and horses

[dusky linden]

a slight depth map blur is so nice

oh im next to the sea

[summer night]

i'm in the central united states, aka one of the "fly over" states

[dusky linden]

netherlands here 🦾

[summer night]

i have another friend who lives there

also, what really is the point of civitai green
you can just adjust the content filters on the main website can't you?

[dusky linden]

i use green at work

and .com at home lmao

[summer night]

i wonder how well the swarm mobile ui will work on phones

it seems like something that would probably work better on a tablet

[dusky linden]

not that nice atm

tablet yeah

[summer night]

i already have tailscale set up on my pc, laptop, and phone so i can use things like sillytavern, koboldcpp, and stable diffusion webuis from anywhere

[dusky linden]

oh thats nice

[summer night]

i can also just remote desktop to my PC from my laptop

but my home network doesn't have good upload speed so i only do that if i need to start something like koboldcpp

[dusky linden]

swarm works within your LAN pretty well

builtin support for it even lol

[summer night]

if it works on LAN i can use it over the internet through tailscale

tailscale is a mesh VPN kinda like what people used to play minecraft multiplayer with their friends back in the day

[modest fog]

Hamachi worst best thing to use lmao

[summer night]

the download took so long that the generation request timed out

[summer night]

this might push my 12 GB of vram

[dusky linden]

I got 8

Takes like 20-30s a image

[summer night]

maybe it offloads layers onto cpu then?

like textgen when you can't fit a whole model in your vram

well it's doing something now

not generating an image though

maybe it needs a different sampler

i changed the sampler to euler and the scheduler to normal and it works now
blurry result though

[modest fog]

Try other schedulers

[summer night]

flux stinks

53 seconds for a blurry image

[modest fog]

[summer night]

i'm done with flux for now, back to reforge where i'm comfortable

[dusky linden]

Did you use the proper scheduler for flux?

Cfg scale for flux is

1

No negative prompt

[summer night]

i was using euler sampler, normal scheduler, 2 cg and 10 steps at 1024x1024

i also had the flux guidance scale thing enabled at 3.5

either way it just took way too long to generate a single image

[trim current]

i use flux def on my 3060 12gb on comfy with wavespeed first block cache and get 1.62s/it

some info to consider if you dont want to use the distilled versions.

[summer night]

i was using the Q5_1 gguf and it was taking almost a minute per 10 step image
on my 4070 super

[trim current]

yeah i dont use any gguf or any

wavespeed is sufficient enough for me on comfyui, allows me to use dev model

25 steps

35*

1min

[summer night]

i'm just going to keep using SDXL models for now

i like being able to generate an entire quality image in 8 seconds

[trim current]

oh wait misread

xl is definitely good bang for buck for the 12gb vram

youll want the 4090 if u want those speeds for flux

its like 1.65it/s

[summer night]

not sure a 650 watt PSU will put up with a 4090

[trim current]

nyo

[desert island]

it will not 😀

[trim current]

youll need 850

[summer night]

what i need is a whole new computer

[trim current]

lol

turn the current one into a nas storage

boom

validated

perfect reasoning

[summer night]

CPU: i7-8700
GPU: RTX 4070 Super
MB: B360 chipset
RAM: 32 GB DDR4 2666 MT/s
650 watt PSU

it originally had a GTX 1070 GPU

[trim current]

https://tenor.com/view/jfk-clone-high-i-like-your-funny-words-magic-man-jack-black-gif-18659433

[dusky linden]

I need to get a new psu, when i get my 5090

[trim current]

my bestie is the nerd when it comes to hardware knowledge

im just good at using stuff

lol

[summer night]

i7-8700 = old and slow

[dusky linden]

But then i also will get ddr5 ram and then i need a new motherboard and so i need a new processor and all in all im just chaning my whole pc

8th gen isnt terrible but isnt the newest either

[trim current]

any slower than ryzen 5 2600x?

[summer night]

the best CPU i could get on this socket is an i9-9900 and it isn't worth the price

[dusky linden]

Fair fair

Im on a ryzen 9 5000x

[summer night]

i dunno

[trim current]

i need a better cpu

esp for games like poe

[summer night]

i want to build a new PC with a Ryzen 5 7500F
it's like a 7600 but is cheaper and uses less power, but lacks an igpu

[dusky linden]

Why go for an integrated gpu if you got a gpu

[desert island]

I've got you beat my friend I'm on an old FX chip and the old 970 gaming mobo, rocking 10 year old technology but I got a 4080 super

[summer night]

you can get legit ones on aliexpress for around 115-135 USD

[dusky linden]

Pc parts from ali express. Concerning

[summer night]

CPUs are okay if you find a legit listing
GPUs are a no-go

https://www.aliexpress.us/item/3256807886761748.html

this one seems legit

you just need to sort by orders

[supple raptor]

[supple raptor]

for flux you're going to want to look for gguf fp6 checkpoints. you'll be spilling well into system ram and looking at several minutes per image otherwise.

[summer night]

eurotypo said he was using that model on a 3070 and getting 20-30 seconds an image

[dusky linden]

On a safetenors on flux unchained yeah

[supple raptor]

same tdp, just no graphics and slightly lower base and boost clocks. worth it if you can save a decent amount of money, the prices you posted in the next message do seem good, but i haven't looked into current prices

[summer night]

the one i sent the link for is currently $125

[supple raptor]

gguf is typically delivered in a .safetensors format so they could still be running a quantized version. the model is popular enough that i know it's name despite only genning like 12 images so quantized versions existing would not surprise me at all

[summer night]

the model he's using (flux unchained) is fp8

actually the description is a little confusing

"This model was built on top of the flux.1_dev_8x8_e4m3fn-marduk191 version, so fp8 quality, though I have included the full FP16 clip-l and T5 models, as I don't like the quality drop off with FP8 T5 clip. If there is demand for an fp8/fp8 version, I can make one available."

[supple raptor]

normal fp8 != gguf fp8. regular flux fp8 is not recommended for 12gb vram

[summer night]

the file itself says "Pruned Model fp8 (11.07 GB)"

https://civitai.com/models/645943?modelVersionId=768009

[supple raptor]

i used wrong terminology. q6 is what you should be looking for in flux models

which is gguf q6

and will properly fit into 12gb vram

[summer night]

well i was using the flux dev q5_1 model from https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main and it was taking nearly a minute to generate a blurry 10 step image on swarmui

[supple raptor]

could be too much vram overhead on your system. you'll want discord and browser hw acceleration turned off as it uses around 11.5gb iirc at native ~1mp res

[summer night]

fucking DWM

[supple raptor]

if you use nvidia broadcast, that's another source of overhead

[summer night]

DWM was like "ayo you got a gpu with more vram?" and started using exponentially more
it used to only get up to about 300 mb

restarting it lowers its usage to 270 mb but it always goes back up

[supple raptor]

yeah i think there may be a memory leak in dwm. i see some absurdly high vram numbers for it after leaving my pc running for multiple days

[summer night]

i'll probably just close chrome and use edge with gpu acceleration off when i want to use flux because turning it off on chrome makes it screen tear BAD when scrolling and watching videos

[supple raptor]

i use edge with hw accel turned off any time i'm genning and firefox with it turned on whenever i visit a site that actually needs it. 99.9% of sites that i visit run perfectly fine without hw acceleration even on the r5 2600 that i had until the beginning of this week.

[summer night]

i might switch from chrome to firefox when chrome drops support for manifest v2 extensions and kills ublock origin

someone needs to figure out how to make noobai look clean

i can't do it

reforge seems to have memory management issues

if i load a lora it will overload the vram the first time i try to generate something, and then it'll work properly afterwards

i really should switch to using swarmui full time

[dusky linden]

Just like minecraft forge it was based on :^)

[summer night]

forge is still the best mod loader for minecraft and i will never use anything else

i don't care about no fabric or cloth or neoforge

[dusky linden]

Eh neoforge isnt too terrible in atm10

[summer night]

i won't switch until every popular mod switches over to using just one of them

[dusky linden]

But im still on 1.12.2 enigmatica expert 2

But solo play is boring ahh

[summer night]

the latest version of minecraft i play modded is 1.20.1 i think

[dusky linden]

1.12.2 is based

But i gotta say the essentials mod is also goated

[summer night]

neoforge and forge mods on 1.20.1 are intercompatible i believe

[supple raptor]

vintage story > minecraft

[summer night]

TRUE

[dusky linden]

Vintage story is also hella goated

Bought it to play with friends, then they dropped out oof

[summer night]

do you mind if i send a picture of my vintage story house

[supple raptor]

do iiit!

[summer night]

i kept expanding on the same initial house

[supple raptor]

i watercooled my windmill setup because i was bored while waiting for my first batch of steel

[summer night]

the house also has 4 max volume cellars underneath it

[supple raptor]

here's my full house durring winter

[summer night]

they have various different uses

[summer night]

that's a lot of windmills

[supple raptor]

it's changed a little since then, i now have berry bushes filling the space between the two sets of covered animal pens and 3 rows of 3 wide reeds at the back of the farm (closest to camera)

[summer night]

do you play with mods

wildcraft fruits and nuts adds like a crap ton of different berry bushes and fruit trees and you can use them with the expanded foods mod too so you can make syrups and aged wines and stuff with them

[supple raptor]

not really. i have an afk mod that prevents hunger increase if you don't move for several seconds, but that's it

i'm waiting for 1.20 official and mods to be updated for it before i add many mods

[summer night]

i haven't played in a while so i'll probably wait until they get updated to 1.20 too

1.20 is officially out though

it released yesterday

[supple raptor]

oh shit! i haven't played in a few days. last i saw was rc9 a few days ago

[summer night]

https://www.vintagestory.at/blog.html/news/v1200-lore-update-the-journey-r391/

[summer night]

my kitchen would not look this cool without the mods "a culinary artillery", "expanded foods", and "stone bake oven"

[supple raptor]

i need to build a bigger house so i can have a real kitchen. i have a carved fireplace and table but not much more in my kitchen.

2 ovens are to the left of the table

[summer night]

this is an above view of the house and workshop side of my base, and i could only find a map screenshot of my farm

wait no here's one

oh yeah i also have a mod that adds actual rivers and lakes

[supple raptor]

i wish i had screenshots of the chute system in my other base. it's pretty elegant imo. it takes the drops from my pulverizers and querns and dumps them each in a column of chests around 30 blocks away. with pretty much all of it hidden under the floor and within walls

i tried the rivers mod and it seemed cool but it was bugged on rc8 and caused chunks to fail to load so i had to disable it until that's fixed

[summer night]

when i was playing, it wasn't compatible with the fishing mod which made me sad
but i'd rather have pretty rivers than yet another food source

huh, i tried using one of my illustrious style loras on noobai xl and it actually made the output look a little clean

i guess you pretty much have to use style loras for good results

[supple raptor]

tbh my farm puts out far more food than i know what to do with. 3/4 of it ends up just being converted to rot so i can make more compost

[dusky linden]

On noobai yeah

On wainsfw its clean from the getgo

But i like mixing 2-3 style loras

[summer night]

i use the excess vegetables to feed my pigs and the excess grain to feed my chickens

the fruit that i don't immediately use gets preserved as candied fruit or syrup, or turned into alcohol

[wispy canopy]

i might, or might not have cooked on this one. i like it ❤️

[supple raptor]

veggies feed my pigs, rabbits, and bighorn sheep. flax feeds my chickens. berries get turned to wine and then brandy, which leaves mash to also feed the animals

[summer night]

my sheep can eat grass and that never spoils so i just give them that

i need to mow my farm every once in a while with my scythe anyways

[supple raptor]

i do that too, but i still have so many veggies that i don't need to do it all the time

the bees also went nuts this year as i spent basically all summer making steel so i was around the farm to collect it as they made it. currently i'm sitting on 130l of honey

[summer night]

i like how noobai can actually do dark

[summer night]

i save iron and steel making for winter

in the warmer months i explore

[supple raptor]

i spent the winter building my 2nd base near a set of 7 hotsprings that i found

[summer night]

any kind of fast travel mod is extremely handy
translocaters are useful and all but they aren't exacly convenient

[supple raptor]

here's a view from the inside of the mid level of my windmill tower.

[summer night]

i use this one mod that adds teleportable waypoints every 4000 blocks or so

https://mods.vintagestory.at/tpnet

[supple raptor]

i made a compromise with myself because i wasn't happy with how any of the teleportation mods work and i just allow myself to use the /tp command from any reasonably established base.

my distillery

[summer night]

is there an objective "best" scheduler type?

i only just recently got access to this option since old forge didn't have it

i've been using Align Your Steps

[supple raptor]

i like AYS, i've done grids on all of the official noob releases and Euler with AYS consistently performed the best on EPS versions while Euler_A with AYS did best on vpred versions

[summer night]

was it you that recommended i use AYS with 28 steps?

[supple raptor]

AYS is intended for 28 steps so hopefully anyone who recommended it would have recommended the steps as well. I've definitely recommended it in this server a few time so it could have been me

[summer night]

it seems like 28 steps has been the norm ever since novelai

[supple raptor]

i think it's about time to reduce the bnuuy numbers

[summer night]

i still set my ENSD to 31337 as a hold over from those old days

is there even any real reason to set this

[wispy canopy]

no, there is no best scheduler

it all comes down to the combo of scheduler, noise (type), sampler, model

[dusky linden]

The most popular one is euler a or dpm++ 2m + karras no

[supple raptor]

the only reason there has ever been to set this is because NAI used it for their service so if you were using their leaked 1.x model you needed to use it in order to replicate images generated on their site. today, there's no reason to do so considering nobody uses that model anymore

[summer night]

i see

[supple raptor]

karras is a very silly scheduler and is particularly bad on vpred models.

[wispy canopy]

if i'm not mistaken, it has pretty high sigma values for a long time

[supple raptor]

use exponential if you want an aggressive scheduler like karras that actually makes sense.

dpm++ 2m is also not great on sdxl in general. dpm++ 2s_a is much better. i like ddim_uniform scheduler with it but it works well with several schedulers.

[dusty rock]

What is this?

[supple raptor]

yesh, here are some graphs (credit goes to Yoinked in touhou server and animeai) https://www.desmos.com/calculator/18uvk2cec4

[supple raptor]

base in Vintage Story

[wispy canopy]

i've done a lot of experimenting with it myself XD

basically, karras is weird because it decreases sigmas slower, but then it just keeps dropping it like mad, so you end up in the very small sigma range very fast, giving very little time in the middle region -> not all models like that, and also similar with samplers, might not like it, i suppose best used on non-ancestral samplers

[dusty rock]

It looks like Minecraft

[desert island]

The style and resolution of these is just so damn good. 👍

[supple raptor]

it's sort of like minecraft advanced. very similar voxel based game, but with a native subvoxel system. crafting often interacts with the subvoxel system either through needing to make clay molds to pour low tier metal tools or smithing higher tier metal tools on an anvil. world gen is very, very different from minecraft. biomes aren't a thing, instead it uses a banded simulation of regions so the world consists of bands that range from polar-equator-polar and generally repeats multiple times. it also has a simulated weather system that takes climate systems into account.

[wispy canopy]

yeah, i love the style for real, and it really showcases my upscale workflow like crazy

[desert island]

Whenever you post I have to always click on the damn image twice so I can see the full resolution version of the file. 😄

[wispy canopy]

the worst part? i'd love to see this will all the really working artists in there -> no way in hell i have the time to do that xD ... i only changed the artist name btw, seed/prompt are all the same. no lora

[keen hollow]

You know.. as much as I want a 5090, I wonder if I should just get a 5070/5080 for games instead. And use some cloud service for AI gens. I could pay for a cloud instance for probably years before the uplift cost of a 5090. 😛

[wispy canopy]

@keen hollow that's a very valid take...

[summer night]

the 5070 will definitely be able to handle at least SDXL pretty fast, and the 5080 could probably do flux at a fair speed

[wispy canopy]

this one, i wont'have to upscale as much, i suppose

[keen hollow]

yeah but that sweet sweet 32GB of VRAM... 😛

[wispy canopy]

for me, increased gen speeds tho, 3090 has some issues with that vs 40xx series

[keen hollow]

oh?

[summer night]

with the 4070 super it takes about 8 seconds per SDXL image, i imagine the 5070 will be around 4 or 6 seconds

[desert island]

Yeah I don't know if the style warrants the level of detail you get in the other styles.

[keen hollow]

I mean, I also want to play modern gen games at 3840x1600. 😛

[wispy canopy]

yeah, there's some things you can do on the 40xx cards that speed up generation, i think it's --fast in comfy?

[supple raptor]

personally i'd probably want to pay less than $550 for the performance that the 5070 is likely to give due to it only having 12gb vram. i'd rather buy a used 3090 at around the same price with double the vram even if it means gens take a little longer.

[wispy canopy]

<- 4k oled monitor masterrace representing xD

[keen hollow]

for resolution support?

[summer night]

take longer and use more power

[wispy canopy]

more vram -> bigger models

[keen hollow]

I use a single ultrawide.

[keen hollow]

ah touché

[supple raptor]

sdxl is pretty much the limit of what you can realistically use on 12gb vram without either offloading or distilled and/or quantized models.

[summer night]

after the 5070 releases you'll probably be able to get a used 4070 super for a fair price

[wispy canopy]

that's why nvidia already stopped production. fuckers.

artificial scarcity

i hate it

[summer night]

i imagine the 5070 probably performs a little better than the 4070 super and costs less msrp wise

[keen hollow]

They went with 16GB on the 5080. Had that been 24, I think I would have just been happy with a 5080.

[wispy canopy]

they really learned their lesson from the 1080ti xD

[wispy canopy]

now why do you think they did that xD

[keen hollow]

pff, to make you buy the biggin'

[wispy canopy]

and they know it

[keen hollow]

which you won't be able to buy

[supple raptor]

yeah, 16 is fine for gaming, but it's still a bit lackluster for ai

[keen hollow]

I refuse to buy from scalpers. Just a personal thing I guess.

[wispy canopy]

basically all my workflows are based around being able to use 24gb

[summer night]

hey, you can already do a lot with 12

[keen hollow]

I feel like it's showing its age on games already. Some are already approaching that upper limit as a requirement.

[wispy canopy]

so i couldn't even go back down at this point

[dusty rock]

Do you have any link about it?

[wispy canopy]

3 second google would also help...

[keen hollow]

But still, I could probably rent a high end runpod or whatever for 5 years... 😛

[summer night]

just get 2 5070s

[keen hollow]

tape 'em together

[wispy canopy]

at that price, i can just buy a 5090 anyway lol

[desert island]

or be a cool kid and get the Digits box in May. 😄

[dusty rock]

I don't know the name of it

[desert island]

Vintage something I don't remember it scrolled by a while ago.

[wispy canopy]

vintage story

https://www.vintagestory.at/

[dusty rock]

Thanks

[keen hollow]

Eface, I like how you organize your models 😄 😄 😆

[wispy canopy]

?

[keen hollow]

The folder structure shows in the meta 😛

[dusty rock]

It's weard that I didn't hear about it at all

[wispy canopy]

xD

[keen hollow]

got a whole folder dedicated to models for... nevermind

[supple raptor]

that's just reasonable business practice. distributors stop ordering them because they don't want to be stuck with them and have to sell at a loss. doesn't really matter anyways since buying a new previous gen card withing 6 months of the next gen launching is always dumb

[dusty rock]

Is it a kind of fork of fork of Minecraft?

[supple raptor]

buy used, preferably right around the next gen launch or around a major coin crash

[dusty rock]

There was another big project supposed to replace Minecraft, but I don't remember the name

[summer night]

i don't get why nvidia gpus don't drop in price, i don't think you can even get a NIB GTX 1070 for less than $400

[supple raptor]

i believe the project started as a minecraft mod. i don't believe it shares any code with minecraft anymore though

[keen hollow]

i didn't think mining was a big draw anymore for snatching up all the gpus

[supple raptor]

not nearly as big as it was in 2020, but for the top tier cards, you'll still see a flood of used inventory any time a major coin crashes

[dusty rock]

Why nobody speak about this game?

[supple raptor]

because it's not distributed by steam, microsoft, or epic

[summer night]

i bought a used 3070 from a miner last april and it stopped working after 6 months

[desert island]

off to bed. 👋

[dusty rock]

There are micro-blocks and mechanism according what I already seen

I'm looking for a full introduction of it

[summer night]

the only way i'd ever buy pre-owned GPUs is if they are ebay refurbished or something
those come with a 1 year full coverage warranty and are backed by ebay's money back guarantee

[wispy canopy]

i bought my 3090 preowned, has been purring along for at least a year now

[summer night]

newegg refreshed is probably also a decent option but those only come with a 90 day warranty

[supple raptor]

yep, here's a fireplace that i chiseled and inlayed other materials into, along with my underfloor, watercooled mechanical setup (watercooling isn't actually a thing, i just thought it would be fun to fill the space with water). last is a shot of my entire farm including the windmill that powers all the mechanics.

[summer night]

back when i was deciding on if i should get a used 3070 or a new 4070 super, this was $300, which i think was a really good deal

[wispy canopy]

this model is weird when it comes to swords -> just puts them everywhere xD

[summer night]

the image looks oversaturated to me

dynamic cfg / rescale cfg is actually really handy for that

[supple raptor]

it'd be hard to convince me to buy an 8gb gpu even just for gaming, i wouldn't even entertain the idea for ai usage

[keen hollow]

Can it do an ethereal forest princess under volumetric lighting, surrounded by the spark of fireflies, wearing a whispy green dress and a sword on her back?

[summer night]

my now broken 3070 could handle SDXL on forge at about 15 seconds per image which was bearable to me

[supple raptor]

i'd much rather have a 3060 than a 3070 due to 12gb vram

[keen hollow]

I'm also struggling with how I want to setup my storage in the new build. I have a significant amount of stuff and need to figure out how to consolidate some of this crap.

[keen hollow]

I'm in the 3060 club 😄

[summer night]

you could just get a new 4060 for not that much more than a used 3070

[wispy canopy]

probably artist combo, i know those combined make for some pretty strong colors

and i could use apg, but that breaks if i want to do batches

[summer night]

is this with the same settings or did you use some kind of dynamic cfg

[wispy canopy]

same settings

[supple raptor]

4060 is a bad card due to too little vram as well. 4060 ti 16gb would be a good choice though.

[summer night]

i just decided to get a 4070 super because i figured it would be fairly future proof

but with the way games are going now i'm not sure

devs are all relying on resolution scaling and fake frames now

[wispy canopy]

hmm, perhaps i should give apg a go after all, i could always build around a switcher for that

[supple raptor]

4070 super is pretty good. the price is around what you could pay for a used 3090 though which makes it a bit tougher of a choice.

[summer night]

i think 2k textures should be the default for games and 4k textures should be an optional download
that saves performance and storage space for people who weren't going to use them anyways

[keen hollow]

all those illustrious models, in my opinion, are oversaturated

[supple raptor]

what cfg are you using?

[summer night]

4-5 cfg is good for illustrious models

[keen hollow]

usually around 4-5 for those

[keen hollow]

hey that turned out pretty well!

[supple raptor]

5 for base illustrious and early noob versions. 4 for noob versions after ~0.5 in both eps and vpred

[keen hollow]

needs to be darker so those flower light thingies shine more 🙂

[summer night]

noobai is the only model i've seen that can do darkness

[supple raptor]

the power of vpred

eps (likely every SD model you've ever used) averages the entire image to 50% brightness. that's the main problem that vpred solves.

[summer night]

i hope the quality of base noobai isn't indicative of all future v-pred models

[supple raptor]

naid3 is also vpred

[summer night]

i haven't used it because i don't want to give anlatan money

[summer night]

it looks decent as long as you use a lora

[wispy canopy]

or just start with a black image, convert that to latents, and then diffuse xD

[wispy canopy]

ok, gotta go, turned out pretty sweet ❤️

[keen hollow]

nice

she's got a mini sword too

[supple raptor]

i virtually never use a lora

[keen hollow]

ooo images

[supple raptor]

works somewhat on eps models if you want the entire image to be dark, but if you want any bright points they will likely be very muted

[summer night]

how did you get it to not have muddy lines and details

[keen hollow]

That 2B image is famous in here.

I didn't know wlop made cameras. 😄

[summer night]

pretty much whenenver i tried to use noobai with style tags and without a lora, the image turned out oversautrated with muddy details

[supple raptor]

that 2b image is pretty new, idk if i even ever posted it here. i've posted many images from a similar prompt though

[supple raptor]

CFG: 4
Sampler: Euler A
Scheduler: Align_Your_Steps
Steps: 28
Negative: worst quality, low quality (only recommended for vpred 1.0)

I run 1.5x upscaling with yandereneoxl upscaler on refiner pass at 0.5 creativity and 56 steps (28 actual steps due to creativity value)

[summer night]

do you use masterpiece, best quality, newest, absurdres, highres, in the prompt?

[supple raptor]

masterpiece isn't trained, throw that out. quality tags for noob models that i use are best quality, amazing quality, very aesthetic, absurdres, very awa placed at the very end of the prompt.

[summer night]

what is very awa

[supple raptor]

very awa can mess with some styles so sometimes it gets dropped

[keen hollow]

i dont believe you... lol

[supple raptor]

aesthetic tag for later noob models. iirc it's like the top 5% aesthetically ranked

[summer night]

everything but the hands looks good

[summer night]

it is consistently doing bad hands

[supple raptor]

hand quality can vary depending on the artist tags used. in general i don't have many problems with hands

[summer night]

i was just testing it out so i wasn't using any artist tags

[supple raptor]

those hands are easily good enough to fix with a quick inpaint at low denoising so i wouldn't complain about them.

[summer night]

it is doing a bit better with artist tags

[supple raptor]

@keen hollow this may be the 2b image that you were thinking of. this one is from a very similar prompt on naid3. still among my favorite images that i've genned 9 months later

i've definitely posted that one here multiple times

these are 2 of my other favorites from naid3

[keen hollow]

yeah that could be it!

I like the newer one even better

[supple raptor]

i've also posted other ones from the same and similar prompts on different models

[summer night]

i hate my keyboard
backspace works normally like 1/3 of the time
1/3 of the time it doesn't work at all when i press it
and the other 1/3 it double inputs
don't ever buy keyboards from steelseries, unreliable less than a year after it's out of warranty

i now believe the people that say steelseries makes expensive e-waste

[summer night]

can i get that artist tag mix or is it a secre

[keen hollow]

Dead serious, I finally ditched the fancy mechanical keyboard world and just went back to a basic membrane keyboard.

My formative years were on a membrane keyboard and so my speed and accuracy is just unmatched on one.

Had to ditch the hype train.

[dusky linden]

I went from membrane to mechanical and its the best switch i made

[summer night]

next keyboard i get is going to be a hot-swappable rgb mechanical keyboard

[dusky linden]

And i annoy my coworkers

No rgb tho

[summer night]

i lurk in the dark in my room so i need a backlight

[supple raptor]

it looks very different on noob, but it's (quasarcake:1.1), (ebifurya:0.7), (wlop:0.9) here's an example of that artist list on noob vpred 06. the list from the one where she's in the field of lunar tears is (ciloranko 2020:0.95), (ask \(askzy\):0.95), wlop, hiten \(hitenkei\), ningen mame, manuga \(manuga430\), modare, (tianliang duohe fangdongye:1.05), (fuzichoco:0.95)

[summer night]

even though i don't need to look at my keyboard to type, i still need it in my peripheral vision

[summer night]

dang, that 2nd one is a lot longer than anything i ever used

i recognize that 2nd to last name

[keen hollow]

noob always looks like watercolors to me

[supple raptor]

yeah, that's a complex one. most of the lists i use are only maybe 3-5 tags. you may need to add a BREAK (in a1111 or forks of it) or <break> in swarm after the artist list on that one for certain prompts because it's so long.

[summer night]

oh yeah this is something i never understood

whenever i'm about to go "over" it, it doubles

i thought BREAK was used for regional prompter

[supple raptor]

ok so this is going to require a little lesson on how the model works

[keen hollow]

here comes a nubby lesson

get your textbooks

[summer night]

i know it's the token count, but that's about it

[keen hollow]

this will be on the final exam

[summer night]

idk why there's a limit on it or why the limit increases

[keen hollow]

Any benefit to storing the AI files (models and binaries) on an NVMe drive versus an older SATA SSD?

[summer night]

i imagine they would load negligibly faster

[keen hollow]

Yeah I wonder how much real-world impact it would have.

Trying to determine if I really need the new NVMe drive on the new build, or if I can just continue using one of the many SATA 2.5" SSDs I have.

[summer night]

maybe like 0.2s

[keen hollow]

naw it'd be more than that, dontcha think?

think like, flux model

Maybe I could store models elsewhere and move them over when I'm using them. That could work.

[summer night]

delete the models you don't use as much

[keen hollow]

delete? what's that?

nubby still typin'.... haha

nubby, you gotta save that lesson somewhere! you've given it multiple times!

[supple raptor]

CLIP is the text encoder of the model. it processes the text that you input into a "language" that the UNET can understand. this is in the form of a tensor which is something like 0,175,236. the issue here is that CLIP can only process blocks of 77 tokens and the first and last are start and end tokens, leaving 75 tokens for your prompt. what we do to get around this is split longer prompts every 75 tokens, then concatenate the results before sending it to the UNET. lets say you have a prompt that is 90 tokens long, the first 75 tokens becomes something like 1,20,126 and the remaining 15 tokens are padded out to 75 and then becomes something like 0,10,14. those are concatenated to 1,30,140 which is sent to unet. unet is happy and you get an image. this is also where BREAK comes in. the default function of break is to give the user control of where the prompt is split. this is useful for a few reasons, the most common is to prevent a situation where a multi token tag is split between blocks. for example, if brown hair falls at tokens 75 and 76, brown will be in the first block and hair will be in the second block giving them basically no association with each other, so you will probably get brown in the image and hair, but brown hair is up to luck. add a BREAK before it and they get sent in the same block and now you get brown hair consistently.

regional prompter just hijacks the BREAK command to use it for it's own function that isn't related to the original purpose at all. they really should have used their own command instead

[summer night]

so when the limit increases, it's automatically setting a BREAK?

[supple raptor]

yes, all of the modern uis effectively automatically add an invisible BREAK after every 75 tokens

the alternative is for everything after the 75th token to be simply ignored

[summer night]

i think on regional prompter you can use AND instead of break

you have to check an option though
or maybe AND just doesn't do anything in regional prompter if you check the "disable convert AND to BREAK" option

[supple raptor]

AND is another unrelated command that they hijacked

iirc AND processes the text before and after it through both text encoder and unet separately and concatenates those results somehow into an image. that function is at least almost always useless so it doesn't annoy me as much that they hijacked it.

i'm less confident about that explanation as it's never really been a useful feature so i never paid a lot of attention to it

[summer night]

good

my favorite war criminal princess never looked so great

https://tenor.com/view/misono-mika-blue-archive-gif-27497693

[supple raptor]

[wispy canopy]

@supple raptor seems i CAN get good output from noobai v-pred if i try

[gloomy scroll]

Nubby is extra smug today

[keen hollow]

welcome back

need moar

[wispy canopy]

so, tends to be overly dark

dunno if that's because of perpneg or smth

nope, scheduler

holy hell vpred is sensitive to that

but, getting somewhere

[wooden sail]

||bartolomeobari|| <

[modest fog]

Too close

[modest fog]

Ultra super giga contrast model

[wispy canopy]

Wrong scheduler tho, so it's kinda ffed

[dusky linden]

Euler A or dpmpp_2s_ancestral + align_your_steps