✨｜sdxl

[hardy cipher]

still having a bit of trouble figuring out the refining of images in sdxl. I've seen so many different approaches and tried a lot of things, but haven't found consistency yet

[fresh path]

can anyone share their workflow to upscale sd1.5 images with sdxl

[hardy cipher]

like this is almost on point I think? but I don't know, kind of feels like it could come together better

[upbeat summit]

nice layers of fluidity

[static prawn]

which one u like more?

[trim orbit]

differences are to subtle to determine a winner really

one has better contrast and one has better details

[static prawn]

yeah i feel like the highresfix always delete kinda detail

😦

[hardy cipher]

Hmm. I'm not completely familiar with that term, but I really put a lot of effort into making things look proper. which is of course subjective

[thorny frost]

Many cats and so few dogs. What injustice. 😦

[hardy cipher]

if in doubt just take the pixel value of the two images and then blend it back into one of them with dodge or burn or something

*pixel value difference

[upbeat summit]

yes, you've been posting some very interesting and unique creations. very exciting! I was talking about the fluidity of the paint and colors, how the layers of paint / different colors have merged together and feel three dimensional because of the lighting and shadows.

[hardy cipher]

thanks! I like to experiment a lot with prompts. they get a bit wordy but they make cool things

[upbeat summit]

yep 🙂 same

[hardy cipher]

I've read tutorials that say that people should keep the prompts concise. but I respectfully disagree, lol. well at least with what I'm making

might not want a huge prompt if you're trying to make other things

[thorny frost]

I think that is for 1.5. with SDXL the longer the better

[upbeat summit]

yeah. having a sleek prompt build is also nice, but I tend to layer concepts, art styles, weighting and mixing tokens in my adventures to explore the latent space 😉

[hardy cipher]

well I don't start off with a big prompt. when starting something new it'll normally be a couple words

then I build on it

[crystal forge]

prompt length also depends a little bit on the goal, sometimes you want something loose to explore different things (shorter prompt better), sometimes you want someting very specific and only have small details change (longer prompt generally better)

[hardy cipher]

and these days I have different things I'll recycle just to see what happens. totally unrelated concepts, lol. but they work

[upbeat summit]

yeah absolutely

[hardy cipher]

wow, that's the kind of style I go for there

[upbeat summit]

for me it's one of the most exciting aspects of it. using different styles, concepts, loras, fine-tunings in general and use them like droplets / ingredients to create something interesting

[crisp owl]

I once got criticised for having a "disjointed" prompt. I just laughed and asked what a proper prompt looks like

[hardy cipher]

I need to et back on this workflow. kind of got bored with it but has potential I think.

never really felt like they were finished though

[hardy cipher]

well I think a lot of people are focused on making or mimicking things that already exist, but I could just look at those things. I want to make new things that no one has ever seen.

[crisp owl]

Yup, that's exactly what I did and he said he liked the image, but still criticised the prompt. I said go ahead and then replicate the image and tell me my prompt is wrong.

[upbeat summit]

yeah. I have some really long and convoluted prompt builds that probably only confuses and overwhelms the model's text encoder.

"don't do this"

but if you like the output nobody can tell you what is right or wrong with this. if it makes you happy you can use any technique you want.

[hardy cipher]

meh, I won't criticize anyone's prompt unless they're looking for critiques. I don't know the right way to do any of it. I don't know that there is a right way

I just focus on putting the important stuff first

so frame the basics of the image structure at the outset. then get weirder as it goes on

and not for the sake of being weird exactly. but it adds layers of complexity to the image

[crisp owl]

If someone says their way is the right way, they're wrong hahaha
Perhaps to get to their result, but that's the nuts thing with all this, just how many damn variables there are. Move one thing accidentally and get 100% different results, or different styles with the same theme, etc etc.

[static prawn]

i dont like the horror pics in sdxl, they always look exactly the same xD

[jolly creek]

Like even commas and periods can affect the output

[hardy cipher]

do they?

[crisp owl]

I was messing with this accidentally yesterday, realized I had the wrong prompt connected to clip L than I thought, so I corrected it, but then didn't like the results anymore hahaha

[static prawn]

i always get those white, bone like characters

[hardy cipher]

lol

[soft zealot]

they're called skeletons ;o)

[static prawn]

oh crazy, never was able to generate those

for me it always tend to white bone like stuff

[hardy cipher]

Uncanny children with shiny cracking porcelain skin, like that of an antique doll, living in an alternate universe that is nothing but beans. They stare through you with dead eyes. The thousand yard stare. They've fallen through a wormhole into another dimension. And there is nothing but beans. They're buried in beans. They're eating the beans like ravenous animals. No love, no parents, no hopes no dreams. No home, no shelter, Just a perpetual living nightmare as the singular component of this dimension is beans. In this existential nightmare they ravenously eat the beans, but it's not as though they need to. They don't need to sleep or eat, they never get full. They spend their perpetual neverending bean nightmare, every moment agony, trying to fill a void that can never be filled. With beans. Only beans. Just beans

[crisp owl]

at least he's trying to eat healthy ^

[hardy cipher]

children of the bean

[jolly creek]

so yea my point is that variety is doable

[static prawn]

probably have to work more with prompts, i still dont get how to get better results, my images always look generic

[hardy cipher]

I like the comic. really illustrates it well

but I got beaned out so no more existential bean nightmare images for a while

[jolly creek]

working double shifts gets you tired

[crisp owl]

is that mark zuk?

[hardy cipher]

without his skin

yes

[crisp owl]

Damn, so the theories were true

[hot rover]

Still tinkering with this process. I love Sytan's workflow, it just works. What I didn't like was allowing the latent space to change things or guess as much. I thought eyes and facial details always came out blurry, cooked or warped in my tests. I added a face refiner group of nodes and a sharpening node before the upscale does it's thing and I think helps the "Upscale Mixed Diff" node work just a tiny bit more predictably because it introduces more "good" detail to work with. My example images had some of the levels adjusted to compensate for some contrast, but overall it seems to work as expected and it works for weird faces too (zombies, ghouls etc.)

[trim orbit]

https://research.nvidia.com/labs/par/Perfusion/ if i knew how to code ML stuff i'd do something about this paper just sitting here, looking at me

[urban fjord]

The promises there is what you find in every paper since textual-inversion.

[trim orbit]

this one seems more demonstratable. google's got their new method too which is waiting on code

[cedar tusk]

Question about SDXL lora training: I have a set of captioned images I used to train a 1.5 lora previously. Many of these are likely not at 1024^2 resolution (or any of the SDXL resolutions), and some might be 512 on the short side. Will those still work for SDXL lora training or is that a bad idea?

[zealous horizon]

For your OCD Cable management.
I call them microdots. They allow you to lay your cables as you desire.
works with default reroute nodes

It helps me understand what goes where on complex workfows

[upbeat summit]

So your nodes extensions are now a must have for my workflows. Thank you for making these QoL features. Really awesome and keep up the great work!

[zealous horizon]

they are just extensions of the UI, "not custom nodes". so far I was able to just do that with the default reroutes. Therefore is is compatible to any workflow with reroutes.

[upbeat summit]

that is what I meant 😄

yeah the microdots sound really interesting. I will try them out shortly

[zealous horizon]

there are 3 modes

[boreal bough]

I'm assuming your were using an ancestral sampler?

[ionic gulch]

results with refiner steps and upscaler steps are so different, i think i have to build a workflow that generates 6 variants of an image to compare and pick the best one. this one for example is refiner 10% instead of 20% - much better than 100% base or 20% refiner:
https://i.imgur.com/MZsWjaj.jpg

[alpine pine]

its the relative amount of steps the refiner vs base do

[ionic gulch]

@vernal ravine calculating the end_step for base = start_step for refiner. so i can enter 80%, 90%, etc. - my new workflow will drop that and create 80%, 90% and 100% base every time - same for the upscaler - that one needs at least two different strengths to compare - so i get 6 results. and if you do it in in a way that uses previous results you don't have to wait 6x longer, but still 2-3x longer than one variant.

[upbeat summit]

looks like a great way to organize the links in more complex workflows.

so if someone doesn't have the web extension installed and opens a workflow that used them, do the links follow the normal flow?

and when it is saved again on a comfyui installation that doesn't have the microdot extension, will the settings be lost? or is it a global setting like your other extensions?

[zealous horizon]

I add metadata to the properties (that info is saved in the file),
Without extension the reroutes are just default. (because they have a weird resize force in them)
You can see that if you try to resize a reroute. It will always jump back to it's original size. I hacked that part when dot or microdot are active.

If someone saves the file without the extension, the logic will remain unless a reroute is deleted (obviously)

[upbeat summit]

nice - great solution 🙂

[echo pine]

are the best settings/workflows that are found with the bot open sourced?

[upbeat summit]

There is one thing I think would be a great QoL addition. Do you think there is a way to increase the resistance of the snapping? So it takes a bit more resistance when pulling something out of a snapped grid position.

I know that is a very delicate thing to tweak but coming from other applications and from doing UX/UI design it is something I think could improve the overall user experience.

[zealous horizon]

threshold? yeah, I can check if the movement can be overwritten.

[upbeat summit]

Maybe there is a parameter in ComfyUI's source code I haven't found

[upbeat summit]

yeah, exactly

I mean maybe I'm the only one that feels this way, but in other applications the threshold is mostly higher so the snapping is more obvious.

Also elements are snapping more obvious closer to snap positions - magnetic like. I know they are couple of parameters that play into it.

I'm not 100% sure how it would feel in ComfyUI though without seeing how it affects it.

[unkempt fractal]

so what we saying, sdxl + hires fix + refiner or without hires fix

[indigo carbon]

me personally; SDXL->last few first pass steps with refiner UNET->hiresfix->last few hiresfix steps with refiner unet

there is an A1111 extension that allows you to implement this in a very efficient way.

by far my favorite way to use SDXL

[barren bluff]

Is there a way to use tiled vae with ComfyUI?

With the SDXL workflow, I mean

[unkempt fractal]

UNET < ?

[upbeat summit]

Hi. I'm sorry I have not yet got a good upscaling workflow going for myself, so I don't know what makes most sense. but you can look at the ComfyUI workflow from Sytan. He recently released a new version (1.0) that includes an upscaling pipeline:
https://github.com/SytanSD/Sytan-SDXL-ComfyUI

For my upscale tests I mostly use Topaz Photo AI

[sonic furnace]

since they moved back to 0.9 vae im noticing the red squares on images alot more 👀

[indigo carbon]

the UNET is the crucial part of the refiner. by swapping just the UNET you can have a massive performance gain without having to unload and reload the base and refiner after every gen. the extension allows you to keep both base UNET and refiner UNET in CPU RAM, and when it's time for the refiner parts or the workflow, it's swaps the UNET's using the ones it keeps in CPU RAM with the ones it keeps in VRAM. this way you use the refiner efficiently without effecting performance.

[unkempt fractal]

I have the refiner addon in A1111

this does that, doesn't it?

[indigo carbon]

yeah, that's the extension that allows this efficient workflow.

[west breach]

morning

[heady vale]

morning

[indigo carbon]

'morn

[upbeat summit]

greetings .)

[west breach]

Had my PC running gens all night with this upscale workflow and prompts from sandbit's file. Interesting to see what it's strengths and weaknesses are. Can't wait for a tile controlnet to help the with upscaling!

[heady vale]

that would be nice

[heady vale]

you got XL working?

[vital ermine]

yes, but I can't do 1024 sq I oom

[vital ermine]

ddim no workie in XL in auto either

[heady vale]

yeah I remember ddim not working in A1111. COmfy is so much better in almost everything though

[vital ermine]

I can't stand comfy's ui as I tried it and just no way

especially how I work it is far too cumbersome

[upbeat summit]

look who's here 🙂

[heady vale]

pretty sure you wouldnt OOM with tiled vae in comfy

[lean garden]

sdxl with no refiner in auto1111, can't see the point of using the refiner when you can already get details like this without it.

[vital ermine]

yeah, that 32 bit float VAE does me in

[heady vale]

with Sytans workflow and tiled vae nodes I did 8k on a 3080

[upbeat summit]

the refiner isn't necessary for all images, styles and settings. it really depends.

[lilac raven]

I installed the refiner extension on a1111. It uses double the memory but it's possible to generate 2 images at once
(by the way, how do I know if the refiner is working on a1111?)

[lean garden]

That's what i was thinking. I haven't generated anything yet where i thought i was missing something.

[vital ermine]

Well, comfy isn't going to work out for me as the reverse mouse wheel, the lack of a sider, and when you use 4-6 loras, hypernetworks, etc... per prompt you begin to feel the issues

[lean garden]

I'd just not bother with it altogether

[upbeat summit]

Hey there 🙂 Long time no talk. There are couple of custom nodes that allow you to insert loras in a prompt editing a1111 way.

It's all experimental and new but it will work eventually

[west breach]

Maybe someone will make an extension to use comfy as the backend?

[vital ermine]

Heyas, I saw you in here but my card I never tried XL. Getting used to it as I plan on getting a 7900XTX end of Aug.

[upbeat summit]

nice!

[vital ermine]

@upbeat summit the two big ones for me in comfy is how I can't just open up a menu to get to my embeddings, loras, etc... and the reverse mouse wheel (that I complained to comfy on day 1) never had the option to allow me to reverse it really is bad.

[untold totem]

Hiya. I'm using vlad's automatic for SDXL, and all seems good. Except the Variation Seed doesn't seem to work? Is that a known problem, or should it be working?

[lilac raven]

The refiner does influence something, I don't know if I want that or not (time to img2img to get proper crossed arms)

[vital ermine]

without vs with refine 0.4 denoise

[pallid path]

zombies

[west breach]

Try using refiner only at 512x or 768x resolutions. I get the feeling it's a 512 model based on being able to compose images best at lower resolutions

[static temple]

Any moderator in here?

[pallid path]

infally, drip pope francis, but open source

[vast narwhal]

There's a way to discover the exact seed of an image when doing batch generation? All pictures seems to have the same metadata

[vital ermine]

I can't actually go larger than 960x536 anyway

[static temple]

i think in xyz plot seed> 1,1,1,1,1,1,1,

then find the image and press same seed

[vast narwhal]

I will try. ty

[static temple]

free upscaler and does a good job and you can do batch images all at once https://www.upscayl.org/

[hard fractal]

Is this person in here? https://civitai.com/models/120763/filmic-vibes-episodic-sci-fi-style-or-sdxl

Doesn't seem like it

[autumn forum]

damn someone beat me to uploading 😦 sad

[hard fractal]

They're definitely not passing in the right conditions

[autumn forum]

maybe mine will be better

[hard fractal]

They're probably doing 4096x4096

[vital ermine]

Thank you

[hard fractal]

Cause SDXL doesn't make doubles at widescreen.

[upbeat summit]

I've used nodes to build my own metadata text file generator right inside ComfyUI using a Save as Text node. It grabs the seed of the current image, prompts (even populated dynamic prompts) etc.

It creates a .txt file with the same name as the image. So you have all the metadata you need for that specific image. And the image itself includes the rest of the ComfyUI workflow that you need to recreate the image.

I guess we need some custom nodes that saves the seed and populated dynamic prompt in the workflow. I got some ideas 🙂

[hard fractal]

@high skiff fix your workflow 😝

[high skiff]

ping

[pallid path]

harry got the new drip 👀

[high skiff]

I've already put enough work into it

I tested latent size conditioning and got a very remarkable almost no difference

But I need to test some other things more

[hard fractal]

but you only tested square is what I'm betting

This is gonna break everything non-square:

[high skiff]

no, I tested from 21:9 to 9:21

and so have some other people I worked with. we have all deduced there is not much of a difference, so I am not sure what to say

Tho I will say, if you go below a 1024^2, it gets really weird

I have tested it before, and i will run more tests soon

[hard fractal]

okay

watch

gimme a prompt

[high skiff]

I am busy with other things right now, I don't have the time

[pallid path]

Billy Butcher if he didn't join the boys

[hard fractal]

okay 🙂

[boreal bough]

I did the 4:1 apsect ratios because I can XD all worked out

[cinder aspen]

I think i should lower the width and lenght (not the target ones) from 4096x4096 to 2048x2048. My laptop had a hard time for at least 10 seconds lol

[high skiff]

it affect gen times none

[cinder aspen]

Also some other thingd

[boreal bough]

oh wait, yeah I manually adjust those when I change aspect ratio

[cinder aspen]

things*

[upbeat summit]

In the beginning I was adjusting the sizes according to my aspect ratio. Than I read in the chat?! that 4096x4096 makes sense for all resolutions and aspect ratios.

So I've been doing it right and the last few weeks wrong? Do the target sizes have to be adjusted to the aspect ratio? like 4096x2048 for 2:1?

[hard fractal]

yeah, you have to

[west breach]

does any comfy node dev know how to get the image size from the conditioning object?

[hard fractal]

yeah, they do

make a very tall image.

you'll see.

[boreal bough]

can confirm. changing these to fit your new aspect ratio makes them work. without adjustment you get the 'crop' or outpaint effect. usually not what you want

[upbeat summit]

yeah I was getting duplication and I thought - okay these are the limits of the model (or my bad prompting)

[hard fractal]

nope

the model doesn't make duplicates

[upbeat summit]

Ok I feel stupid now. I should have just followed my first hunch 😄

[cinder aspen]

I blindly applied the setup of a youtuber without considering his PC specs and mine

[hard fractal]

or our official workflow

lol

[brazen patrol]

so I tried to load Sytan 1.0 and all I get is a black screen.

No nodes appear

[upbeat summit]

but 4096 should be the maximum internal length for either width or height, right? depending on your ratio

[hard fractal]

idk

we're testing it now

[cinder aspen]

I might have opened the portal to hell

[boreal bough]

but have you tested my finetune settings? cause the community members that have tested it, made it work 🙂

[upbeat summit]

so do we adjust the target_width and width to the same values?

[high skiff]

general awareness, in the SDXL chat? :O

[cinder aspen]

Just when I said i might have opened the portal to hell...he appears

Hmmm

[vital ermine]

LOL, wth?!?

[cinder aspen]

Fast! Bring the holy water!

[vital ermine]

I will be getting a better card so just playing with it now as 1024 gives me OOM

[upbeat summit]

ok so I'm now going to build a math node setup that calculates the "latent reference" values depending on my image resolutions

[pallid path]

last drip for today

[high skiff]

my 3090 gets here tomorrow

[cinder aspen]

What is OOM and what VRAM do you have

[vital ermine]

Oh, nice

[pallid path]

hopefully not broken this time 😔

[hard fractal]

careful with your conditions! SDXL is pretty good at not making tall people, or duplicates when wide:

[vital ermine]

shhhhh

[heady vale]

wiiiide

[vital ermine]

OOM is Out Of Memory and I have a 6GB 1060.

[pallid path]

cyberpunk?

[upbeat summit]

why I have been using SDXL all wrong (when I used it right in the beginning)

[pallid path]

I thought you had a 3060

[vital ermine]

no, that was Sytan

[upbeat summit]

ok. let's fix it with math :D. I don't want to update this manually because I change resolutions so much

[boreal bough]

input nodes?

(primitives)

[hard fractal]

@upbeat summit – because you're trying out community workflows

and that's great!

[vast narwhal]

Ty my friend.

[upbeat summit]

but I build my own workflow in the beginning using the correct method - and than I adapted something I shouldn't have 😄

[hard fractal]

this is same seed

4096x4096 <> correct conditioning

[pallid path]

excellent

[echo pine]

the 4096 setting kind of adds more detail, sometimes its good but sometimes it might add some stuff that shouldn't be there

[high skiff]

@hard fractalI do see some differences with res conditioning, but nothing... Major, from my tests

just different images, nothing really better or worse

[upbeat summit]

yup time to adapt and improve 😄

[cinder aspen]

So if i lower my width and lenght from 4096x4096 to 2048x2048 (target 1024x1024) i might have less issues with my specs?

[west breach]

@hard fractal create a ML model that outputs the best sampler settings for a given prompt based on the bot data

[high skiff]

this is 1024x1024x conditioning

this is 1700x600 conditioning

[hard fractal]

That's a too-wide tiger

Yeah

There ya go

[candid walrus]

phat

[boreal bough]

chonker tiger 🤣

[candid walrus]

don't be so jugdy llooll

[upbeat summit]

16:9 works great with the "wrong" setting 😉

[autumn forum]

🅰️ | 🅱️ | 🇨 | 🇩 ???

[high skiff]

Like, I do see a difference, but its not as big as I think its being made to seem

[cinder aspen]

Mr. Emaaad...we saw you typing 😉 😛

[hard fractal]

#✨｜sdxl message

[echo pine]

try 3400x1200 conditioning, does it add more detail? or maybe double or quadruple the conditioning for the upscale?

[high skiff]

I am only going off of things I can repeat myself, personally

[hard fractal]

Try 9:21

like I said, there are several things that are incorrect about your workflow from a technical perspective

[alpine pine]

Looks like OP updated the description and is here

[high skiff]

feel free to share the info rather than just expecting me to figure it out, if thats the case. I am open to construstive criticism

oh lord

[hard fractal]

Your conditioning is wrong for anything but square.

ddim-u is the least-preferred schedule.

[hard fractal]

Now, trying conditioning it with the correct size

[high skiff]

that is conditioning lol

[hard fractal]

What size?

[high skiff]

600x1700

[autumn forum]

well i guess no one cares enough to vote on my post cuz im a nobody. lol. yall couldve just said they are all shit😂

[candid walrus]

here's some phat tigers...

[hard fractal]

that's not 1MP

[high skiff]

this is not conditioned

[ionic gulch]

part 1 of my multi-workflow is done:
https://i.imgur.com/oI0w6PH.jpg

[hard fractal]

they're all conditioned. it just matters what size.

640x1536

that's the right 9:21 size

[upbeat summit]

I think this looks correct

[high skiff]

its 0.98MPX, its close enough

lmfao

also, thats not 1MPX

[hard fractal]

can you share me your workflow?

[high skiff]

thats further off than mine is

[hard fractal]

fair!

you're right

[boreal bough]

sketch, a drawing of the town from shaun of the dead
+sketch LoRA
4096x4096 | 2048x512 | 512x2048
(all on seed 2)

[hard fractal]

what's your tiger prompt?

[echo pine]

try doubling the condition size from image size, does that help?

[high skiff]

its the stock prompt in my workflow

[boreal bough]

results vary I guess? in this case 4096x4096 actually won XD

[hard fractal]

which is, hwat?

[high skiff]

not conditioned

conditioned

I prefer not conditioned here

[hard fractal]

okay hold up

[high skiff]

both look rough tho

[hard fractal]

let's get the terminology straight

those are both conditioned

one is conditioned at 4096x4096

the other, at what?

[candid walrus]

workflow be like...

[high skiff]

one is at 1024x1024, the other at your res

[cedar tusk]

Question about SDXL lora training: I have a set of captioned images I used to train a 1.5 lora previously. Many of these are likely not at 1024^2 resolution (or any of the SDXL resolutions), and some might be 512 on the short side. Will those still work for SDXL lora training or is that a bad idea?

[high skiff]

one is aspect conditioned, the other is not

[alpine pine]

so.. are we supposed to set the latent size to 2x the resolution?

I am confused

[cedar tusk]

Basically, can SDXL LoRAs be trained with images that are below the standard resolutions of SDXL images

[ionic gulch]

using 4096x4096 or something like that gets problematic if aspect ratio is not 1:1 from my experience. results often get stretched in that case

[boreal bough]

sketch, a girl walking down the street
2048x512 | 4096x4096

[upbeat summit]

are you really telling me that my "wide shoulder and head trauma" is because of my wrong settings? omg

[high skiff]

I see that sometimes as well, but I also see that conditioning at the aspect ratio also has the same issue, just on different seeds

[boreal bough]

4096 only works if you don't mind the 'outpainting' effect

good for towns, forests, landscapes. bad for anything where quantity of objects matter

[candid walrus]

it's been done, you can give it a go and then try to compare, i seen some people have success in civitai, though i'm sure it'd be better if it was 1mp

[cedar tusk]

How should the base clip and target res be related? Same ratio but scaled (e.g. 4096 and 2048)? Should target res match generation res?

[upbeat summit]

that minimalism is amazing! when can I play with it? 😄

[boreal bough]

especially on closeup of a person, where multiple noses may be a bad thing

[echo pine]

it will work, I'm training on a dataset with like 40 large images and over 300 pretty small images, it does work but the images do look pretty lo-fi, the training images are pretty lo-fi too tho so it's not a great test, I think it's better to have high quality images if you can find them though

[high skiff]

oh my lol

[cedar tusk]

@candid walrus @echo pine Good to know, thanks. I'll see if I can even train locally first with my 2080 haha

[boreal bough]

will be a while till it's online, since I'll post them at the same time the guide goes up. will take a while as I'm essentially posting 14 loras + full walkthrough on how to make all of them, as well as how not to make them

but I don't mind sharing them ^^

[hard fractal]

click into that

[upbeat summit]

wow - thank you so much. I've been following your really helpful tips in #🔧｜finetune because I'm a noob with training and want to get started 🙂 thank you!

[candid walrus]

i'm really looking forward to your guide

[upbeat summit]

no chonkers

[high skiff]

I do see differences, for sure, I just don't think its as big of a deal as some may think

[hard fractal]

okay!

it's totally up to you

[alpine pine]

1. res * 2
2. 4096
3. 1024

The side view on the lora breaks regardless of latent size

[candid walrus]

to be honest, i would rather have the left outcome any day of the week

[hard fractal]

that's what's fun about this – some people love the gritty kind of look... some people don't mind stretchy tigers.

etc

[high skiff]

I can't implement it into my workflow in a math based way anyways unfortunately

[alpine pine]

Res * 2 seems to be the most distorted, so should latent stay square?

[hard fractal]

why not?

[high skiff]

if I was able to, I would

[west breach]

are we supposed to multiply the dimensions by 4 for the base? I thought it was only for the refiner

[high skiff]

no stock nodes to do the math with

[hard fractal]

ah, true.

[candid walrus]

i get the vibe of dumping tons of energy into the workflow and releasing it tho, and looking at it one more moment would be the end of me

[autumn forum]

get comfy to add em lmao

[upbeat summit]

I don't want deformed stuff if it's not part of the aesthetic

[high skiff]

if there were stock nodes, I would implemenet it cause I don't see why not

but as of now, i can't

[hot rover]

I mean, can we see your workflow? 😛

[echo pine]

maybe do a test with image resolution for conditioning for base
and then image resolution * 2 for conditioning for refiner
and then image resolution * 4 for conditioning for refiner upscale?

[candid walrus]

yeah comfy is so cool with how you can 'nearly program' with nodes, but its a labor almost more deep than programing if you don't have the right nodes/plugins

[alpine pine]

the exact image res looks to produce better than the 2x res

[high skiff]

I wish it had built in math nodes, that would make my work for the community much easier

[candid walrus]

same

is there even a 'switch' node?

that would be epic

[echo pine]

yeah but I mean exact same res in base and 2x res in refiner or did u try that

[high skiff]

there is not

tho I did pitch a switch node to comfy

but the poor guy is run so thin

[hard fractal]

careful with your conditions!

[candid walrus]

right

[alpine pine]

ah! I am not using refiner atm, but the upscale res def needs to be 2x as it distorts then

[upbeat summit]

on the 2nd image - the fixed one - is it correct that the negative target sizes are not adjusted like the positive ones, @hard fractal?

[high skiff]

these look a little more impactful

weird that the post processing is very different

[echo pine]

okay i'll try it tomorrow when my lora finishes training

[candid walrus]

can you be detailed on the math? i'm guessing there is some type if input latent space that needs to conform to the output aspect respectfully?

[west breach]

are we supposed to multiply the dimensions by 4 for the base conditioner? I thought it was only for the refiner

[hard fractal]

you're telling the model to make a specific image of a specific size

that's why the default comfy node just does it

now

you can trick it

1024x1024 is one kind of image.

4096x4096 is another type

and the latter looks better.

this is an active field of research for us

so, would love all your input!

[boreal bough]

fyi, if anyones not using refiner, move over to swarmui with comfy in the background. it's by far the best non-refiner experience you can get.
(it does do refiner as well - but if you go no refiner, it's just the best option)

[hard fractal]

maybe 2x framesize is best... maybe 4x...

maybe you change it in between

[candid walrus]

yes, but can you maybe post your full workflow and highlight the areas that respectfully need to change

[hard fractal]

all are valid.

[hard fractal]

I was doing Sytan

[high skiff]

screw it, I am gonna make myself an aspect conditioned workflow

[hard fractal]

that workflow

[candid walrus]

i think it would all a better understanding of what you're getting at

[astral jay]

Hi 😄

[hard fractal]

so, that's just switching to the default nodes

these

[west breach]

what's the difference between width and target_width?

[echo pine]

I noticed it adds a lot of detail when increasing the framesize with my tests from a couple days ago, that's why I suggested putting it on the refiner and upscaler, maybe that will keep the original image undistorted but add a lot of detail

[high skiff]

I am not looking to give up any granularity/functionality, so i will just make my own personal workflow with the math nodes and such for myself

[gloomy barn]

I cant get a handle on how sdxl treats tokens, the prompt...I completely remove something, but it still generates that aspect

[hard fractal]

sure! that's why I can't share mine

there are nodes we're not ready to release yet

because they're being tested on the bot

#1100170312106127410

et al

[candid walrus]

ah

[boreal bough]

being able to separate the two clips genuinely gives a lot of room for improvement. please dont take clip_l from me ❤️

[high skiff]

I think its time for me to work on some SDXL things for myself for once haha

[hard fractal]

width/height is the size of the original images as they were being trained.

target w/h is the sie of the bucket

[upbeat summit]

@hard fractal can you please clarify if in your image the negative target values are correct and should not be equal to the positive target values? or does it not matter?
https://cdn.discordapp.com/attachments/1089974139927920741/1136101501224353914/cond2.png

[high skiff]

ok, is there not a way to keep --listen on with oobabooga. This is pretty annoying

I am just trying to setup a LLM form some prompt mods

[hard fractal]

oh, I messed up one target there. oops. should've matched.

we think

that's also being tested 😄

[upbeat summit]

gotcha 🙂 thank you

[hard fractal]

good catch, mass

[astral jay]

Also these

[hard fractal]

try 0,0 on the negative for really funky stuff.

[upbeat summit]

let's just compress space and time and see what it does

[cinder aspen]

Enough of things i cant comprehend right now xD Time for reading a few pages of comic and then sleep xD Good night everyone 😄

[spark bear]

trying out some upscalers with ultimate upscale, NMKD supercscale on left , siax 200k on right..close, I think ultimate a bit better detail and contrast.

[boreal bough]

just don't start dividing by 0

[upbeat summit]

https://tenor.com/view/galaxy-explosion-space-gif-24958497

[azure oxide]

what if ai actually helps us figure out how to do so tho

[candid walrus]

then we don't know what nothing is

[hard fractal]

it can even go negative.

so can CFG.

-7 CFG

is literally the opposite of what you asked for

[candid walrus]

meh, just give me some pixels

[azure oxide]

have y'all messed around with that?

[astral jay]

I think i'm getting unprecedented results :v

[west breach]

@hard fractal so this would be the recommended config for a 1024x1024 image?

[azure oxide]

i remember someone messing around with negative values for controlnet weights, results inconclusive

[upbeat summit]

how much processing goes into this if I may ask? loras, 1.5 model diffusion to fix stuff? I mean it looks really good 🙂 nice work

[astral jay]

Straight up just SDXL and upscale, no 1.5

[boreal bough]

part of me thinks its a picture and I am being trolled

[astral jay]

And the Lora is trained on sdxl as well

[boreal bough]

so I'd say yeah. good result XD

[upbeat summit]

but hands how... did you find some magic prompt tokens I've been hunting for weeks? 😄

[astral jay]

Thank you! 😄

I was amazed honestly

T'is not me, T'is the magic of SDXL

[high skiff]

oh look, sexualized anime girls. how original

:p

[astral jay]

Trial and error mostly

[high skiff]

I agree, although i do still feel like these images look like just basic 1.5 models, personally. They have that sort of fake/plastic look to them to me

but I will say, it is very clear that you were able to make a LoRA that changes the way SDXL looks, which is cool on its own

[boreal bough]

having done enough photography. I can say this is what photos taken at conventions genuinely look like 🤣

[upbeat summit]

this explains a lot

[astral jay]

Even this?

[hard fractal]

https://tenor.com/view/ohhh-alexander-cora-the-family-business-woah-oh-really-gif-26262990

[high skiff]

I have done photography for years as well, and I truly believe that they look like flat out 1.5 gens, personally

[high skiff]

especially the left

[upbeat summit]

glad that my 2:3 and 4:3 images weren't that effect by it:d

[astral jay]

Zoom in if youd like

[high skiff]

I did

[hard fractal]

@astral jay super high quality stuff! I'm a fan.

the eyes especially:

@high skiff – please be kind

[high skiff]

I feel like its basically the same as this 1.5 model, personally

Something in teh tones and textures for me

[hard fractal]

this is early days of SDXL

we're still all trying to find the best settings

[candid walrus]

i think they are better than 1.5 but i wouldn't be able to say by how much, they are not photos, but they are better than 1.5, the eyes being round is a testament for sure if they are coming out that way consistently

[astral jay]

It's probably due to the dataset using makeup, I do agree, there could be more texture

[high skiff]

I am not being unkind. I am respectfully stating my opinions on what they did. Not everything from me is a dig, Joe

[crystal lily]

time to start the wait until someone makes a fork that supports AMD

[azure oxide]

that was a dig

[astral jay]

More blemishes, more imperfections

[high skiff]

I'm not saying it's not impressive, and I'm not saying that it's poorly done, I'm just saying that I feel it mimics the style of 1.5

[boreal bough]

real photo XD

[vale eagle]

with lora

[upbeat summit]

you are doing a latent hires fix like upscale? because the eyes are really good and I've been making portraits the last few weeks with SDXL and it's a hit or miss

[astral jay]

I appreciate it, without criticism there's no way to get closer to perfection... one day

[astral jay]

Yep, upscale using NMKD4x

[candid walrus]

a real photo has the ablity to like be out of focus and in focus on a very minute level to me, something i'm not sure that sd knows how to do yet

[barren bluff]

That's what I do with SD mostly lol

[astral jay]

But interestingly, with this last Lora I've been getting surprisingly good results with just txt2img

[candid walrus]

with faces at least

[astral jay]

Not nearly as realistic but decent

[upbeat summit]

hmm... so you figured out a real nice workflow. without me knowing I would definitely say there are a couple of other ingredients beside just base SDXL. congrats :D. really cool

[astral jay]

It's all thanks to @hard fractal

[high skiff]

That's what a lot of people do haha

[west breach]

My sparkling lora was trained on outputs from a 1.5 model, so it does get 'that face' if you don't be more specific with the prompt. You know the face from a lot of the 1.5 fine tunes like azovya etc

[astral jay]

And his help, couldn't have done it without him

[boreal bough]

good boi snek ❤️

[high skiff]

Yeah, they all look the same haha

[astral jay]

The key was real regularization images

[high skiff]

It's for sure a type

[spark bear]

surprised it can do snakes well, I dont want to know how mangled that would come out on base 1.5

[vast narwhal]

The condition should match the image size or should be larger but with the same ratio?

[upbeat summit]

the Sparkling lora is really nice 😉

[west breach]

but SDXL has it's own type too, high cheek bones and square jaw look

[high skiff]

Seems like that info is being researched still

[hard fractal]

that info is being researched still

[candid walrus]

that's the wording i was looking for ^

[boreal bough]

forgot to mention this earlier. it obviously has effects, regardless of setting - but this is the originally envisioned setting

base only

[astral jay]

Also very decent flexibility

[upbeat summit]

I only know regularization images from training. do you mean an init image for you generation? or am I misunderstanding?

[vast narwhal]

🤣

[astral jay]

For training the Lora, I use real images

Instead of generated

All of it being UHD

[upbeat summit]

my metadata .txt file "component" build with just nodes

[astral jay]

And the dataset was carefully and painstakingly prepared using Photoshop Beta with the help of creative fill

[upbeat summit]

so you are using a lora

[astral jay]

Yup

[upbeat summit]

sorry, I didn't catch it

this explains a lot 😄

I thought you were just using the base model without anything else

fine-tunings etc

and my prompting skills are just bad

[astral jay]

Imperfect skin is possible, just difficult

[hard fractal]

hm hm

[upbeat summit]

the quality your lora adds is really good - eyes, overall look. SDXL is amazing.

[astral jay]

SDXL is a miracle

It's the beginning of a new era of technology

[hard fractal]

❤️

[spark bear]

sdxl does macro photography really well

[urban fjord]

I hope 3D scenes won't be too far away.

[high skiff]

Same here

[boreal bough]

sketch, friendly snake swimming in a lake of lotus flowers

[astral jay]

And all Open Source! What more can we ask for?

lol

[hard fractal]

Super glad the community has adopted SDXL

[candid walrus]

so much potential

[boreal bough]

tiny derp ❤️

[candid walrus]

it's hard to look away

[high skiff]

I am excited to see the increased quality we will have when people who know better how to train it release models

[astral jay]

I have this idea, of training SDXL on stereoscopic images... when Dreambooth supports sdxl

[hard fractal]

I love that idea!

feels doable.

[urban fjord]

I thought there was dreambooth support for SDXL..

[spark bear]

love his expression of anger and confusion

[high skiff]

Who knows, maybe when my 3090 gets here, I'll make some big LoRA's

[astral jay]

In theory, it would be more like emulating human eyes, not a camera

[hard fractal]

❤️

[high skiff]

I'm pretty sure you could achieve that with just a single stereo processor post-processing script

[astral jay]

Thus, better realism

[high skiff]

All it needs is a depth, segmentation and then to shift over the image very slightly

[hard fractal]

Gonna go watch Harry Potter with my son – but thank you all!

[urban fjord]

Are you sure you don't mean Controlnet?

[high skiff]

I do parallel photography all the time, you really only need about two to five inches of shift to get the effect

[hard fractal]

Excited to see your tests with resolution conditioning!

[astral jay]

Well, not for perfection. You need real stereoscopic images

[high skiff]

A control nut could likely do it, but I think it would be way faster and more efficient to just generate one image and then run a simple Python script to segment it and offset it.

It should be a lot faster and a lot more reliable

[urban fjord]

Can't you use steroscopic ones generated from a 3D-engine?

[astral jay]

Sure, but in my head, if you want real train real

[high skiff]

I have lots of stereographic pairs I've made on my DSLR, I could potentially contribute images to anybody who wants to try and find tune that

*finetune

[astral jay]

My plan is to get my hands on 3d movies and extract like mad

[urban fjord]

SDXL should be able to generalize so you can supplement real data with synthetic data.

[high skiff]

I still think a simple depth processor with a slight airlines effect would achieve just as good results but faster, but I'm interested to see what the community comes up with

[candid walrus]

^ yeah it's just a math calc to sdxl

[high skiff]

*parallax

Voice typing is failing me

[candid walrus]

it really have to be stable for 'video' frames

[astral jay]

How about object occlusion?

[high skiff]

I don't think that will be a major issue, there are already scripts that can do that

[astral jay]

No, I mean, real 3d can hide certain details in one frame to the next one

So any script wont be able to just add detail that doesn't exist

[high skiff]

Typically, they will separate the subject and foreground, then crop in a little bit to where there is additional wiggle room in the background from the miss alignment of the scaled down background, so even if they shift over the background, and still displaying information generated image

[astral jay]

Sounds like a good way of training a green-screen like effect into the model

But hey, who knows

maybe I'm wrong

Anywho, cheers and g'night.

[urban fjord]

A steroscopic effect is just a gimmick and far from being able to create actual 3D scenes.

[high skiff]

Yeah, I just think that people were looking to achieve some level of stereoscopic effect

[astral jay]

No

[urban fjord]

I thought that were possible a long time ago.

[high skiff]

I take stereoscopic images all the time, and they are very fun to look at if you're capable of aligning your eyes parallel

[urban fjord]

I'm not able to allign my eyes like that at all.

[hard fractal]

oh

but

[astral jay]

I don't wanna create the effect, if you were to train stereo into the model, correctly, you would be able to achieve unparalleled realism.

[hard fractal]

do these make sense now?

haha

[urban fjord]

Not actually.

[hard fractal]

the condition isn't 4096x4096

[astral jay]

First two are incredible

[hard fractal]

literally just this:

[astral jay]

Widescreen? Full on training movies now

[hard fractal]

that's the base model

[astral jay]

😮

[heady vale]

nice

[astral jay]

Only A list celebs get to be in both hollywood and sdxl 😎

[cedar tusk]

Anyone know the syntax for Comfy nodes such that you have a string input but without a text box by default?

[cedar tusk]

Is that a batch of 4 or one image that looks like 4?

[urban fjord]

Sorry, I just don't get what you're trying to show off. I think widescreen has been much of a problem for me.

[upbeat summit]

some images made with Sketch - an upcoming LORA by @boreal bough
sketch, the cookie monster buying cookies in the supermarket

[urban fjord]

But yeah SDXL can create some nice wide images.
This was with img2img to give it a bit of a push, but managed most of the composition itself. Though probably need a upscaler step to get the quality up to par.

1904 x 392 isn't exactly the supported resulution of SDXL.

[high skiff]

Sometime soon, I should be able to share a collection of high quality stereographic images from my trip to Hawaii

They were taking on a DSLR with an f 1.8 lens, so they have extremely strong depth fall off, so I'm not sure if that will be ideal for the types of stereographic image training the community wants to do

[urban fjord]

It is at least a proof of concept.

[chrome flicker]

now sdxl can use input image as reference like midjourney? Midjourney can use input image and generate many creativate img related to input like style, constructure, object.

[autumn forum]

i uploaded my lora. might be ass but ya know whatever lmao https://civitai.com/models/120845

[urban fjord]

I guess I should try it.

[chrome flicker]

We need a way to use input image in sdxl like the same way in midjourney, I think sdxl'strong performance can do the same thing but don't know how to achieve that.

[boreal bough]

time to upvote and get you that 4090! ❤️

[autumn forum]

4090 would be fantastic. a100 would be even better tho lmao. but nobody aint got money for that

[high skiff]

I would love to try and create some loras for the competition when my 3090 gets here, however I do feel as though going from a 3090 to a 4090 is a little bit of a strange move haha

[visual glade]

for inference the 4090 is a way better choice unless you don't care at all about price

[urban fjord]

I don't think this was the right prompt for your LoRA.

[autumn forum]

i mean it was trained on specifically people lol.

[tribal jackal]

Retrained my LoRA! Size reduced from 1.7GB to just 19MB!

[boreal bough]

😂

[autumn forum]

snek

[urban fjord]

You need more creepy puppet scenes in your training data 😄

[vast narwhal]

Anyone that uses this node knows what is that? wasn't there yesterday. Isn't in the github page.

[tribal jackal]

Also, I don't think Discord strips metadata... y'all should drag one of my pics into ComfyUI if you want to sample LoRA's with 6 prompts at once (no upscale for expediency)

[sudden cliff]

can anyone tell me why this particular image is so ultrascuffed?

(pipeline included)

[autumn forum]

those arnt creepy enough?? they look pretty terrifying to me.

maybe i should add some stills from the muppets show. hmm

[upbeat summit]

some portrait images made with Sketch - an upcoming LORA by @boreal bough

[heady vale]

looks great

[urban fjord]

The right one is without the LoRA.
Do you have any with and without LoRA grid for your model?

[upbeat summit]

yeah - Caith did a great job with the Lora 🙂

[heady vale]

I havent tried Loras in XL yet

[autumn forum]

i do not, im not sure how to make a grid in comfui yet. or at least i dont have one setup

[boreal bough]

without/with lora
(went all out on prompting)

[autumn forum]

ive also noticed just attaching the lora to the pipeline changes the output so i have to do 2 seperate ones.

[urban fjord]

Having comparison images are important when creating a LoRA so you know that it is the LoRA creating the results and not the base itself, I've tricked myself on that before.

[autumn forum]

everytime i generate a image i do a base gen and a lora gen to see the difference.

[ionic gulch]

i guess it runs a few refiner steps in front of the base steps, as some workflows do manually with separate nodes.

[autumn forum]

but yes i agree, if i could do grids like the ones you posted i would!.

[urban fjord]

I'm trying more prompts but I do feel like they are blurrier than base output.

I should look for a dataset of weird puppets. There is two that comes to mind...

[vast narwhal]

It's possible

[boreal bough]

@autumn forum did you include any anatomy words in your dataset?
feels like anatomy got touched, which is causing the typical issues

[autumn forum]

if you get me the names of some movies ill add em!. i found a better website for movie stills so ill definitely be adding higher quality pics in the future.

[boreal bough]

without/with
(Movie Still:1.2), in sweet November a girl falls in love under the the setting sun as her life is also about to set

[autumn forum]

no. should i ? if so how so? like head, shoulder, mid shot, waist?

[urban fjord]

Not sure if I can find high res version, sorry.

[boreal bough]

no the opposite, those words are considered taboo unless you're specifically aiming to train them

"face", "body", "waist", "legs" <- words like those

positions/angles/poses are all fine

[autumn forum]

ohhh okay

[urban fjord]

movie still of Tom Cruise as Captain America
(XasY is my own, blank is no LoRA)
Yeah this is sadly not working for me.

[autumn forum]

thats okay! thanks for the feedback.

[autumn forum]

i clearly need to work on my lora skills

[boreal bough]

the effect is strong and good though
aesthetic of 90s movies peeks through on about half my prompts

maybe go all in on that?

[upbeat summit]

your dog has so many adventures!

[autumn forum]

thanks! and maybe i should stick with a particular style of movie instead of a broad range. maybe that would help with the consistancy.

[pliant vigil]

Im... kinda lots on where to start as far as downloading, like I have the huggingface website, but I can't find where to download from there .....

[upbeat summit]

or just a couple of concepts

[pliant vigil]

can someone point me in the right direction?

[boreal bough]

broad also definitely works - you just need to watch that no particular style accidentally trumps the rest

[autumn forum]

without and with

[tribal jackal]

He really didn't want to leave the pixel art land

[boreal bough]

but you can do one style first with 200~400 images. then scale and add more styles, each with 200 more images

[lean granite]

What are you looking for exactly?

[pliant vigil]

I have stable diffusion, trying to get xl, i have xl's huggingface website but Im not sure where to go from there?
Maybe a good video about instalation of stable diffusion XL?

[autumn forum]

gotcha, okay i will do that. thanks for the help!

[urban fjord]

You are training at 1024x1024 with training images with the right resulution right?

[autumn forum]

i trained them on 1344x768 all images were that

[urban fjord]

Alright, so I should really test that resolution.

[sharp robin]

under files and versions

[lean granite]

You gotta pick between getting a1111 which is a "simple" ui or comfy ui Which is node based

You'll need either to run the models

[soft bone]

i'd love to hear more. can you send your training config perhaps?

[pliant vigil]

i have a1111

oh its the lil ⬇️ 🙃 wheee thank~

[boreal bough]

changed from hit & miss, to 100% success rate with that resolution

did add blur XD

but it works!

[lean granite]

I'm not sure if xl runs well on a1111, unless they fixed it 🤔

[autumn forum]

yyeah the blur is definitely an issue lmao

[urban fjord]

What prompts are you running Caith...

[boreal bough]

yep. the resolution got baked in

umm, I mean aspect ratio

essentially it only does that aspect ratio well now

(which is fine for a movie still lora)

[boreal bough]

Movie Still, in sweet November a girl falls in love under the the setting sun as her life is also about to set
clip_l: movie still, bittersweet, strong colors
Negative I use pseudos

[urban fjord]

I've no idea what he uses.

[lean granite]

Do we know if controlnet is out for sdxl yet in comfy?

[autumn forum]

to be fair this is my second lora ever made sooo im happy with. "it works" 😂 i will and can always improve in the future with more testing and time.

this is actually a really pretty picture

[urban fjord]

Keeping it more focused always help.

[upbeat summit]

some portrait images made with Sketch - an upcoming LORA by @boreal bough

[boreal bough]

kay. got his permission

polymorphic, washed-out low-contrast (deep fried) watermark, cropped, out-of-frame, low quality, low res, poorly drawn, (mutated hands and fingers:1.4), mutation, mutated, ugly, disgusting, blurry, amputation

is by far the best negative prompt that currently exists
credit goes to pseudoterminalx whos not on this server XD

[autumn forum]

yes, i use this religiously

[upbeat summit]

yeah why did pseudoterminalx leave. he was here daily 2 weeks ago

[azure oxide]

r i p pseudo 😭

assisting us even when hes not with us

the mans we dont deserve

[boreal bough]

add that negative to the sketches, and see what happens 😄

[high skiff]

For various different reasons. He is part of my research group tho

[lilac raven]

I think it's better than before. Crossed arms + pixel art was a bit of a challenge

[high skiff]

I talk with him daily, he is very smart, and knows a lot more about SD/SDXL than most

[candid walrus]

that's awesome

[chrome flicker]

pixel art anime looks good, but anime image is really bad in sdxl

[upbeat summit]

thanks for sharing it! I'll check it out. I was building negative prompt stacks intensely for SD 2. SDXL works of course differently and needs way less negative prompting or even none at all.

I did some studies into what each negative prompt token does in SD2 and it revealed a couple of interesting behaviors to push the fidelity of images.

[autumn forum]

i wonder what would happen if i trained a lora on the entirety of a single move like pulp fiction. lol

[boreal bough]

not sure what I was expecting, but hey! this is what happens when you use my sketch lora with -1 strength 🤣
with/without lora

drawing in sweet November a girl falls in love under the the setting sun as her life is also about to set

movie still, bittersweet, strong colors

[urban fjord]

Testing is the only way to figure stuff out. I guess every person will be replaced with Samuel L. Jackson

[hardy cipher]

these are quality images, guys. very clean

[tribal jackal]

ooh I really liike this. A little less contrast, blur, and red curve and it'd be stellar

[hardy cipher]

does anyone actually know what pos_g, pos_l, and pos_r actually mean or what impact they each have? the searge prompt encoder uses them and I really don't know what to put where. even moreso with neg_g, l, r

[vital ermine]

I know and that is base XL so not sure what happened?

[boreal bough]

for science, I can tell you that my LoRA trained on 'the entirety' of the turbo killer music video was a huge failure 🤣
turns out dataset size does matter

[tribal jackal]

wow weird

[autumn forum]

okay lmao thanks for the knowledge, i probably shouldnt do that lmao. but i wanna

[hardy cipher]

I mean, how would I even split the negative prompt into 3 parts? lol. just seems like a lot. it's not like I have 3 distinct categories of negative words

[vital ermine]

I wonder if I could tone down the red with a neg?

[hardy cipher]

yes

you can

[vital ermine]

which word(s)?

[hardy cipher]

I'll put things like "excessively red" in the negative

then adjust weight accordingly

[boreal bough]

movie will definitely work XD was just joking that a 3 minute music video isn't enough

[vital ermine]

let me try that

[boreal bough]

but I wished it would

[alpine pine]

The filmic lord from earlier is based on a single episode of Andor

It gives a decent result based on the small dataset. Tho andor has a very muted and real color range.

[sharp robin]

g is a new clip, its more stronger, l is like old clip, SDXL uses both, g for foreground and l for background -ish

[hardy cipher]

ahh, thank you. and then r is refiner I guess?

[autumn forum]

hmm interesting.

[vital ermine]

dang :/

[alpine pine]

https://civitai.com/models/120763?modelVersionId=131346

Only Issue I have found is that you end up getting the motion blur and compression artifacts in the generations. Which I personally love, but many dont

[autumn forum]

so quick question. for captioning, is it best to run vit-h and leaf it with all of its words, or manually caption would be wayy better?

[hardy cipher]

something weird going on here

[tribal jackal]

I still can't get over reducing my lora from 1.7GB to 19MB

[urban fjord]

3 minute of video sounds like a ton of training images.

[tribal jackal]

nahhhh 144fps 5120x1440 that shit

[urban fjord]

Davonair: Yeah using a high dim is not necessary for SDXL.

[autumn forum]

oh yeah i saw that earlier. i thought damn someone beat me to it. but im kinda going for a more early 80's realistic movie approach. ill have to try it out

[vital ermine]

Trying my 2,1 neg minus my embeddings

[tribal jackal]

Not sure what a dim even is yet but I'll read up. Further improvements always welcome.

[hardy cipher]

I still haven't figured out what models and/or loras I like. they take sooo long to load on my computer, lol. so when I start using one I feel committed and haven't done a lot of direct comparisons

[boreal bough]

might be better to use the lazy method.
no captions at all, only 1 or 2 words on the folder

it won't train the concepts of the movie, but it will capture the feeling (and the colors/contrast)

not recommended for many types of lora, but in this case, it should work well

[autumn forum]

hmmm ill try both. just to see. it only takes an hour to train. sooo ill give it a shot

[vital ermine]

Something has a mem leak

[boreal bough]

for no captions, expect it to train for a whole lot longer - as overfitting doesn't happen till way way later

I usually go x5

[alpine pine]

How do you know when it over fits?

[urban fjord]

You test it.

[vital ermine]

I killed auto1111

[tribal jackal]

oh "dim" short for "dimension rank"

Yeah I trained a model at rank 8, but the quality suffered so much compared to the 1.7GB model I just scrapped it and went a different route.

[boreal bough]

way way overfit = "cat" turns into woman that resembles a woman in your dataset XD
mild overfit = it generally only does the same poses/background/compositions as those found in your dataset

bigger datasets cause this to happen much much slower

[vital ermine]

pagefile is 20gb

[boreal bough]

(there are rare cases where you want overfitting to occur, so for those you use 50 images or less.)
or even 1~4 images if you're a madman like @urban fjord

[urban fjord]

I have success with ridiculous datasets.

[vital ermine]

I can't train XL due to its bulk.

[boreal bough]

'success' is putting it mildly when you improve the whole damn sdxl model with 1 image 🤣

everyone should try this model from Diodotos
https://civitai.com/models/119388?modelVersionId=129692

it's a miracly this lora is even possible

[urban fjord]

You can do LoRA with 8 GB VRAM I think. But yeah if you have less than that you need to train in the cloud.

[vital ermine]

Colab I can barely do BS2 with 2.1

XL 1024 would be forget it without checkpointing and that doubles the time

[upbeat summit]

@boreal bough's Sketch Lora + a drop of Beksiński

[urban fjord]

You can get decent results in less than 100 steps so time isn't that big of an issue.

[gritty breach]

anyone else feel like the previews of the generated image always look better than the final product?

[urban fjord]

I'm not using previews as it slows down generations.

[gritty breach]

ah. i just like to see its progress so ive kept it on.

[heady vale]

inspired by the theme

[boreal bough]

wth sdxl XD I appreciate your improvement. but this is not 'the thing'

[vital ermine]

It would appear the latest auto has a very bad pc mem leak

[trim orbit]

https://tenor.com/view/yes-rick-morty-snap-gif-5747281

[vale eagle]

wow, @urban fjord you should also make blue background and green background.

[vital ermine]

This is weird but I am getting a lot of stuff like this from base, or no heads at all

[urban fjord]

What's your prompt + parameters?

[vital ermine]

parameters does't matter as I have changed them all around. I wonder if it is the 16:9 aspect ratio?

[urban fjord]

I can't create a LoRA for every colour sadly. It would take 127 years on a decent GPU to train a LoRA for all 16 million RBG colours.

[vale eagle]

I mean blue and green used to be background removing tech for further processing in video or image. That the thing when I saw the white background

[urban fjord]

With a clean background it should be a lot easier to do background removal and then you can just replace your white with whatever you want.

[upbeat summit]

interesting lora, @urban fjord 🙂 nice

[urban fjord]

I might try 00FF00 or 0000FF to see if that works.

Alright, I'll try now.

[vital ermine]

should work

OH, I see where the pc mem leak is coming from when you switch the models back and forth

[queen basalt]

Sorry to bother you guys with this but I get a runtime error when I try to load the base model on automatic1111. It's on an rtx 4090, 64 GB of ram, I7-13700K. Refiner model works (still figuring it out) but base model will act like it's loading, fail, and then revert to whatever model I was using last.

I'm too dumb to intuitively understand what's going wrong.

[vital ermine]

redownload the base model it is corrupt

make sure to get the one on hugging face

I dunno this XL is being very weird

[viscid pelican]

sd_xl_base_1.0_0.9vae.safetensors IS the stable diffusion model, it goes in the stable-diffusion folder, not the VAE folder

[vital ermine]

everything seems to be overcooked

[viscid pelican]

The fixed vae is baked in.

[urban fjord]

Unless you share your prompt + parameters it is hard to give you any feedback.

Are you using the refiner btw?

[vital ermine]

that is what I am telling you it could be a dog running down a street or a cat getting some tang on a fence top it doesn't matter what I prompt

everything has been opver cooked

*over

[heady vale]

close auto and use comfy

[queen basalt]

There's no prompt, I can't even load the model. I redownloaded it (from hugging face) and even did a fresh install of auto1111 (no extensions)

Yeah comfyui may be the only answer for us lost ones

[urban fjord]

When you refuse to give your prompt + parameters we cannot help you at all.

[vital ermine]

nope, can't do it I tried but it is too little for what I need it for AND it doesn't work with IA3 or DyLoRA

[lilac raven]

Did you try generating a 1024x1024 picture to check if the same problem happens? Smaller pictures than 1024x1024 had a lot of problems to me

[vital ermine]

I can't I run out of video ram

[queen basalt]

Well did you try xformers or medvram arguments?

[lilac raven]

How much do you have? I'm being able to use SDXL on 4 GB VRAM

[vital ermine]

I have 6gb of video ram and use --medvram and xformers since 1.4 was released.

[urban fjord]

There is no need to run higher than batch size 1 if that's what you're doing.

[lilac raven]

You could try using --lowvram

[urban fjord]

Try the fp16 vae.

[vital ermine]

I see the overcooked in that one too

[vital ermine]

do you have a link as all I found was fp16 1.0 and it refused to load it was for [diffusers]

[heady vale]

yeah some subjects are like that I found with XL, using the same cfg across all images. I will try lowering it

[heady vale]

that was 6 cfg

[vital ermine]

ahhhh, I thought so. Thank you

[urban fjord]

Is that base or a LoRa to create giant dogs?

[tribal jackal]

Just a LoRA of my dog that happens to sometimes have happy little accidents with scaling because I'm new lol

[lilac raven]

The white background LoRA is so useful. I suffered so much to remove backgrounds in the past (past is like yesterday)

And now I'm not sure I can go back to 1.5

[urban fjord]

I'm happy some people find it useful.

[boreal bough]

post an image (as a review) to civitai if you wanna support him! XD gets him entered into the competition if enough people post

[urban fjord]

It takes longer to generate test images for my blue LoRA model than I used to train it...

[hardy cipher]

lol

[queen basalt]

So I reread what you said more carefully. You mentioned the "vae" model going in stable-diffusion and not the vae folder. But I'm still having the same problem even after doing that.

The model will not load. It gives me that error about missing keys in state_dict

[hoary saddle]

SDXL with image2image is a gamechanger

[hoary saddle]

excited to see the loras and finetunes

[lilac raven]

I was generating something on the laptop but I forgot to actually copy the LoRA. 5 minutes wasted

[urban fjord]

Here is the results, not as clean as the white LoRA sadly.

[upbeat summit]

only 16,699,998 colors to go

[hardy cipher]

what is the deal with comfyui on colab? I copy paste that ip address/password that I'm told to enter, and I always get this

"Error: endpoint IP is not correct. Please try again or contact whoever gave you this link for the correct public IP."

[echo pine]

@high skiff some stuff I did for testing the conditioning sizing

just drag it into comfy

[hard fractal]

That's looking great!

[spark bear]

thats some extreme background blur even for a connesour of shallow dof that is sdxl

[hardy cipher]

I mean, it doesn't have to be blurry

[gloomy barn]

Controlnet for SDXL is under development? https://www.reddit.com/r/StableDiffusion/comments/158n1td/controlnet_sdxl_10/

[hardy cipher]

nice

[vocal rapids]

pencil art lora is good , https://civitai.com/models/119771

[hardy cipher]

I guess I just can't use comfy on colab, lol. this is silly. I don't understand why it just gives me a bad password every time

[hard fractal]

Working on it, yeah

[vital ermine]

Any comfyui people know what this is all about?

[peak dove]

What is "resolution conditioning"?

[hard fractal]

Is that for XL?

[urban fjord]

Yes, but it's not working well enough atm. With and without LoRA barely makes a difference.

[peak dove]

Yeah, u can safely ignore it

[vital ermine]

ty

[visual glade]

something that's not in code I can fix

[visual glade]

it's a harmless warning

[vital ermine]

It produced this so something is wrong though

[urban fjord]

SDXL can do decent pencil drawing out of the box.

[hardy cipher]

comfy! could you offer any insight on why I keep getting this when I try to load the ui in colab?

"Error: endpoint IP is not correct. Please try again or contact whoever gave you this link for the correct public IP."

[peak dove]

Looks like too high cfg?

[hard fractal]

Looks like a sampler issue

[hardy cipher]

it's a persistent thing and I can't figure out what the issue is

[hard fractal]

What's your sampler?

I'm guessing SDE?

[hard fractal]

ye

[hardy cipher]

mine looked like that when I accidentally used a 1.5 lora with xl

[hard fractal]

try ddim as a test

[urban fjord]

Wait, can you actually use 1.5 LoRA with XL?

[visual glade]

_2m needs exponential or karras to be happy

[thin nova]

Though, loras can help make prompting easier, like a TI

[vital ermine]

I was using the same one that auto has only his doesn't say gpu

[hardy cipher]

in comfy. it's not something that really should be done though, lol

[vital ermine]

oh, good catch

Is there any single way to not save an image until I manually save it? I had to do that in auto last year.

[visual glade]

I ran it just to make sure and it worked

34.87.2.137

that's what the "password" should look like

[echo pine]

yeah, still have to do some testing for how many steps I put the first pass of the base model with normal res conditioning and second pass with higher res conditioning and if the conditioning of the refiner is that important

[vital ermine]

using karras so I lost the man and got that

[vocal rapids]

yup SDXL reli unleash the power of that....

[vital ermine]

Using exponential instead of karras

[hardy cipher]

hmm. thanks. I'll restart it and see what happens. what's weird is it worked when I tried it on my phone. only problem is I really couldn't do anything, lol

[elfin flare]

Man, LORAS make SDXL nuts.

[hardy cipher]

just want to run these sdxl models quicker. they take their time on my computer

[urban fjord]

Wait what, how does ComfyUI manage to use SDXL LoRA on 1.5?

[elfin flare]

Pixel art lora moment

[visual glade]

it's only going to apply the lora weights where the shapes match

[vital ermine]

how do I do the refiner in comfy in one go?

[vocal rapids]

any sample of using img2img in ComfyUI with lora? I wanna change the style of input image using lora..

[hardy cipher]

I don't know how it actually works. but there's a lora loader and a checkpoint loader. and you can load 1.5 or xl or whatever in them

[visual glade]

https://comfyanonymous.github.io/ComfyUI_examples/sdxl/

[hardy cipher]

alright, here's the real question, how to get 90 degree angles and straight lines in comfy?

I get so OCD about the paths going under nodes, lol

[vital ermine]

Ugh, this seems very complicated. What happens if I have 6-10 loras+embeddings that I use and swap in and out etc... per prompt?

[hardy cipher]

use the efficiency loader?

or whatever it's called

[vital ermine]

No idea what that is

[hardy cipher]

well, it's on the internet. I can show you

[vital ermine]

please

[hardy cipher]

you can also find it in the manager

[astral jay]

Obvious inspiration for this widescreen

[vital ermine]

wait, there is a manager?

[hardy cipher]

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

[astral jay]

Can't for the life of me get a filter so I cheated and color corrected in PS

[heady vale]

https://github.com/ltdrdata/ComfyUI-Manager

[hardy cipher]

should be on the right side of your window

below save and all that

[vital ermine]

I book marked both of those and will grab them for sure as they look tight

Thank you both

[astral jay]

No cheating on that one

[vital ermine]

there are two things that are driving me away from comfy. The lack of a scroll that I can tell to go the opposite way that it does now (customization is a very good thing) and the lack of prompt editing.

[hardy cipher]

you can change the scroll direction

[vital ermine]

let me guess an extension?

[hardy cipher]

and there are all sorts of different nopdes for prompt editting

man, if you don't like it don't use it

I'm just trying to help you

[lilac raven]

The last time I use Comfy on mobile it was unusable, not sure if it has improved since then

[visual glade]

it works enough to generate images on mobile but it's not a good mobile experience

[vital ermine]

What? No, I was saying it appears everything is an extension when such things like this should have been in a settings option.

[hardy cipher]

it had a long delay for me, and it'd do a bunch of stuff all at once

ahh, my bad

[hard fractal]

Wouldn't that bloat it?

I feel it's a bit similar to a Linux distro

[hardy cipher]

well tbh I don't really know what sort of prompt editting options you'd like. I guess I'm out of the loop on what the options are. I just type prompts in

[vital ermine]

Settings should never bloat anything tbh.

[hard fractal]

You can get the barebones, but you can also get the managers and the packages

[vital ermine]

a simple config file so no bloat

[hard fractal]

General, give SwarmUI a try maybe

[hardy cipher]

be mindful of what you install. I really goofed my install buy willy nilly clicking all the nodepacks

[visual glade]

what's wrong with the scroll direction?

[vital ermine]

I think I heard of that before.

down is up and up is down

[hardy cipher]

comfy, it's backwards from what I'm accustomed to, but only takes a few minutes to acclimate. it's like inverting up and down in a first person shooter

[vital ermine]

or in a flight sim

[hard fractal]

My trackball, it's backwards from what I'd expect

[astral jay]

Gotta say, I love widescreen

[heady vale]

I scroll up to zoom in. I prefer it that way

[visual glade]

right now it's like on google maps

[vital ermine]

I am talking about on menu selecting like models

[hard fractal]

usually clockwise zoom in, anti-clockwise zoom out for me

[brazen patrol]

It is not actually a scroll IMHO. THat is the menu going up or down

[hard fractal]

Oo, and most node programs, it's middle click to pan around

[alpine pine]

I believe Unreal’s node editor also uses the same scroll direction. Or atleast I’ve switching between the two without thinking…

[vital ermine]

Middle mouse button scroll wheel

[alpine pine]

(With rmb to pan)

[hard fractal]

nice that there's an extension to reverse zoom

[visual glade]

yeah I think litegraph is inspired from unreal

[vital ermine]

is there one for the menus too?

[hard fractal]

You can right click, open properties

keeps a floating window up

[hard fractal]

Ah, thought that had the drop-down there

[astral jay]

Is the comfy name at all related to LilyPichu?

[visual glade]

no I don't know what that is

[astral jay]

Damn

[hard fractal]

something to consider adding. would be nice to make a persistent prompt text box always on the bottom

[ionic dragon]

Do these 3rd party controlnet work with xl?
https://huggingface.co/TencentARC

[vital ermine]

I am talking about when you click the model to select a new model it is backwards scrolling and no scroll bar either

[upbeat summit]

he's going to fix all your UI problems