✨｜sdxl

[lusty moss]

Ah meant in the same image, seed etc.., to see the difference. But no problem.

[bright valley]

here's it instead of pixel art

[lusty moss]

Little more pop. Iterate between the two gives a little more depth in my experience. basically 3d pixel art.

[bright valley]

Yeah it's just that pixel art is a huge activation word, with the three types

so I think it works best along with it if youre going for that style

[lusty moss]

Figured voxel would pull from the model. Pixel from your lora.

[bright valley]

yeah i guess you could say that that's exactly what happened

[lusty moss]

Not sure if you are familiar but the way i formated it i believe it switches between each one every other iteration.

[bright valley]

ye logical or

[lusty moss]

Love these two iterations because of it. Both great but feel like when you mix them you get a lot of depth with pixel art.

Seems like it wierdly pulled up the saturation/contrast for yours as well.

[bright valley]

Dude I love pixel art style so much

This one goes crazy too

[bright valley]

(I'm trying to make new preview images)

[lusty moss]

For v2

[lusty moss]

https://tenor.com/view/tardis-doctor-who-whovian-gif-10033462

[bright valley]

just at all

[bright valley]

the ones i have arent even upscaled and its on the front page

but I did finally finish a training job ive been trying to all day/night

[bright valley]

1st gen

ll take that as a good sign

[peak dove]

Is this an update on yesterday's Harrologo?

[peak dove]

I'm actually getting it to
spell right! 😄

[peak dove]

Cool!

[peak dove]

Harrologos LoRA

[peak dove]

An actual correct spelling! I am in clover ...

[bright valley]

not yet but hopefully soon, after im done testing it

[bright valley]

Nice dude that stuff is amazing!

Good sign so far with the update im testing

2/2 what is tihs?!?!

[lusty moss]

Definitely making some really cool realistic images.

[bright valley]

these are dynavision arent they

[lusty moss]

Yea I was running batches of 10 to get 1-3 good/acceptable spellings.

[peak dove]

Good guess!!! 🙂

[bright valley]

That's what I use just about all the time, i can tell that smoothness/quality from anywhere

Torcello I love these!!! Do you ever upload to Civit?

[peak dove]

No, I am more of an artist than anything else.

Do you mean that I can upload my art to Civit?

[vital ermine]

yes

[peak dove]

OK, well I will go to Civit and get some art onto the site 🙂

[peak dove]

SpaceX Harrologo LoRA

[bright valley]

yes

then it will be displayed among the rest of the artwork in the gallery on the model page

[peak dove]

But I must generate it at Civit?

[bright valley]

Nope! you can use whatever you want

I showed people on the Civit discord one of your images

[lusty moss]

One more generation then actually going to bed tonight lol.

[peak dove]

I have placed several pictures at Harrlogos ... 🙂

[bright valley]

i saw they were beautiful ty

[bright valley]

says every person ever

I seriously cannot believe what it just gave me

The R and O made a heart

that's it, v2 it's already better than I am at making it

[peak dove]

IPAdapter - is there a video totorial as to how to modify the strength of each individual input photo - I am using a 3-photo IPAdapter - but one photo seems to rule the roost all the time. How do I peg it back and make it less of an influence?

[wet glacier]

How about this? https://youtu.be/mJQ62ly7jrg?si=x8kG3Tzgk_USC62M

[peak dove]

Yes, I think its in there - I'm now stuck - cannot get past NULL for IPA Model Loader

It's stuck - there are no other files for IPA Model Loader to Load ... ?

[zinc cargo]

isnt this bcz you dont have a model in your folder?

[ionic gulch]

microsoft paint:

https://i.imgur.com/fIh9Azh.jpg

[peak dove]

I agree - but i cannot locate the folder to place any model? Have you any idea? 🙂

[zinc cargo]

also here for plus\

[native knot]

Origami alien?

[pure crystal]

(origami:1.5), $SUBJECT, propaganda photo, Bernie Wrightson photography, 1986 cinematic lighting by Denis Villeneuve good prompt

[peak dove]

Thanks, found it!!!

[zinc cargo]

may your generation be swift and smooth

[stone fossil]

Awesome I realy need a fortnite skin changer and valorant aim script for SD , hook me up.

😉

[noble shoal]

Yeah, I could use one too.

Didn't someone already pinged the mods? Maybe they need a fortnight skin changer too

[pure crystal]

friendly neighborhood cheat dealer

[stone fossil]

Im more intrested in the fortnite skin changer and valorant aim script for SD.

[noble shoal]

Imagine the possibilities.

[pure crystal]

I'd love to be having a blast

[stone fossil]

😄

[noble shoal]

He posted in every single channel

That's dedication

[stone fossil]

Nah thats hacked.

[west breach]

autoblock any new person that starts off their with Bro

[stone fossil]

Bro why would you do that?

[noble shoal]

Bro, nice Idea bro

[west breach]

thanks bro, happy to help bro

[indigo carbon]

dude, that reminds me of those pro-hamas bots on social media that always spam the same misinformation

[noble shoal]

Did you meant: "Bro..."?

[stone fossil]

Also where can I download Sony Vegas?

I need this bro.

[pure crystal]

what specific feature of vegas? just wondering

[noble shoal]

Bro, DaVinci Resolve bro. Free

[stone fossil]

Idk the one he said? 😛

[indigo carbon]

gambling

[pure crystal]

hahha my eyes didnt even see that. I look past spam with cruelty

[stone fossil]

I love crap like that its so much fun. 🙂

[indigo carbon]

BRO, that shit frfr

anyways, I bet we can add a bot to this server that detects people like that

[stone fossil]

Screw you guys, im downloading Sony Vegas. 😄 👌🏽

[noble shoal]

It would be easier for them if they would post SD related stuff like "Bro, check this out, 100% speedup in ComfyUi"

[ivory solstice]

PIXART-α devs claim to reduce carbon emissions with their project. Are they even serious? Some kind of joke?

[indigo carbon]

that actually exists though, but will be a part of the official ComfyUI soon

[stone fossil]

It might be so that if you add more bro into your prompt it might speed it up.

[noble shoal]

Let's hope nobody will announce it starting "Bro..."

[indigo carbon]

wtf SDXL?

I pasted whatever that scam bot said into ComfyUI, it just made these weird DJ chicks

[noble shoal]

If he just posted his prompt, he might be a fucking genius

[stone fossil]

😄

[west breach]

Somebody get this man some free nitro

[floral island]

bro

[stone fossil]

Nice bro! 🙂

[floral island]

🤜

oh no, it fizzled 😮

[floral island]

aight, imma go have to hurt my model more

again, thx to thaevil1 for his amazing model ^^

[floral island]

instead of longcat, longcow?

[west breach]

Took the limo cow to the prom

[stone fossil]

xD

[indigo carbon]

yeah, I merged that model with the model that I did and it made a pretty awesome model imo

[stone fossil]

Hey np. 🙂

Enjoy. ❤️

[zinc cargo]

extra longhorn

[stone fossil]

Eeeh.

I also offer other services.

[indigo carbon]

vegas?

[stone fossil]

Oh shit.

I need to download that one bro.

[floral island]

this on your model?

[stone fossil]

Ofc.

[floral island]

if so, you stole it pretty well ❤️

dang

[west breach]

time to carve the turkey

[noble shoal]

Bro, can you download that for me too? I can pick it up later, bro.

[stone fossil]

Ofc bro ofc bro I mean how can I let a bro out bro.

[noble shoal]

K

[pure crystal]

https://civitai.com/models/176555/harrlogos-finally-custom-text-generation-in-sd

[noble shoal]

Tidooes?

[pure crystal]

Tid-Ds

[floral island]

somehow, i dont feel this is completely realistic...

[noble shoal]

I don't see a single problem with that image.

[floral island]

it has the best car ever tho 😄

[noble shoal]

Well I was on a Dutch highway a few hours ago, so I would give it the name: "Dutch driving". Man what I was going through. 😢

[floral island]

here have a big t goth succubus gf, with veiled face, to boot

[native knot]

Gene was sick, so TayTay had to step in.

[vale eagle]

Testing 25% fine tune with descriptive captions

[native knot]

Something interesting to experiment with in a model, if anyone hasn't thought about it, is to allow the CFG to drop as far as you can get it while keeping results true to your prompt, then cranking up the steps to give it enough to try to converge. I find that it helps to keep the prompt simple in that case, but doing something like a 1-3 word prompt, dropping CFG down incrementally, but slamming the steps up into hundreds, if not a thousand can give it enough time to converge. (Also depends on the sampler's behavior.) You can get some really interesting results. (Can't post the one I'm doing now because it's NSFW, though.)

[vale eagle]

You might share some sfw result in the same way

[native knot]

Yeah. Was going to run one in a bit.

[indigo carbon]

it probably won't really matter if you're tuning on natural language prompts. what's limiting SDXL from having a complex language understanding is the text encoder

[vale eagle]

I don't think so

I think it could increase the prompt following in current architecture

[indigo carbon]

CLiP's LM doesn't have enough parameters to understand humanoid language. you can certainly improve the prompt following, but CLiP will limit it's capacity

I'm sure it can be improved, but just to a certain degree

[vale eagle]

I agree with that

[wispy sedge]

Hi everybody,

I was playing with IMG2IMG to try and create watercolor images based on real photos, and I got good results. I used Controlnet Canny model to match the shapes of the source image, and I am pleased with the results, but one thing is always off - the color. For example, the roofs on the buildings in the final image would be blue or green instead of red or the opposite; if the building in the original image has a green roof, the final image would have a building with a red roof. Or, for example, the green tree in the original image would turn into an orange tree in the final image, like it's autumn.

I tried selecting "Apply color correction to img2img results to match original colors," but this option is not producing good results. Is there another way to tell Stable Diffusion to transfer the colors of the elements in the original to the final image?

Thank you!

[vital ermine]

any idea which node package has "simpleeval" in it?

[native knot]

For instance, in my workflow, the positive for this is "dandelion plasma ball". The first example is pretty standard output one might expect for 40 steps at CFG 6.0. (All of these use seed value zero for the comparison.)

But...

Drop down to CFG 1.0 and crank up to 1000 steps:

Takes quite a bit to get the level of detail, but at that many steps, you can see a lot. However...it's not a great image. So...move to 1.5 and 800 steps:

But now, no glass. 😦 So, move up again to CFG 2.0 and 600 steps and you get something interesting and usable:

I'd probably tune in how many steps it takes to fix the definition of the middle finger under the ball, but that wouldn't be too tough.

And quite frankly, when I had the thought of a dandelion plasma ball in my head, that last image is a lot closer to what I pictured, rather than the first.

So I just find it interesting to try to see what the lower boundary of it is that I can get to adhere while pulling enough definition and convergence out of the model. (JuggernautXL v6 in this case.)

Also, in some cases, I'd probably take the outcome of the 3rd run at CFG 1.5 because that may have been what I had in mind in another instance.

Also, as a side note....Juggernaut really wants to turn the word Daisy into a woman instead of a flower. 🤣

[vital ermine]

I hate how all of them want to change cyborg into Cyborg and make them the Marvel character now.

[native knot]

DC, you mean?

[vital ermine]

Yeah, sorry

Bit me in the ass a fews days ago.

[native knot]

Yeah...ran into that last night when I was working on stuff.

[vital ermine]

prior to XL cyborg was not a DC character

the problem is it has no context and does not understand capitilzation so oof.

[vale eagle]

That means Juggernaut trained with the DC character

[vital ermine]

base too

[native knot]

I think the standard XL is as well.

yeah

[vital ermine]

it is very much systemic

Not sure how you unlearn something

[native knot]

It also irks me that anything moderately feminine weighs really heavy into making something a female...some models are worse than others for that.

[vale eagle]

You learn another thing to replace the existed thing

[vital ermine]

I had a few models gender swap on me and nothing else changed.

I was training on cyborgs and it did not unlearn as they all came out black cyborgs but some female some male.

I could not force them caucasian only close

so I gave up on it.

[narrow cradle]

Oh, for the love of Sleipnir, how did you prompt the 8-legged horse? That has been my quest sinve midjourney v4

[crisp owl]

Repeated comment? Lol

[narrow cradle]

honest question, still seeking

[crisp owl]

Oh, I did answer you on what I did below your q haha

[narrow cradle]

my bad, didn't see the tag thought it wasn't a response, I had written the q and left in a busy life at that moment, I guess. Thanks for the answer kind sir XD Sorry for the trouble

I'll have to decypher, when you mean, latent pushed, is there a way to pause a render ... then I'm having noob experience to '...push to a refiner'... I'm such a newb. I don't understand the process currently.

[crisp owl]

This was the original I eventually got that I figured I could work with. I was getting some multi legged horses, but nothing super amazing. So I ended up taking this into photoshop, duplicating the legs, bringing it into my img2img workflow to help further refine it.

It did take a while, but I needed a decent image for my project and after trying probably 100+ different generations, decided to go that route

Eventually when I get a better GPU, I'd love to make Lora's or a checkpoint of all the norse characters

[narrow cradle]

solid man, ok, so post in PS... thanks. I, too, had many attempts to get 8-legged, without success, now that I know how to use controlnet and ip2adapters, I'm going to retry with another piece of artwork I found of Sleipnir

[crisp owl]

Here's my current folder where I store them. (not all folders have images in them yet)

I haven't tried the IPA route, but that could have good potential

[narrow cradle]

'dats da bomb! I love that organization tree, I aspire to get there, I'm in the middle of using "Shade" and hope to gain such structure

[crisp owl]

Will have to look at "shade", haven't heard of it, but looks interesting

[narrow cradle]

warning, they might upgrade you to the middle tier above free 5k indexing, to get you used to it, but 10$/m might be worth it, another great discord community/devs, as they are responsive and working diligently to improve their product/service.... I'm already finding it useful in that the indexing is able to find 'similar images' ...

but the free month still applies

no card info given

[crisp owl]

I'll take a deeper dive later, but thanks for bringing it up. Could be useful

[narrow cradle]

stable-beluga

[narrow cradle]

do you use heterochromia as a prompt?

[vital ermine]

No, I never found anything but trouble when I tried it.

[narrow cradle]

I found it useful in different instances, but, yeah, it is a hit or miss operation

[vital ermine]

Yeah, really unstable

[noble shoal]

@stone fossil

[ocean ermine]

General! You always seem to be up to date with things. You know of any good art-centered SDXL model out yet? Other than the standard dreamshaper, I'm not seeing anything on civitai.

[vital ermine]

Not yet. Best I could do is grab some Anime ones.

What was out there was rubbish and only in lora not ful models.

[noble shoal]

Sure, here is my support and my review.
(in case the original message gets deleted, it was an Advert)

[ocean ermine]

Gotcha, thanks. Alas, my persistent addiction to generating Baldur's Gate style fantasy portraits will have to cope for now, heh.

SDXL does ok with art and art styles compared to 1.x--I've just been looking for that extra artist name awareness and style cohesion that comes with further training in that regard.

[vital ermine]

Well, can you not train?

Many can't now.

anyone get that glitch where a node refuses to move in comfy? Not locked and switching locked states, or reloading, works.

[vale eagle]

you might missed some input or output

[vital ermine]

No, they work brilliantly I just can't move them. Others have ran into this as well.

[vale eagle]

try to unpin

[vital ermine]

it is not locked

WOW

This is crazy. for giggles I hit the pin button (no unpin did I see) and it unpinned

Thank you but this is backwards.

Ahhhh, I see the label in the context menu does not change from pin to unpin. Thank you again.

phew

@vale eagle Happen to know of any S/R text nodes? I have one from Mikey but I never did figure it out and no info on his github.

[floral island]

time to shoot some waifus

[vale eagle]

Input X/Y prompt something like that?

haven't touch much

[vital ermine]

That one worked

[ocean ermine]

Haven't got into that yet, but certainly don't have the vram for it with SDXL using my 3080 I knew back before AI ever became a big thing that 10 gb was just not enough for what I was paying for the card.

[vital ermine]

Yeah, and neither is mine and I paid out the ass. took me six months to save for it. I am not doing this again, and if I had to do over wouldn't. My hope is Intel can do something as the route AMD is going with CUDA emulators is not for me.

[ocean ermine]

Wow, 6 months sorry to hear that. What're you running?

[vital ermine]

4090

[ocean ermine]

Wow. Yeah, I can see your disatisfaction.

It's insane prices

[vital ermine]

I had birthday and christmas money to help offset or it would have been a year. 1700USD for a vid card is robbery

My hope, overall, is that we get ASIC, or something, and the vid cards can go back to just being vid cards again 300 more mid range.

[ocean ermine]

And they're saying the prices are going to rise even further soon.

[vital ermine]

Yes

the reason I had to dip into savings

I saw it coming

[ocean ermine]

Good foresight then. Crazy times.

[vital ermine]

good night

[ocean ermine]

Going to bed? G'night then. May your dreams be easy and light.

Oh, I think you read my message wrong, lol. I was saying good "foresight" not good night, haha.

https://tenor.com/view/anime-anime-blush-embarrassed-shy-chuckle-gif-17723725

I think the only cards to recommend getting for AI use are likely workstation ones at this point. Especially if one is wanting to use it for text gen. 48 gb minimum in my mind.

[vital ermine]

Oh, lol

I lay odds the future is 600 1100 2k for what is now entry level x060 range cards. Intel/AMD/Nvidia.

[steady grove]

Intel is bringing machine learning instruction sets to their CPUs so we may have some nice capabilities there with memory requirements

nvidia has stirred them

"actual competition? what?"

workstation gpus will continue to be good for media work including generative images video audio

[vale eagle]

https://arxiv.org/pdf/2311.00618.pdf

image2text from Google deepmind

[steady grove]

if it comes from alphabet ai labs, there's little reason to beleive its democratized and runs on home PCs, or released at all

open source AI is killing alphabet and they hate it

[vale eagle]

It said code coming soon.

[steady grove]

their own top researchers have left and declared alphabet's AI strategies to be evil so i don't think they're very relevant. if anything i'm scared of them continuing any research

[vale eagle]

I think whether it open source or not. Many other ways to achieve the same idea

[steady grove]

they're the worlds largest advertising company so they intend to use these systems to manipulate people with machine precision

[vital ermine]

Lots of old movies about that and here we are. Yeah, Google is far from any person's friend though they self subscribe as woke far left they are just whatever it takes to manipulate us all until they could afford to purchase entire planets.

[steady grove]

google lobbys right and acts left. they enjoy very little regulation because they've groomed politicians to their favor

[vital ermine]

Are there any face fix nodes for ComfyUI besides this one? Just doesn't really work and does more damage than it ever fixed.

[ocean ermine]

That face detailer? I was just about to try that node

[vital ermine]

I think it is but I have no cues as to what it is doing and it has never really worked.

Might be made for larger faces while something I own does it so small and just works.

[ocean ermine]

Hmm. Looking at your settings, I'd up the denoise to 0.4

[vital ermine]

Photo AI is nice

[ocean ermine]

If it's anything like Adetailer for Auto1111

[vital ermine]

I had it at 0.4 it tore it all to hell

[ocean ermine]

Only the face section or the entire image?

[vital ermine]

I tried steps, denoise level, etc...

only faces

so I know it is it

too bad e can't get a box over what it found for faces

*we

[ocean ermine]

Yeah, I love that addition to Adetailer for auto1111.

[vital ermine]

yep, photo AI has it too

[ocean ermine]

Dunno man. I'll be following a guide online for when I set mine up. If I get it working correctly I'll send you the guide or be able to tell ya what you're doing wrong maybe, heh.

[vital ermine]

Cool. The problem is so little info out there even on the githubs

[ocean ermine]

For as useful as an implement as it is, you'd think there'd be more info.

It's just plain needed imo.

[vital ermine]

100%

[vale eagle]

Just build your detailer workflow

[steady grove]

if this training i'm trying keeps erroring out again, i'll load up comfyui an get a face detailer workflow functioning. i had a lot of success

[vital ermine]

I got it to semi work it picked up 1 in 5 and four had faces

[carmine sluice]

Hello, everyone

[steady grove]

https://github.com/ltdrdata/ComfyUI-Impact-Pack this is the one i use for facedetailer. it doesn't look like yours. there's all sorts of outputs for the mask and refined portion specific

[carmine sluice]

What are you doing?

[soft bone]

it can do this

face detailer

[bright valley]

@peak dove Hey, I saw that you uploaded a bunch of those images to Civitai, and they are prominently displayed in the model gallery now! Here is a token of my appreciation:

[ocean ermine]

Neat. I'm currently in the process of setting it up in my workflow. We'll see how it goes.

[high skiff]

@bright valley I am now at 3 companies interested

🤞

well, 3 companies, and 1 individual

[ocean ermine]

Ok, followed this guide and everything seems to work properly: https://www.youtube.com/watch?v=z4JzuuFyr1Y
The only changes I'd make to the settings is to up the sam_dialation and sam_bbox_expansion to 15 by default, to ensure that it'll always get a bit more of the face (in fact I might increase it even more in the future, we'll see.)
I also set the denoise at 0.4 instead of 0.35, and the max size at 1024 (I believe that number is the biggest size it'll try and redo the face it detects.)
.

With that being said however, though I put it together manually in my workflow, I saw a workflow in the comments section of the YT vid that I want to try moving to, as it's doing nearly everything that I'm already doing with my own workflow, but looks a ton neater and has a better system than mine. It's here: https://civitai.com/models/119257?modelVersionId=179750

[bright valley]

Let me know when you see a dollar

[high skiff]

my latest tests are heavenly lol

left is base SDXL, middle is my new 3.5 LoRA, right is realistic vision

[ocean ermine]

An example of the end result, and the face before/after the detailer.

Before/After

[high skiff]

thats a solid improvement

usually face detailers suck ass

[vital ermine]

I see sam and only see Surface to Air Missile. :/

[ocean ermine]

Lol

Well, you are a General after all.

[vital ermine]

LOL

@ocean ermine It see the face but honestly it is not doing all that much with it.

[ocean ermine]

Set the bbox_dilation back to 0. Set the sam_dilation and sam_bbox_expansion to 15. Also, I keep the guide size at 256--that may or may not change the outcome.

It also depends on the gen. You're going to see greater benefits the further away someone is.

[ocean ermine]

Yeah that looks right.

[ocean ermine]

Doesn't look too bad. I can see the issues with the left and left middle face though. I usually try to prompt for the detailer specifically and/or with wildcards to help it along. Most often with the art keywords that I use for the main prompt. E.g. "fantasy art" or "victorian style" etc.

[ocean ermine]

In the end though, it will always be limited by three things: The model you're using, the base generation quality (before face detailer), and the fact that you're essentially automatically inpainting in faces after the fact. This last point means that, since it's only looking at a face when it's inpainting, it doesn't have any context for the rest of the scene or body--so you can get too small or oddly positioned faces, or childs faces on adult bodies, etc.

The best result we can get atm is a decent quality base gen, that you then have face detailer come in and perfect.

Come to think of it, you're not upscaling before face detailer, so that likely is the difference between mine and your generations. With my generating, then upscaling 2x, there's better surrounding context and base generation quality that it can work with and then come in with the face detailing afterward.

[vital ermine]

no, this does

[ocean ermine]

Oh, you're not using SDXL?

[vital ermine]

of course I am

[ocean ermine]

That image isn't upscaled though.

[vital ermine]

nope

[ocean ermine]

That's what I'm referring to. You gen, then upscale, then face detailer for the best result.

[vital ermine]

yeah, but 4x is insanely huge. Look at my shrooms above that took forever and is astronomically large.

[ocean ermine]

Yeah, that's why you only upscale 2x.

Ignore the spaghetti.

Using an upscale model that's 4x, you put that through and scale it back by 0.5 with the "Upscale Image By" node, then pipe that through your KSampler.

That's the proper way to "Hires fix" an image with a simple 2x upscale. As long as you're using a decent model and an upscale model, 2x shouldn't be a problem.

[bright valley]

mine too actually

[high skiff]

Doppeee

[high skiff]

Yup, that's one of the steps in all my upscales too

[vital ermine]

We need new upscalers and some 2x ones too

some of these date back to before SD

[ocean ermine]

4x works just fine imo. It's better to be able to scale back then not scale enough.

[vital ermine]

seems stupid to scale at all. 4x then cut half away when you could just eliminate the middle man and wasted resources.

[ocean ermine]

With some 1.x models, I've gotten away with scaling up 3x without issue. I believe this was one.

[high skiff]

True, but you also get the benefit of down sampling, which cleans up the details and anti aliases them

[vital ermine]

true

[ocean ermine]

Here's my recommended models: I recommend fatal_anime and HugePaint the most.

Though you'll probably balk at the 8x factor, lol.

[vital ermine]

never seen fatal or huge or fatally huge he said

[ocean ermine]

fatal_anime is my go to. It slightly increases the contrast of the end result, but it works fantastically on realistic, stylized, anime, etc. It really gets the details right at 0.4-0.5 denoise.

[vital ermine]

where can I grab them at?

[ocean ermine]

Lemme pull the link

[ocean ermine]

They moved sites so it took a second to find. Go here: https://openmodeldb.info/

Fatal anime is direct here: https://openmodeldb.info/models/4x-Fatal-Anime

Hugepaint is here: https://openmodeldb.info/models/8x-HugePaint

Remacri is here if you wanted it: https://openmodeldb.info/models/4x-Remacri

It used to be my go to, but I've found fatal and hugepaint better serve my purposes.

[vital ermine]

yeah, I was using remcari

it is so old now

[bright valley]

do you ever img2img with the same parameters to upscale anything?

[vital ermine]

much has changed

[ocean ermine]

Always use a different seed. Other than that, sure.

[vital ermine]

I never i2i in comfy as I just don't care for ir

it

[bright valley]

probably the best AI video/animator I know, Tyler from CivitAI, he runs all his frames through A1111 batch instead of standard upscale, or gigapixel, etc

May be an animation thing I guess but just having that light diffusion with the upscale can really produce great stuff

I have yet to try it personally, I've just seen him do it, and his output is next level

[vital ermine]

what is the hugepaint one for?

[ocean ermine]

Yeah the difference between something upscaled to an algorithm and upscaled with SD is night and day to my eyes.

[bright valley]

But still most people upscale i the extra tab, using the ones you're posting

[ocean ermine]

Art and anime, generally. It blends where it needs, and keeps detail where it needs as well.

[bright valley]

including me, currently lol

[ocean ermine]

Here's my HugePaint comparison I made a while back: #🍥｜anime message

[vital ermine]

I use either photo ai or the ones it fails at I go to extras tab to clean up

[bright valley]

i was having a HELL of a time finding a comfy workflow that would produce high quality upscales of images with my LoRA

Latent upscale/Hires fix was just murdering everything no matter what I tried

[vital ermine]

hugepaint is 8x, shit

[ocean ermine]

LOL #✨｜sdxl message

0.25 rescale man. Easy peasy.

[bright valley]

Base gen, latent upscale

[vital ermine]

8x it found this, but that is too much so is that the bbox crop?

[ocean ermine]

Not sure what I'm looking at here.

[vital ermine]

the two face bbox found

*faces

that is far too large than it was before the 8x

[ocean ermine]

Are you upscaling it actually at 8k?

Fully 8 times larger?

[vital ermine]

hell if I know. I think 1.25x

[ocean ermine]

How are you rescaling it?

[vital ermine]

I changed from 2 to 3 bbox crop

[ocean ermine]

Which one is your preview image coming from?

[vital ermine]

I have many. The ones I have been showing you is cropped_refined

1.25 changed to 1.3 but is better

[ocean ermine]

Ok. Then that's not representative of what's actually being changed I believe.

[ocean ermine]

This however does look like it. It's very different from mine. lemme gen something and get you an example of what mine looks like.

[vital ermine]

those are the faces being changed, yes

Ultimate upscaler I dunno that thing goes on and on and on forever

[ocean ermine]

Yeah, ult upscaler is pretty slow.

[vital ermine]

then when it is done it gets sent to detailer

[ocean ermine]

This is the preview image I get from cropped_enhanced_alpha

[ocean ermine]

Yeah, that looks how it should.

[vital ermine]

figured that is what you were after the mask. before I did bbox 1.3 (down from 2) they were too big

[ocean ermine]

bbox what?

bbox_crop_factor?

I have that at 3.

[vital ermine]

I had it at 2 but 3 I might as well not have it all

sucks in the entire man

[ocean ermine]

What's your guide_size?

[vital ermine]

was 384 now 512

[ocean ermine]

Pull it to 256.

[vital ermine]

k

omg

ugh

[ocean ermine]

Lol. Put the crop back at 3.

[vital ermine]

that was insane, lol

[ocean ermine]

Yep. Seems to be working great for me. It started wanting to skip facedetailer but setting force_inpaint on helped. It seems 256 is (at least in my still rather brief testing) a good all-around size for the guide_size.

[vital ermine]

wait, mine is on

enabled

[ocean ermine]

Yep. I'd leave it on.

[ocean ermine]

RIP her right hand though.

[soft bone]

max size is the longest it will render the longer side of a bbox segment, it is very useful to go over 1024 in many cases.

also, there is a way to set up the sizes so that it does in fact have context of the surrounding image to use during inpainting, without actually inpainting anything outside the face.

[ocean ermine]

Informative! And fantastic news regarding the context. How do we make that happen?

Ah, it's the crop factor.

Deets here: https://www.youtube.com/watch?v=KvZ8ucBqyqw

"Setting the crop_factor to 4 means that it provides up to four times the area of the mask as context to be used during image generating for inpainting.
As the crop_factor increases, it allows for generating images that better blend with their surroundings. However, it also leads to a slower image regeneration speed."

[vital ermine]

well, yeah

what it said. :/

[shy kelp]

Adetailer does amazing face correction, I never have to deal with images with messed up faces now

[vital ermine]

@ocean ermineWhat would make it keep messing up? latest is left orig then right is face fixed.

It seems to change facial hair into objects

[ocean ermine]

That's weird. I dunno. Try bringing the denoise down to 0.35 or so. Your original looks very good so that should only help with good base gens.

[vital ermine]

Yeah, the previous it made the lip to chin hair (going into a beard) into a cog

try 0.35 now

if this works like i2i denoise 0.36 is the point I mostly used as it left the original alone

can't seem to get the same gen so will keep an eye out

[vague gyro]

d

[vital ermine]

@ocean ermineI have a model ultimate upscale hates.

I am also seeing seams now and reducing from 64 to 32 didn't help

Yeah, 32 to 128 did nothing

[ocean ermine]

Yeah, seams are tough to work around. I haven't found a way to get rid of them in the gen process. I just have to photoshop them out as best I can after the fact.

Obviously the goal should be as few tiles as possible, but sometimes it doesn't matter--the seams see all.

[vital ermine]

MOHAWK V4 ultimate upscale washes out

so far it is the only one it does that to

denoise fixed the seams

Changed from none and 1.00

it did a ton of passes though

[crisp owl]

yeah seamfix makes it take much longer. Sometimes seamfix sucks, othertimes it fixes issues

[vital ermine]

yeah, figure it is one of those cases of using it if you see seams

1.00 worked as well so it iwas the seam fix mode

[crisp owl]

Yup, I never used it be default due to the extra time especially

[vital ermine]

I dunno but this detailer is pissing me off

orig vs face detailer

thing is here is the mask

notice, the ear is not part of it, so wth?

Now I see. not it

the ultimate upscale zapped his ear

after upscale and before face det

LOL, do not use ddim

[crisp owl]

isn't ddim basically euler now?

[vital ermine]

can't be because they act different in auto and comfy

different results

[crisp owl]

https://github.com/comfyanonymous/ComfyUI/commit/a268a574fab025deed91f5201910ac052132c42c

There's the commit I read

[vital ermine]

No wonder ddim reacts differently in his thing

I liked the old ddim

Now I haven't updated mine in about a week (just too sick to do anything) so I guess mine is still the old one?

[crisp owl]

possibly unless he made other changes prior

[vital ermine]

ddim is a standard so why is he fucking with a standard?

[vital ermine]

A steampunk-meets-steampunk-meets-steampunk-meets-steampunk-meets-steampunk-meets-steampunk-meets-steampunk-meets-steampunk-meets-steampunk-meets-steampunk-meets-steampunk-meets-steampunk-meets-steampunk-meets-steampunk-meets-steampunk-meets-steampunk-meets-steampunk-meets-

[vale epoch]

so im training my first lora, yay. i was able to make one but when i use it to render stuff it doesnt work, i'm also using the trigger words. they are images of hands. shiny black hands to be precise. i'm not sure what im doing wrong. my dataset has images and text with the same name (filename wise). im also training it at like 5k steps. but every time i try to generate one using the trigger word, it renders a real hand

[vital ermine]

5k steps for a lora is over cooked

what is your batch size?

[vale epoch]

dang i even did half that and same issue.. not sure if i had a trigger word tho i might try that again.. batch_size:2

[vital ermine]

5k steps with a BS of 2 means you did 10k steps

[vale epoch]

i love that ppl doing this stuff call it undercooked or overcooked lmao

oh yeah that sounds right sorry i mispoke

[vital ermine]

no worries

that is about 10x more than is needed

[vale epoch]

gotcha, ill reduce and see how that goes, ty!

[vital ermine]

800-2400 is the spot it will be best at

your issue transcends that

yu using comfyui or automatic1111?

[vale epoch]

fooocus

[vital ermine]

I'm out, lol

[vale epoch]

oh is that bad, its the only one i could get running t-t

[vital ermine]

no, not bad just I never used it so know zip about it

[vale epoch]

oh gotcha

im using some google colabs to make my dataset and train

[vital ermine]

I did that for a year

never again, lol

I wanted to kill myself as colab and I just no no. I don't like clocks then they take forever to do anything that locally takes seconds

[vale epoch]

makes sense

[vital ermine]

I also, honestly, had no successes with colab but locally have had a lot

just me on that one

I still do not have a lot of success with loras

I prefer to dreambooth for over a year then extract the lora, though I miss extracting locon but for sdxl not avbailable.

[vale epoch]

yea i wanted to use dreambooth first, but after having issues, read about lora and then found civitai and my mind was blown

i wanna use it for work stuff, but theres like a bazillion licenses i gotta read through lol

[vital ermine]

thing is with DB the world is your oyster. can use it directly as a model or strip it naked as a lora. best of both.

model merging too and lora merging it gets deep

[vale epoch]

yea fooocus has it where you can use multiple loras

[vital ermine]

all do, and you can mix them to make one lora or a new model.

[vale epoch]

yea thats cool

[vital ermine]

really is

it sounds like a foocus issue for your problem

[vale epoch]

im trying 2650 steps

[vital ermine]

are you sure it is even loading? does it have a cli window?

[vital ermine]

now, be aware trainers use S/R to find the best lora as they will train say 10k steps and find where the last step is at before it explodes

[vale epoch]

👀

whats sr?

[vital ermine]

sr?

[vale epoch]

S/R?

lol

[vital ermine]

search and replace

[vale epoch]

oooh

[vital ermine]

it allows to easily gen all your loras to find which one is the last good usable one

the screencap you showed is just the trainer

[vale epoch]

yea

oh you wanted an ss of foocus?

[vital ermine]

yes, its cli window

I want to see if it is complaining about the lora you loaded

[vale epoch]

ohh

[vital ermine]

it should comlain if something isn't right

[vale epoch]

not seeing anything 👀

[vital ermine]

Can you not get to its command window?

the cli?

command prompt

[vale epoch]

oh

[vital ermine]

I see bhand, and the lora called bhand, but what is your trigger word?

[vale epoch]

bhand

[vital ermine]

did you set it?

[vale epoch]

yea

its the first word in all the text files

and i have it enabled, and training an object based lora

[vital ermine]

should work if foocus is loading it right

even if it goes crazy it should do something at 10k steps

[vale epoch]

only errors i see

[vital ermine]

it is showing it loaded a model but is it considering the lora a model or sdxl itself because those errors are no bueno

[vale epoch]

i think sdxl itself

[vital ermine]

btw looks you have a nice broken lora

[vale epoch]

lol

hey at least i got stuff generating.. thats half... the battle LOL

[vital ermine]

It genned something so sdxl is loaded which means it tried to load your lora and it is broken or not created for xl

I have only ever seen errors like that on loras not made for my currently loaded model version.

did you tell colab trainer this was for xl? Should have a checkmark or a toggle.

then you have to use the xl model as source

[vale epoch]

ooh good point yea it needs to be sdxl, lemme see if the trainer supports that

[vital ermine]

most don't for colab now 😦

[vale epoch]

orly

[vital ermine]

free colab at least

can't use webui on free tier of colab

I suspect your lora is not trained for xl

[vale epoch]

yea im not seeing anything mentioning xl

koya ss for training then?

[vital ermine]

I say this due to training on colab for a long long time and the speed I saw is not indicative of a T4 training XL

yes

a T4 training XL as BS2 is secs per it not it per sec

once you train for XL I think you will be fine.

[vale epoch]

groooovy imma go try that

[vital ermine]

gl

[vale epoch]

ty so much!

[peak dove]

I'm using Nerdy Rodent's IPA Workflow - triple input is cool!!!

Not sure how to tweek FreeU (Free Lunch!) but I'll give it a go!!!

[crisp owl]

I wasn't impresed with FreeU. Seems more of a "tinker this scene" type of thing where you get a gen you like, lock the seed, then tweak settings as you go

[peak dove]

Scott Detweiler seems convinced 😉

But I will see what it can achieve - I might even like it!!!

[crisp owl]

The code is the tiniest bit of code.
I'm not saying it's bad, just that it doens't seem to be doing anything special for an "every generation" type of thing

my opinion on it of course

[glad grove]

his music seems convincing

[peak dove]

Must expand things and use Controlnet as well ... I have disabled UpScale as I don't have until the next Millenium to wait for output!!! 😄

[peak dove]

These triple-input IPA's are fantastic!!!

[zinc cargo]

anyone here with automatic Vlad and can tell me if Loras even work?

[slender coral]

Anyone got suggestions for a prompt on removing this webbing?

[crisp owl]

strings in negative?

[slender coral]

No negatives, in the actual prompt.

[crisp owl]

why no negative?

[slender coral]

I don't have negatives in my workflow.

[crisp owl]

well, that's a limiting factor

[noble shoal]

I think solesbeedude means you should use strings in the negative. It was not a question if you have that.

[vital ermine]

I am still at a loss when I train loras as to what I am looking for to know that is the one. They are all so different and none ever implode so I just never know.

[slender coral]

Yes I don't use negatives in my workflow.

[vital ermine]

you need to

XL is more forgiving but sometimes it is all you can do

[slender coral]

Yeah was hoping there was a work around.

[vital ermine]

Short of a full on training negs is the way to go

[noble shoal]

Is this a religious, political or technical decision?

[vital ermine]

LOL

Trump said negative prompting was yuge. Welp, no negs for me.

[kindred shoal]

Yeah why don't you have negatives? They're kind of important.

[slender coral]

I've rolled out a custom NLP parser, don't want to invest a ton of work into updating to support negative tokens if it's not a successful product.

[slender coral]

I Supported Trump.

[vital ermine]

high fucking five. I support MAGA. 'nuff said.

[noble shoal]

I mean in 95% of the times i also have no negative. But if i really don't want to have something in my picture i negative it out of there.

[slender coral]

Yeah If i was using a workflow that supports it I would 😄

[vital ermine]

Although, to be fair, XL is a sob for negs as I have negs running at 2 and 3 weight sometimes you are not budging it. Never before have I ran negs so high.

[noble shoal]

You have a sampler without negative input? 🤔

[slender coral]

Testing a app I've written for android.

[vital ermine]

did you know there is a port, or w/e you wish to call it, for negative prompt training. OT dev told me but unsure how to work it or what, if anything, it will do.

I got excited when I he told me that

yeah, since 2.0 I have begged for neg prompting for training.

[noble shoal]

So, your actual question would be: "Testing a app I've written for android. Is there a way to get rid of this without using negatives?"

[kindred shoal]

Quick question for you sdxl users, is there a way to get a 1660ti to run sdxl or is that a dream that'll never come true? I keep finding REALLY cool SDXL models I wanna try but uh...1660ti limit here.

[slender coral]

No, my question is still, what would be the best way to remove the webbing with only a prompt.

[noble shoal]

Whats the original prompt?

[slender coral]

Asteroid in space.

Some are good, and some are dogshit:

[noble shoal]

Are those cropped images?

[slender coral]

yeah

[noble shoal]

I can't reproduce this webs in my tests. But do you consider putting a prompt styler in your app? I mean, SD has no idea what kind of image it has to generate. Is it a photo, a painting, a mosaic, something a guru painted with his feed in the dirt? Don't guiding SD can also lead to weird artefacts.

[kindred shoal]

Why is sdxl taking so long to start sampling steps in comfyui?

on a rtx 3060

Oh god, why?

Oh. It was the image size. woops.

[peak dove]

IPAdapter SDXL Guitaristas

[static prawn]

which one of the sdxl base versions do i need?

the 0.9vae one?

[peak dove]

Second one, as the first one had some artefacts (lines) produced

1.0_0.9vae yes

[stone fossil]

These are not healthy numbers for GPU, looks like you are running on CPU.

[kindred shoal]

I had the image size way too big, so it was not happy with me.

[stone fossil]

Ah ok. 🙂

[half cedar]

Dalle3, remember remember the 5th of flavember

[velvet shore]

😭💀

[peak dove]

SDXL ComfyUI Unstable Diffusers model, using 3 input IPAdapter - Guitaristas!!!

Other instruments

[fallow fox]

hi, dudes? Is SDXL for free?

[zinc cargo]

yea

free until these guys come around...

[fallow fox]

is it better than automatic1111? i want to aim for GIFs. and short movies in long term.

i tryed deforum but each image is like with LSD

[zinc cargo]

sdxl is just the version of the model, not the way to create images.

i guess you refer to ComfyUI, which is also free.

[fallow fox]

great, your work? its clrear, no noise around, coherent gif

[zinc cargo]

yea, it's vid2vid with sdxl hotshot

(i swear it's not jibrish)

i just love this set

i should do more

[fallow fox]

we all are miles from making a AI movies, but we are moving close.

[icy brook]

https://civitai.com/models/190242/aether-watercolor-and-ink-lora-for-sdxl

Simple little thing that does watercolor, ink and sketch really well!

[zinc cargo]

really really really well

[floral island]

weirdest ken ever

[zinc cargo]

@icy brook you nailed it with this one

[floral island]

my new model merge does it better tho

[fallow fox]

i hope oyu get job as advertiser and marketer

[floral island]

best milkshake ad ever?

[fallow fox]

yy, wqith your self portrait its vile

[timid garnet]

Hey

Posted a model on Civitai, should i shill it here ?!

For creative and colorful content.

Anyway, it's humbly called UltimateDiffusionXL
Still the very first version.

[icy brook]

Thanks! 🤗

[vital ermine]

https://civitai.com/models/190501?modelVersionId=213923

[shy kelp]

ipadapter is a model i take it?

theres so many samplers, are there any that you all would never use? like its hard to pick one, but maybe its easier to remember the few not to use

[peak dove]

Image 4 Prompt Adapter - u can add more than 1 image as a prompt - it gets really good consistency in faces etc

[pure crystal]

trypophobians beware

[kind pewter]

my new UWQHD background :D

[shy kelp]

oh lol ok. i thought ipadapter like 'ip' for intellectual properties aka big popular franchise type model haha

[vale eagle]

pixart-alpha-leaked: 1, it could fit into 24GB in inference. 2, it is not enough when decoding the image. 3, a prompt from the paper: A Pikachu with an angry expression and red eyes, with lightning around it, hyper realistic style. (Not int8). PS: It is from a leaked weight might not represent the final release performance.

[noble shoal]

With the 512 or the 1024 model? Not that I have 24gb anyway.

[lusty wolf]

Good Night!

[bright valley]

Dude

Guy named Dever from the civitai discord just posted this

😂

[high skiff]

@bright valleywow, so far out of all of the companies I have talked to, NightCafe has been hands down the worst. I will never consider doing work with them again lmao. Everybody else has been super nice and responsive, and night cafe started off with basically "so who the hell told you we fund people?" followed by demanding their names, and the name of our research group, while being super aggressive and passive aggressive. gets a 10/10 yikes for me

"a simple question you can't seem to answer for the third time"

Like bro, what are they trying to do? Scare me into silence? I told them I wasn't even sure if they fund people, yet they also won't tell me if they do or they don't

It was a simple slip up lmao, not that serious man

[vital ermine]

CCP ran company?

[high skiff]

no idea honestly, but man it was just terrible lmao

every other company has been great

[vital ermine]

I have heard of them so I bet so

When they act like that then they have a lot to hide.

[high skiff]

I guess so, like damn

all the other companies had staff reach out to me within the hour

[vital ermine]

Next a van arrives, hits you over the head, and away you go to some bad place.

[high skiff]

super nice, super pleasant, really interested in my stuff

While they just basically tried to crucify me lmao

[vital ermine]

put you into one of those CCP metal restraint chairs like a modern version of the Iron Maiden and you get grilled for 23h.

[vale eagle]

1024

[vital ermine]

I see control net and eff nodes hate each other? Now I have to redo my workflow and ditch eff nodes.

[kindred shoal]

Are these the nodes youre using? https://github.com/LucianoCirino/efficiency-nodes-comfyui

They look..really cool actually.

[lusty moss]

Use the new repository.. just switched over to a new person. Uninstall the old ones and install the new ones.

[kindred shoal]

Who's the new person? I might get these

[lusty moss]

https://github.com/jags111/efficiency-nodes-comfyui

[vital ermine]

I think I do have the new ones but CN refuse to work with them

[lusty moss]

Haven't tried CN with them, was more for Quinzel.

[kindred shoal]

CN refuses to work well with me anyway, always messes with image quality.

It only works well in a1111, comfyui is like trying to pull teeth depending on the model.

[lusty moss]

Maybe report an issue at the GitHub for CN?

[vital ermine]

Well, no choice I must redo this and remove those node and unstall them.

[lusty moss]

What is the error?

[vital ermine]

it would require reporting on his and frankly I got introuble on Github for reporting so I don't do it anymore.

[lusty moss]

I was having an issue with floats, had to redo my setup which I'm currently doing.

[vital ermine]

overbaked, horrible.

[kindred shoal]

Overbaked?

Like deepfried or something?

[lusty moss]

Gotcha. That sucks. Adjusted steps and CFG?

[vital ermine]

yeah, this works on other nodes so must be the eff. I never really liked them anyway

[lusty moss]

Can recommend ComfyRoll if you are not using those nodes, might be worth a shot.

[kindred shoal]

If it's oversaturated, that's an issue I face on certain models with cnet on SD1.5.

i have to turn down the start and end usage to nearly the very beginning

0.2-0.3

doesn't make the poses look nice, but it gets them alright.

[vital ermine]

yeah, this is crazy as it isn't related to the model it is just something odd in here

[kindred shoal]

Yeah I have that issue on sd1.5. The colors look so bad sometimes, while others look passable.

[vital ermine]

yes

[kindred shoal]

I don't have any examples on me since I moved them to my flashdrive, but they are...awful.

[vital ermine]

the background looks streaking and the sat is off the scale

[kindred shoal]

I have no idea what would be wrong with my workflow to provide that kind of issue, as it passes through controlnet then an upscaler, then ultimate sd upscaler... ...could it be my cutoff prompter?

[glad grove]

probably the upscaler

[kindred shoal]

Hmmmm.

I'll have to experiment. this sucks xD

So cnet could be arguing with the upscaler, so I need to figure out how to use ultimate sd upscale without running it past an upscaler first...

[glad grove]

best way to do it would be make img with just controlnet then use another workflow to upscale

[steady grove]

if you're using ultimate upscaler with controlnet, make sure the settings have "allow other scripts to control controlnet" or whatever it says

[kindred shoal]

how do I do that on comfyui?

[glad grove]

yea a1111 has that setting also the other setting in a1111 is disable controlnet in high res fix,idk about comfy

[kindred shoal]

Oh it's a1111, okay

[glad grove]

yea in comfy idk probably had to do a bunch of spaghetti nodes

[kindred shoal]

That's what I fear, I like working on one workflow, as I have no idea how to send a batch of images into USDU without doing one image at a time

it's probably super simple and I'm just daft

[steady grove]

ultimate upscale script is fro webui, so i don't know.

[kindred shoal]

Usdu is also for comfyui, so i understand the confusion.

[glad grove]

if u use 1.5 just do it in a1111

[steady grove]

i'd put the controlnet nodes on the connections leading into the upscaler node

[kindred shoal]

I would, but a1111's --no-half argument makes my 1660ti chug like mad, and comfyui doesn't do that.

but at this point, with all the issues i've been having with comfyui on my 1660ti, i might as well switch back to a1111 and just deal with the chug.

[glad grove]

how many it/s u get in comfy compared with a1111?

[kindred shoal]

About the same I think?

[vital ermine]

Here, this is what I mean.

[kindred shoal]

so like.. 1-2s/it

[vital ermine]

wtf is up with that? turn of CN and fine

*off

[glad grove]

dont use -no-half for 1.5 since u have 6gb no need for it

[steady grove]

i grew tired of the it/s measure. it was useful to make sure my 4080 was being used the best. i like seconds per image now.

[kindred shoal]

Oh that is awful 😆

[kindred shoal]

I have a 16xx, it's mandatory according to the community guides here.

[glad grove]

try without it to see how it goes

[kindred shoal]

Alr, I'll go fetch the laptop and see what happens. I think I had to use --no-half-vae or else it'd give me a float point error

[steady grove]

1660ti has proper fp16 support so i dont see why you'd need the switch

[glad grove]

if u get NaN try to change vae use clearvae for example to see if it fixes your problem

[steady grove]

or --no-half-vae maybe

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Troubleshooting#a-tensor-with-all-nans-was-produced-in-the-vae

there should be songs praising the manpages

[kindred shoal]

never heard of --upcast-scaling... hm.

[vital ermine]

got it

you can't use the normal way you have to use the fat stack way

[steady grove]

💵 💵 💵 💵

[vital ermine]

hehe

[kindred shoal]

downloading everything for a1111 now...

this is gonna be a minute

[kindred shoal]

aye nice! Glad you got it working!

[slender coral]

Hey lads, if I wanted to move over this node to another workflow how would I go about it?

[kindred shoal]

Is there a copy paste function? I feel like i read about that somewhere.

[heady vale]

just ctrl c and then ctrl v

or ctrl + shift + v to keep connections

[slender coral]

That works holy shit, thank you.

[slender coral]

hmmm, do I need to select all connections? this one isn't workign for me

[heady vale]

yeah the last one is if you have multiple nodes to paste

[slender coral]

Select all nodes,

Got it

[lusty moss]

So has anyone had the problem where comyui just starts que on boot? Never had this happen until just now.. but it isn't actually working throught the pipeline?

[kindred shoal]

So um. My computer is nearly completely frozen while it's collecting torch, and it's been collecting torch for 5 minutes.

[heady vale]

click the command window and press enter. you might have paused it

[lusty moss]

Had to disable and re-enable "Extra Options" in the menu.. weird..

[kindred shoal]

pressed enter and now I'm waiting for it to do something

[slender coral]

So I'm trying to understand these nodes and I don't if I wanted to easily add img2img to an existing comfy workflow is there a quick way?

[lusty moss]

Can send you my workflow it is txt2img>UltimateSD Uplscale>After Detailer. Just refactored to it to be easier to understand. Would have to download my CSV Loader custom addon but also could re-adjust that part since it is just the prompt. Or steal whatever portion you want and paste it into your workflow.

[slender coral]

I have the workflow I'd like to use A and B, I just don't fully know how I would input my Img2img before the vae:

There's no input nodes like my A workflow

This is the first "Image save" but there's no output like my second image:

Do I need to code this in python?

[lusty moss]

Whole lot of all over the place so can't tell much what is happening honestly..

[heady vale]

before the base Ksampler - latent image node, you want an upload image node then a vae encode.
You can switch manually between either img2img or text2img

[lusty moss]

Mine if you are interested.

[vital ermine]

Yeah, it dawned on me since I saw a special point on it

[slender coral]

Am I on the righ path?

[slender coral]

Thanks

[heady vale]

then image goes to vae encode, then into latent image

[slender coral]

Do I need to grab a vae?

[kindred shoal]

There is so much I'm looking at, but goddamn that's so organized

[heady vale]

yep

[kindred shoal]

you should have an option on the model with "VAE" and you can hook it up that way

that's what I do anyway.

[heady vale]

yeah either from the model or separate, whichever you use

[lusty moss]

Is it easy to get a custom node added to the comfyui manager node downloads? Want to add my CSV Loader but unsre how to do it. Do I just make a clean ComfyUI repo on my PC and push the change?

[slender coral]

I'm so close, do you know what I'm missing?

[heady vale]

image is probably too small?

grab a 1024 image

[lusty moss]

Not sure if you can use this to steal my workflow if anyone is interested. Can upload the JSON. Not sure how discord handles the pictures.

[slender coral]

Thank you so much, this is why I hated blueprints and visual scripting. I should have put together that the empty latent image was being generated in the prev node.

Could you suggest a repo or anywhere I can learn all of these? Or is it learn by fire?

[lusty moss]

https://tenor.com/view/chiquichico-gif-26004262

[slender coral]

Final question, do I need to pass this through an upscaler to manually set a size?

[lusty moss]

More learn by discord lol.

[heady vale]

either use the workflow from other people or try using the examples on the github page. you will learn it slowly
https://github.com/comfyanonymous/ComfyUI_examples/tree/master

[slender coral]

Appreciate you lads.

[lusty moss]

Anyone that uses After Detailer have a problem with it closing the nostrils for faces?

[glad grove]

no problem at all with that on a1111 idk about comfy

[lusty moss]

Just seems to gloss them over.

[slender coral]

Final question, I have a 1024x1024 that I want to turn into a 1024x2048, the workflow example just explains the latent image. https://github.com/comfyanonymous/ComfyUI_examples/tree/master/img2img

would I need to rescale my image in Paint and run it, or can I do it easily with nodes?

Just so used to

[lusty moss]

After Detailer settings. If anyone can see something wway off let me know, not super familar with this node yet.

[glad grove]

try to reduce feather to 3

also tried it on a1111 and no problem with face

[lusty moss]

@bright valley V2

[lusty moss]

Still smoothed them over, will try to put closed nostrils in the negative and open nostrils in the positive and adjust some settings..

[glad grove]

try to reduce denoise to 35

[lusty moss]

These the standard models used?

[glad grove]

or 30 even

yea

[lusty moss]

Using Art Nouveau style?

[vital ermine]

yeah, testing it out

[lusty moss]

Tried .30 , no luck. Should the mask look completely black and white or more like a depth map?

[steady grove]

trying to figure out how to load workflows in comfy so that stable swarm ui gets them. i want to use it but i want things like face detailers or high res passes iin the ui

i get as far as all the custom nodes installed in the comfyui folder that stableswarm installs, but it just hangs when it restarts. loading forever

[lusty moss]

Can try these artists to push it further. Art Nouveau, Alphonse Mucha, Gustav Klimt, Louis Comfort Tiffany, Aubrey Beardsley. Use that style in my CSV Loader a lot. Like the results as well.

[bright valley]

i think people arent seeing the rainbow one

[lusty moss]

Just downloaded, was going to try after I figure out my after detailer issue. Any changes in prompt setup?

[bright valley]

which sucks

Nope just some minor additions

they're in the prompt notes

[lusty moss]

Cool, setup a little workflow to more easily swap colors so wanted to make sure it would be the same.

[glad grove]

try to change the sampler from spm sde to dpm 2m karras

[steady grove]

if i remove all the custom nodes , the backend server restarts fine. lame

[lusty moss]

Trying it now

[lusty moss]

Think changing the sampler helped.

Thank you for the help!

[lusty moss]

Does ComfyUI Manager automatically add your custom nodes to the database? Looks like mine is already listed there, went to add it and it was already in the JSON file. Curious if someone else added it or if it does it automatically..

[nimble heart]

think the manager author adds them as they start to show up in workflows/searches

[lusty moss]

Awesome, saved me some work and glad to know it is available via the missing nodes or if anyone wants it or uses a workflow i put out.

[craggy agate]

made a jax lora

[steady grove]

its supposed to, but many custom node authors completely change the way nodes work instead of freezing their model early in development. as a result, if you pick up a workflow older than a month, it'll likely be built for an older version of the nodes and the manager won't help at all

it always installs whatever the newest version of the node is, which could be wildly incompatible with older layouts

this wouldn't be so much of a problem if authors just froze their node models from the start and not providing legacy versions . it's just such a messy development culture.

[steady grove]

dunno who that is but it reminds me of jiren from dbs

[slender coral]

Could someone help me out with telling me how / where I would put a denoise to change the image weights on my prompts?

[crisp owl]

denoising is just where the steps start at vs where it either ends at, or what input it's given. Using the advanced nodes you just have to play with the steps start to determine what works best for you.

Say the 1st ksampler ends at 20 steps, 2nd ksampler you can start it right at 20 steps, or alternatively to denoise more start it at step 15, or to reduce denoising you can start it at step 25.

Alternatively, you can just go with the standard ksampler and use the denoising % option and not deal with fiddling with the steps start/end etc

[slender coral]

A okay so just so I understand, if I want more of my image start the sampler later?

[crisp owl]

Yup

Other things you can do to fiddle with that some more, but in essence, yeah

[lusty moss]

Just found out about this, might help you @slender coral they also have a discord.
https://comfyworkflows.com/

From the ComfyUI Manager interface.

https://comfyworkflows.com/workflows/320ac429-28cd-4192-a7c3-da1ed3fd1893

Was able to share mine on there for downloading.

[nimble heart]

fun test image
high resolution dslr photograph of pink roses in the misty rain -w 1256 -h 840 --seed -1039012146 -g 6 -s 30 -G 0.8

[high skiff]

Does anybdoy know how to merge LoRA's? I could really use that for my project I am working on

I trained the same dataset 2 different ways, and they are decent alone, but excellent when run in series

[arctic bloom]

short answer is possible with kohya-ss sd scripts (same with SDXL and pre-SDXL)
https://github.com/kohya-ss/sd-scripts/blob/dev/networks/sdxl_merge_lora.py

[high skiff]

incredible, thank you!