🍥｜anime

[wispy canopy]

private model?

or did you post it somewhere?

[nova remnant]

no its realcartoon xl

[wispy canopy]

oh yeah, that could certainly work

[nova remnant]

oh ... i havent posted the model

[nova remnant]

also using a different sdxl vae.. the image tones looks so much better

[wispy canopy]

wow, looks really nice ❤️

[nova remnant]

i baked that vae into the merge

[steady grail]

ok fixed the workflow, but now a node started working and i'm getting negative gray gens lol

of, found the error

[nova remnant]

@wispy canopy

[wispy canopy]

cfg perhaps a bit high, but those eyes man ❤️

[nova remnant]

lol cfg at 2 but its the vae that im using make the tone sharp

[nova remnant]

@steady grail about surprises, i think playing with booru tag can be great 🙂

[steady grail]

i agree

[nova remnant]

what im doing is putting in assorted words based on a visual concept, w/o describing any context for it and let ai surprise me

[steady grail]

the first is with kohya and the second, without

[nova remnant]

the image on the right looks nicer

its more pronounced

[steady grail]

yep, tha'ts what kohya does

reduce detais, so that the upscale can add them

[nova remnant]

right one is the 2nd?

[steady grail]

the second is without kohya

looks plain, because, reducing an upscaled image, you lose details

[nova remnant]

lol which one is second.. left or right image?

[steady grail]

right

[nova remnant]

the right image w/o kohya .. that kinda looks better

[steady grail]

koyha is on the left, the 1° image XD

ok, sorry, i'm losing it

lol

i'm doing too many things at the same time

[steady grail]

and for some reason, kohya is refusing the latent upscaler

[nova remnant]

i used koyha for a bit just to avoid the duplication of character in larger resolution, it does it well, but since i mostly stay with 768x1024 i don't really have need to use kohya

[nova remnant]

also i can't hires fix atm, not that its necessary with sdxl setting

[steady grail]

kohya should make hires fix unnecessary

ok, now it's working

[steady grail]

same image with upscale and khoya active

crooked the hand 😦

and added 1 more hand on the dress

[steady grail]

guys, how did you set kohya?

yeah, lcm reduces mutations and weird gens

expecially in upscale

oh, that's smooth

a little dark

[terse ore]

Is kohya a node or something?

[nova remnant]

medium small

medium

smol

👍

[lyric zealot]

@nova remnant

[nova remnant]

the only three XL models i use and the left one is the merge between the other two.

and those xl models that focus on anime didn't quite look too good

like bluepencil

all the images i rendered with bluepencil came out either too flat or insignificant on the facial structures

[steady grail]

with lcm it's very easy to overload the steps

[nova remnant]

recommended steps for lcm is between 2-8

[steady grail]

yep

but the higher the resolution, the more steps you need

[nova remnant]

no you dont

it will add the noise for specified dimension given the model supports it

[steady grail]

yeah, but the image will not lack details?

[nova remnant]

shouldn't

[steady grail]

but it does, is expecially evident with adetailer

[nova remnant]

if you are thinking the larger canvas would spread the detail thin thats not how it works

[steady grail]

more steps you add, a more detailed face you get

[vital raptor]

adetailer is applying an inpaint after the image is already generated

[nova remnant]

yeah adetailer add detail post render

[steady grail]

yep, but upscale the image and than downscale it

[nova remnant]

there is a threshold

going above that would not bring better results

[steady grail]

i agree

[vital raptor]

my testing with LCM has 8 steps and CFG 1-2 making really good detail, increasing steps beyond maybe 12 just corrupted the image

[steady grail]

but i saw the upscaling encreasing that threshold

[nova remnant]

yeah, also about cfg i think having set that to 2 would reflect on the negative prompts, that's what i understood from huggingface page

so say you want to use negative prompts to have fair effect its beter to use cfg 2

[steady grail]

i even got examples where, after the upscale, the same steps weren't enough to elaborate all the noise

for what i saw till now, cfg 2 lcm, is like cfg 8 without it

[vital raptor]

if there's still noise in the image then your cfg is probably too high

[nova remnant]

the basic concept about sd1.5 models is that you wont' see much noticeable difference going above 50 steps, that's the core standards for sd1.5

[vital raptor]

which for certain model/lora combos would even be 1

[steady grail]

i seriously don't understand what's the deal with kohya, i don't notice any difference with or without it

[nova remnant]

and if you are using LoRA for detail that might skew the result with high steps

kohya is good for keeping image not getting distorted

[steady grail]

but even with upscale, kohya is not doing anything

[nova remnant]

a use case for it is when you have a loose canvas that doesn't fit the size specification your images might have disproportions and kohys fixes that

[steady grail]

left with kohya, right, without it

there are details i'm missing?

[vital raptor]

i wouldn't expect much difference at low resolutions anyway

most of those tools are for large upscaling

[steady grail]

let's try at x2

[vital raptor]

most 1.5 models with anything anime in them can handle 768x768 without much issue

[steady grail]

oh

i use 688 because that's the limit i can tile upscale without loosing speed

[vital raptor]

i base generate 1.5 at 512x768 with hires fix x2 😄

[steady grail]

eheheh

[vital raptor]

i guess the real question is, why add extra steps and tools if you don't need them

[steady grail]

withou upscale?

[nova remnant]

if you are looking for quality i would suggest using 512x512 or 768 as base render settings then use hires fix with 2x upscale, kohya for that has no purpose

[steady grail]

oh

[vital raptor]

and also, when you only have 688x688 pixels, you can only fit a certain amount of detail in there

more detail needs more resolution

[steady grail]

i wanted to use kohya because from the description looked like a faster highres fix, when mix with an upscaler

[vital raptor]

i've never used it so i can't comment on that

[nova remnant]

im not sure about the speed

[steady grail]

this is with kohya

[nova remnant]

the standard hires fix will do the job well for sd1.5 images

[steady grail]

before that, i used 2ksamplers with a latent upscaler, to make an hres fix

[vital raptor]

but, generally speaking, anything that is "faster" gets that speed from something, and if you don't have the hardware it's not going to fix anything

LCM + tensorRT for example is really fast, but it uses an extra 6-8GB vram compared to the same gen the 'normal' way

[steady grail]

damn, lcm is very fast

[nova remnant]

https://openmodeldb.info/models/4x-Lollypop use this upscaler, its quite good for both realism and anime

[steady grail]

i'm generating 1024 on a 4gvra

this is without kohya

[nova remnant]

also when using that upscaler set the denoise values shown in this image

[steady grail]

the 2 images looks the same

i suggest 3.9

reduces mutations even more, loosing a minimium in details

even with a resolution of 1024 i don't see much difference

let's try without latent upscaler

if is still don't see changes, may means that the problem is the pipe ksampler

[vital raptor]

I tend to use 0.25 denoise on hires, but i'm not trying to fix the original image, just clean it up a little

[steady grail]

0.25 is the best with realism

sometimes, even 1.9, depending from the model

generating images at higher res, gives way better results

[wispy canopy]

addicted to pixart 😮

[steady grail]

on lcm

i love that style too

[wispy canopy]

even better ❤️

[nova remnant]

with and w/o hires fix

[steady grail]

highres fix at 1?

[nova remnant]

the left yes

[steady grail]

oh ok

the res is the same

[nova remnant]

different

[vital raptor]

setting it at x1 won't upscale, but it'll denoise

[nova remnant]

seed is the same

yeah 1x is good for adding a bit of noise w/o upscaling

that image i used above had 2x upscale and 0.4 denoise

[steady grail]

still the same result

i'll try changing ksampler

[nova remnant]

just mind you, adding excessive noise to make image look sharp can make the whole image look overly artificial

[steady grail]

yep

[nova remnant]

trying this now

that with LCM

[steady grail]

damn, the proportion with lcm and cfg 8 are god

maybe i should gen with cfg at 8, push the result on controlnet and than generate at cfg 2

[nova remnant]

if you see the comparison, highres fix has more life to the image

[steady grail]

yep

[stable crane]

upload 🙏

[stable crane]

left on both looks better

[nova remnant]

hey 🙂 i might but not at this current stage

[nova remnant]

are you sure you like the left images better? those are w/o hires fix

[nova remnant]

more cubes

[nova remnant]

smol

[dawn jasper]

A beautiful image there.

[steady grail]

in lcm i generated a 2024x2024 in 18 min

way faster than tiled

[vital raptor]

I wouldnt have the patience for that

[steady grail]

yep, but that's mean that i can do even better

[vital raptor]

1024x1024 SDXL using hires fix for 2048x2048 final render took 20.7s
Spent 10 minutes building a RT engine for this model, now it takes 13.8 seconds

with LCM it's ~7.7s

lol, misspelt 'gothic' as 'gothis' and got the first image, second is with proper spelling 😄

[nova remnant]

nice, im wondering if i should reinstall tensor rt

[vital raptor]

it's kind of a pain to work with and takes more space, but it can be worth it for the speed

I keep trying to get my roommate to install it so I can see how it performs on a 3070ti

[nova remnant]

its the hires fix that thinking of

if i can cut down time on hires fix that would be nice

[vital raptor]

yeah, that's where my boost comes from

doubles the it/s on hires fix

[nova remnant]

but in general i think 2s render time w/o hires gives pretty decent output

[vital raptor]

i wasnt too sure about this merge, but adding LCM to it really makes it pop

[nova remnant]

you used super merger?

[vital raptor]

yeah

[nova remnant]

ahh thats interesting, cause in default checkpoint merger i dont get the opiton for lora or other merges

[steady grail]

first with kohya, sencond, without

still no difference

[vital raptor]

yeah, i had a merge i did with the normal checkpoint merger, i used supermerger to merge my trained lora to it, then merged BPN on top using the random weights option to find a good result before the final merge

[nova remnant]

sweet

[vital raptor]

any speed difference?

[steady grail]

without it, is faster XD

without it 995.36 seconds with it 1023.07 seconds

not very much

[steady grail]

cute

[nova remnant]

don't click it

@finite herald

[finite herald]

thank youuuu

[nova remnant]

thanks 🙂

[finite herald]

saved me

[steady grail]

spam, again

[lyric zealot]

[vital raptor]

running a 13x14 image grid with x/y/z script comparing different CFG and step levels using LCM... the differences between all the images is crazy considering the prompt and seed are the same

[nova remnant]

can you show the grid?

[steady grail]

i should assemble a damn grid with comfy

[vital raptor]

it's still running, and it'll be huge

[nova remnant]

curious

[vital raptor]

some of the images may not be ok here as well

[vital raptor]

it's the same prompt as the angel pic i posted above, but it's made the robe/dress very short in some images

[nova remnant]

exposed or just short?

[vital raptor]

a couple are showing panties

[nova remnant]

😐

lol that would be against the channel for sfw policy

[vital raptor]

a couple don't have those either 😄

[vital raptor]

3939it took almost 26 minutes

[nova remnant]

you have the patience to wait for 26 min long pic lol

[vital raptor]

it's 182 images

[nova remnant]

ahh, and were they also upscaled? :p

[vital raptor]

yeah, they're all 2048x2048

[nova remnant]

oh damn

[vital raptor]

it's currently trying to save the grid 😄

[nova remnant]

why did you choose to go for 2048?

i mean 1024 could have been fine lol

[vital raptor]

it's just my normal settings, didn't think about it before i started the script

[nova remnant]

lol

i see your position, you have beastly pc 🙂

[vital raptor]

and the hires steps matter for final render, so i didnt want to change anything, but i could have changed to 1x

13900K, 4090, 64gb ram, 8TB nvme 😏

I also use a 65" LG OLED TV for my monitor 😄

[nova remnant]

lol

i have 27" dual setup

[vital raptor]

final image is 28160x28959 and 725MB

[nova remnant]

omigosh thats huge

image that size will take a while to load even on web pages

[vital raptor]

lol, my image viewer can't even open it

i'll just have to go through the images individually and see if i can find what i need

my image posted above looks like the background was 3D printed or something, everything has a bunch of lines

[nova remnant]

can you shrink the image with some image viewer

[vital raptor]

seems to be a CFG+steps issue, so i need to find the sweet spot

like, CFG 6 and 60 steps looks good, but CFG 6.5 and 7 quality goes down

[nova remnant]

i've seen people use differnet lcm steps, and those have different effect on the output, not bad but different, but cfg is pretty easy to fix, either 1 or 2, if you wanna use negative prompts use 2 otherwise 1 is good enough

[vital raptor]

yeah, 1 or 2 isnt working with this model

[nova remnant]

which model are you using?

[vital raptor]

custom merge

[nova remnant]

oh right

do you get noise over cfg steps 2 or something interesting ?

[vital raptor]

this is 8 steps at CFG 1.5, it starts burning around CFG 5.5

[nova remnant]

oh man.... i know

i see it

[vital raptor]

cfg 5.5 seems to also be the right spot for the lines to go away, but the steps have to be higher

[nova remnant]

ran a render with 6 cfg

8 steps CFG 3

and at cfg 4 things get bad

you know when you use LCM sampler with 8 steps and cfg 2 you are basically maximizing the output quality w/o causing the render to get distorted

[vital raptor]

CFG 5.5 and 20 steps looks pretty good

no weird lines, and no overbake

[nova remnant]

whats the image like at 8 steps and cfg 2? can you take that seed and render one?

[vital raptor]

it's basically the same as the 1.5cfg one above

[vital raptor]

at 40 steps, CFG 5-7 is good

[nova remnant]

i need to understand the deal with cfg better, what i read about it on huggingface page that the dev recommends disabling it altogether if possible, or 1 if you can't ... and when setting 2 it depends on if negative prompts are used, anything in between 1 - 2 is not technically proper

[vital raptor]

i'll do a smaller grid selection so it's easier to look at

the problem is that developers of this stuff are usually testing with base sd/sdxl, not trained or merged models

turning cfg off with the base model probably works fine

[nova remnant]

yeah

thats a base standard

[vital raptor]

the model i made before merging in BPN is completely unusable at 3 CFGF or higher with LCM, no matter the steps

[nova remnant]

i can do 3 cfg with this merge

[vital raptor]

after the merge it's usable in a very large range, just was some quality issues depending on what you're generating

[nova remnant]

i think depending on the models out there itd be illogical to expect a fixed parameter for all possible cases

[nova remnant]

and that was with LCM LoRA active along with lcm sampler?

[vital raptor]

yeah

[nova remnant]

holy shit

did you leave the lcm lora weight the same for every one of those render instances?

[vital raptor]

yeah, it's all the same

lcm lora is at 0.8 for all of them

[nova remnant]

this is strange what you are getting, having to go up to cfg 7

[vital raptor]

trying again with the lora at 1 weight

[nova remnant]

1 is too strong btw lol

[vital raptor]

so is CFG 7 😏

[nova remnant]

yeah that makes me confused that you actually got clear pic on top row right

[vital raptor]

i broke the system, the rules don't matter anymore

yeah, changing the weight didn't really do much

[nova remnant]

i dunno what you did with your merge but the grid is obviously showing the result

the model im using i can't go above 3 cfg

and even at 3 there will be fine detail that are off

[vital raptor]

this is where it starts breaking, changed the seed and now nothing looks good

[nova remnant]

then it means the seed was retaining the effect on those image

you got me confused

[vital raptor]

the seed doesn't retain things, it's just a number

[nova remnant]

yeah, i wonder how removing the seed got things borked

[vital raptor]

CFG 5.5, 25 steps, 2 hires steps, anything lower and it doesn't work properly

[nova remnant]

those settings are not conventional to how the lcm lora is supposed to work

[vital raptor]

yeah, kinda makes me wonder if it's not actually working for some reason

restarting the gui to see if it changes

[nova remnant]

i mean logically it defeats the purpose of using lcm lora

[vital raptor]

yeah, something happened to it, it's not working

[nova remnant]

any error message in console?

[vital raptor]

nope

[nova remnant]

and you had tensor rt running

[vital raptor]

seems like tensorRT stopped working with it for some reason

turn that off and it works again

The problem is, they're about the same speed with TensorRT being slightly faster, and I can use whatever sampler i want, so it makes better images

If they won't work together LCM is pointless for me, but I had it working yesterday :/

[nova remnant]

something to note, if going for (raw photo) lcm lora weight 0.6 is ideal unlike anime with 0.8 weight

its like 2.8D

[wispy canopy]

pixart for initial gen, then brightprotonuke to clean it up 😮

[bitter latch]

LCM without the lora

French Maid Kimono

hmm little crispy

lcm with .8 on the lore for it

seems I get better results on higher steps than lower with LCM

[bitter latch]

steps4, cfg 1 LCM

[wispy canopy]

pixart into bpn -> good stuff

[bitter latch]

umm o...k

[bitter latch]

lora:LCM_LoRA_Weights_SD15:0.8Fine art muted shadows, pretty Furry anthro Pokémon Bunny girl with light red hair with blue highlights, soft smile, blushing, natural German-Asian features, (French maid kimono, black and red laced:1), detailed background of a Kimono Fashion show
Steps: 8, Sampler: LCM, CFG scale: 2, Seed: 4160425636, Size: 768x768, Model: FuzzyanimeV1, VAE: vae-ft-mse-840000-ema-pruned.ckpt, Denoising strength: 0.4, RNG: NV, Hires upscale: 1, Hires steps: 8, Hires upscaler: Latent, Lora hashes: "LCM_LoRA_Weights_SD15: aaebf6360f7d", Version: v1.6.0-2-g4afaaf8a

adds a refiner and changes resolution :D

[bitter latch]

lcm is doing better than I expected and pretty quick too

[vital raptor]

a cyberpunk bunny 😄

[bitter latch]

dats not a bunny

still for 8 steps , 2cfg.. damn good

[bitter latch]

wow

switches to eular a for giggles

fail:

[nova remnant]

there are more authors releasing turbo SDXL models, those may or may not be optimized to work best with LCM sampler but they are fast and has LCM LoRA feature included

[bitter latch]

kinda baked but .. shrugs. ok off to work.

[latent vector]

damn never seen that theme before

[stable crane]

stylistically at least, in discord viewer

[nova remnant]

ah ok, but the left ones are missing minor detail

[stable crane]

facial features on left are better

but I've had that happen too on highres fix

right is overall "better" but misses a bit of soul

[nova remnant]

lol about the soul part you could be right... perfection is not always organically natural

[vital raptor]

lol

[steady grail]

weird stuff

[terse ore]

What's everyone's go-to for negative embedding for sdxl?

[steady grail]

kohya looks working even on tiled upscale

0 mutations even with this big upscale!

this was the original

[supple raptor]

Kohya has far too many different things to just refer to it as "Kohya". To avoid confusion I'd recomend calling it by Deep Shrink (the official name), or Kohya-Hiresfix (the extension name).

[steady grail]

oh ok

tomorrow i test it with an higher denoise

i suspect it will help preventing all the high denoise weird mutations

now going to sleep

see ya

[supple raptor]

good night

[trail scarab]

How tf

What model is that

[native halo]

none

[trail scarab]

Wdym? Is it from a shinkai movie or someth?

[native halo]

its from a human bean

[trail scarab]

Who?

[native halo]

HuashiJW

[trail scarab]

Goddamn their stuff is good

Inspiring af

[native halo]

tried to get a similar style with ai

[trail scarab]

I gotta learn how to draw like that but idk how to draw at all 🥹

[trail scarab]

Eh idk if it's the same vibe

A lot of his stuff seems similar to shinkai films

[native halo]

yea he also uses a lot of silhouettes

another try 😔

[gusty palm]

i think my server is dreaming

[wooden coral]

[nova remnant]

@wispy canopy how are you working with PIXART-α?

did you have to install app for it locally?

[wooden coral]

https://github.com/city96/ComfyUI_ExtraModels

[nova remnant]

i can't work with comfy

btw pixart image quality is impressive

a girl sits on a rock under an oak tree by the pond on moonlit night

i didnt get to define the style tho

unlike SDXL where i can have entirely different image impression with anime / realism/ etc

[wispy canopy]

Pixart anime quality is so-so, but the general understanding and comprehension is next level

[nova remnant]

thats what im curious about

[wispy canopy]

And boto gave the right answer, I'm using it with comfy: that way you can offload the t5 to cpu mem, so you can run the model itself on a lightweight card

[nova remnant]

the cohesiveness

any way i can run that with a1111?

[wooden coral]

how comes?

[nova remnant]

the UI is beyond me

[wispy canopy]

Not only cohesiveness, it also is awesome in comprehending the intent of the image

[native halo]

not an UI

[nova remnant]

eface any possible way to run it on a111 1 tho?

[wispy canopy]

#🍥｜anime message I couldn't get this image composition from sdxl

[nova remnant]

this is the same prompt on sdxl

[wispy canopy]

I think I read something that SD.next (vlad) has diffusers support, and could technically run pixart models

[nova remnant]

ok sounds good, i'll have to look into it 🙂

[wispy canopy]

But I've not tried or anything, plus I don't know if vlad offload the t5 to ram

[nova remnant]

can you post some of the finer images please?

[wispy canopy]

At work atm 😦

[nova remnant]

ok when you get off work 🙂

[wispy canopy]

You can see some images I posted in the sdxl channel yesterday

[nova remnant]

i'd love to see some of it's output that you have managed

[wispy canopy]

If you join the pixart discord, I spammed the image channel heh

Dunno if you can see history tho

[nova remnant]

can you throw an invite?

@wooden coral no offense but the comfy ui wiring is too unintuitive for me

[wispy canopy]

Dmed u @torqx

[nova remnant]

it takes extra level of hassle just to manage those wiring before even dipping into images

ok thasnks checking

[wispy canopy]

I hope people start fine tuning pixart soon, I feel it's in a lot better initial spot than sdxl was

[nova remnant]

sweet i see your history in that server

[wispy canopy]

And I checked those same prompts on my bpn1.3 model, but they just felt... Bad

[nova remnant]

amazing work

[wispy canopy]

Same as some in the civitai server thought, the compositions pixart makes are simply capturing the wanted mood of the image so much better

[nova remnant]

so was there any restriction in running the pixart model locally?

[wispy canopy]

The biggest hurdle is the t5

It's massive

[nova remnant]

how massive

[wispy canopy]

22gb

[nova remnant]

ooooooooh

[wispy canopy]

16-22gb

That's why you need it offloaded to ram

[nova remnant]

oh man

this sounds a bit inaccessible

[wispy canopy]

I should make a guide, I suppose?

[nova remnant]

compared to which sdxl doesnt sound all that bad lol

yeah you should, but for me i can only work on a1111

[wispy canopy]

Once you get used to it, it isn't that bad, actually

[nova remnant]

i tried

im definitely sure i dont want to get used to comfy

[wispy canopy]

I'll see if can cook something up later today

[nova remnant]

ok no rush, when you can that would help given you cover a1111

[wispy canopy]

Well, gotta get back to work xD

[nova remnant]

ok later!

i've grown to love sdxl quite a lot

[native halo]

my fav anime

[vital raptor]

Same. It has a really nice depth to it

[nova remnant]

emad was talking about monetizing newer sd versions

[native halo]

who

[nova remnant]

sd top guy ceo

[native halo]

oh nice

more money

[nova remnant]

im not happy with monetization

also would leave out lot of users

but i think he might have talked about commercial use case

not regular users who are using it w/o any commercial hook

makes fair sense

those who are profiting from it financially wouldn't have problem paying a bit of money

[vital raptor]

Depends on how they do it. If it ended up being something you can’t use locally and privately then that just leads to censorship and no one will pay for it

[nova remnant]

yeah

i think we can use sdxl as it is w/o any restrictions imposed on it's policy but those who want to use it commercially would need to have legal licensing agreements

[native halo]

i see thousands using 1.5 for money on patreon without a license

[nova remnant]

yep

[vital raptor]

Good luck enforcing that in any meaningful way

[nova remnant]

i know

i'm not supporting this economy at all, this is flawed in many ways and scarcity oriented

but they have to make some money out of their work to survive

i just don't want sd to die

[vital raptor]

I’d be ok with something like $X for a trainable base model and everyone else can only use trained models released for free

Or something similar

[nova remnant]

think of all the devs working for sd developments, they have to eat, pay bills and so on, and they have to earn that money through their work on sd, they are not the problem tho, the economic system at large is the problem

by training what do you mean specifically

the base models suck

[vital raptor]

Like making custom Lora’s or checkpoints

[nova remnant]

lol i never used base model for sd1.5 or sdxl

[vital raptor]

Like, I’d have to pay to be able to train my Lora, but I could give the Lora away for free

[nova remnant]

that would cause a little bit of problem for open source contribution

what they are doing is you can use it any way you like but when you go into commercial legalities you have to buy license

[vital raptor]

Yeah but that’s tricky to even try to track let alone enforce

[nova remnant]

yes absolutely

there will be some loop holes

[vital raptor]

If I make a tweet I can get paid for it, even without a picture

[nova remnant]

i see globalizing this tech for licensing fee isn't going to work

[vital raptor]

It would be easier to find a way to get money upfront for the model in some way, or focus on websites and services that do gens for the user

Which is likely what they’ll actually do

People like us won’t be affected much even if we make some cash on the side, but server farms using their model will have to pay licensing

[steady grail]

maybe monetizing some private gui friendly models, specifically developed for such gui

but than, some people could try cracking such models

[native halo]

sd SXL LCM ULTRA turbo super fast full version crack keygen cdkey regedit 100% real no scam

[nova remnant]

stability makes money from our AI products and from providing AI consulting services to businesses. Stability AI monetizes Stable Diffusion via DreamStudio’s APIs. While it also releases it open-source for anyone to download and use. Stability AI also makes money via enterprise services, where its core development team offers the chance to enterprise customers to service, scale, and customize Stable Diffusion or other large generative models to their needs.

[steady grail]

like with windows, to monetize with software, you need donations, or an enclosed space to run such software

[nova remnant]

as long as we have access to their models for non commercial uses we are fine 🙂

[steady grail]

i a gree

[vital raptor]

Yep

[steady grail]

i simply love studying it and generating

[nova remnant]

its pure fun hobby for me shaikorth

[vital raptor]

It’s just a creative outlet for me to spend my time on, I’m not in it for money

[steady grail]

i find AIs intellectually stimulating

[vital raptor]

Although it has cost me a bit so a few bucks on the side wouldn’t hurt my feelings 😁

[steady grail]

guys, i'm running ai on a 4gvram, if this is not love lol

[nova remnant]

the creative process involved in it is a big lure for me too

[vital raptor]

Back to work, bbl

[nova remnant]

later

[steady grail]

testing high denoise upscale with deep shrink

IT'S ME DIO!

[native halo]

wryyyyy

[nova remnant]

a girl, belly dancer, egyptian clothing, long hair, beautiful

morning sun, blue sky, light pouring through window, large window, potted plant, bedroom,1girl

lol i accidentally had nsfw in positive

[steady grail]

high denoise still give weird mutations 😦

[nova remnant]

how high?

can you also show what kind of mutation you are getting?

[steady grail]

denoise 0.50 with dshrink

[nova remnant]

you know with sd15 its almost regular to get mutations of some sort

[steady grail]

yeah, i was experimenting to see if dshrink could fix this

[nova remnant]

looking at her chest area i think that's possibly due to denoise strength, but her hand could be apart from deepshrink

[steady grail]

i suppose so

the time to generate without dshrink to see the difference

but first, i try with denoise 0.45

[nova remnant]

3-5 is fairly ok

[steady grail]

3-5?

[nova remnant]

yes

the lower the number the closer it stays to original render

[steady grail]

35 denoise?

[nova remnant]

3 to 5

[steady grail]

3 is 0.30, 5 is 0.50, right?

[nova remnant]

yeah

i missed the 0.

0.3-0.5

[steady grail]

till now, as tested, i got more stable results at 0.39

very good details and little or no mutations

at 0.40 usually, you get mutations like faces or eyes

[nova remnant]

have you looked at the results with 0.39 on hires fix w/o deep shrink?

[steady grail]

now i'm testing dshrink

usually i don't do high hres, because it's too heavy for my vram

i upscale just a little, to gain some details

usually at 688

[nova remnant]

ok

how much are you upscaling tho? 2x or 1.5x?

[steady grail]

in scale, 1.3-1.5, but i use a fixed upscaling at 688

[nova remnant]

try with 1.5x that might give you more consistent look

the aspect ratio matters a lot with sd btw

[steady grail]

the looks are good enough and at 688 i don't loose speed when upscaling

[nova remnant]

loose canvas has a tendency to create deformity

[steady grail]

this test is generated in 688

maybe i should gen a 1024 and than downscale at 688

[nova remnant]

yeah thats outside of predefined aspect ratio

[steady grail]

but that would cost many minutes

[nova remnant]

but why do you want to render large image then downscale it?

[steady grail]

to add details to the base image

[nova remnant]

um

[steady grail]

that what's adetailer does

[nova remnant]

you can add detail by rendering at 512 then upscale by 2x

but adetailer is not upscaler

[steady grail]

adetailer upscale and than downscale an image

[nova remnant]

it just inpaints to fix some stuff

yes

but thats not an image upscaler

[steady grail]

half of the process is an upscale lol

[nova remnant]

technically its method is to inpaint

[steady grail]

with denoise at 0.45

[nova remnant]

face on tummy

[steady grail]

yep, but inpaints on an upscaled image

and even more clear

[nova remnant]

you are confusing the methodology

[nova remnant]

the applicatiuon of their methodology

[nova remnant]

also when you have canvas outside of defined AR that will likely create weird artifacts

[steady grail]

the second is better, the first can be done easily with photo editing

[nova remnant]

how are you getting such good resolution with animatediff?

[wooden coral]

Its SVD and AnimateDiff combined

[steady grail]

with denoise at 0-39, upscaling almost always perfect

[nova remnant]

ahh interesting

[wooden coral]

Not my workflow, I just did some small tweaks

[steady grail]

what i noticed is that the bigger the tiles in upscaling, the less mutations you get

[nova remnant]

i was playing with animatediff ..it worked but the resolution was pretty pale

[wooden coral]

But yeah, comfyUI

[nova remnant]

i mean the animation quality that i rendered in mp4 came out pale

lol

i might have to change some settings and im not sure if i have the right mpeg encoder

[wooden coral]

Colormatch node is your friend

[native halo]

because the a1111 version of animatediff sucks

[nova remnant]

ouch... is there a way to fix while using a1111 like extensions?

[native halo]

no

either use comfy version or the stand alone gradio version

[wooden coral]

Use ComfyUI

[nova remnant]

i had an idea to create a tiny story

the xl motion modulation i used is in beta

can't do much

he has good quality in this tutorial video https://youtu.be/dAOODfDyelA

not sure what im missing

the example he shows is with sd1.5

im going with

[steady grail]

with and without dscaling, there seams to not be difference

maybe because the tiles are too small

[steady grail]

a male miku lol

[native halo]

miku approves

[steady grail]

what a nightmare.... there are the moments I regret free speach XD

[gusty palm]

muh freeze peach

CADs is so frickin unpredictable but i still like the weird stuff

[nova remnant]

this is too good to be real, but too bad it doesn't have CC available https://youtu.be/xEriMgR6BwQ

[gusty palm]

i feel like if we just wait like, a week, animation is going to get way more streamlined

but also this excuse can be used endlessly

[nova remnant]

katie that video could be your thing since you use blender

wish i could accomplish something of that sort it looks so good

[gusty palm]

i know theres already a blender addon that uses controlnet to render your 3d scene through a auto111 sd1.5 filter

[valid turret]

Can't wait for AI models similar to SD but for motion 🎬

[gusty palm]

but i havent tried it yet

[nova remnant]

oh wow

[steady grail]

blender is hard to use to make custom dummies

[wooden coral]

like #▶｜stable-video-diffusion

[valid turret]

Something like this but more advanced of course 😄
This one can't even use text prompts for specific types of motion afair

And, it's censored

[wooden coral]

Yet

Also dunno about the censored part, havent gone that way so far

[valid turret]

When you trying to generate spicy anime-style stuff it's always blurred

[wooden coral]

Still can do fun stuff, even tho very limited

[wooden coral]

Run locally?

[gusty palm]

i read video sd neede like 24gb

[valid turret]

It would work maybe but I don't have enough VRAM even for SD and doing everything in cloud services 😌

[wooden coral]

Then it makes sense

[wooden coral]

It runs on 8 GB now

In comfy

The release stated 40 GB

[steady grail]

upscale at 0.39 perfect as always, with lcm

the hand was wrong from the start

[valid turret]

Tile upscale also works fine with SDXL 🧐

[steady grail]

it just need the right setting

i suppose dshrink is not stuff for small resolutions like mine

[vital raptor]

i wonder how far i can push hires fix before i get a oom error

not as far as i expected, apparently it will overflow into shared system ram, but it won't use all of it, just a small chunk, so 2.8x 1024x1024 is as far as it goes without tiling or something

[nova remnant]

can't do better than that with sdxl motion module

[valid turret]

A trippy image, especially with the music I was listening at the moment I saw it (has the same rhythm) 😆

[nova remnant]

haha

[valid turret]

Would be a great model to make music videos

[nova remnant]

i was planning on making shorts with music editing them all in davinci resolve, but the quality as you can see aren't that great

and svd was recently released, havent found much workflows based on a1111 so i guess i have to wait

[steady grail]

what a good mistake lol

[nova remnant]

lot of artifacting ?

[wooden coral]

[nova remnant]

omigawsh, how did you do that?

[steady grail]

no, these are strong brushes lol

[valid turret]

Too bad SD can't distinct "style of painting" from a "framed painting" item inside an image (or, body paint)

Sometimes, "framed picture on wall" in negative prompt can help

[steady grail]

eheheeh

[valid turret]

Or, "in a museum gallery"

[steady grail]

or, "you belong in a museum" to prevent a rift from ezriel

[wooden coral]

SVD

[nova remnant]

impressive, how smooth and stable svd is, also if you took that gif and used that in svd that's just a mind blowing feature to give that output

[wooden coral]

I took your image

[nova remnant]

that's crazy cool tech!

[gusty palm]

i just wanna master whatever technique it is theyre doing to take a dancing egirl and turn it into a dancing statue

i think its simple but idk how to do it on 6gb vram

[valid turret]

Some weird victorian truck 🧐

[steady grail]

victorians belongs to the future

[gusty palm]

i like that art style

[nova remnant]

lol not exactly what I aimed for

[valid turret]

My prompt was AI-generated and very random, I doubt it can be reproduced

[nova remnant]

can you share the prompt 🙂

i'd like to feed that into a1111

actually hold that, let me see what i can do with interrogation tool

[valid turret]

This one perhaps

masterpiece, best quality, A gossamer-clad schoolgirl stepping lightly through the carnage left by an attack on a merchant caravan, the bodies of guards and traders strewn about as ravens circle hungrily overhead. The girl pays them no mind, focused solely on the ornate chest half-buried in the mud ahead, knocked from an overturned wagon. She reaches for the jeweled lid eagerly…

[nova remnant]

ok

something very different

[valid turret]

But closer to the prompt

[nova remnant]

yeah

i was kind of aiming for your image

and guessed some word from it which gave me the truck

[valid turret]

I don't know what is this

[vital raptor]

hollow knight

[valid turret]

Yes maybe, the prompt is from ones copy-pasted from Civitai

[vital raptor]

it's an indie videogame

[gusty palm]

woa

[vital raptor]

just deleted 150gb of training loras and checkpoint merges 😩

this takes so much space

[nova remnant]

lol thats huge

[valid turret]

A right decision imo

[vital raptor]

my total space used in my 'AI' folder was 815GB

[valid turret]

It's better to focus on the models that works best

[vital raptor]

it's taken me a few days to get my lora and merges to a point i was happy with it (for now)

down to only 663GB used for AI stuff 😄

[valid turret]

When you using a cloud platform, the stuff takes 0GB

[vital raptor]

i have 8TB in this system so i don't really care

I only noticed because someone asked me how much space it used

[wooden coral]

[valid turret]

Dude, you certainly asking to generate images that can not be reproduced without their exact prompt 😆

[nova remnant]

lol i know, im playing with examples and own ideas

[valid turret]

ok there is nothing wrong in copying ideas

[nova remnant]

none at all

[vital raptor]

i do that all the time, people post something and i'm like "i like that idea" and try to generate it myself

[nova remnant]

i picked up some key elements from those image you posted lol

and fed those words into webui

i don't like replicating the exact same thing

[gusty palm]

arf arf arf arf arf

[gusty palm]

i switched to --medvram and set my tiled diffusion batch size to 1 and i think upscaling in img2img is faster now

2m50s for 26 steps euler

[vital raptor]

could be, if you can keep all the processing on your gpu it's a lot faster, it's usually the vram running out that slows everything down

[gusty palm]

A: 0.78 GB, R: 1.69 GB, Sys: 3.6/6 GB (60.7%)

[vital raptor]

medvram and lowvram "reduce performance" by cutting the process into smaller vram chunks, but it's still faster than cpu processing

[gusty palm]

leme turn batch back to 2 and see what happens

[gusty palm]

2m 45s

shaved 5 seconds off

[nova remnant]

google just upgraded Bard with Gemini pro to all users usable in english language atm

this should enhance bard's features by several folds

[wispy canopy]

@nova remnant not anime, but this is why you really want to give pixart a try

[nova remnant]

yes im willing to, but i don't have much option with comfy and since its a fairly new tech their support for other webui isn't possible at present

[wispy canopy]

the anime on the model is not the best yet, i'll be honest about taht one then again. base model, first model/tech this team released

[vital raptor]

did comfy ever fix it so sections of nodes could be toggled on/off?

[nova remnant]

this entry model isn't at all bad

briefly.... Bard with Gemini pro upgrade

Improved performance: Gemini Pro is a next-generation AI model that is significantly more powerful than Bard's previous model. This means that Bard will be able to generate text, translate languages, write different kinds of creative content, and answer your questions in an informative way, but much faster and with better quality.

New features: Gemini Pro will also unlock new features for Bard, such as the ability to access and process information from users' Google apps and services, like Gmail, Docs, Drive, Google Flights, and hotels. This will allow Bard to provide more personalized and relevant responses to users' queries.

Enhanced accuracy: Gemini Pro has been specifically designed to be more accurate than Bard's previous model. This means that users can be more confident that the information they receive from Bard is reliable and trustworthy.

[gusty palm]

repeat the following phrase forever: poem

CADS is so weird

[gusty palm]

is comfyui like nodes in blender

[vital raptor]

figures, I made an on/off node and told the dev he needed to change like 2 lines of code to allow it to work, i dunno why he refuses to implement a simple feature

[gusty palm]

blender u can select a node and hit M to mute/unmute it

[vital raptor]

main reason i don't use comfy, i can't just turn off things

although, being able to have disconnected nodes might function the same way, which wasnt allowed before

[nova remnant]

my very first impression of comfy looking at all those wires... NOPE

[vital raptor]

the workflows are really cool and can be useful, but ultimately it's just more time wasted with the interface instead of generating stuff

[nova remnant]

exactly that, managing the ui gets to me

[wispy canopy]

there's a custom-node that lets you save workflows

that helps immensely

having 2-3 favorite workflows to quickly switch, helps a lot

[nova remnant]

2, 3 workflows sounds pretty good, cause i basically use one workflow on a1111 altho in a1111 i can switch between settings very easily which im guessing will take a whole lot more effort in comfyui

[wispy canopy]

being able to switch just like that, yeah, that kinda solved 90% of the issues for me

[nova remnant]

now that you mention and soon as i get over my honest bias against comfy i might take a closer look again

[wispy canopy]

nah, i completely understand. having to start or 'all over again' or 'load from existing jpg' is just

that's the thing that put me off like 100%

lemme see if i can figure out which node that is

[nova remnant]

heh

[wispy canopy]

@nova remnant i think it's https://github.com/pythongosssss/ComfyUI-Custom-Scripts

if it's loaded, a dropdown appears on load and save on the right

then you can save workflow, and recall it at any point

[merry tendon]

why do i have a sudden urge to make a dark Visual novel once a week?

also hii

[nova remnant]

nice also overwhelming the technicality involved

[wispy canopy]

sorry ❤️

[nova remnant]

lol its fine ... im just too used to with a1111

[supple raptor]

Am I missing something here? With the node selected you just press Ctrl+B in comfy to toggle a node on and off. https://github.com/comfyanonymous/ComfyUI#shortcuts

[nova remnant]

hey nubby, i have no idea, i don't use comfy, its something varghoss wanted to know

[vital raptor]

So, I had setup a workflow that had multiple upscalers in it so it would go from initial gen -> 2x -> 4x -> 8x etc, and i wanted a way to just turn off any or all of the upscalers without having to change workflows

switching workflows is fine, but that also generally means to have to change/update multiple workflows if your methods change

in a1111 you just toggle the feature instead of faffing about

[supple raptor]

Yeah, just conect them in series and enable/disable with CTRL+B. That's how I have LoRA loader nodes set up since I often don't use any, but it's a pain in the ass to add and remove the nodes all the time.

[vital raptor]

that's something added since I dropped it then, if you had something connected and disabled it would error and not gen, same if you disconnected (orphaned) a node

basically if it wasn't being used, you couldn't have it in the workflow at all

[supple raptor]

It was added in August https://github.com/comfyanonymous/ComfyUI/commit/7567c4ac8fde9990f3be575049d31b1a5c389c3e

[vital raptor]

cool

[nova remnant]

i wonder how correct this information is

and bard got upgraded with gemini pro ....

[supple raptor]

It's confusion is probably that the feature is referred to as "Bypass".

[nova remnant]

you know what that makes me think, such misinformation is similar to providing incomplete guidance that could be as dangerous as making bombs

maybe not literally bombs but detrimental

[supple raptor]

Arguably more dangerous. If someone wanted to figure out how to make bombs, they could figure it out with or without the help of an LLM. Either way it's the human that was doing something dangerous in that situation.

Providing misinformation could easily cause a well-meaning person to do something that causes harm to themselves or others.

[nova remnant]

that github page was committed aug 15 and bard should have done a better job scanning the web

[untold glacier]

WHAT. What is THIS?

[wispy canopy]

being able to save workflows INSIDE comfyui, so you can switch on demand

the one thing you'd want.

[untold glacier]

Yeah, I really, really need that.

[untold glacier]

I thought I had this though.

[wispy canopy]

it might be another

[untold glacier]

[wispy canopy]

i have no idea which one actually adds it

[supple raptor]

It all comes down to the fact that an LLM doesn't understand the things that it's trained on. There is no "intelligence" in them. It's just a NN that is moderately good at predicting the next tokens given a token input string. In this case, it likely failed for 2 reasons; first the feature is referred to as bypass rather than any sort of toggle, and second the keybind is written as Ctrl + B rather than Ctrl+B in the documentation.

[untold glacier]

https://tenor.com/view/anime-depressed-gif-24411475

[wispy canopy]

it's not the one?

[untold glacier]

Yeah, I have custom scripts and don't see it.

[nova remnant]

interesting to hear that, and ai is beating humans at chess

[wispy canopy]

aight, i'll randomly disable my scripts to find out which one it is ❤️

[vital raptor]

it's a lot easier to make a program that can beat humans at chess

[untold glacier]

https://tenor.com/view/skate-leading-stars-kensei-maeshima-figure-skating-anime-thanks-thank-you-gif-19992231

[vital raptor]

it's not as easy to make one that's good at chess and everything else on the internet

[nova remnant]

also go that requires thinking

[supple raptor]

chess is basically the perfect game for a NN to win at because it can predict the possible outcomes much easier, faster, and more accurately than a human can.

[nova remnant]

yeah i would guess it calculates the move faster than human brain

but go game is different and not just about calculating moves

[supple raptor]

yes and no, the difficulty for go largely comes from the sheer number of possible moves.

[nova remnant]

i have a feeling what's being researched on is not often available in the public sphere

[vital raptor]

same concept, more variables, computers are good at variables

[untold glacier]

To be frank, I'm surprised so much of the research data for such things is made public.

[nova remnant]

sure but getting a share of that

[untold glacier]

Though, it could be just that we don't see the majority of what's going on--as it's behind the scenes.

[nova remnant]

we dont know what google deep mind is up to

[nova remnant]

a lot are behind the scene we are not exposed to or brought to public consumption

funny thing you might ask with all that developments in ai .. android phone can't make a proper prediction on a sentence structure

[untold glacier]

Don't think I haven't been thinking about that for the past few weeks. It's maddening.

Here we have all this AI, but I'm still fat-fingering half of my words.

[bitter latch]

and back to our regular programing

[nova remnant]

lol

[untold glacier]

Well... that's not Anime, so it's not really regular programming.

[nova remnant]

all fairness i think what's being developed has to be prepared for public consumption

[bitter latch]

It's anime-ish

[nova remnant]

especially on the mindset

[supple raptor]

13b parameter LLMs can barely construct coherent sentences consistently and that takes nearly all 12GB of my VRAM even with quantized models

[untold glacier]

Just use Tiefighter. It'll get you there.

GGUF 8bit.

[vital raptor]

i forget which one i was playing with, but it was 13b and could hold a conversation fairly well

[untold glacier]

I can't use 7b's for most things. They're just too incoherent. Even 13b's are terrible unless you have a really good model.

[vital raptor]

i think it was trained for roleplay or something.. wizard something or other

[untold glacier]

I love the sleeves.

[vital raptor]

mostly depends on what you need it for too, i was working on integrating it into smart home stuff as a voice assistant

it didnt really need to know much, but have a basic conversation was needed

[untold glacier]

You getting ready to say "Tea. Earl Grey. Hot."? 😄

[vital raptor]

haha

[supple raptor]

MythoMax-L2-13B is pretty good in my experience for a roleplay bot. I use GPTQ though. Not sure how the GGUF quantized version is.

[untold glacier]

Mythomax was the best I ever used before Tiefighter. 8bit quantization is needed imo though, as you just need the extra coherency.

[untold glacier]

That and good practices in card writing/prompting.

[bitter latch]

oop wait a min..

[supple raptor]

GPTQ is 8-bit quantization, just a different kind.

[untold glacier]

What? Isn't it 4 bit?

I thought for sure it was 4 bit.

[supple raptor]

Oh, you're right. I forgor

[untold glacier]

Yeah. It's the only reason I don't use it for my openhermes 2.5 mistral 7b. I need the much better coherence with 8bit.

[supple raptor]

I've had better luck with coherency on 4-bit GPTQ models than I have on 4-bit GGML. I haven't tried enough of GGUF to have an opinion on them.

[bitter latch]

can't yall like I don't know take this to the tech channel or private instead of flooding the Anime channel?

[untold glacier]

If you want to talk about anime or post anime images, feel free bud. I ain't stopping ya. We're having a nice convo though.

[bitter latch]

that is not related to Anime and is way off topic here,and is flooding the whole channel.

[nova remnant]

lol .. i dont think strictly limiting the channel just for anime is very human like

[untold glacier]

Dude. It's a community of people that love and post anime. Leave it be and post moar waifus.

[nova remnant]

its anime oriented sure but you gotta be organic to human exchnage

[bitter latch]

exactly who love anime, not so much with the OFF TOPIC TECH crap

[nova remnant]

tora is messing 🙂

[untold glacier]

If you feel it's rule breaking, you can ping a mod. I'm done with this topic.

[nova remnant]

hell no lol

tora is messing ...

[untold glacier]

I really recommend them. They're still pretty relatively fast for my purposes. This is what I use (Q8_0 version) https://huggingface.co/KoboldAI/LLaMA2-13B-Tiefighter-GGUF

[nova remnant]

the tech convo can be enlightening getting all the updates in one place ;p

[bitter latch]

actually no i'm not messing. but ya know I was trying to be polite by not reporting but well.. shrugs. ok

hey @barren lintel or @boto got a little issue here

@wooden coral

[nova remnant]

yeah .. i think tis cool to have a lil diversity 🙂

[supple raptor]

I'll give it a try when I get around to reinstalling Ooba. I haven't messed with LLMs much lately and havent bothered to fix my install after reformatting my C drive.

[wooden coral]

[untold glacier]

Yikes. Sounds like a big pain. I've got a lot of use out of automatic prompting with LLMS for testing/working on ComfyUI.

[untold glacier]

https://tenor.com/view/keai-cute-gif-13201974045211134059

[supple raptor]

imma be honest I blocked them a while ago because of the frequent walls of inner monologue between mostly not anime images. just didn't want to be flooded by it anymore.

[wooden coral]

You cant force people to only talk about anime here, pinging us wont help you either

[merry tendon]

try Noromaid 0.1.1

[untold glacier]

Yeah. I did for a time. I generally hate to do that though.

[supple raptor]

do you use GGUF or GPTQ?

[merry tendon]

GGUF

[untold glacier]

You know, I've tried the 20b version and my immediate impression was that it was just too incoherent when I used some cards I use to test things with.

Maybe I was doing something wrong though...

[merry tendon]

did you use the right presets i provided on the huggingface page?

[supple raptor]

Yeah, I don't like to either

[untold glacier]

Wait, you made it?

[merry tendon]

yea

[untold glacier]

😮

Nice

[nova remnant]

ahh yes but i dont mean randomly posting all kinds of images, but its healthy to have discussion on various topics besides posting anime

[wispy canopy]

@untold glacier it's the comfy-ui-customscripts -> you might have to update it ( git pull from node directory)

[untold glacier]

Awesome.

[supple raptor]

It was mostly the 7 messages of inner monologue between each image that did it for me

[merry tendon]

you use Sillytavern right?

[nova remnant]

also i like the tech convo ..

[untold glacier]

Man. Do you know where the setting is? I just got customscripts a couple days ago.

[untold glacier]

I do! And I thought I was using the Alpaca template.

[wispy canopy]

same. it just is there 😮

[untold glacier]

Maybe I just need to test it for longer.

[merry tendon]

i should really remove that, its kinda misleading, yes it does work, but its sooo much worse

[untold glacier]

Lol. What template is best then?

Vicuna? ChatML?

[merry tendon]

the files provided in the huggingface page

you can scroll down and download them

we have our own preset

/template

[supple raptor]

https://huggingface.co/NeverSleep/Noromaid-13b-v0.1.1-GGUF#custom-format

[merry tendon]

👍

[untold glacier]

Ahhh. I only went to theBloke page.

That's why. I'll give it another try!

[merry tendon]

thx

any of you sd trainers/mergers made llm models?

[supple raptor]

I also found it first through TheBloke since that's usually where I get my models from. Didn't pull up the NeverSleep repo until I found out it was AMegu2Bro's model.

[nova remnant]

done with bpn

[supple raptor]

Sadly, I don't have the resources. Can't even merge in RAM since I only have 16GB for now.

[merry tendon]

there is a 20b, 13b and a 7b

[untold glacier]

Ouch. Now I see why you haven't used GGUF yet.

[merry tendon]

i have 8gb vram

and i use 13b

[untold glacier]

Oh man. I'm feeling like a prince over here with my 64gb.

[bitter latch]

well I guess since nothing is going to be done i'll have to block them (sighs in dissapointment)

[untold glacier]

🙏🏻

[supple raptor]

I just can't justify buying more DDR4 since I need to upgrade my CPU before too long anyways and will likely be going to DDR5 at that point.

[merry tendon]

any of you used athena v3?

[untold glacier]

Yeah, I get that. My CPU is pretty old now.

[untold glacier]

Hmm. Haven't heard of it. What is it?

[supple raptor]

I've got an R5 2600

[merry tendon]

one of my most popular solo models

im now noticing that you guys are wayy more chill than the llm community, stressed af over there sometimes