✨｜sdxl

[tribal jackal]

Gotta find a good realistic xl model. This is all just the base sdxl so far.

[rich coral]

I am using dreamshaperXL10 and loving it. it can do more realistic than the 1.5 version did

[tribal jackal]

hmmm, the gallery does show off its versatility even if the sample uploads were blurry

I kind of like my current workflow but I shouldn't avoid it out of laziness

It is a little tricky to avoid sometimes yeah

[livid basin]

I've also noticed

[heady vale]

Messing about with img2img

[livid basin]

Using the refiner gives photos a very painting look @trim orbit

lol @heady vale

What's your workflow

[glad flint]

Does anyone know a method get seamless tiling textures in comfyui the same as it is in automatic?
There's a seamless texture node in WAS node suite but that's just a simple gradient overlap algorithm used as post processing that will muddy the picture.

Whatever auto tiling checkmark is doing, affects the generation every step of the way. I keep trying to figure out how it works but I get lost trying to follow function calls around 😅

[tribal jackal]

Still new with comfyui, but could you copy paste the node and put each copy at a different setting then output to multiple preview images?

[livid basin]

AFAIK, AUTO sets the unet to use convolution padding to circular

So the network always gives you seamless textures

[soft zealot]

I delete the outputs folder in side Comfy and then create symlinks to the folders comy expects to the actual locatins.

Have a .bat file I can run when I update COmfy to restore the symlinks

NB of you aren't familiar with symlinks these commads need to be run inside an elevated permissions cmd window

[peak dove]

Triple Workflow Refiner-Base-Refiner https://drive.google.com/file/d/1LLIY5IGwGFQ7fwiAg1VpysPD6g2Xorpo/view?usp=sharing

[high skiff]

Just submitted my 1.0 workflow for testing by Comfy :p

[spring fulcrum]

So are you saying its still not public 😦

[peak dove]

I really don't understand that Symlink diagram ... ? 😦

[high skiff]

I said the goal was late today or late tomorrow (its 10Pm for me)

if all goes well, it will be up in a few hours

[vocal rapids]

can't wait and thx @high skiff

[ornate elk]

this is what I get when I ask for a space station orbiting neptune 🤦

[spring fulcrum]

Its all good I'm just being funny.... I really appreciate all the work you have been doing. 🙂

[rich coral]

I find SDXL still likes word salad

[soft zealot]

https://www.maketecheasier.com/create-symbolic-links-windows10/

[ornate elk]

it like, super duper likes to make a space station OUT OF neptune

[glad flint]

asdhfkasdfjkdah accidentally loaded SDXL 0.9 in auto, now I have to wait a long time again

[high skiff]

I just need Comfy's confirmation, and to find the sample prompt I wanna ship it with

[peak dove]

If I have my models at C:\Stable-diffusion\models\Stable-diffusion - what would a Symlink look like to ComfyUI B:\ComfyUI_windows_portable\ComfyUI\models\checkpoints

[hoary saddle]

@visual glade wake up. You've got mail.

[high skiff]

Hes testing it lmao

[spring fulcrum]

lol... we the people... are hungry for ai art workflow

[high skiff]

this is the current default prompt

I wanna change it up a bit, maybe

[heady vale]

do you have a pic of architecture or vehicles etc

[high skiff]

I can generate some

[soft zealot]

Noafter making sure the comfy folders are empty (as the frst line deletes them the format for you would be

FYI if you're doing thi sfor the first time on an exisiting install I would rename the folder in comfy from say checkpoints to OG_Checkpoints first

rmdir /s /q "B:\ComfyUI_windows_portable\ComfyUI\models\checkpoints"
mklink /D "B:\ComfyUI_windows_portable\ComfyUI\models\checkpoints" "C:\Stable-diffusion\models\Stable-diffusion"

[livid basin]

I dig it

A space station orbiting neptune orbiting earth

[peak dove]

I'm trying mklink in an elevated command window

[ornate elk]

neptune has been ensmallened. or earth embiggened

[high skiff]

It does vehicles really well

just checked

crop of the low res

vs the upscale

this upscale is insane, I can't believe how well it works, especially with how fast it is lol

[urban breach]

i tried to upscale and add more detail. and i reduced detail or unfocused the background or added weird new body parts to the subject lol

[high skiff]

yeah, I had to work on a whole different way of doing it to get it to work well

[urban breach]

waiting for the experts 🥳

[heady vale]

Looks good. Im already getting similar results but yours is probably much faster

[high skiff]

I increased the quality, full image gen at 2048x is taking about 45 seconds on my 3080

let me try architecture, tho from what i have seen, its probably gonna be amazing

this is high res fix after all, it fixes a lot of deformities

[peak dove]

This worked a treat - thank you 🙂

[high skiff]

@heady vale Tested on some sleep and angular arch

*sleek

[heady vale]

Looks good. Nice work

[high skiff]

you can also control how much it upscales, from 1x to 4x, as well as how much "fixing" it does

so you can choose to stay faithful, or change things

[urban breach]

Thanks it's a nice workflow, pics are looking pretty good. I like the batch display.

[high skiff]

my fix is going to ship at 33% fix

[peak dove]

Here's some weird stuff using R-B-R - Prompt = Pointillism magic surreal grand guignol peruvian arpillera venice carnival psychedelic Avant garde art in the style of yoh nagao Eric madigan heck yayoi kusama

[glad flint]

sdxl results set to tiling in automatic looks mangled compared to 1.5 🤔

I haven't done a lot of tests though to compare because auto is crashing too much and it takes several minutes to load the model each time

[peak dove]

I have come to the conclusion that A1111 is getting left behind (to ComfyUI) in the SDXL Stakes - but that's just me!!! 😄

[trim orbit]

sdxl tiling will probably take a bit to rejigger for the 1024 resolution. inpainting at that scale can be different i believe

[high skiff]

Automatic is genuinely just a way worse experience for SDXL, especially now that I have a working high quality and reliable high-res fix workflow, of which automatic is not capable of implementing

[spring fulcrum]

That seems spot on... A1111 just can't seem to do the same things

[high skiff]

And the reason I say it's not capable of implementing it is, it is also based off of my mixed diffusion, which automatic does not have support for

[glad flint]

tbf i'm using the dev branch and next time I launch auto it will be to 1.5 inpaint a bunch of my SDXL images

[midnight shuttle]

It was easier to include SDXL in Comfy because the interface is already all about arbitrary workflow. It's harder for A1111 because that was designed around a single model workflow not a dual-model.

[high skiff]

The fact that automatic doesn't have any support for mixed diffusion puts it at a huge disadvantage for SDXL

[peak dove]

... and of course, Comfy himself actually works at Stability AI 🙂

[midnight shuttle]

And it was not designed in a maintainable way so a lot needs to be undone in order to make it dual-model.

[high skiff]

Pseudo and I worked together in order to port my mixed diffusion pipeline into diffusers, and it improved diffuser outputs considerably

To the point where the developer for diffusers has implemented it as his default method now

[ionic dragon]

What about the styles prompt?

[high skiff]

You can very much use the same prompt to cross both text encoders, however my research for my workflow has concluded that proper separations between terms for the first and second text encoder can result in even higher quality images

[trim orbit]

i'm using auto right now just playing around with the new hires stuff. i'm not using the refiner as i'm moving away from using it in primary passes now. i'll bring it into play for situations where i want it. i've found that the base model scales up very nicely.

just doing a straight 2x upscale pass in auto like i would before has been working quite nicely. just the vae takes SO long to cook and there are a few memory leaks. i wouldn't say incapable. Comfy is still better for SDXL simply because of the much more required optimizations in place.

[ionic dragon]

@primal vault how would the styles prompt of this bepost-apocalyptic survivor overlooking abandoned city, overgrown ruins, hazy horizon, muted tones, artstation, concept art, detailed textures, art by maciej kuciarawould it be muted tones, artstation, concept art, detailed textures, art by maciej kuciara

[high skiff]

I have plans to release a stripped down version of my workflow that only uses the base model, and one positive text encoder

[ionic dragon]

Oh ok, so 3 prompts aren't necessary?

[trim orbit]

the new stuff is just switching to a different model for the upscale pass. but i haven't been doing that anyways. i'm thinking the refiner feels more like noise offset, only sometimes needed

[ionic dragon]

How improved were the images while using 2 text encoders?

[glad flint]

automatic tiling for SDXL gives me stuff like this lol

[high skiff]

yeah, while that is true, Auto also doesn't have support for some functions that enable all three of my SDXL tricks to work

all three of my functions (Mixed diffusion, fractional offset, and SDXL high res fix) are not compatible with Auto cause it has no support for advanced sampler control

and i don't see it coding in all of that any time soon

[trim orbit]

the new branch seems to.

[peak dove]

post-apocalyptic survivor overlooking abandoned city, overgrown ruins, hazy horizon, muted tones, artstation, concept art, detailed textures, art by maciej kuciara - done using R-B-R

[high skiff]

have support for advanced sampler control?

[trim orbit]

different samplers on the second pass anyways. i'm not sure the advanced granularity that we see in comfy nodes are really an end user requirement. i've not seen a use case

[high skiff]

They are required for any of my 3 tricks to work, of which, all 3 are very beneficial and efficient

[midnight shuttle]

Comfy and A1111 will probably never have exact feature parity and that's fine. A1111 does some things that Comfy doesn't and vice versa.

[high skiff]

if Auto were to add the features I need to implement my fixes into it, I would be willing to, but it needs to add a lotttt of things, especially for my high res fix to work

[clever fable]

which upscaler are you using?

[high skiff]

my own custom workflow, which should be released in a few hours on the comfy wiki

[clever fable]

nodnod, thanks ^^

[high skiff]

no worries

[peak dove]

ComfyUI is ahead of A1111 as you can lift the hood on ComfyUI - so to speak

[hot rover]

Had some time to fiddle with settings today, Day 2 of trying to wrap my head around comfyui. I'm blown away by the detail some of the outputs are giving. 🤯 Spent tonight trying to do a face refine workflow for wonky human faces and got a little closer, but still not sure if I'm doing it right.

[trim orbit]

guess i'll have to wait to see the 1.0 release. all i hear is hype so far. your fractional offset trick honestly is a fail more often than it wins. i was whelmed by it. it's a neat novel trick that can be done with the way nodes work. not sure of it's use case still though.

[midnight shuttle]

Code quality and maintainability is a lot better. But there are some workflows (meaning processes, not node layouts) that A1111 handles a lot better.

[high skiff]

I have some friends testing the crap out of my upscale right now, and its working very good across different styles, so I am happy for that

[ionic dragon]

what does hires fix mean in a1111?
does it pass the image through the ckpt again?

[high skiff]

yeah, my fractional offset is the least important, but the mixed diffusion and especially the high res fix are a big deal, which need the sampler control, otherwise they are not possible in auto

[peak dove]

A1111 has Latent Mirroring; and a great Image Browser... !

[high skiff]

altho, I will say, I found some more use cases for my fractional offset, but I will be keeping those to myself, as I can't give away EVERYTHING I learn :p

they are untested at the moment as well

or I should say unproven

[midnight shuttle]

I hope A1111 can somehow get updated to work well with SDXL and that Comfy also keeps getting better. They are different tools for different people. Not competitors.

[warm rivet]

The staff already the best parameters found on the bot? I am waiting for it 😦

[trim orbit]

freckles i thought it'd do well for but prompting for them works better

[high skiff]

not that I know of, and I wouldn't really listen too much to their results either

[clever fable]

SD.Next moves quite a bit faster than A1111 and the maintainer is more responsive. it's pretty good in the meantime

[high skiff]

its data from chery picked pre-prompts, from people who were looking for aesthetics over prompt adherence, running in a less efficient way, with excessive steps, and tons of randomized parameters

It was all too inconsistent to draw any major conclusions from IMO, especially compared to the way my workflow runs, as my workflow is optimized for different samplers than theirs

[midnight shuttle]

I haven't tried it because I heard it has a lot of problems and crashes. But maybe that's not true.

[high skiff]

:>

walph

[trim orbit]

they don't use the data raw like that. they evaluate the results themselves and the votes are just clues for them to hone in on what people are seeing

[high skiff]

fair enough, tho I will say that for my workflow, I have already proven that their data is not applicable to the way I do things, as I ran some small scale tests to compared, and found very different conclusions to them in this community (specifically on samplers and step counts)

[peak dove]

A1111 trashes my 8Gb VRAM RTX2070 in SDXL - ComfyUI is just that - my GPU is Comfy in SDXL!!!

[warm rivet]

Do you run different cfg and samplers frequenly? 🤔

[clever fable]

it's worth a try for sure. you'll get updates a lot quicker

[trim orbit]

my latest auto1111 image with the hires pass on the bass model. i honestly think it's fine. The obvious ai artifacts come more from a phoned in prompt approach than anything else

the worst part of the auto process is the VAE takes minutes

[high skiff]

no, but I did test them for the core part of my workflow, and found that most samplers were not compatible with my workflow as a whole (I believe 11 of them all crapped themselves), and then the ones I didn't pick were inconsistent

the sampler I use is reliably solid, whereas some of the others have a tendency to either be great or trash

[tribal jackal]

I wonder if reducing the contrast is better as a negative text prompt or an offset lora added to the workflow

or both

[high skiff]

I would say it consistenly gives you 80-90% quality results, where the others that worked were more like 40-100%

[midnight shuttle]

This is because Comfy loads and unloads models as needed. A1111 crams them all into VRAM at once which is faster if you have enough VRAM but a LOT slower if you don't. You can use --medvram in A1111 to make it behave somewhat more like Comfy in terms of memory.

[spring fulcrum]

Is there any way to remove the ones that don't work completely?

[warm rivet]

What is your current parameters now if you dont mind? I am usign dpm 2m keras with 20 steps 7 cfg most time

[high skiff]

yeah, I omitted them from my tests

[peak dove]

I have tried --medvram and --lowvram - but it pushes my GPU to its limit - spitting ERROR messages as it goes 😄

[spring fulcrum]

Did you list the ones that worked well?

[high skiff]

yeah, its been close to the same since my 0.5 release

DDIM, Normal, 25 steps, 7.5 CFG is the go top rock solid base for my mixed diffusion base

[midnight shuttle]

Something else might be wrong then. I've looked at the code so I know what those options really do.

[trim orbit]

think i'm going to run it with --med-vram just so it decides to try to avoid doing that

[high skiff]

let me see if I still have the list

[mossy gazelle]

A whimsical 3D illustration of Barbie and children, surrounded by colorful bubbles reminiscent of Pop Mart blind boxes, with a cartoonish style and playful details. Camera: Nikon D3. Lens: 35mm.

[warm rivet]

Dimm? Interesting, I was using it for fast gen in the past. I will run some tests too.

[high skiff]

the ones that didn't have a tendency to explode were:

DDIM
Euler A
UniPC
and DPMPP SDE

[peak dove]

I saw that movie - DONNIE DARKO?! 🙂

[high skiff]

DPMPP SDE did have higher quality results some times, but consistenly had worse ones as well

[trim orbit]

honestly is a trip trip movie. i was reminded more of aeon flux / matrix white rabit

[high skiff]

so it was a little all over the place hit or miss

[ornate elk]

I didn't reference that!

[south igloo]

Wish comfy could do high-res fix pass it seems far superior to the refiner then normal upscale

[high skiff]

I am preparing my reddit announcement for release right now

[high skiff]

my workflow getting ready to drop can do damn good high res fix

[royal fern]

I did this yesterday... not sure if it's of any help. 240 styles in one picture

[peak dove]

Scott Detweiler's favourite model DPMPP SDE GPU

[trim orbit]

it can. just build the node chain up

[south igloo]

Awesome can you please ping me when it drops I really need a workflow like that!

[south igloo]

I'm too dumb to do it sadly lol

[peak dove]

... and who knew that Karras actually works at nVidia!!!

[trim orbit]

i know. me too man. me too.

[high skiff]

like, damn good

first image is crop of base 1024x, second is my high res fix at 2048x

[peak dove]

Scott Detweiler has demonstrated a hi-res fix

[south igloo]

Yeah that's so much better

!!!

Not sure why but high-res fix just makes the images so detailed

[peak dove]

Is "Hi Res Fix" just a three word phrase for "Sharpening?!"

[high skiff]

for my high res fix, the final image is 4x the pixel count

so it goes from 1024x1024 to 2048x2048, while also fixing and refining details and deformities

[peak dove]

So Topaz GigaPixel and Topaz Sharpen AI - ... 🙂

[high skiff]

no lol

thats a way worse version of what I am doing

[midnight shuttle]

I think A1111 Hires Fix isn't really "fixing" anything but it is just using multiple passes with the model and then stiching those together to make a higher resolution image.

[high skiff]

ah yeah, that has the crusty pixel upscale look

[midnight shuttle]

SDXL VAE uses a lot of VRAM.

[high skiff]

static in a lot of areas

I had that issue when I was starting my SDXL venture

its a cool image in general tho

I like it a lot :p

[ornate elk]

if you guys like cyborg rabbits here are more

[high skiff]

love

[trim orbit]

naw thats just bad prompting on my part. i get that effect without any hires pass when i phone the negative in

[high skiff]

ah, fair enough then

[trim orbit]

it comes from my magic word habits. "fujifilm"

[ionic dragon]

the man with beard looks better on the left side(a lot more depth and texture) but the eyes are off

[heady vale]

so that was without refiner pass?

[trim orbit]

or the other good one i use "magazine centerfold" but then you get that crusty offset print look

[trim orbit]

no refiner

all about that base

[ornate elk]

please make the dog the lead singer

[heady vale]

base can do a lot of good gens

[high skiff]

wait a sec

did somebody steal my workflow and pose as me on my own repository?

WTF

bro what lmao

[ionic dragon]

how can they?

your acc got hacked?

[peak dove]

Tries to envisage what "posing on my own repository" might look like as a prompt?!?!? 😄

[high skiff]

oh, they added a low quality upscale ontop and then reused my whole read me linking to their donation pages instead of mine

wow lmao

that is scummy as hell

[ionic dragon]

they forked yours?

[trim orbit]

https://github.com/SytanSD/Sytan-SDXL-ComfyUI/blob/main/LICENSE

[high skiff]

like, they did credit me, but they took all of my writing and just slotted their name into it

[peak dove]

Cyborg Rabbits in Post-Apocalyptic City

[high skiff]

they legit took all my writing and just put their donation pages in its place lmao

[trim orbit]

you literally gave it a gpl

[peak dove]

Get that Shinguular guy banned!

[high skiff]

I know I did, for people to add to it, not pose as me

[trim orbit]

you don't really get to decide how people use the gpl though. that's a different kind of restrictive license

[midnight shuttle]

Unfortunate side effect of GPL and many other open source licenses.

[ornate elk]

looks like he might be chewing on it

[high skiff]

alright, then 1.0 will be released more strictly, cause I am not cool with people copy pasting my work and posing as me pretending like their donation pages are mine lol

[tribal jackal]

I'm sure some of that was involved.

[meager canopy]

Almost 😄

[ionic dragon]

damn, ppl can really fall into the trap, can you send the link to their repo

[civic sigil]

Omg I didnt realize how big of a difference the negatives make

[high skiff]

https://github.com/ShinguLari/SinguLari-SDXL-ComfyUI

They made a fork, and just reused all of my test with their socials and stuff

it does say at the very end that it originated from me, but they even filled in my donation links with their own name and stuff

(we inch closer to release :p)

[spring fulcrum]

What happens if I report the repository?

[ionic dragon]

for hires fix should we use both base+refiner or just refiner?

[high skiff]

no idea, I would say don't worry about it

[high skiff]

its all preset in my workflow

[ionic dragon]

oh ok

[trim orbit]

whats wrong with it? seems to follow the gpl don't it

[ionic dragon]

will check

[trim orbit]

yeah. if you want to maintain control of how it's used afterwards, you need a less permissive license.

not trying to be xeno, but i'm not surprised that fork is chinese

[civic sigil]

I thought that was japanese?

[trim orbit]

maybe i'm being a little xeno

[civic sigil]

😂

[peak dove]

China has the tradition of "stealing everything" - and that's not blatant prejudice - that's a fact!

[high skiff]

I didn't realize that the GPL would just allow people to pose as me and use my about me to promote their own donations and stuff

[trim orbit]

double edged sword these copyleft licenses can be

[peak dove]

Prompt - Posing On My Own Repository!!!

[civic sigil]

I mean they are translating the readme and releasing tutorial videos in their native language

[high skiff]

and substituting my name for theirs, and my links to my donation pages to theirs ._.

[civic sigil]

I cant read it so its hard to say if they are claiming its stheres but yeah a little sketchy for sure

[ornate elk]

oh I gotchu fam

[high skiff]

oh well, I don't care much, their workflow seems to have been slapped together with no real knowledge of what they are doing, and they have gotten little attentionf or it, so I will just let it be for now

[trim orbit]

FOSS is freedom as in beer, but more importantly, free to do what you want with it

[mossy spindle]

What is the difference between base model and refiner? Which one do I use?

[peak dove]

All my CC0 (License) photos are for sale over in China ... but CC0 lets them do it without theft!

[trim orbit]

refiner is an optional second model for adding detail and refinements to finalize an image

[ornate elk]

refiner: it's not very useful. send your images generated with base to img2img and run the refiner at a ~~cfg ~~denoise of around .2. But it depends on what you're doing.

[high skiff]

oh man, that is not at all how you should do that haha

CFG 0.2?

[mossy spindle]

Then Should I use both of them?

[high skiff]

hmm

[spring fulcrum]

is there already an updated upscaler model to use? other than 4x-UltraSharp?

[ornate elk]

sorrry denoising

[high skiff]

my new workflow comes with one I found works just a litttllleee better, it will be on my github

I am preparing my reddit post for launch

[ornate elk]

at least for postapocalyptic cyborg rabbit illustrations, it's not very effective

[median isle]

I swapped from Voldy's WebUI to ComfyUI and I'm having real big troubles with the eye department. Is there a workflow for Comfy that would automatically find faces and redo the face area?

[high skiff]

ah shit, NVIDIA ICAT doesn't work on images at a lower res

aha, fixed it

dayum

[trim orbit]

on the right seems cockeyed

[ionic dragon]

@high skiff you workflow didnt include hires fix

[high skiff]

dude

its not out yet lmao

[ionic dragon]

wait

oh

[high skiff]

I have been saying this for so long lol

I keep telling you lol

its releasing soon

like, within the next few hours soon

[ionic dragon]

i though in the one on github

my bad

[high skiff]

damn, what am I supposed to do here

hmmm

I need them scaled together

[civic sigil]

Wow very clean

[ornate elk]

why does SD do this?

happens with darks too

[nimble heart]

was asleep at totally regular healthy human times didn't see this

yesterday I updated the Principled node a bit. Highlights include refiner clip/model being optional and an experimental N-Pass upscale that turns your GPU into a space heater
https://github.com/Beinsezii/bsz-cui-extras
kinda been neglecting my two non-principled workflows, might go back and update them some more sometime soon
Also despite refiner being optional, it's still not usable for 1.5 due to it using the SDXL conditioning nodes internally

I've updated every day so there might be more depending when you last looked
https://github.com/Beinsezii/bsz-cui-extras/commits/master

[civic sigil]

Couldn't you normalize the colors to fix this?

[ornate elk]

yes. I am asking why it often comes out like this

[civic sigil]

I just checked one of my gens and it doesnt seem to have that problem but I am using a lora so idk

[ornate elk]

which one?

[civic sigil]

Custom

Did you try that offset noise lora that shipped with xl?

[trim orbit]

the base model is trained without noise offsetting. it averages pixels towards a middle point unless you use a tooled lora to offset it towards darker images

[ornate elk]

that example is missing lights

[trim orbit]

mb. probably the same reason. it's trying to average the picture towards the midpoint. noise offset shifts it away from center

[civic sigil]

It does look a little washed out

[trim orbit]

does this work?

heyyyy it works

[ornate elk]

yeah that's deliberate. it can be hard to see this with the naked eye, just provided the image as an example

[trim orbit]

that file is the offset lora

[ornate elk]

ah is that what that is

I have that

sometimes I use it, but I never notice it...but I am usually not looking at histograms just mashing generate like a crazy person

[spring fulcrum]

@high skiff What was that voice software you were trying the other day... RCV or something? I can't remember... I figured I could play around with that while I patiently go crazy waiting for your Comfy Release to drop?

[peak dove]

https://drive.google.com/file/d/1LLIY5IGwGFQ7fwiAg1VpysPD6g2Xorpo/view?usp=sharing

[trim orbit]

yeah i hear that. you can often see it when it really rears it's face and you're looking for it. some prompts benefit better from offsetting so when you see that flat washed out look, try it

[ornate elk]

actually I was more curious than anything. not looking for hyper fine technical quality, mostly happy surfing the latent space for interesting stuff

[high skiff]

yeah, RVC

[spring fulcrum]

Is this the one?

https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/blob/main/docs/README.en.md

[high skiff]

I believe so

[peak dove]

Can I do Donald Trump with this?

[glad grove]

yes

[high skiff]

yes

[ornate elk]

woah cool, I am looking to make a short written by ChatGPT, Voiced by elevenlabs, art directed by SD, and shot by runway. Was just looking for something I could use for music.

[high skiff]

I have a trump model lol

[peak dove]

Woo hoo

[ornate elk]

btw elevenlabs is amazing

[high skiff]

elevenlabs is cool. but RVC can achieve way higher quality

[ornate elk]

Does is have emotional control?

[spring fulcrum]

It must be FANTASTIC... I mean The BEST model EVER... Truly AMAZING (in trump voice)

[ornate elk]

I find elevenlabs to be nearly indistinguishable from reality for some voices

[high skiff]

oh, you have to put voice into it

[ornate elk]

ah ok, that's quite diff from 11 labs

[high skiff]

This one is also RVC

[ornate elk]

so this is you singing?

[high skiff]

no, they are real son covers

I didn't make these

these are examples

[ornate elk]

so you put in a song and it transforms the voice to another artist?

[high skiff]

I did make this one tho lol

[high skiff]

it can, if you use it for that

[ornate elk]

hmm

[high skiff]

it can work on anything that has a pitch really

[ornate elk]

I will poke at it. I need to do more orchestral stuff

cinematic

[high skiff]

and it doesn't just have to do voices

[civic sigil]

Like its only for singing?

[high skiff]

no, you can run it as a real time voice filter

on my 3090, it ran super good

[civic sigil]

Huh thats surprising

[high skiff]

yeah

[civic sigil]

I might have to try that on my new build

Lol Rihanna is killing it

[high skiff]

those singing ones, like the ballin, only take about 20 seconds to render on a 3090

[ornate elk]

this was me asking a rude chatgpt to recite the events of a very nasty film. then I had elevenlabs read it in attenborough

[high skiff]

that took about 2 seconds on my 3090

[civic sigil]

Wtf

[ornate elk]

it'd help if I knew what any of these things were lol

[high skiff]

Also did death singing Boss Bitch by Doja Cat, this one is supposed to be bad lo

[peak dove]

Me listening to RVC

[ornate elk]

same thing with scarjo

[peak dove]

Do one by Awkwafina?

[ornate elk]

I would have to train a model for her

[high skiff]

there is likely one already

there are thousands

[peak dove]

Eau keigh 🙂

[ornate elk]

I mean on elevenlabs. they don't let you share

[high skiff]

oops, nevermind

I was thinking of the wrong person

you can also use RVC for non voice sounds

[ornate elk]

you guys know Karoly from two minute papers?

[high skiff]

like, if any of you guys have ever heard meow synth, somebody made an insanely good meow synth RVC lol

[ornate elk]

he's got an ultra weird accent and it nailed him

[high skiff]

Meow synth singing ballin lol

[ornate elk]

lol every single thing you have mentioned except rihanna I don't know

[peak dove]

Is that Mickey Mouse singing?! 😄

[high skiff]

Its meow synth!

[peak dove]

^..^<

[high skiff]

https://www.youtube.com/watch?v=JhAa_y0CyC8

meow synth

[ornate elk]

thank you, I have been edified

[high skiff]

now you can see how damn good that RVC us haha

gary from Spongebob lol

one of my favorites lmao

[ornate elk]

david attenborough quoting mike tyson

[high skiff]

jesus fuck that is vulgar

wow

[peak dove]

Some of this stuff is "too toilet"!!!!! 😦

[high skiff]

speaking of, I have the sound of aa toilet flushing singing rap god by eminem lmfao

[ornate elk]

sounds more like em rapping into a toilet

[glad grove]

luckily u havent heard ambatukam covers yet

[thin sinew]

Iron Mike: When your best choices are getting your ear bitten off, or cussed off. Do not engage.

[high skiff]

I have the spectrum dial assistant doing the low tier god rant lmfao

[glad grove]

[peak dove]

U mene that whoosh is the toilet? Nawww 😦

[high skiff]

@glad grove

lmao

[peak dove]

Me, hanging with my homies ...

[high skiff]

ok, i deleted that lmao

I don't want that to be taken out of context lol

[crisp owl]

"let me transfer you"

[glad grove]

LTG if he was a receptionist and not a streamer

[high skiff]

Minecraft Level singing ballin lol

[ornate elk]

aw man, my elevenlabs sub has run out

[median isle]

WHAT A TIME TO BE ALIVE!

[ornate elk]

What a time to fourscore and seven years ago!

[high skiff]

ahem

They indeed have an RVC of him lol

[ornate elk]

can RVC do text to voice?

[high skiff]

no, just voice to voice

well, maybe there are addons

[candid walrus]

i want to train my own voice then pipe audio books through them for my kids, but i haven't gotten around to training yet

[ornate elk]

ah boo

[ornate elk]

with elevenlabs it takes about 3 minutes.

[high skiff]

tho, some people suggest using Bark into RVC, as Bark can do text to voice, but poorly, so yuou use RVC to fix it up

[ornate elk]

yeah there's quite a few that do it badly

[peak dove]

I use NaturalReader to make my own AudioBooks - just feed it vanilla text - and ogg/wav/mp3 comes out!

[ornate elk]

ima show this to my mate, he's making a sarcastic assitant thing

[ornate elk]

it's cool but a flat delivery would get old quick I think

[peak dove]

U get used to it

🙂

[ornate elk]

nahh. I listen to a lot of audiobooks, and I can't stomach a bad reader

I dont finish those

[spring fulcrum]

@high skiff I think I have RVC installed. How do I run it?

[candid walrus]

you could say the same about being blind, but i like to see 😛

[high skiff]

Somebody made an RVC for the "DAMN DANIEL" and the "AR AR AR" part of this video lol

https://www.youtube.com/watch?v=kJBsOj-TKSo

[peak dove]

I have listened to the complete Moby Dick twice (19 hours each); and then Don Quixote (another 19)! 🙂

[ornate elk]

bro there are very good narrations of both of those

[peak dove]

No matter 🙂

[candid walrus]

these two @spring fulcrum

[ornate elk]

yeesh

[peak dove]

There are totally free vanilla texts at gutenberg.com

[candid walrus]

the go-web would be where i'd start, you need to get some voice data tho

[spring fulcrum]

Tried those and got this

[copper rose]

Just finished my first XL lora 😄

[candid walrus]

did you download the .7z from huggingface?

[copper rose]

They take forever to train and don't seem to have a very fast learning rate

[glad grove]

how big is it?

[copper rose]

800mb

[peak dove]

SDXL is pointing towards a sea-change in our desktop hardware - or more renting GPU time! 😄

[copper rose]

yea about to rent one tbh

[high skiff]

I have trained a few SDXL LoRA's, they are pretty easy honestly

just slower

[copper rose]

took 7 hours for 16000 steps

[high skiff]

jesus crist, why 16k steps?

what is the LoRA for?

[copper rose]

phyrexian horror

[high skiff]

hmm..

how many images?

[copper rose]

83 I think?

[high skiff]

I see

phyrexians are dope BTW

[copper rose]

🙂 i know had to do it

[candid walrus]

@spring fulcrum if you're on windows you can download from here (i think the RVC-Beta.7z) i have a previous version ending in 0528 https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main, then i believe it will run with all the dependencies they packaged together https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main

[high skiff]

oooo

[copper rose]

pretty cool results

[candid walrus]

but then you're going to need some vocie libs

[peak dove]

I suspect when we get to SuperSDXL (or whatever it'll be called when it drops) 24Gb VRAM might not just be enough 🙂

[candid walrus]

or they'll use new tech so it will

[high skiff]

My favorite card in all of MTG is Vorinclex, Monstrous Raider

[thin nova]

Can you share your training params for that lora?

[livid basin]

Right now, with SDXL 1.0 and pytorch compiled unets (base and refiner), I'm using 23GB ram at fp16

[copper rose]

I tried training in batch 4 and it required 40gb of vram

I had to stick to 1 batch

[candid walrus]

i mean didn't some dev peeps say we'll have 30 images generated per second one day soon lol

[copper rose]

I prefer training in bigger batches, the results are better tbh

[high skiff]

oh, thats not good

you must have used non optimized settings

[copper rose]

yea definitely not good at all

[high skiff]

on my 3090 I was doing BS 12

[copper rose]

WTF

[high skiff]

yeah

[copper rose]

ok can you teach me pls?

[candid walrus]

all hail sytan

[high skiff]

I don't have my 3090, so my SDXL LoRA tests were cut short

[copper rose]

I have 3090

[high skiff]

but @boreal bough Is a god at LoRA's

hes really good, much better than me ATM, as my 3090 was DOA

[thin nova]

i need these settings before I run this training session overnight

[candid walrus]

yeah, i want to get better, i would just rent a 3090 24 for doing deambooths from the terminal

[high skiff]

you need to make sure you have gradient checkpointing on, cache latents (to disk) on

[civic sigil]

For what rank?

[high skiff]

I don't remember at the moment unfortunately

the LoRA cam eout shockingly good

[glossy patio]

Hey guys might be a really stupid question, but does it cost to train a model with Lora? I'm a real noob when it comes to this

[alpine pine]

If your renting a gpu, yes

If you use your own, only your electric prices and the gpu wear

[candid walrus]

@boreal bough posted this last night and i copied it down... it was a few people back and forth...

accelerate launch --num_cpu_threads_per_process $num_cpu_threads_per_process $repo_path/sdxl_train_network.py
--pretrained_model_name_or_path=$pretrained_model_name_or_path
--train_data_dir="$image_folder"
--output_dir=$output_dir
--logging_dir=$log_dir
--output_name="${train_name}-${versionname}"
--train_batch_size=$train_batch_size
--unet_lr=$ss_unet_lr
--max_train_steps=$max_train_set
--lr_warmup_steps=$lr_warmup_steps
--use_8bit_adam
--xformers
--mixed_precision=$mixed_precision
--persistent_data_loader_workers
--network_dim=$network_dim
--network_alpha=$network_alpha
--shuffle_caption
--keep_tokens=1
--caption_extension=".txt"
--lr_scheduler $lr_scheduler
--min_snr_gamma=5
--network_train_unet_only
--resolution=$max_resolution
--min_bucket_reso 512
--max_bucket_reso 2048
--enable_bucket
--save_every_n_epochs=$save_every_n_epochs
--save_model_as=safetensors
--save_precision=$save_precision
--seed=$seed
--network_module=networks.lora

It seems like I can do batch 8 with gradient checkpointing now
--cache_latents --cache_latents_to_disk --gradient_checkpointing --mem_eff_attn --bucket_reso_steps=64

accelerate launch --num_cpu_threads_per_process=2 "./sdxl_train_network.py" --enable_bucket --pretrained_model_name_or_path="A:/models/SDXL 1.0/sd_xl_base_1.0.safetensors" --train_data_dir="A:/Datasets/Concepts/nier/2B" --resolution="1024,1024" --output_dir="A:/Datasets/Concepts/TRAINING" --logging_dir="A:/Datasets/Concepts/TRAINING" --network_alpha="1" --save_model_as=safetensors --network_module=networks.lora --unet_lr=0.001 --network_dim=8 --output_name="2B_v1" --lr_scheduler_num_cycles="40" --learning_rate="0.001" --lr_scheduler="constant_with_warmup" --lr_warmup_steps="50" --train_batch_size="8" --max_train_steps="35800" --save_every_n_epochs="2" --mixed_precision="bf16" --save_precision="bf16" --seed="1234" --caption_extension=".txt" --cache_latents --cache_latents_to_disk --optimizer_type="AdamW8bit" --max_data_loader_n_workers="0" --keep_tokens="1" --bucket_reso_steps=64 --mem_eff_attn --flip_aug --shuffle_caption --gradient_checkpointing --xformers --network_train_unet_only

[high skiff]

some results from my LoRA I did before my 3090 died

obviously on the Na'vi from Avatar

very small data set, so it would be way better with a full one

[ionic dragon]

wait, your 3090 died?

[spring fulcrum]

That did work but nothing is in english... lol

[high skiff]

few more tests I ran as well

[high skiff]

where have you been qwerty

yes, my 3090 died

[copper rose]

Gah im really loving this lora It just took ages

ill copy his code and run it with my next one tonight

[ionic dragon]

i have been off for 3 days

[high skiff]

yeaahhh, thats the right level of fucked up and messy for phyrexian horror for sure

[ionic dragon]

what happened exactly?

[candid walrus]

@spring fulcrum you'll need to get some trained models/files here's some, there are 2 types here, so you'll need to find RVC ones... https://huggingface.co/QuickWick/Music-AI-Voices/tree/main

[high skiff]

I don't care to get into it. I am being refunded as we speak

[spring fulcrum]

any way to change the interface to english?

[high skiff]

very very very short story, seller sold me the GPU with missing parts, tried for several days to get me to lose the money for it, blamed me, then the mods in the server destroyed him, and he apologized and he is now refunding me in exchange for the GPU back

I have another new 3090 coming in on wednesday, and its faster, highr quality, and comes with a 275 day EVGA warranry

(3090 FTW3 Ultra)

[candid walrus]

@spring fulcrum ha mind isn't like that i don't think there might be a setting somewhere for that hmmm... but when you get to loading in models you can follow here... https://youtu.be/qZ12-Vm2ryc?t=160

[high skiff]

Thank you for helping scorp, I genuinely forgot all of the setup after doing it like a week ago lmao

SDXL has had an iron grip on my psyche

[candid walrus]

@spring fulcrum ha guy in the video just right clicks and hits 'translate'

[high skiff]

that is such a good super reaction haha

one of the best I have seen

[tribal jackal]

lol thanks

[buoyant axle]

hmmm... i really dont get his CLIP_G / CLIP_L stuff ... tried several combinations between prompt/style plugged to CLIP_G/CLIP_L, but the only working combination that gives me good results, is to put prompt+style to both... is there any use case to pass different values to them (and still giving good results)?

[spring fulcrum]

This is what I got when running the file you suggested downloading...

[ionic dragon]

thats better,
for the same price?

[high skiff]

$100 more, but arguably an even better deal than the first

[tribal jackal]

Now he's just showing off...

[candid walrus]

haha

[boreal bough]

kohya gui, main branch, and use this config file
for anyone wanting to make loras

epoch and max epoch need to be adjusted. 40 for normal loras, 80 for very complex loras (faces/anatomy).
(also obviously adjust batch size, to whatever your card can handle)
repeat on dataset folders = 1

[high skiff]

$700 shipped with tax for an EVGA 3090 FTW3 ULTRA with 275 days waaranty left

[ionic dragon]

700?

[high skiff]

thats a $1900 3090

one of the best on the market as a whole

[spring fulcrum]

nvm I ran the web.bat and i got the webui... Its mostly in english

[ionic dragon]

dude, thats a great deal

[boreal bough]

there'll be a guide eventually on what exactly all settings do, but tl;dr. that config will work as long as your dataset isn't too bad

[candid walrus]

jelly

[peak dove]

Did that rogue GPU work after all?

[candid walrus]

@boreal bough i'd love to get your setup and see how easy i can get like a runpod/vast setup

[livid basin]

Just FYI, the 3090 has overheating ram issues that were fixed in the 3090 Ti

Since the 3090 has ram on both sides, but with 3090 Ti they fixed it and placed ram on just one side

Currently with a 4090, running stable diffusion in fp16 for 1 hour, I'm at 81C temp

[candid walrus]

it might just be as easy as git clone diffusers, or which ever lib you use, but gettin a linux machine to run it quick for easy training sessions would be nice

i only have 12gb locally so i have to reach out to rent

[primal vault]

If you put muted tones, artstation, concept art, detailed textures, art by maciej kuciara in the style prompt, remove those keywords from the main or secondary prompt. The style prompt is mixed into both of them and you don't want it to be applied twice (unless you do)

[solemn scroll]

Am I the only one getting this feeling of "low quality colors" ?

[high skiff]

Add the negatives

Desaturated, hazy, washed out, low contrast, foggy, dull colors

To your negative

[tribal jackal]

I mean if y'all are buyin pc's I can throw my 2080FE back in this tower, sell the unit as a whole and upgrade to a case that can fit my 40 series

[crisp owl]

https://tenor.com/view/ken-jeong-community-too-small-to-read-read-reading-gif-5494204

[tribal jackal]

I def already have an upgrade wishlist picked out lol

[alpine pine]

does the batch size of 8 run on 24 gigs?

[high skiff]

I just recently upgraded my computer for the first time in 6 years lol

[livid basin]

https://www.coreweave.com/gpu-cloud-pricing

80GB A100 for $2.21 an hour, dirt cheap

[tribal jackal]

I'll even refurb the gpu. Got fresh paste and a pad for it

[solemn scroll]

thx, kind of lost with the new prompting thing

[high skiff]

All good

[tribal jackal]

not that anyone here wants a 2080fe

[high skiff]

I mean, it's a fine GPU for AI

[livid basin]

A 4090 is $1600, so if a A100 (last gen, released with 3000 series) is $2.2, you can do 700 hours of work to make the 4090 worth it

A cup of coffee ever 1.5 hours

[boreal bough]

yep. batch 8 to still use your pc. 10~12 is max limit if you're going to sleep and not touching it anymore

fyi, batch 3 only runs a small bit slower - but then you can use comfy even while training

[high skiff]

I've ran them at BS 4, 8, 10, and 12

[boreal bough]

so that makes testing your checkpoints during training much easier

[high skiff]

I was doing that while training 1.5 checkpoints on my 3090

It was extremely useful to be able to just pull out one of the checkpoints and test it while the training was still going

[boreal bough]

just set your lora output folder as an additional lora folder within comfy, then you don't even have to move anything

[ionic dragon]

so the main and primary prompt should look like post-apocalyptic survivor overlooking abandoned city, overgrown ruins, hazy horizon

[boreal bough]

extra_model_paths.yaml

[alpine pine]

With epoch 30, is that just like a skyhigh value, so you can stop the lora when its ready?

[high skiff]

ok guys, I need a little community help to speed up my SDXL 1.0 worklfow launch

I need to find a program that can allow me to compared 2 images of different resolution side bys ide and export it as a collage

[ionic dragon]

but you dont know the flexibility of owning a gpu and running it locally, you dont need to switch into the website and activate and deactivate or wait for any server probs

[boreal bough]

it's not high - it's cause we're running with repeat 1
expect the lora to finish in like 20~30min

[high skiff]

I am struggling to find something that can do that ATM

[ionic dragon]

so a grid?

[alpine pine]

ohh! that makes sense lol

[boreal bough]

also, usually epoch 20 is the 'final' checkpoint. i run till 40 to see if I want a small bit of overfitting

[high skiff]

just one image next the other

[boreal bough]

faces/anatomy are still hard to get right, but overfitting solves that for now until I write a full guide on how to do it properly

[peak dove]

... but my yaml file is already called extra_model_paths.yaml - so why wouldn't it load? 🙂

[high skiff]

I tried ICAT, but it can't do images at different resolutions

[candid walrus]

are you expecting the lower res image to be upscaled so that the heights appear the same or do you want it 1:1

[boreal bough]

extra_model_paths.yaml.example
remove the example

[high skiff]

I am expecting just 1:1

[peak dove]

Its OK though, I've used Symlink instead

[high skiff]

I need my 1024x images next to my 2048x images at the same scale

[peak dove]

Oh, I didn't see the hidden extension! 😦

[high skiff]

like this, but it can't export them like this

[candid walrus]

so the 1024 upscaled or the 2048 downscaled?

the 1024 up

[high skiff]

yes, 1024 up

[primal vault]

sounds about right

[candid walrus]

how many image pairs do you have?

[high skiff]

I wonder if a standard collage maker could work

[high skiff]

trying to do 10 demo images

[alpine pine]

you could potentially do a photoshop script, though idk how it works

[high skiff]

so 20 images, 10 pairs

[candid walrus]

i'd probably just use https://www.photopea.com/

[ionic dragon]

you can use photoshop

[high skiff]

I don't do photoshop #ScrewAdobe

ah, I have used photopea before

[alpine pine]

I respect that, in that case I would recommend using a video editor

[ionic dragon]

should I do it for you?

[alpine pine]

import upscaled and original as image sequences into resolve
export as a png sequence

[high skiff]

ok, i think I found a site that can do it!

[boreal bough]

also, just in case anyones curious about file size. the full details & nuance of a face fit in dim 1 - we just use 8 cause we want the 8:1 ratio of dim to alpha.
setting it higher usually won't give better results, as dim 8 is big enough that I've fit the concepts of 100 complete dresses + faces + anatomy into it. So unless you're doing a full finetune level of lora with more than 5k images, dim 8 will be good enough

[candid walrus]

@high skiff you could make an 'action' where you record upscaling it to help quicken it up

[nimble heart]

gimp cult gimp cult

[high skiff]

I don't use gimp either :p

[ionic dragon]

which site is that?

[high skiff]

photojoiner.com

let me see if it saves at the full high res

cool, it doesn't

[high skiff]

completely useless :D

[ionic dragon]

you can create a project of 1024x2048(assuming the res of smallest image 1024x1024), and paste 2 images side by side, resize the larger one to 1024x1024

[high skiff]

no, thats the last thing I want

I need the bigger image to be full res

alright, I think I know what I have to do

I am just gonna run the 1024x images through area projection upscaling in comfy

[candid walrus]

in photopea you'd use edit -> image size -> 200 percent & nearest neighbor

[nimble heart]

what an amazingly roundabout way of doing it

[candid walrus]

on the 1024

[alpine pine]

area projection?

[candid walrus]

then you could place it on th 2048 with a 200 percent canvas increas from the edit menu

[high skiff]

0 pixel filtering

[civic sigil]

Not sure what you're trying to do exactly but you can probably use Chainner

[ionic dragon]

then, use 4096x8192 and make the 1024x1024 bigger by uncropping it

by dragging

[civic sigil]

Can batch resize, stack, and caption images

[high skiff]

your face is a roundabout

[midnight shuttle]

This sounds like a 5 line Python script using Pillow.

[high skiff]

there we go

did it

[nimble heart]

magick mogrify -scale 200% lowres.png && magick montage -geomtry -0+0 loweres.png hires.png grid.png made your grid that'll be $40

[civic sigil]

Really comfy should have nodes for that stuff

[high skiff]

there

yes, the high res fix for whatever reason messes with the white balance, I have not sorted out why

[peak dove]

Right side more texture

[candid walrus]

@nimble heart yeah if he said like 100 i would have said something like that

[nimble heart]

upscaler mayhaps?

[high skiff]

maybe, tho I think its a byproduct of any img2img, as it also raises the black levels, even with no upscaler

[peak dove]

Texture mutes contrast

[civic sigil]

Don't you think it's a fairer comparison to use bilinear / bicubic filtering for the smaller image?

[nimble heart]

the world isn't ready to hear our truths of magick

[candid walrus]

but you can see in the background that the sky is ever so slightly different like he's saying

[high skiff]

Meh, I don't reallt caew honestly

I am jsut trying to get this out to people

[civic sigil]

Fair enough

Results look nice

[candid walrus]

@nimble heart i mean to be honest i have to look up the answer for 5 minutes, so if it's 10 images i just do it by hand, same with like ffmpeg stuff lol

[vale eagle]

Still waiting for Sytan workflow release

[ionic dragon]

@primal vault why are these highres and lowres images the same?

[peak dove]

The more texture you add the more contrast shifts/mutes. Add texture with contrast compensation?

[nimble heart]

without really going into details the contrast thing is way worse normally. there's a filter to fix it in the workflow

[high skiff]

using nearest is actually the most fair, cause its displaying the pixels at their actual size

[nimble heart]

it's like the gamma is + 0.2 points on the upscaled image

[civic sigil]

I disagree, they are texels not pixels

[nimble heart]

so he gotta filter it back to somewhat normal

[high skiff]

I have a post processing node for after the high res fix that you can mess with to fix the black levels

[civic sigil]

If they were actually displayed at 1 pixel per texel then it would be "fair"

[alpine pine]

whats a texel?

[candid walrus]

there is no texture map

[alpine pine]

since he is doing a 2x upscale, there should be 4 pixels for every 1 1024 pixel, so the neighbor upscale will be perfectly accurate to the original assuming your image viewing software doesnt have an upscaler

[civic sigil]

Well its basically a pixel of an image lol, but if you are displaying it at a size mismatch with your monitor you would typically use bilinear filtering

[primal vault]

why wouldn't they? the highres version is an upscale of the lowres version

[civic sigil]

If you displayed the image at native size it would in fact be a pixel

[candid walrus]

Nearest Neighbor can be used on continuous data but the results can be blocky. Bilinear Interpolation uses a weighted average of the four nearest cell centers. The closer an input cell center is to the output cell center, the higher the influence of its value is on the output cell value.

[crisp owl]

at first glance I thought that was two images of two geese staring each other down

[high skiff]

lol

I can see it haha

[candid walrus]

i don't think he's looking for an interpolation

[high skiff]

I am specifically not

[peak dove]

A texel is a kind of sheep from an island in north Holland/Netherlands

[candid walrus]

if you stare at those cross-eyed they'll bite you

[nimble heart]

you get used to it.....
I have the ffmpeg high quality gif filter sewn into my heart
ffmpeg -ss 00:00:20 -i vid.mp4 -copyts -ss 00:00:20 -to 00:25 -vf 'split[0][1];[0]palettegen[p];[1][p]paletteuse' out.gif

[civic sigil]

Images are continuous, you would interpolate them unless they are specifically meant to be pixel art. If you are intending to resize an image then you are agreeing that it is interpolatable

[candid walrus]

he's wanting 'pixel' art to remain true to the data

[civic sigil]

If you wanted to do that badly enough you wouldn't rescale the image

[candid walrus]

potato, potato

[nimble heart]

I read both of those like "potato"

[ionic dragon]

isnt it highres fix?

[peak dove]

My neighbours out and about ... 😄

[primal vault]

not the kind of highres fix that you know from 1.5, no. it's an upscale of the output image to double the resolution

[copper rose]

runpod has em for 1.69$ per hour

[ionic dragon]

oh, so in the workflow the upscaled image isn't being run through anoth ksamp?

[primal vault]

no, that's something you can do with the img2img workflow

[nimble heart]

damn I totally recognize you from somewhere

[ionic dragon]

but we should again paste the image, it would be great if you could add a highres workflow by passing the txt2img through ksamp again

[copper rose]

sometimes I get these wonked out images, is this because my computer is shitting the bed?

[civic sigil]

I am wondering if lora effects Clip-G and Clip-L equally... does anyone understand how that works?

[nimble heart]

a1111 settings?

[copper rose]

auto1111 is what im running on yea, settings 30steps, high res fix x2, latent upscaler, the usual

[livid basin]

Oh wow, amazing price

[nimble heart]

what's the denoise on latent upscale

[boreal bough]

@copper rose #🔧｜finetune message

[copper rose]

0.4

[nimble heart]

make it 0.6

0.4 is too low for latent

[boreal bough]

posted it a bit more descriptive there

[livid basin]

Oh wow, seriously those prices are amazing

[copper rose]

Tysm caith

[livid basin]

How tf

[copper rose]

yuuuuh they they real ones

[nimble heart]

I'm not caith :(

[livid basin]

Do they allow baremetal ssh access? @copper rose

Persistent disk etc

[copper rose]

LMAO mybad clicked reply on wrong person

[boreal bough]

we can all be caith ❤️

[livid basin]

Cuz at those prices

[nimble heart]

we should all be Horatio

[livid basin]

I would be stupid to not use cloud A100s

Even if they are 3000 gen release equivs

[nimble heart]

multiple $ per hour for a gpu is wild

guess those are spendy boys though

[livid basin]

Well, if you aren't using tensorcores, 4090 is faster in fp32 than the H100

[candid walrus]

you can keep your files for a few per min/hr

[livid basin]

But the H100 has more mem, and much faster tensorcores

[candid walrus]

and yes you can ssh in

[livid basin]

Nice

I wish they had persistent disk

Like 100GB for $10 a month or whatever

[copper rose]

any other upscalers I can use at a lower denoising strength? Latent gets kinda wonky at 0.6 with my lora

[nimble heart]

pixel based ones

so non-latent

[candid walrus]

runpod does have something like that, but the deal is that if someone else is using that rack of like 4 gpus or whatever they have, then you have to wait to get a gpu, but can always access without gpu

[copper rose]

recommended?

[nimble heart]

for simple just bicubic I guess

there's also those ai upscaler models like esrgan

[copper rose]

Latent go brrrr

[livid basin]

Ah I see, that seems like something AWS offers that runpod does not

But I guess cheap spenders can't be choosers

With AWS you can persist with no GPU compute currently running

[nimble heart]

yuh. some images work well with latent, others I use pixel then like 0.2 denoise to just sorta cram some extra details in

refiner is good for that as long as your lora isn't too distant

[candid walrus]

i mean you could do that too, run the cpu all the time without a gpu if you wanted

[copper rose]

my lora is a little to complex for refiner I think

[livid basin]

I have access to 16 V100s, kinda old but it is what it is

RTX 2000 release equivs

[nimble heart]

also @candid walrus you can view image generation paramters from a1111 and comfy using magick identify

[livid basin]

So from like 4 years ago

[livid basin]

If the rumors of the 5090 next year has 32GB memory are true, it would be nice

[nimble heart]

no man's sky RTX ON

[nimble heart]

$2000

[high skiff]

here are 5 of the 10 demo high res fix images for my release. It should be out very soon

Left is base 1024x, right is my 2048x high res fix pass

[still dove]

??

[upbeat summit]

getting some Annihilation vibes 🙂

[nimble heart]

no man's sky is a game

[still dove]

😄 😄

[upbeat summit]

deconstructed model shoot

[candid walrus]

i'll check it out first thing in the morning, it's just going to be your main branch on your Sytan-SDXL-ComfyUI github right?

[nimble heart]

Annihilation was such a weird movie. Cool body horror for the first half then just anime bs for the ending

[high skiff]

yes, and on reddit

[still dove]

previous shot

[high skiff]

and comfy is releasing it on the example wiki timorrow

[candid walrus]

nice, looking forward to it

[still dove]

before the crash

[molten gull]

yeah, she looks a bit headless to me 🙂

[nimble heart]

she's running headless to reduce vram overhead

[candid walrus]

haha

[ionic dragon]

@primal vault can I dm about your workflow?

[upbeat summit]

is this your dog? 🙂

[tribal jackal]

Yep!!!

One of'em anyway

[upbeat summit]

really like your images. lots of dynamic actions and atmosphere

[tribal jackal]

Aw thanks!

[candid walrus]

is there a lib for comfy to generate forever (not add to queue), but something that on completion askes for another and just keeps going?

[candid walrus]

they're all like, 'this pos spaceship, got to push this thing again to get it started'

[tribal jackal]

"Use the force"

[still dove]

yeah! it broke in the last frame

[candid walrus]

haha

[nimble heart]

i swear comfy had that in a right click menu somewhere

this mysterious unlabled checkbox generates forever

[candid walrus]

yeah i know a1111 has a right click for it, but i've just been loading up the queue before i step away, but it really starts to lag once i get past like 300 to add them

oh snap!!!

[nimble heart]

uncheck it while it's running to stop the gen forever

[candid walrus]

the "checkbox" of mysteriousness speaks

that's "almost" the best part of tonight

[nimble heart]

next you're gonna tell me you didn't know you could use control and arrows to adjust the weight of words

[alpine pine]

Does anyone know of a good program for captioning image sets manually?

[candid walrus]

actually i stumbled on that by accident 2 days ago

i was trying to select like a whole word or line or something like i do on my mac and it wasn't working, bceause windows, and suddenly it starts ticking up

[woeful patio]

Is it me or does it take around 9 hours to train a lora on a 3090?

[candid walrus]

i'm like damn, i didn't know comfy weighted, this is GREAT

[nimble heart]

damn. oh well, at least you already know you can comment out parts of your prompt with /* C markers */
// like this

[alpine pine]

follow Caith's guide in the finetune channel, it should take 15-45 minutes depending on size

[candid walrus]

ha i didn't know that one

[icy brook]

@upbeat summit I have dm’d you

I have a question for you guys: any specific styles or subjects that are not very well represented in SDXL 1.0?

[nimble heart]

damn. I almost didn't know you could use brackets to {get|random|chosen|tokens} into your prompts, but I did

[woeful patio]

legend, thank you 🙂 #🔧｜finetune

[candid walrus]

that's random right, damn comfy even does that out of the box

i had to install an extention for that on a1111

that's going to pair nicely with 'forever gen checkbox'

[nimble heart]

yea all it's missing is easier grid and inpainting workflows and it can completely replace a1111 for me

ah shit I think the new controlnet poser is specifically a1111 too I'll have to look

[civic sigil]

What is that

Oh you mean to create openpose images

[nimble heart]

specifically this one https://github.com/huchenlei/sd-webui-openpose-editor has hands and face support

I know there's a 3d standalone one but it's the buggiest fucking software I've ever used in my life and I installed Windows 8 the day it came out

[charred beacon]

Great success!

[woeful patio]

stupid question but is a 100 steps per image still the norm when training loras?

[nimble heart]

there's a full lora faq thing somewhere Caith always links to it

[still dove]

they got out safely

[molten gull]

is there a way/node in comfyUI that lets me generate a batch of image at once, with one variable changing ? like cfg 4,5,6,...,10 ?

[nimble heart]

if you set cfg to an input and attach a primitive node you can set it to increment each generation

[still dove]

yes, like randomised seed, you can use randomised cfg for every render

[nimble heart]

then you can set "batch count" in extra options to like 10

Don't think "batch size" on latent input will work since the counts only randomize/increment/decrement every new batch

[molten gull]

i made it a input node

and connected to a primitive

i got there value and control_after_generated as parameters

[nimble heart]

control-after-genrated increment

[molten gull]

i can change that to increment, yes. how do i tell it to increment by 0.5 ?

[nimble heart]

cfg has a step of like 0.5 by default I think

[molten gull]

how do i tell it to increment by 0.4 ?

[heady vale]

or just manually change the cfg and press Queue each time, they all go into a queue and gen

[nimble heart]

that's the neat part, you don't

[molten gull]

why is that the neat part ?

what i want to do is run one generation with one seed from 4.0 to 6.5, and then do the same for the next seed

[nimble heart]

https://tenor.com/view/thats-the-neat-part-you-dont-invincible-gif-27194608

[nimble heart]

its a meme

[nimble heart]

first.

[woeful patio]

damn I'll get you next time

[nimble heart]

can't catch this

[molten gull]

or even better, as a matrix, i wanne do all combinations of cfg 4.0 to 6.5 (+0.5) and start_at_step 5 to 15 (+3)

[nimble heart]

I just rewatched it so it's fresh in my brain

[nimble heart]

nope.

mcmonkey or someone said there was a semi-official-kinda grid extension coming for it

[molten gull]

so its not possible right now, no ?

is it possible to start the whole thing with parameters somehow from the outside ? like an api-call or something ?

[nimble heart]

yes

[molten gull]

so could i realize it that way ?

[nimble heart]

Probably

[molten gull]

how do i do this?

[nimble heart]

https://github.com/comfyanonymous/ComfyUI/tree/master/script_examples api examples

think it's basically just sending a serialized json over socket to the comfy daemon

[molten gull]

so i gotta export my comfyUI thing into a json first, yes ?

[tribal jackal]

He goes by Barksy

[nimble heart]

yea you could probably just make the node graph in the actual ui then just use the python standard library's json lib to easily edit the values you want before sending it to comfy

[molten gull]

how do i export my comfyUI thing into a json ?

[nimble heart]

never done it personally so may be more steps

the big "save" button

every image you make also has its params embedded as json data under the prompts meta field

[tribal jackal]

baller

[molten gull]

can i see in the graphical UI which node has which id ?

[nimble heart]

probably easier to just ctrl-f and search by title in the json

[molten gull]

ah, that may work, yes

[nimble heart]

alternatively if you wanna be really really lazy you could just copy-paste the whole json into your python file and use an f-string to just replace the specific parts. no json deserialization needed

gonna have a 30 kb python file though

[upbeat summit]

did you make a fine-tuning of images of your dog?

[short marsh]

Put myself in the blade runner universe

[rustic surge]

Can WebUI use dual models?

[trim kiln]

hi all - does anyone know how to perform a "re-imagine" of a room?

[icy brook]

anyone?

[lean garden]

SDXL Is Amazing!

[delicate grotto]

its just called img2img
or you can inpaint

use lots of noise, but try to stay clear of 0.9, use lower values, unless you use it on a single color area

[woeful patio]

do we still use the 100_(name) folder when training loras? or is this outdated?

[delicate grotto]

wings on objects, seems like the model cannot put wings normally, it always looks cropped

[icy brook]

oh ok!

[delicate grotto]

and 100% of the time misalligned, on the back from all sides, if ya made a lora/emb id really apreaciate it

man this sucks ahahahaha why can't i replace my 6gb vram to more than taht

im even afraid to solder more vram because hell knows what will happen to the laptop

[trim kiln]

thanks ill give it a try

[delicate grotto]

and vram isn't that expensive anymore! freaking hell
~50% of the gpu's cost 5 years ago was vram(less on lower)
now its like 10$(expensive type) per gb

and they still only go with 6 to 12

[south igloo]

cant wait

[south igloo]

why batmans sad?

[delicate grotto]

sad? that's his happy face

[indigo vine]

I asked for a smoking batman so I imagine he's in need

[delicate grotto]

it seems a little odd, can you try to do a 2 stage k-sampler img?
add the comics style in the 2nd one

[delicate grotto]

i noticed that you lose a lot of quality with the comics if it exists in the first few steps, it tried to purge details sometimes

[civic forge]

Tuxedos are also called "smoking" in some countries.

[indigo vine]

sorry, I don't understand what you ask me to do. I am using searge reborn comfy setup: https://github.com/SeargeDP/SeargeSDXL/tree/main/example

[vital apex]

Wang ling images

#1072220168534642768 Wang ling

[tribal magnet]

They say a watched pot never boils, but i'm 90% sure SDXL speeds up if i stare at the UI.

[west breach]

Just posted my first two loras 'Grumpy Style' and 'Sparkling' to civitai https://civitai.com/user/humblemikey/models

[tribal jackal]

lol yep

[tribal magnet]

i have prompts in which the grumpy one would fit nicely.

[honest flint]

sort of interesting even if it wasn't what i asked gpt4 ```Here's a high-level mapping of the changes:

Text Embedding and Projection: Both models use text encoders, but the new model has a larger embedding size (1280 vs 768) and includes a text projection layer. This suggests that the new model might be better at handling larger vocabularies or more complex language tasks.
Transformer Architecture: Both models use transformer architectures, but the new model seems to have a more complex setup. The new model includes a 'conditioner' component, which might be used to condition the model's outputs on some input, such as a text prompt.
Image Generation and Diffusion Models: Both models include components for image generation, but the new model seems to have a more complex setup with a 'first_stage_model' and a 'diffusion_model'. The 'first_stage_model' includes an encoder and a decoder, which are typical components of autoencoder architectures used for tasks like image generation. The 'diffusion_model' might be used to add noise to the model's outputs, which can help generate more diverse images.
Increased Complexity: The new model has more layers and larger tensor sizes, suggesting that it is more complex and potentially more powerful than the old model. This could mean that the new model is better at handling complex tasks, but it might also be more computationally intensive.
Attention Mechanism: Both models use attention mechanisms, but the new model seems to use them more extensively. This could mean that the new model is better at focusing on relevant parts of the input when making predictions.``` sd 2.1 vs sdxl

[delicate grotto]

mmm well, nvm then

[static prawn]

did someone have success in auto 1111 with the refiner?

[peak dove]

Something beautiful using the R-B-R triple-process in ComfyUI

[uncut steeple]

Any advantage in that setup?

[peak dove]

Its the only setup I know - but its very clear and sharp

https://drive.google.com/file/d/1LLIY5IGwGFQ7fwiAg1VpysPD6g2Xorpo/view?usp=sharing

[honest flint]

yeh? just need to interupt the base generation when its about 80% complete then img2img it

[static prawn]

really? oh crazy

it might work with steps too? like do 35 steps in base, then move to img2img and do another like 10 steps or sth like that

did u try that?

[honest flint]

yeh it also works pretty well just doing normal img2img as well

[peak dove]

Detweiler's R-B-R Triple-Process uses 3 steps Refiner, then 9 steps Base; then a further 8 steps Refiner - 20 steps overall.

[static prawn]

normal img2img erases my details with the refiner

base

[honest flint]

yeh it has a tendency to do that a bit which is why you cancel early to leave some noise in the base image

[static prawn]

refiner, a lot of image loss

[static prawn]

oh ok , maybe i can play around with steps too, will give it a try 😄

[boreal bough]

HydrusNetwork = best program. but takes a bit of effort to learn the first time. scales well from super small, till giant datasets for fast manual tagging.
https://hydrusnetwork.github.io/hydrus/index.html

FastCaption = super easy to use. 0 learning curve. (does take me 4x as long as on hydrus network, for a dataset of 40 images though - but that's cause I've properly learned hydrus)
https://github.com/lukemoore66/FastCaption

[static prawn]

dunno i just generated a lot of images with all the same settings (different seeds) but the overall quality and prompt understanding was always way better in auto , i have no idea why

[honest flint]

i have doubts on your giant dataset claim

[uncut steeple]

Dont click if you dont like spiders

[honest flint]

but cool if true

[static prawn]

its just a personal feel, never would say thats "data"

[boreal bough]

20~40 is the new average, if you use my settings. depends on how good your tagging is.

[static prawn]

to clearify that 😄

[peak dove]

Spoiler-the-Spider

[honest flint]

(giant datasets to me are 1 billion + images, large are 1 million, medium are 100k, small is everything below that)

[uncut steeple]

Spider Warning ⚠️

[static prawn]

my few images are not even small 😄 , its just a feeling haha

[boreal bough]

I feel slightly offended that my 70k cosplay images dataset is considered 'small' 🤣

also rip my miniscule raccoon dataset. only 6k images ❤️

[honest flint]

my last scrape i deleted 900k images heh i really need to spin up a database to store latents, wasting a day everytime i extend my dataset is so dumb

[boreal bough]

that moment when you need a stack of A100 stacks to finetune your model. rip. XD

[static prawn]

hopefully auto will implement the refiner so its usable, sounds not really like an option stopping at 80 percent

u have to time it by hand , right?

or is there any option to do it automatically?

[honest flint]

yeh have to do it manually unless you patch the code

[honest flint]

i think you need a higher denoise strength

unless that is the look you are going for

[boreal bough]

you can't. a1111 doesn't support keeping the noise.
you can stop early (without noise) - then pass that on.

but if you're already making concessions, might as well go base only on a1111.
dpm 2m sde + karras + cfg 6 + 25/50 steps

[static prawn]

this was made by comfy, was more a reaction that we have to do anyhing manually 😄

[static prawn]

i feel like Base and highresfix giving similar results to the refiner

if u upscale to like 1,2x

[boreal bough]

sytan will post his workflow today - highres fully working and looking damn nice.
even has better lora support than any other base+refiner setup, due to how he implemented it

[honest flint]

oh i think its finally clicking why some people are so fanatical about comfy.. it allows them to become experts

[uncut steeple]

My workflow is decent too now id say

At least im getting the results im expecting mostly

[turbid garnet]

How to deal with token length in SDXL which is now restricted to 77?

[honest flint]

the token length is the same in sdxl and sd

[boreal bough]

^ this

[honest flint]

you chunk and average the chunks

which is why you usually want to stay below 75 tokens

[boreal bough]

(or go all out into 200 territory)

abuse the average! 🤣

[honest flint]

serious, serious, serious, serious, serious, 1girl

[bold maple]

Great job on SDXL! 🙂 We're have an open source project that runs automatic web searches and then evolves solutions to problems using GPT-4. We're then using GPT-4 to create image prompts for each solution and creating 400-500 images using SDXL API - it works really well! This is a use case that could never be accomplished with out AI images! You can check out he project and the generated images here: https://policy-synth.ai/projects/1/

[honest flint]

cool project but most people would consider that spam

[bold maple]

oh

is there another channel that would be better?

[honest flint]

off-topic maybe?

[pallid sail]

Tried to make a comparison of the R-B-R (left) workflow versus B-R (right). Prompt was a tapestry of a cat playing volleyball, masterpiece, detailed, masterwork, excellent craft, same seed for both. R1 (3 steps), Base (9 steps) R2 (8 steps) versus Base (12 steps), R1 (8 steps). Ran using Diffusers

[honest flint]

i dont see a promotion channel on this server

[bold maple]

well it's an open source project run by a nonprofit foundation, we're not really promoting anything

[glad grove]

then here its ok since its related to sdxl

[bold maple]

just wanted to share this as the SDXL API is working so well for us 🙂

[high skiff]

Alright, preparing some final things. Workflow will drop in a few minutes

[static prawn]

dunno when i wanna do photographic stuff i get a lot of weird eyes

[honest flint]

sdxl 1.0 is still pretty undertrained imo

eyes, anatomy in general are really underdeveloped

[boreal bough]

undertrained and overtrained at the same time

[high skiff]

until you use my res fix :p

@soft zealot You can start sharing your stuff

[honest flint]

yeh you can nudge it to do the right thing but thats just a bandaid

[static prawn]

im guess im not able to use a high res fix on my hardware anyway

[soft zealot]

you have to compare apples to apples

SDXL1 is already far superior to how SDXL1.5 was at launch

You shouldnt be comparong to Community trained & fettled 1.5 or 2.1 models

[static prawn]

😄

[boreal bough]

cause those areas were passed on to the refiner. refiner is overtrained in exactly those areas

[honest flint]

i would say its superior to 1.2 comparing apples to apples... 1.5 was pretty good for the model arch

[raw cedar]

chonk

[boreal bough]

good chonk ❤️

[honest flint]

i do appreciate how accurate sdxl 1.0 has been with cat breeds... that was unexpected

[hollow hare]

hi guys, i have a question, does anybody ever tried SDXL lora training using learning rate 4e-7? can I know if there are any differences between 4e-7 and 4e-4 in terms of output artifacts and overfitting speed?

[honest flint]

makes me want to go scrape wikipedia and yoink all of their images/captions

[high skiff]

ahem...

[boreal bough]

4e-7 is intended for full finetuning - which is not LoRA
for LoRA you can use a lot higher settings

[high skiff]

https://www.reddit.com/r/StableDiffusion/comments/15ed0o8/sytans_sdxl_offical_comyfui_10_workflow_with/

ITS UP

[wintry zephyr]

Hi guys! Is comming a specific model for inpainting? Actual sdxl-1-base give weird results

[high skiff]

@spring fulcrum

[boreal bough]

#🔧｜finetune message
and yes, there are differences, but they depend on a lot of variables. training speed only matters in context of the other settings

[high skiff]

ahhh, I have so many butterflies in my stomach haha

hundreds of hours of work now, all in once place!

[full idol]

Is it possible to use SDXL loras on Auto1111?

[honest flint]

congrats on shipping!

[high skiff]

I am pooped lol

[soft zealot]

Many thanks to @high skiff for all his work (https://github.com/SytanSD/Sytan-SDXL-ComfyUI) and I'm pleased to report that his method of upscaling (at least on my 1080Ti) runs around 100% quicker than the method I was using with no loss of detail.

Previously E2E generation of a 1024x1024 plus a 2048x2048 would take 500-600 seconds, using this method that is now 200-250 seconds(depending non denoise value selected, I use between 15% & 20% so minimal changes from the original) for the same/better quality.Those of you on newer hardware YMMV as to the percentage gains seen.

My previous workflow was based on Sytans V0.5 and I have now adapted it to include his HRF/Upscale method from V1 adding a few tweaks of my own such as a box you can type the denoise % in directly rather then calculating manually in the Ksampler and also a universal seed input to all relevant nodes.

My workflow also includes a preconditioning step,Lora loader( I like to use the example offset LORA at a max of 0.6 from https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main ) and a feature to write the Prompts data and seeed (still working on other details) to a .txt file alongside the generated images (matching file names).

Will be posting some quick & dirty examples in various style along with prompts used to generate them shortly.

[spring fulcrum]

Hi

[high skiff]

1.0 workflow is out, just above

[icy brook]

what do you guys to get realistic skin and not just plastic, airbrushed looks? out of the box it pretty much looks the same to me as when earlier versions was up on the bots: