✨｜sdxl

[hard fractal]

Most are h264?

[upbeat summit]

4k streaming gives you 25mbps while blu-ray gives you 40mbps

[hard fractal]

I'm sure there's a 1x RealESRGAN that cleans up those artifacts

[upbeat summit]

so way less compression

[boreal bough]

depends. some streaming services are highly optimized - where they have ultra clean keyframes - so you'd just have to grab those

[visual glade]

blurays, etc... are encoded in yuv420 so only the luma channel is really the full resolution, the chroma channels are half the resolution

[boreal bough]

more relevant for animes/cartoons though. real life movies - no streaming gonna do it

[hard fractal]

https://github.com/rlaphoenix/VSGAN

[upbeat summit]

it's still highly compressed and the vectors are optimized. the blu-ray will be in every case much sharper. not saying you can't use streaming as a source (look at arcane fine-tunings) but blu-ray is the cleanest source you can publicly get.

[boreal bough]

absolutely agree ^^

[vital ermine]

any workflow out there so I can test G/I?

[heady vale]

grab Sytan's one

[vital ermine]

I heard that sucker was complex

[autumn forum]

nah

[vital ermine]

I just need something simple for now

[autumn forum]

ill make you one lmao

[vital ermine]

oh, nice, thank you

Once I see it I can add on and stuff

[heady vale]

with the XL offset lora, do I adjust the model or clip strength?

[pallid sail]

beautifully horrific

[autumn forum]

here, hope this is simple enough! edit oops hold on

[autumn forum]

okay @vital ermine here you go

[vital ermine]

Thank you

Dang, this will take a lot of getting used to

[nimble heart]

also @vital ermine if you wanna know where ROCm for the XTX is these places are good to check. Linux based but shows what consumer rocm can and can't do yet:
https://github.com/evshiron/rocm_lab/issues/2
https://are-we-gfx1100-yet.github.io/

[vital ermine]

Very interested in that but will rocm ever do it all?

[nimble heart]

They're kind of perpetually playing catchup to nvidia since there's a chicken egg situation with 3rd parties making rocm compatible libraries, but in general it seems like the big libraries will at least work on RDNA 3 "eventually"

[vital ermine]

Nvidia is clamping production to keep supplies low and even stopping A100 so, yeah, they are going where I don't wish to follow.

My fear is they will do rdna4 and 3 will just get skipped over

[vital ermine]

did this just do 2048x2048?

[autumn forum]

no, 1024 by 1024

[vital ermine]

shows 2048 though

[autumn forum]

the clips are at 2048x2048 not the empty latents

tbh im not sure why they are that but it seems to make the image better in most cases,

[vital ermine]

Ahhhh, well I can't do 1024x1024 in Auto for sdxl and with cpu mem leaks etc... I tossed it to the side but in Comfy this worked

[nimble heart]

they've put a lot of sauce into 3 so I don't think so personally. 2 and 1 definitely are to some degree

[autumn forum]

becasue comfy is awesome!

[hardy cipher]

ooh, I think I might have figured out how to do cfg scheduling

maybe

[visual glade]

AMD has a lot of issues with their software

[vital ermine]

They always have since they were ATI

Point is someone, somewhere, needs to step up and stop Jensen and any offspring it may have.

Intel, or AMD doesn't matter something

[midnight shuttle]

I hope they will compete and not "collude" (not in the legal sense).

[hardy cipher]

don't they always do that though?

[vital ermine]

Well, the issue is MI300 for AMD and H100 for Nvidia

Also Lisa su being Jensen's cousin doesn't help

[midnight shuttle]

Always compete? Not always. Sometimes they look at the market and realize they can make more profit by keeping prices high than by undercutting the competition.

[vital ermine]

Oh, and AMD does not want to be #1 ever

[midnight shuttle]

Better to sell 10 $30,000 cards than 100 $3,000 cards. Less work for same money.

[hardy cipher]

no, I meant collude

[vital ermine]

Exactly, Light

[midnight shuttle]

They don't collude in the legal sense. But they look at the market and consider whether keeping prices high is better than one company lowering prices and starting a price war.

AMD could lower prices 30% and win over the whole market. And what do they get? More complaints about bad drivers. More RMAs. More problems with limited supply. Or they can keep prices high and still make the same amount of money overall. The choice is easy.

The only real way to influence the market is to refuse to buy. And no gamer or AI enthusiast will go without a GPU even if the prices double.

[vital ermine]

double one more time and they will

[midnight shuttle]

Some will. Maybe 35% will. Do the math and it's a win for Nvidia.

[vital ermine]

3-4k per 5090 not many can do that

I know

[midnight shuttle]

Gamers need their games.

[vital ermine]

Meth addicts

[midnight shuttle]

Almost.

[azure oxide]

i paid $3k for my 3090 near release and would do the same again

luckily wont need to since itll take a loooong time for a 3090 to "need" to be replaced

[midnight shuttle]

For games that's true. For AI that's next year.

I've heard rumors that one ChatGPT message uses something like 400 GB VRAM.

[heady vale]

to be fair the 3090 and 4090 were nearly the same price on release

[shy kelp]

is that SDXL with refiner or just base?

[visual glade]

I'm sure AMD wants to be number 1 but their entire software stack needs a lot of improvements

it's too bad because their hardware is great

[hardy cipher]

I need to figure out how all these math functions and things work in comfy.

with the samplers, can input values vary over the course of the steps? or do they get their input and then do their things on their own?

[vocal rapids]

what is different for using FP16 vs BF16 while training lora?

[boreal bough]

bf16 has better math, and therefore better accuracy. but as for how much difference that accuracy makes - now that we're using pruned models to begin with - needs to be tested

lets put it this way - bf16 is tested and guaranteed to work well

[midnight shuttle]

FP = floating point? BF = ????

[visual glade]

bf16 has no precision but won't overflow like fp16

[midnight shuttle]

Brain floating point? That's a new one to me.

[visual glade]

bf16 is fp32 but with a 16 bits chopped off from the end

[autumn forum]

@boreal bough getting some good hq pics for the chappie lora, this was made in blender, not ai so sorry for posting here lol

[boreal bough]

noice!

[hardy cipher]

basically want to do something like this. or even just a very simplified version of this

and then curious if this is still a thing in xl https://www.crosslabs.org/blog/diffusion-with-offset-noise

I guess there's the noise offset lora though

[vital ermine]

one made yet for XL?

[hardy cipher]

noise offset lora?

[vital ermine]

yeah

[hardy cipher]

yeah,, I forget where I downloaded it. either huggingface or civitai

[vital ermine]

good to know.

[hardy cipher]

haven't got the hang of it yet. I just really miss the cfg scheduling. it really adds a lot of control to what you're doing. maybe it's a thing in comfy already, but I've asked about 10 times and no one has ever responded

[heady vale]

cant you sorta do that with different cfg for base and refiner?

[vital ermine]

I never liked that cfg scheduling and removed not long after

[hardy cipher]

I actually might have figured it out, but I'm not quite sure what I'm doing, lol

well simple linear scheduling, but I have to test it a bit more

[vital ermine]

I give up trying to figure out how this G/L workspace works.

[hardy cipher]

g for full sentence type stuff, L for descriptors

L is what was in 1.5. G is more conversational

[vital ermine]

Nah, I know that I meant all these nodes

[hardy cipher]

as I understand it

ahh

[heady vale]

G/L workflow does get a little complicated

[vital ermine]

I am totally lost

[hardy cipher]

what does G/L even mean?

[heady vale]

I just know in G you type a sentence and L you type tags like old gens

[hardy cipher]

I meant in reference to what you guys are saying

G is a newer clip model so smarter I guess

[heady vale]

yeah General is trying to build it out in a workflow to understand, I guess

[vital ermine]

yeah, I have one here and all the new node module boxes my head explodes

cave man..prompt goes here come out there me like no like repeat

[hardy cipher]

I just never considered splitting the negative up into different prompts. from what I understand it's not really necessary, but still

[crisp owl]

Oh nice!
Have you tried bows in 1.5 🤣

[vital ermine]

2.1 too

[crisp owl]

everytime I tried in 1.5 was an absolute hot mess lol

[vital ermine]

NGL, I miss inpainting

[crisp owl]

Currently pretty rudimentary in that regard in COmfy yeah

[nimble heart]

screw you and your cringe I got 4 bit working now. 30B on 7900 XTX all in mem

[vital ermine]

I use it for fixing stuff

btw, that was my prompt from a 2.1 training I was working on with G being new

[hardy cipher]

did you ever get it working properly?

[vital ermine]

Nah, it was never good for the stuff I did

[hardy cipher]

well I think I have something happening here.just wish I knew definitively how this math node works. but searching the comfy install folder for the node's name doesn't give me any results

so that's pretty cool

[vital ermine]

How does one stop a gen mid way?

not what I was envisioning

[nimble heart]

'cancel'

[vital ermine]

I didn't see that as an option. let me try again

[crisp owl]

view queue

wow lemme edit that again haha

[vital ermine]

thank you

[heady vale]

view history and scroll to the bottom will be your previous gen

should be at the top but...

[hardy cipher]

could anyone explain how this works exactly?

[midnight shuttle]

It looks like it subtracts two numbers and then gives you a choice of outputting it as a float or int.

[vital ermine]

can't fix the faces with a convoluted inpainting so I hope an extension comes

[vital ermine]

https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html

SDP is currently available for Navi 3x, but among the three underlying implementations of SDP: Flash Attention, Memory Efficient Attention, and the math impl, Navi 3x can only use the last one, which is just invoking PyTorch methods from C++ and does not offer substantial optimization.

The current development of Flash Attention for ROCm is focused on CDNA, and I don't know when RDNA will truly be able to utilize Flash Attention. All I can say is that there is potential.

[nimble heart]

yea it uses a fallback method which is actually worse than most other attention optimizers in auto/comfyui afaik

[vital ermine]

That sucks

[nimble heart]

according to that git thing I linked, there's work for flash attention on RDNA 3 already in the ROCm composable kernel it's just not part of the main release yet

[nimble heart]

so my guesstimate was this fall once the 7900 XTX gets "official" rocm support instead of just "it works because the CDNA cards work"

[nimble heart]

sounds like someon tried hacking it in and failed

We'll see once it lands proper in the CK

[vital ermine]

That someone - evshiron/rocm_lab. I am feeling discouraged again

I am not sure why it keeps making unrealistic images.

[nimble heart]

tbf they're not an expert. They got an AutoGPT-Q port working but they couldn't get bitsandbytes working

[hardy cipher]

man, I don't know. lthere's definitely a difference between say cfg 20, and then setting that math box to add and putting in 5, 15. not exactly sure what's happening though

they're close

[vital ermine]

that makes me feel a bit better as bitsandbytes has rocm support now

[nimble heart]

?

where

[vital ermine]

I had a link

[hardy cipher]

first one is 5, 15, second one is 20

[vital ermine]

from one of those you gave me

[nimble heart]

I have a patched 0.37.2 build but it NaNs on everything but RWKV

[midnight shuttle]

I was having trouble the other day trying to connect primitives to various width/height settings in nodes. Turns out the data type on the Empty Latent Image width/height isn't the same as the data type on the SDXL Clip nodes. That's why it wouldn't connect.

[vital ermine]

This one? https://github.com/agrocylo/bitsandbytes-rocm

[nimble heart]

the patched AutoGPT-Q works fine though. I get 17.3 it/s on 30B vicuna-wizard

[midnight shuttle]

That looks like normal non-determinism caused by xformers to me. Much more likely than 15 + 5 not equaling 20.

[nimble heart]

yea

[hardy cipher]

lol, I like when I've gotten NUMBER != number

[nimble heart]

all NaNs

[vital ermine]

damn

[nimble heart]

try it yourself maybe you have more luck

I use python 3.11 which is apparently buggy with some ML stuff

[vital ermine]

Oh, 3.11 is even for nvidia

[nimble heart]

I haven't noticed any bugs but I've just heard so

maybe the bugs are just windows related

[vital ermine]

same, but I moved on to 3.18 I believe

[hardy cipher]

you may be right. but I can replicate both results using the same methods. but I dont' actually know enough about xformers to say what's going on

[nimble heart]

?

you mean 3.10?

[vital ermine]

yeah

[hardy cipher]

what's interesting to me is the 20 seems a bit more burnt than the 5-15. but that could just be coincidence

[midnight shuttle]

That's more unusual then. Are you using the INT or FLOAT output? xformers introduces some level of non-determinism, meaning the same parameters won't produce the exact same image.

[hardy cipher]

float

init won't even connect to cfg

int

[vital ermine]

Python 3.10.9

[midnight shuttle]

Which I guess makes sense since CFG is a float value. Hmm. I think the node knows how to add so there is something else going on.

[hardy cipher]

yes. I just don't know enough to know what

[midnight shuttle]

Is this a built-in node or did you download it?

[hardy cipher]

I just turned cfg into an input and then looked at all the nodes it would accept connections with

and went through them until I found one that worked, lol

[nimble heart]

if you want some more hope, compared to his posted numbers, using that same autogptq repo newly compiled I get 32 it/s on 13B and 17.3 it/s on 30B

so it's getting there. eventually. in a few months

[hardy cipher]

it's just hard to find definitive information on that node since I can't seem to find it in my comfy install

did a search for the name but didn't give me any results. and I really don't have any idea what I"m doing, lol. I'm just persistent

[midnight shuttle]

So how did you originally get it installed? If you can find the source code it will be easy to find if it is written correctly.

[hardy cipher]

yeah, I'll go through the node groups and try to figure out where it came from. if there's a way to do that from the node itself I haven't figured it out. looking at properties doesn't reveal anything

[midnight shuttle]

If it's not built-in you had to install it somehow? You don't know where it came from?

[vital ermine]

Yes, I saw that and saw hope but I do not do any LLM stuff so not sure how that would related speedwise to SD potential.

[hardy cipher]

like I said, I just looked at the lost of all the cfg input options, and it was on the list

[midnight shuttle]

Have you installed any "node packs" or anything like that?

[hardy cipher]

it pops up by name if I search. but doesn't tell me where it's from. I imstalled too many node packs tbh. I'll figure out where it came from

[vital ermine]

my thing is for just SD I am not buying used from some schmo I have never met so that leaves only new and 7900XTX for 900ish or 4090 for 1730ish.

[midnight shuttle]

I can't find it in the Comfy source tree so it must be an add-on.

[hardy cipher]

yeah, I wish I knew of an efficient method. but I'll let you know when I find it

[vital ermine]

I sure get a lot of this even if I neg

I don't get it

I put cartoon, animated, unrealistic in the neg

[crisp owl]

Are you using preconditioning? (refiner before base model)

[vital ermine]

no, I am using G/L

then refiner last

[crisp owl]

What's going where then? Hard to rell without seeing

[midnight shuttle]

What are your G/L prompts?

[crisp owl]

L should be supporting terms, G should be your main prompt

[vital ermine]

hmmm

[vital ermine]

I thought L was wd1.4 like tags?

[ionic dragon]

are you on 4090?

[midnight shuttle]

That's been disproven. G is the newer CLIP model that supports more natural language. L is the old CLIP that is in SD 1/2.

[nimble heart]

7900 XTX

[midnight shuttle]

Both prompts should describe the image but using the proper format for the given CLIP encoder.

[ionic dragon]

I never thought amd cards would perform that great

[vital ermine]

Ahhh, I am going back to one prompt and forget this G/L nonsense

[nimble heart]

Me neither lol

[midnight shuttle]

It's not nonsense. It's how the model was trained. It was trained on both and the prompts are combined in the encoder. Concatenated really.

[nimble heart]

that's a 3rd party patch for autoqptq too so its probably suboptimal

[civic sigil]

But with only one prompt...

Right?

[ionic dragon]

isnt that 30b llm's req 48g of vram?
7b req 12gb
13gb req 20gb

[midnight shuttle]

No, the model was trained with CLIP-G captions and CLIP-L captions, and then the weights are concatenated. You can see how it works in the Comfy source code.

[nimble heart]

4 bit mode

that's why it needs autogptq

[midnight shuttle]

The model was trained with both prompts. G and L.

[ionic dragon]

whats that?

[civic sigil]

But where did they source the prompts?

[midnight shuttle]

From the CLIP models. Showing CLIP the image and getting the caption from it.

[nimble heart]

also I have no idea where these numbers are from. 13B requires 32 GB @ 16 bits or <16 for 8 bit

[midnight shuttle]

At least that's how I understand it.

[nimble heart]

way to run models @ 8 and 4 bit

also 3 bit I guess but it's kinda weird

[vital ermine]

you do you as far as I am concerned I am not making money from this so I will take a path of least resistance.

[nimble heart]

Sytan's 3090 performed like garbage @ 3 bit so I don't even wanna try it

[midnight shuttle]

Least resistance is not to use a new techology and wait for it to be fully mature. 😆

[ionic dragon]

llama v2

[nimble heart]

me stuck perpetually using torch nightly because normal torch doesn't support my GPU

[ionic dragon]

so its basically like pruned model

[nimble heart]

I guess

the actual safetensors files are way smaller so basically

[vital ermine]

wait for that and be waiting past the grave

[midnight shuttle]

I don't think so. In a couple of years there will be easily installed and fully documented image generator tools.

[nimble heart]

dunno what that is. im just using wizard-vicuna cause sytan told me to

its pretty cool. downloaded the 30b version and asked it to give me an example of famous Eldritch smut. It did. kinda wish it didnt

[vital ermine]

I went back to one prompt and refiner

[hardy cipher]

shouldn't there be a master list of nodes?

[midnight shuttle]

Impossible because anyone can create custom nodes.

[hardy cipher]

I meant within my install

[midnight shuttle]

I think if you search with no text it will list all of them but I've never tried.

[hardy cipher]

but it's alright. I don't feel like spending all night figuring it out. I just know it's one of several nodes with "_AS" as a suffix

[midnight shuttle]

I don't really have a use for many custom nodes. It seems like exceedingly complex Comfy workflows will be the SDXL version of 20 line word salad prompts on SD 1/2.

[ionic dragon]

asking it stable diffusion prompts?

[hardy cipher]

yeah, I'm not trying to put everything in. I just got curious about all the different things since with a1111 I wasn't really able to learn a bunch about the inner workings of things

[pallid sail]

Updated to the new base and refiner models that use the 0.9 rollback VAE and now RBR doesn't work anymore

[vital ermine]

If I put nighttime in refiner I don't get any difference

[hardy cipher]

I tested the refiner with a bunch of different words and saw nominal differences at best

[vital ermine]

I tried all sorts of words

[nimble heart]

dont think it knows what stable diffusion is

[hardy cipher]

literally like a little bit of shadowing on an ear would change

[crisp owl]

refiner isn't gonna vastly change the image

[vital ermine]

Yes, what I am getting too

[midnight shuttle]

That's what the refiner is for. The tiniest of changes to the existing image. Not to add style or objects.

[ionic dragon]

yeah, then what are you doin?

[hardy cipher]

well it doesn't really change anything. at least the prompt

[nimble heart]

?

[midnight shuttle]

Look at the SDXL paper for the with and without refiner images. The difference is very slight.

[vital ermine]

I can say 1, or 2, or fire, or ice, greass, etc... and no real change

[hardy cipher]

yeah, I do realize it has an impact on the overall image. just thought I'd see ore for some reason. also, I'm sure if I bumped itup from 20 percent of the steps it'd do more. but not sure that'd work otu well

[vital ermine]

then we should skip refiner I guess

[hardy cipher]

accidentally ran the refiner as base and refiner earlier. those images were not that pretty. not terrible abominations, but not good

[midnight shuttle]

The refiner is only to add the finishing details. It is not a dramatic change. It is like going from 9.9 to 10. Not from 5 to 10.

[vital ermine]

Now I did have one image it refined very well

[crisp owl]

the refiner does exactly as it's labeled. Just refines the already determined image and finetunes some details

[vital ermine]

ahhh, good, I can zap that then for most gens

[astral jay]

That's with ref

😛

[hardy cipher]

well I get that, but typing in "sharp" didn't seem to visibly make anything sharper. but then maybe that's not a great descriptor or something?

[midnight shuttle]

The refiner isn't a sharpener. If your image is blurry work on the base prompt.

[vital ermine]

exactly, which is what I thought it was for

[midnight shuttle]

The refiner is for fixing tiny details such as hair, ears. Look at the paper and it will show the examples.

[nimble heart]

sometimes adding things like "dslr" or "cinematic" makes it blurry cause it turns the bokeh up to 11

[hardy cipher]

I think the refiner just refines no matter what

I did go over that paper some, but not all of it

[vital ermine]

I tried it just now without anything in the prompt and the same outcome

[midnight shuttle]

You only have to look at the pictures in the paper to understand the refiner.

[vital ermine]

what is the link to the papers again?

[crisp owl]

https://arxiv.org/abs/2307.01952

[vital ermine]

thanks

[hardy cipher]

tried to have gpt-4 summarize, you know, one of those things it's supposed to do. but it did a terrible job. and also forgot the conversation we'd just been having

[midnight shuttle]

Directly from the paper, about the refiner: We note that this step is optional, but improves sample quality for detailed backgrounds and human faces, as demonstrated in Fig. 6 and Fig. 13.

[fleet harness]

try out https://claude.ai/login?selectAccount&returnTo=https%3A%2F%2Fclaude.ai%2Fchat%2F659d1ad1-7d9c-4c99-8659-fa4df600ae73

[hardy cipher]

asked it my initial question again and asked it to please keep in mind the question this time. and it just rattled off the same response asd had no recollection of my initial questoin

[midnight shuttle]

GPT has no recollection of anything. The only "memory" it has is from resubmitting your entire conversation each and every time. Did you click "new conversation" or did you use an existing conversation?

[hardy cipher]

same conversation, and repeated the question

[ionic dragon]

@midnight shuttle how to divide this prompt?anime manga robot cat tattoo, cyborg cat, exposed wires and gears, fully robotic cat, manga in the style of junji ito and naoko takeuchi, cute chibi cat, tattoo on upper arm, arm tattoo

[midnight shuttle]

But you started with brand new conversation, right?

[hardy cipher]

believe me, I used to use it very extensively, I got pretty good with what it was

[midnight shuttle]

That looks like an L prompt to me.

[vital ermine]

This paper used the same prompt for both

[hardy cipher]

well I mean, why would a new conversation go any differently? I reiterated everything I'd said and it did the same thing in the same conversation

[ionic dragon]

then how can the g prompt be?

[midnight shuttle]

That seems to be the most common way, with the idea that if you can properly structure the G and L prompts you can get better results.

[hardy cipher]

it's memory is 4096 tokens or something

[midnight shuttle]

You can put whatever you want in any prompt. It's just tokens after all. No magic.

[hardy cipher]

but every single thing pushes something else out. no discernment of importance

[ionic dragon]

oh ok

[vital ermine]

yeah, after a lot of playing with and without I was finding I might as well just ship the same prompt or leave it empty even

[midnight shuttle]

Meaning that only 4096 tokens can be submitted in a query. In the ChatGPT interface, if the conversation is over 4096 tokens it will start removing older messages from the context. But there is no memory. The whole conversation is resubmitted every time.

[midnight shuttle]

I would use the same for both. That's closer to the way it was designed to work.

[hardy cipher]

man, I probably gave it about 50 tokens

but the pdf blew it's mind

[vital ermine]

alright

[midnight shuttle]

It can't discern importance. All it does is resubmit your entire conversation and then try to guess what comes next.

[hardy cipher]

it's tools are useless

[midnight shuttle]

That PDF is well over 4096 tokens.

[ionic dragon]

@midnight shuttle is it a g prompt?A robot with lights around its body facing towards the camera in a dim room and holding a lightsaber in hand

[hardy cipher]

well I think that's inherent with the transformer model structure right?

they can't differentiate

[midnight shuttle]

Yes I think that is more like a G prompt.

Doesn't mean it will work or produce what you want. But it is the structure of a G prompt.

[midnight shuttle]

You mean that the whole conversation has to be submitted every time? Yes.

It has no memory.

[hardy cipher]

well I'm about ready to cancel my membership because I can't find any use for it anymore. a few months ago I was asking it questions all the time

but all those guardrails, and I'm fairly certain it's been very dumbed down resource-wise. millions of new users and it all the sudden responds in 3.5 speed

[crisp owl]

gpt3.5 does well enough for my random questions

[midnight shuttle]

The context limits and model remain unchanged. But more guardrails have been added to avoid saying inappropriate things and to avoid inappropriate uses such as legal/medical.

[hardy cipher]

yes, the guardrails have crippled it completely

[midnight shuttle]

OpenAI has to stay out of trouble.

[hardy cipher]

I realize the model itsslf is still just as capable

I think they're in trouble either way. not legal maybe. but their product is crippled

[midnight shuttle]

But if you just look at the pictures in the paper you can understand the refiner. No AI GPT needed.

[hardy cipher]

true

[midnight shuttle]

You will see how subtle the refiner is, but you will also see how it makes the images look better overall.

[midnight shuttle]

Their product is fine for the real use case, which is for businesses to use it in specific purposes. ChatGPT is just a demo.

Even paid ChatGPT is only a demo, but they need to charge some money to limit usage.

[vale eagle]

Somehow, I think putting the merged prompt to both clip g and clip l is better than clip g for clip g, clip l for clip l.

[hardy cipher]

I'd considered that. plus they can just push a slightly improved gpt-4 as gpt-5 and people will be all over it

[midnight shuttle]

You mean same prompt in both? That way seems to work quite well. But you understand the difference between the CLIP models so you also understand why the prompts could be different.

[ionic dragon]

ok, and l prompt is just supporting tags like 8k, uhd, intricate, cinematic and stuff?

[midnight shuttle]

No, it describes the image but in the format of a SD 1/2 prompt, so more like tags than sentences.

[vale eagle]

yes. Something like 'a cinematic photo of a cute dog at evening in 2018, 8k, surreal, etc' and put it to both clip g and clip l

[midnight shuttle]

For example G prompt: A close-up picture of a large gray cat that is sitting in a chair next to a television set with books next to ie. L prompt: large gray cat, photo, close-up, television, chair, books

[midnight shuttle]

That will work. Same prompt is both seems to be common and produces good results.

G prompt is sentences like normal language. L prompt is tags like SD 1/2 prompts.

Because L prompt is using the same CLIP encoder as SD 1/2.

[ionic dragon]

oh ok, in l prompt it gives details of the scene?

[midnight shuttle]

No. It is just the SD 1/2 prompt. Both prompts describe the scene, but in different ways because they use different CLIP encoders.

Don't overthink it. Both are prompts. But the G encoder is more sophisticated.

Both are full prompts.

The example I chose was too complex for any CLIP encoder right now because it has too many relationships between objects. I was only showing the sentence structure.

[ionic dragon]

why not use same prompt in both

[midnight shuttle]

You can. Most people do.

[vital ermine]

Weird, I can't get away from images like this so I am genning now with a different seed.

Sigh

[midnight shuttle]

I don't even know what that style is called. What prompts are you using?

[vital ermine]

Made up ones I used to do in 1.5 and 2.x

I can change them to anything and I get that same style

must be something in this workspace I would think

[midnight shuttle]

Only the prompts are used to generate the weights. But the G prompt seems like it doesn't describe an image as much as a poetic scene. The G encoder might not produce meaningful weights with that.

And the L prompt doesn't describe the image but appears to be a collection of style words.

[vital ermine]

As I said I described the zombie scene above and got the same style

Don't get me wrong I like this style but not intended

[midnight shuttle]

Did you also change the L prompt?

[vital ermine]

I have changed even the neg

[midnight shuttle]

Try putting a non-poetic description in both the G and L prompts. Same for both.

[vital ermine]

apparently it is ignoring the neg

You should see my 2.1 prompts they are like that for some really wild gens

I lost all my old 2.1 gens a couple of months back so hmph

[soft zealot]

for testing reverted to Sytans basic rather than using my normal tweaked flow

Sampler might have an impact as well BTW, I used DDIM & DDIM_Uniform

[vital ermine]

did you see my message to you?

I mentioned to you about comfybox. I think that thing is doa

[soft zealot]

No I hadnt noticed, TBH as I said to Alex its notfro me as I'm happy with ComyUI although I can see a use for it and wished him luck

removingthe negatove

[vital ermine]

Well, I want a UI like that to try but his shit doesn't work it serves me the folder contents and not a UI

I noticed he doesn't answer tickets so figure it is dead

[soft zealot]

describing it as "shit" is a little OTT don''t you think?

There are merits to it and I can see where he is going and I get the concept, it's just not a concept for me.

Hangon sorry Im getting confused, havent had COffee yet.

ComfyBox not SwarmUI

Doh

[vital ermine]

Yeah, not swarm, lol

I said his shit because I have no idea if good or not it doesn't work and it becomes "shit" when a dev seems to not really be answering his tickets or taking care of stuff. I think I saw the last update was 4 months ago. btw, his instructions were way off and even in the readme his port is off.

[soft zealot]

Thisis why I'm not bothered about either Swarm or ComfyBox

[vital ermine]

swarm has merit

Way too overly complicated and my card can barely handle XL as it is.

I like that workflow though as it is super clean. Your baby or is it available publicly?

[full idol]

Lower is better no?

[vital ermine]

no, unless s/it but lower it/s is bad

[soft zealot]

"overly complicated? This is the simple bit that broken down into Zone

Orange- Image Previews
Purpl - Progress
Pink - Prompt Input
Yellow - Image Settings

And what card do you have ? Im running a 1080Ti

[vital ermine]

I am on a 1060

[soft zealot]

layout of workflow has no impact on performance. Calling it overcomplicated when all you've seen is the basixc settings screen is a little "harsh" isnt it ;o)

Heres the rest of it lol

[late marsh]

Morning of monsters today. Hello! 👋🏻

[vital ermine]

For me, and I am speaking only for me, it is overly complicated. I am only on day 2 with this XL stuff.

[heady vale]

yep. takes a few days then you will be hungering for more custom nodes

[vital ermine]

so far I am fighting going back to 2.1 tbh. I just liked the workflow better, but everyone is different AND why I was hoping comfybox would work.

[soft zealot]

so a screen layout that hides the inner workings and displays the interface in a structured format (a nit like a A111 or other WebUI interface) is "overly complicated"?

Yes It makes harder to follow through the workflow but thats not the point, This is my daily driver interface that has what I need in one screen loaded wlike this when it opens.

THis is why I'm not so bothered about a wrap such as Swarm of COmfyBox.

Sure I'll give them a whirl and see what theyre like but...........

[midnight shuttle]

It seems like SDXL still gives good results using 2.1 style prompts. I think you can even get away with using the old CLIP conditioning nodes. But I think those only feed to the L prompt.

[royal fern]

the suggested workflow doesn't use the specialized SDXL nodes

and indeed it's easier to use

[fleet harness]

Pulling prompt ideas from the news... Post Malone has the One Ring !

[midnight shuttle]

Would be interesting to trace the code and see which tokenizers are used. But not today.

[soft zealot]

you are you are confusing UIs with Models.

UIs such as COmfyUI , Automatic 1111 etc can all use the various SD models whrtther that s 1.5 , 2.1 or XL

When you say "going back to 2.1! I presume you actually mean "going back to A1111 (or whatever UI you were using prior to using COmfyUI"

[ionic dragon]

are the images grainy?

[vital ermine]

Yes, and no. A111 is a dog turd now with a pc mem leak from hades itself (128GB and growing) and the latest update is so sluggish if I went back to 2.1 the model it would be back in the old Feb version of A1111.

[midnight shuttle]

I haven't seen any memory leaks with A1111. Maybe it's caused by an extension?

[royal fern]

are you upscaling the latent?

[soft zealot]

If you're more comfortable with 2.1 or 1.5 prompting why not build a workflow in COmfyUI to use 1.5 or 2.1 models?

[vital ermine]

A PC mem leak and others have it too.

[midnight shuttle]

I haven't seen a steady increase in system RAM or VRAM with A1111 1.5.1.

[soft zealot]

Comfy UI is not restricted to SDXL you know

[vital ermine]

I did as soon as I used SDXL

[ionic dragon]

idk, using sytan's workflow

[vital ermine]

2.1 it was just so sluggish I felt like I was on an old core2duo again

night and day

[midnight shuttle]

I've used SDXL in A1111. But I don't really use many extensions. So that could be the difference.

[vital ermine]

Probably

[royal fern]

comfyui is pretty much officially supported by SAI at this point

[hardy cipher]

2.1 just never had the loras and other tools. and also they tried to lock too many things out

well 2.0

[vital ermine]

one extension I had to turn off it was a native one too as it was preventing xl from genning

[hardy cipher]

maybe 2.1 was a bit better

[ionic dragon]

lemme share the workflow

[vital ermine]

prompt magic I think it was called or something like that

[midnight shuttle]

I don't see why 2.1 couldn't have been waifu trained just as much as 1.5 was. I think the bigger reason people hated 2.1 was because they were dependent on artist names to get styles and a lot of artist names were removed in 2.1.

[vital ermine]

I really didn't miss them after a few days in 2.0

main reason people stayed with 1.5 was artists then the lack of waifu titties

lack of nudes in 2.x did cause anatomy issues, ngl

[midnight shuttle]

Wouldn't showing 2.1 enough nudes in a LoRa training set have solved that?

[vital ermine]

Not sure but I never visited that part of the web as most of that seems to be 1.5 based

[royal fern]

I'd need to test it a little bit better but it's like it adds a little horizontal/vertical texture. can you try a photorealistic picture? I don't have the image batch plugin from that workflow

[vital ermine]

Oh, here is a 100% as a trainer I can tell you 1.5 training was so dang easy with the exact same dataset and 2.1 junk. So hard t train 2.1

[midnight shuttle]

Basically I don't think 2.1 was given a fair chance and I think artist names were a bigger factor than most people realize. But now we have SDXL so it's all good.

SDXL can be waifu trained and most people have learned how to prompt without having to tag Greg Rutkowski in every image.

[vital ermine]

lack of movie stars to be replaced with body doubles didn't help

[midnight shuttle]

They also took out celebrities? I didn't realize that.

[vital ermine]

you know I started to do 2.1 gens then inpaint 1.4 model for movie stars just to get them.

[hardy cipher]

no celebs, no nips

[vital ermine]

1.4 or 1.5 depending on the star

yep

I recently retried 1.5 due to Sytan and my first 3 gens on base were all nudes. LOL. In 2.1 I hadn't seen a nude unless I had a model/lora with them.

[ionic dragon]

2.1 is censored, i dont think they trained with NSFW pics for 2.1

[vital ermine]

As emad said long ago you can have children or nudes but not both.

[ionic dragon]

for 2.1?

but I think ppl have made loras for that too

[vital ermine]

for 2.x or going forward. I hope he didn't change his mind as that was in response to possible lititgation if they did both

you may just they can't as a company

[royal fern]

I don't really see the problems since it can be trained... most of the 2.1 models I use can do boobies

[ionic dragon]

yeah, too many legality issues for them

[vital ermine]

you don't see? They aren't directly responsible so no case will last against them. Wise move for SAI

[ionic dragon]

90% models have used NSFW images to train

as NSFW are the most generated using SD

as MJ is censored

[visual glade]

You can now bypass nodes in ComfyUI by using: CTRL-B. This will act like if the node was removed and the links connected across.

[vital ermine]

yeah, and some are darn hard to not get nude from them too even with nude, nsfw, etc... some I had to add the word corsette to them so it wouldn't be a nude.

[ionic dragon]

soon, we can expect nudes from sdxl

[vital ermine]

Of course.

[hardy cipher]

civitai is on it

[ionic dragon]

civitai is flooded with nudes

[vital ermine]

I had to NSFW civit after my first signup to be met with big dongs and other bad things that would even make pornhub blush

lucky I wasn't at work

[hardy cipher]

those lads are thirsty

[thorny frost]

Civitai Is flooded with a lot of crap henrtai loras

[hardy cipher]

yeah, it creeps me out tbh

[vital ermine]

Dr. Octipus no thanks

[thorny frost]

I can get ok nudes in SDxl already with a couple of loras

[royal fern]

yeah especially considering the extremely young age of those fine girls

[ionic dragon]

yeah, its mostly hentai so there isnt too much threat, if ppl start training lora on celbs and nudes, then civitai might be banned

[royal fern]

it honestly makes me sick and the less I access that website the better

[thorny frost]

Fortunately there are filters you can apply to not clutter the site with so much crap

[hardy cipher]

fine girls? I honestly don't get any of that. not trying to be preachy about it, but it's really unsettling to me

[vital ermine]

you can even block authors

[royal fern]

the other day was presented with an uncensored anime/furrie/wolf showing up its p**ssy

[hardy cipher]

well some of those models will be good but then all the images will have random nudity in them. and you can't stop it with negative prompts

[thorny frost]

But otoh it has amazing things too I can't find anywhere else....

[royal fern]

necessary evil

[hardy cipher]

this is the first place I've found that isn't flooded with that stuff

as far as stable diffusion discussions are concerned

even hugging face

good news, everyone. I figured out a convoluted cfg method

using this. seems unnecesarily complex though. at least for what I'm trying to do https://github.com/ltdrdata/ComfyUI-extension-tutorials/blob/Main/ComfyUI-Impact-Pack/tutorial/pk_hook.md

[hardy cipher]

decent

[flint copper]

Hi community, hope everyone's doing good. I'm having an issue whereby some of the workflows, even Sytan's, will not load on my Windows installation of ComfyUI. Some workflows just won't show up at all, while others will. Any insights?

[uncut steeple]

#🤝｜tech-support id say, but also does the console spit any errors?

Getting somewhere

[royal fern]

start with this https://comfyanonymous.github.io/ComfyUI_examples/sdxl/

[flint copper]

Now that you mention it, I'll have a look. Thanks.

[flint copper]

Thank you for that. Just getting onboard the comfyui bandwagon. It's quite promising and the interface is amazing.

[royal fern]

the text_l/g is kinda weird and the suggested workflow doesn't use them

[hardy cipher]

meh, use it or don't, I think there aer arguments for both approaches

just depends on what you're going for

[royal fern]

it's great to get very good results with relatively simple prompts

ala midjourney

[hardy cipher]

yeah, but to me that's like having a self driving car

[royal fern]

but you can't finetune

[hardy cipher]

yes

[flint copper]

Hmm... used to fix hands/fingers with Adetailer extension on automatic1111. Wonder what's the solution here.

[hardy cipher]

but I'm not going to knock what someone else prefers.

[royal fern]

well certain things cannot really be done with the text_l/g workflow

so you can use it up to the point you don't need certain details

[unreal plover]

rundiffusionXL is a very good model

[vital ermine]

is there a web page for comfyui, like auto1111 has, to explain what Command Line Arguments and Settings are used for? -h only tells you what they are not what they are used for.

[hardy cipher]

I'm sure they're listed on the github. but yeah, not sure about explanations

and I don't there are very many

[vital ermine]

there is a few, yeah

[-h] [--listen [IP]] [--port PORT]
[--enable-cors-header [ORIGIN]]
[--extra-model-paths-config PATH [PATH ...]]
[--output-directory OUTPUT_DIRECTORY] [--auto-launch]
[--cuda-device DEVICE_ID]
[--cuda-malloc | --disable-cuda-malloc]
[--dont-upcast-attention] [--force-fp32 | --force-fp16]
[--fp16-vae | --bf16-vae] [--directml [DIRECTML_DEVICE]]
[--preview-method [none,auto,latent2rgb,taesd]]
[--use-split-cross-attention | --use-quad-cross-attention | --use-pytorch-cross-attention]
[--disable-xformers]
[--gpu-only | --highvram | --normalvram | --lowvram | --novram | --cpu]
[--dont-print-server] [--quick-test-for-ci]
[--windows-standalone-build] [--disable-metadata]

[hardy cipher]

I actually don't use any right now. maybe i should though

[royal fern]

comfyui is pretty smart and generally the default config works

[vital ermine]

I have specials wants and like to know

[royal fern]

you may need to change something only with very low spec'ed PCs

[hardy cipher]

yeah, xformers on by default

[hollow halo]

I think there's a similar extension for Comfy but I usually do the ones I really care about in photoshop

[half ivy]

I am here

What did I miss?

[unreal plover]

Have you heard about LK99 ?

[half ivy]

Yeeeaaa

[unreal plover]

so you miss nothing important

[half ivy]

Game changer

Kkkkkk

No SDXL control net things?

[unreal plover]

We are all waiting for this

[royal fern]

yummy

[uncut steeple]

[half ivy]

Thanks for this 🙏

[ionic dragon]

did anyone use searge's workflow v3?

[ionic dragon]

can you please tell why my image is too noisy

can someone please help me

[royal fern]

area conditioning?

[heady vale]

not even totally sure what that is

I guess not

[winged scroll]

Hi guys!
I need your help. What negative prompt can save eyes in my renders?

sdxl 1.0

[stray wave]

Can SDXL make first person images? 1.5 strugged with it

[unreal plover]

near

[royal fern]

how do you mean?

[unreal plover]

new fashon style

[stray wave]

Like Skyrim first person view or Doom

Seeing hands and body first person

[royal fern]

oh I see

let's give it a try 😄

[hardy cipher]

you know, I was just thinking of making something in a video game

[stray wave]

Not bad, but can it generate legs too?

[unreal plover]

csgo ?

[flint copper]

Just a random prompt. One of those "I'm feeling lucky" nodes.

[royal fern]

lol it's actually pretty fun

[meager canopy]

My workflows use something very similar to adetailer.
https://civitai.com/models/119257/gtm-comfyui-workflows-including-sdxl-and-sd15

[royal fern]

you need a bit of luck with the seed though

[meager canopy]

The section on the right has a switch to enable/disable and you choose the detailing model for hands, faces etc.

[hardy cipher]

nice

[royal fern]

in a fantasy setup it doesn't like to put the first person sword/shield/whatever

[unreal plover]

it is hard to do POV view, i will try to make a lora

[lean granite]

What's lk99?

P realistic picture btw, impressive 💪

[unreal plover]

thanks but i just use the stability Photographic presets :
p : cinematic photo {prompt}. 35mm photograph,film,bokeh,professional,4k,highly detailed
n : drawing,painting,crayon,sketch,graphite,impressionist,noisy,blurry,soft,deformed,ugly

[wanton oxide]

Is that a straight-up text prompt with SDXL?

[royal fern]

yes

it works very well with doom-like FPS

anything else is a bit more complicated

[royal fern]

to do something a little more realistic it took some convincing...

[unreal plover]

LK99 is this

[unreal plover]

That is disturbing

[royal fern]

yeah

as soon as I said cop it was easy to put a gun on

[unreal plover]

that fail

[west breach]

fascinating!

[unreal plover]

The imagination of this model is amazing, no need to do heavy prompt

just type a fractal tornado in paris with the photographic style

[west breach]

crazy, I wouldn't have thought it could do something so abstract

[unreal plover]

some tokens are more flexibles than others

[vital ermine]

if I click a node how do I unstick it from the cursor as it acts like TP stuck to a shoe?

[heady vale]

drag your mouse across the carpet 😆

[peak dove]

Bot 8

[hollow hare]

hi guys, i want to ask, between SDXL 0.9 VAE and SDXL 1.0 VAE, which one has better output? I think I saw some ppl said that 0.9 is better, but im not sure too

[peak dove]

Some said VAE1.0 had some horizontal lines artefact...

[thorny frost]

A trick: I ask Bing to generate some images (it uses Dall-E), I say it to feel free to add as many details as he wants and such. The result uses to be very imaginative and variated, very 'human' I would say, but tends to be somewhat imperfect and even ugly sometimes, but then I take it and send it to A1111 for an img2img treatment with SDXL botox, so it is also beautiful. Best from both worlds.

[hollow hare]

i see, even i think the ema prund kinda generate better face than sdxl tbh

does anybody have links to download the SDXL 0.9 VAE?

[minor glacier]

The main model reverted to the 0.9 vae about 2 days ago. So just download the SDXL 1.0 vae from hugginface

[hollow hare]

i downloaded the VAE few days ago this week, but the file upload time is 7 days ago. is this the 0.9 VAE one?

[minor glacier]

Nope, not that one., there a full safetensors model from the page I linked
Or you can get the vae alone from the vae sub-directory
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main

[hollow hare]

ah the diffusion pytorch model is it?

[molten gull]

is the guy who's developing a1111 around here, too, by the way ?

[median isle]

I think Voldy got kicked from here

Voldy being the person developing A1111

[minor glacier]

Yes , stable-diffusion-xl-base-1.0/vae/diffusion_pytorch_model.safetensors
will be the vae on its own, the other one in the same directory is the fp16 version, if you want that, but I understand some people are having issues with the fp16 version (although not sure if that was fixed by the 0.9 vae being restored, it works fine on MacOS using python scripts).

[hardy cipher]

lol, how'd that come about?

[late marsh]

I can die happy now 🥹

[heady vale]

beautiful

[median isle]

There was that big hubbub about Voldy having a Rimworld mod that removed some of the diversity in the game in the same Github account as the SD Webui PLUS at the same time NovelAI leaks happened and Voldy's implementation was too good too quick so they assumed stolen code PLUS I think there was some beef between Emad and Voldy

Just a collection of smaller incidences that lead to Voldy getting unpersoned

[upbeat summit]

Those are amazing! great work.

[hardy cipher]

he sounds like a bit of a drama llama

[naive inlet]

anyone know what might be causing my images to look like this?

[mellow tendon]

You could try upscaling with a 0.3-0.4 denoising strength, maybe use a good SD 1.5 model to do it.

[hardy cipher]

https://tenor.com/view/monday-life-funny-llama-funny-llama-gif-13058459

[hollow hare]

are you using ema pruned vae? last time it happened to me too

[median isle]

aye, but on the other hand I think he's like a relatively young lad

[naive inlet]

no, i'm trying sdxl1.0base and refiner, and dreamshaper8

and dreamshaperXL

all look lke that

[median isle]

so it's that phase I guess

but back to topic, I've been fiddling with face fixing mid gen and I'm getting great results

[hardy cipher]

fair enough. and he did make something that changed the game imo. not sure how much help he had or if he did steal code. but people do dumb things

[median isle]

Oh, he didn't steal any code

[molten gull]

what kind of face fixing did you do ? can you explain ?

[hardy cipher]

i haven't really messed with the face fixing. maybe I should

[median isle]

the implementation was done just too fast so they assumed it

[median isle]

I have a two-pass set up where the first pass finishes with noise (45 steps out of 60 on DDIM), then it's VAE Decoded and fed into an ComfyUI Impact Pack face fix, then it's VAE Encoded and fed into the second pass which finishes the gen

The end result is way better blended in

[hardy cipher]

so should at least one pass be 100 percent? or can you take care of the noise without that? never really figured that out definitively

[median isle]

The second pass starts at step 45 and goes to 60

it basically just pauses mid generation, fixes the face first, then finishes

[hardy cipher]

ahh. I tried that before and it didn't work the way it was supposed to. not sure what I did wrong.

[molten gull]

do you have a link of where to get the impack pack face fix ?

[hardy cipher]

I saw them in the comfyui manager

haven't downloaded them

[median isle]

https://github.com/ltdrdata/ComfyUI-Impact-Pack

but yeah the way to install is via the Manager

[naive inlet]

@hollow hare it was the steps, I had 20 steps (fixed) and 15 end_at_steps (fixed). when i removed those it looks normal

[molten gull]

thanks

what nodes of that do you use for facerestore ?

[median isle]

This is the entire pass, it's really messy still so apologies

Two passes on the face detailer because it sometimes leaves a ghosting chin line

[molten gull]

ok, thx

is there a "hand-fix", too ? or only a face-fix ? or a leg-fix, or a whateverfix ?

[median isle]

There's handfix as well yeah

just have to download the appropriate models

[molten gull]

how do you create a hand-fix ?

it's not a lora, is it ?

[late marsh]

Oh noes ❤️

[median isle]

Uhhhhh I think you load the correct segment model and then use a generic detailer for fixing the hands

[molten gull]

hmm, how would i do this ? i'd like to do my own facefix

[uncut steeple]

this workflow was posted here before, it has the hand and face fix built in so you can look at it

Not mine

[royal fern]

not sure if this was already posted by I just saw this in SAI code

SD_XL_BASE_RATIOS = {
    "0.5": (704, 1408),
    "0.52": (704, 1344),
    "0.57": (768, 1344),
    "0.6": (768, 1280),
    "0.68": (832, 1216),
    "0.72": (832, 1152),
    "0.78": (896, 1152),
    "0.82": (896, 1088),
    "0.88": (960, 1088),
    "0.94": (960, 1024),
    "1.0": (1024, 1024),
    "1.07": (1024, 960),
    "1.13": (1088, 960),
    "1.21": (1088, 896),
    "1.29": (1152, 896),
    "1.38": (1152, 832),
    "1.46": (1216, 832),
    "1.67": (1280, 768),
    "1.75": (1344, 768),
    "1.91": (1344, 704),
    "2.0": (1408, 704),
    "2.09": (1472, 704),
    "2.4": (1536, 640),
    "2.5": (1600, 640),
    "2.89": (1664, 576),
    "3.0": (1728, 576),
}

[molten gull]

but in the end it's a new node, aint it ? the FaceFixRestorationNode or something

how would i make a MalicorFaceFixNode ?

[uncut steeple]

There is an apsect ratio selector node in comfyroll

[median isle]

if you look at the docs, the aspect ratios that are officially supported are

1024 x 1024
1152 x 896
896 x 1152
1216 x 832
832 x 1216
1344 x 768
768 x 1344
1536 x 640
640 x 1536

[royal fern]

it's more than that apparently

there are resolutions that I haven't seen before

[vital ermine]

any comfy users know the convert text to node which node do I use to connect to it?

[uncut steeple]

got an image of what you need?

[vital ermine]

sec

[uncut steeple]

Cause im not sure what you mean

[median isle]

yeah but I've had mixed results with the full trained data resolution list, usually proportions get spaghettified

[royal fern]

right click > convert to input you mean?

[vital ermine]

I can't find a node that will connect to it

[uncut steeple]

So you are asking what to connect to it if you convert it?

[vital ermine]

yes

[royal fern]

me too

[uncut steeple]

Use a primitive

They should adjust to your needs

[vital ermine]

I only have int primitive

[heady vale]

double click the circle for that input

[uncut steeple]

make a new one

[tribal jackal]

idk why I find these so funny

[vital ermine]

Oh, kick ass. Thank you

[split magnet]

Hey there so I have a question, I downloaded SDXL 1.0 on my PC and the file was about 6GB in size, but when I enable it inside Stable Diffusion it suddenly starts to unpack and eats up 20+ GB of space,
what's up with that?

[vital ermine]

XL wants 1024sq after testing as anything less it gets wonky

right one is what it was supposed to be

[uncut steeple]

I agree to disagree

[urban breach]

https://tenor.com/view/spawn-party-crasher-gif-10785692

You're getting a little too good at that. You posted that picture with the flowing black outfit yesterday morning, I was thinking I wonder if that Runway v2 thing would make the outflit flap around in the wind like this scene in Spawn

[unreal plover]

Octopath traveler token give interesting results mixed with some other effects

[uncut steeple]

Almost got the text

[urban breach]

And meanwhile I can't even get it to generate the Battlestar Galactica lol

[vivid tide]

Can anyone here help me with sdxl problems?

[uncut steeple]

Maybe explain the problem

[late marsh]

Here's some salt 😛

[royal fern]

love the style

[vivid tide]

Thanks. So first i had the problem that xformers was missing but i was able to fix that by myself. My second problem is that its unaable to create model quickly(Failed to create model quickly; will retry using slow method.). Just all errors

[uncut steeple]

What are you using to run SDXL?

[vivid tide]

A1111

[uncut steeple]

Also thats a question for #🤝｜tech-support i feel

[vivid tide]

No one answers there :/

[uncut steeple]

You cant expect help within 10 mins

[vital ermine]

Just telling you my problems disappear when I go 1024x1024 and it listens to my prompt.

[vivid tide]

True. Some people say i should downgrade my python

[uncut steeple]

Did you check the pinned message from joe? That might help you

[vital ermine]

Where?

[uncut steeple]

The first pin in this channel

[vital ermine]

I am not using any conditions

[jolly creek]

It could be a problem with lack of ram/vram. Try closing everything on pc and try running A1111 again

[vivid tide]

I have 32g and using 10. i will try that

[uncut steeple]

Its more about target width and height

[vital ermine]

768x768

[jolly creek]

I have 32gb as well but that helped me when I was using A1111. Had the same error

[vital ermine]

no issue at that 1024x1024

[uncut steeple]

set it to the size you want to generate at

[urban breach]

"blood moon" style, i think the last one is from a Devin prompt

[vital ermine]

I think your workflow and mine are different. I am using simple

[vivid tide]

Doesnt change it. Can i send my code in here?

[vital ermine]

the default one

[uncut steeple]

Maybe you should evolve

[vital ermine]

nah, no thanks

[uncut steeple]

Send your workflow ill have a look

[vital ermine]

rather go back to 2.1 as I did evolve and went back to default

[hollow halo]

That would be interesting to try out!

[uncut steeple]

[vivid tide]

Loading weights [31e35c80fc] from C:\Users\Edgar\Desktop\AI\sd.webui\webui\models\Stable-diffusion\sd_xl_base_1.0_2.safetensors

crashes

[jolly creek]

Have you set additional COMMANDLINE_ARGS ?

[vivid tide]

Yes i do.--api --xformers

[jolly creek]

try adding --no-half-vae --medvram --opt-split-attention as well

[vivid tide]

Didnt fix it. IT seems like it crashing at To create a public link, set `share=True` in `launch()`. Startup time: 10.4s (launcher: 2.3s, import torch: 3.2s, import gradio: 0.9s, setup paths: 0.7s, other imports: 0.9s, load scripts: 1.3s, create ui: 0.6s, gradio launch: 0.2s). Creating model from config: C:\Users\Edgar\Desktop\AI\sd.webui\webui\repositories\generative-models\configs\inference\sd_xl_base.yaml

[royal fern]

oookay I tested ALL the resolutions suggested in the SAI code (not only the "standard" ones). They actually all worked without replication, long necks and doubles... I'm actually pretty impressed.

Since there are a lot of them I don't post them here but if you are interested I made a gallery on imgur https://imgur.com/a/5R8s6Jr

[jolly creek]

Hmm that's all from my side, those were some quick fixes that helped me with same issue. But all my launch problems went away when I switched to ComfyUI

[vivid tide]

Oh okay, i will try that then

[royal fern]

this is impressive 🙂
https://i.imgur.com/qVoNPgt.jpeg

[jolly creek]

If you're using windows in github It has this quick install zip archive that all you have to do is extract and you're good to go.

[royal fern]

who needs outpaint?!

I really need a better GPU for these tests

anyone donating a 4090?

[uncut steeple]

Not sure what it is

[upbeat summit]

it's beautiful 🙂

birth of a fire phoenix

[vivid tide]

YOU DA MAAN! After installing comfyUI all my stuff works now

[jolly creek]

Awesome, have fun

[west breach]

Is there a node that connects to oobabooga?

[upbeat summit]

noice: https://github.com/comfyanonymous/ComfyUI/commit/05321fd947947d6122fd2a40520fa8da0d376456

[upbeat summit]

there are couple of nodes that integrate LM... let me check my bookmarks 😄

[uncut steeple]

[upbeat summit]

Wes Anderson Candy-GTA?

[split magnet]

there is actually a pretty nice Wes Anderson themed extension FYI

[upbeat summit]

cool 🙂 yeah, wes anderson is strong in base SDXL for sure but I haven't checked out a fine-tuning in the style

[west breach]

love the wes anderson style. my current desktop

[upbeat summit]

wonderful image!

[split magnet]

Looks like an actual shot from one of his movie where the camera randomly focuses on a lady staring at the main character while they drive a car past the cabbage field

[west breach]

it really feels like she's contemplating life on her cabbage farm. maybe thinking back to the days when she was the winner of beauty queen pageants

[upbeat summit]

so this is supposed to work with oobabooga - problem is it's 4 months old. I haven't tried it.

https://github.com/xXAdonesXx/NodeGPT

most people I know just let it run in an extra window and copy and paste - or I would maybe try writing the LM output in a txt file and load that with a file or wildcard node as an integration workaround

[west breach]

interesting, thanks for the link. I saw the api example was pretty straight forward

[upbeat summit]

you're welcome! or you take a look what you can do with a custom script and ComfyUI's API: https://github.com/comfyanonymous/ComfyUI/blob/master/script_examples/basic_api_example.py

[unreal plover]

I want to play this game !

[upbeat summit]

epic boss fights

[west breach]

Shadow of the Colossus?

[unreal plover]

Not in the prompt but i have "titan" token
Octopath Traveler style,a giant monster titan made of fractal smoke destroying a montain ,film,professional,4k,highly detailed

no montain, it look like the first token is more important

[upbeat summit]

really cool prompt build!

[unreal plover]

so badass !

cinematic photo a giant monster titan made of fractal smoke destroying a montain. 35mm photograph,film,bokeh,professional,4k,highly detailed

[lilac wren]

I'm curious about your method to have so much consistency in this format

[unreal plover]

adding mechanical in the prompt waw

@lilac wren just set the size to 1280x720

DPM++ 2M SDE Karras

[royal fern]

try it at 1728x576 😄

[unreal plover]

negative : drawing,painting,crayon,sketch,graphite,impressionist,noisy,blurry,soft,deformed,ugly

[west breach]

epic

[lilac wren]

No? Really?!

https://tenor.com/view/thanks-obvious-captain-thankscaptainobvious-gif-18783644

[uncut steeple]

My workflow is in the metadata, its comfyUI

[royal fern]

my new preferred resolution 😄

[swift acorn]

its just me or sdxl is worse in nsfw than 1.5?

[unreal plover]

ultrawide

[royal fern]

what GPU do you have?

[unreal plover]

3090

[uncut steeple]

Make me a wallpaper for 5120x1440 for my odyssey G9 and im happy

[royal fern]

I can tell 😉

[unreal plover]

oof

[royal fern]

that's not too bad

[unreal plover]

1080p

[royal fern]

tile decode and you are good to go

[lilac wren]

did this last night

[upbeat summit]

it was a great image last night, it's great now 🙂

[unreal plover]

@uncut steeple Sorry but vertical images are broken

[uncut steeple]

Let me try

[upbeat summit]

this can be fixed. you need to adjust the width/height + target_values in your textencode nodes

[unreal plover]

i use auto

[upbeat summit]

oh okay - I haven't used SDXL in auto yet :/

[unreal plover]

yes i dont use refiner too, so slow to load for each images

[lilac wren]

GUys, how to updates custom nodes? I did a script but i have to add every nodes inside each time. Isn't there a way already present in comfy?
`@echo off

REM Liste des dossiers à mettre à jour (remplacez "dossier1", "dossier2", etc. par les noms de vos dossiers)
set "folders=ComfyUI-Impact-Pack ComfyUI-Manager ComfyUI-QualityOfLifeSuit_Omar92 ComfyUI_Comfyroll_CustomNodes ComfyUI_ImageProcessing ComfyUI_UltimateSDUpscale comfy_mtb Derfuu_ComfyUI_ModdedNodes efficiency-nodes-comfyui MergeBlockWeighted_fo_ComfyUI SeargeSDXL"

REM Parcours de chaque dossier et mise à jour avec git pull
for %%f in (%folders%) do (
echo Mise à jour du dossier %%f...
cd %%f
git pull
cd ..
)
pause`

[uncut steeple]

I use comfy

[upbeat summit]

I've build myself a batch file... let me get it

[unreal plover]

@uncut steeple I tryed the auto installer, it failed !

[uncut steeple]

I do it with comfy manager

[unreal plover]

oh confuy, i was thinking about InvokAI

[lilac wren]

i already have x), but often, you don't know missing extension, so you use "missing custom nodes", and after, you don't know how they are called... painy

[uncut steeple]

comfy does it pretty well

[unreal plover]

What is the workflow o achieve this vertical @uncut steeple ?

[uncut steeple]

Its in the metadata to check

[uncut steeple]

where are the legs

[west breach]

he ate them, hungry

[royal fern]

enhance

[upbeat summit]

This is my super simple start batch file for ComfyUI on Windows. It asks Y/N to update ComfyUI and if you want to update all nodes under /custom_nodes/.

Nothing fancy - no error handling or anything.

• It requires that you are already in your python environment (could be added otherwise) and git is properly installed.

• If you press Enter it defaults to "No". Be aware that sometimes batch files do strange things and it will proceed with "Yes" instead anyway.

• You put it in your /ComfyUI/ folder

Please make backups. Use at your own risk.

[lilac wren]

Not bad, but i prefer mine x)

[upbeat summit]

Yeah of course. Just wanted to share because you asked how to update custom nodes.

[lilac wren]

yeah, ty anyway

[south frigate]

i think its worse yet cuz no loras and good models for it

[lilac wren]

@upbeat summit Still have this error with your workflow, everything is updated, i installed missing nodes with manager

[full idol]

Does controlNET work with SDXL in automatic 1111?

[lilac wren]

Cnet not works atm with SDXL

sadly

[full idol]

damn, when?

[lilac wren]

or there is just Depth

[sleek meadow]

Is gigabot going to update the FAQ to include information about SDXL 1.0?

[upbeat summit]

that is not mine

[lilac wren]

Don't know

[full idol]

Depth controlNET?

[lilac wren]

yeah

[lilac wren]

to be taken with a grain of salt, I saw that on the net

[lilac wren]

HMm i see

[upbeat summit]

but can you right click the missing node and see if it's actually called "Load Lora". that name isn't really descriptive

[zealous horizon]

what you can achieve with SDXL 1.0 from a LoRA trained on 1 image (3 zoom steps)
20 repeats, 10 epochs
Original Image via Unsplash

Used these crops

[upbeat summit]

impressive! thanks for sharing your research 🙂

[upbeat summit]

yeah I'm looking through different custom nodes. almost all packages have a load lora node but I think I know which one it is.. gimme a sec

[royal fern]

what level of denoise?

[upbeat summit]

I think it is the Load Lora node from https://github.com/WASasquatch/was-node-suite-comfyui the pins are matching

[lilac wren]

weird

[zealous horizon]

very basic Comfy setup

LoRA:
model strength: 0.8 - 1.0
CLIP strength: 1.0

Denoise: 1.0 (only using base model)

[upbeat summit]

it could still be a custom version, but the WAS node version does have the same pins. can you check if you have a node with the same name in comfyui? open the node search window (double left click and type load lora). if it doesn't show up, you might not have installed all requirements for WAS nodes

[royal fern]

oh okay I thought the zoom was img2img

[lilac wren]

I did, don't have

[chrome furnace]

hey peepz. did smth change with sdxl or comfyui? can get it to generate images :/
5 days ago i was able to create images.

[zealous horizon]

no. it is an automated preparation step for 1 image.

[lilac wren]

I should reinstall this one

[upbeat summit]

yeah try that

[chrome furnace]

i just have to run the "run_nvidia_gpu.bat" , right?

[lilac wren]

is it the good one @upbeat summit ?
https://github.com/WASasquatch/was-node-suite-comfyui

[unreal plover]

my PC nearly crashed for this image

upscale from 1080p should do a better job

[vivid tide]

Prompt? looks amazing!

[lilac wren]

Worked @upbeat summit

[unreal plover]

Octopath Traveler style,miniature,tilt-shift,voxel syle, a crown protesting,film,bokeh,professional,4k,highly detailed,lora:VoxelXL_v1:0.7

[upbeat summit]

Awesome - have fun!

[lilac wren]

This should probably explain the problems I had with other workflows

And guys, do you know a way do disable nodes without cuting cables?

I don't want to use this 3 lora loaders, but i don't want cut this cables.

[upbeat summit]

ctrl+m for muted. but the signal will not process further. but now there is a bypass option, but I haven't really tried it out

[lilac wren]

Ty

[upbeat summit]

muting a node will only work if it doesn't break the pipeline, so you have to try it out if it works or breaks anything

[lilac wren]

yeah

i have to do 2 roads

[royal fern]

ctrl-b never breaks the pipeline apparently

[upbeat summit]

yeah that how it should work

[royal fern]

ctrl-m would be great if it worked on reroutes

[upbeat summit]

yeah most are doing the jumper workaround

[unreal plover]

You can make youtube miniatures

POV FPS view of CSGO

[upbeat summit]

crazy

[royal fern]

you'll be able to do a full youtube channel soon enough

[stray wave]

Can it make GO-PRO images?

[royal fern]

can it make coffee?

[unreal plover]

GoPro give wide angle

[upbeat summit]

fisheye lens works too

[uncut steeple]

Thats pretty dope

[uncut steeple]

Those have an on/off switch

[unreal plover]

yes nice one

[late marsh]

Whooooooa 😮

[unreal plover]

i want to live in this world

[boreal bough]

cries in 400 regularization images

when a whole work day isn't enough to finish one lora on max speed on a rtx4090 XD

[unreal plover]

2h left

[boreal bough]

yep. doing anatomy + clothing + faces all in one lora is quite taxing ^^'

individually they'd take like 30min

[full idol]

do you find reg images help in training your sketch style lora?

and are your reg images then other sketch styles?

[little gazelle]

Has anyone had any weirdness when generating SDXL images where some colours look 'burnt'

[upbeat summit]

so you should probably get more than one

[boreal bough]

sketch style lora is my poster child for 'easy lora training' XD setup took less than 30 min, and trained in like 1hour

most likely I'll have to supplement it with around 600 high quality images of faces, to make it a 'perfect' lora

training time gonna be 2 days after that x_x

[upbeat summit]

what concept are you training?

[late marsh]

... 🥹

[boreal bough]

most complicated one
which is these combined in 1 lora:
A.) full body training - which while not nsfw, is functionally the same (relevant so that skin under clothing is shown correctly - since I'll be touching on anatomy words)
B.) clothing - which while very easy, gets hard in combination without other things, as it loves to overfit before other concepts such as face are even remotely trained
C.) faces - really really annoying to train without relying on overfitting. needs to be done in ratio of 6:1 with other concepts (a mistake I made in this lora, thinking it'd fine if I just scale up all datasets. hint hint, I was wrong.)

essentially this is a finetune like lora XD

[upbeat summit]

very interesting - thanks for explaining! because I'm a total noob at training this sounds like a complex endeavor. but looks like you got a plan.

[dusk mica]

insane workflow

[upbeat summit]

overfitting you say 😉

[boreal bough]

training is easy as long as you don't do anatomy. faces are easy as well if you're fine with overfitting (which is what all the models on civitai are doing right now)

[boreal bough]

XD

[upbeat summit]

yeah I'm planning to make a deep dive into training soon

Only made one SD 2.1 embed 😉

[boreal bough]

hit me up when you do it ^^ I'll guide you through your first one

[upbeat summit]

Awesome 🙂 thanks for the offer!

[west breach]

what do you think of this for a lora? balloon clothes

[boreal bough]

cough ... latex... cough

[upbeat summit]

very nice! you will have some fans 😄

[west breach]

well yeah I guess it would produce latex clothes too haha

[upbeat summit]

Balloon Clothing - Latex Edition

[boreal bough]

XD

[west breach]

I was thinking of calling it 'Talcum' 🤣

[median isle]

Testing out the Serj SDXL controlnet

[uncut steeple]

[west breach]

couple more from the training dataset I'm sloooowly building

[urban fjord]

Meanwhile I feel I spend a lot of time training if it takes 10 minutes to train.

[boreal bough]

please make more train loras XD it's amazing

[late marsh]

One more of this.. didn't know I could generate 2048x1152 with 3060 😅

[urban fjord]

People standing on trains? Or is there other trains you want?

[soft zealot]

@late marsh why wouldnt you be able to ona 3060?

this 2048x2048 cane from a 1080ti ;o)

[vagrant wasp]

where is sdxl git ?

[soft zealot]

why is sdxl git

[west breach]

I can run sdxl on my 3060 6gb

[uncut steeple]

who is sdxl git

[west breach]

git good

[soft zealot]

there is no git

[uncut steeple]

its on huggingface.co

[soft zealot]

you want the git? you couldn't handle the git!

[west breach]

heading over to git now

[soft zealot]

you checking it out?

[uncut steeple]

hes just cloning it

[soft zealot]

send in the clones, there have to be clones

[uncut steeple]