✨｜sdxl

[molten gull]

chatgpt fixed it 🙂 was a cuda cpu/gpu thingie

what might this here be:

Error occurred when executing FaceDetailer:

'Sam' object has no attribute 'is_auto_mode'

File "D:\AI\ComfyUI\execution.py", line 151, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
File "D:\AI\ComfyUI\execution.py", line 81, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
File "D:\AI\ComfyUI\execution.py", line 74, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
File "D:\AI\ComfyUI\custom_nodes\ComfyUI-Impact-Pack\modules\impact\impact_pack.py", line 985, in doit
enhanced_img, cropped_enhanced, cropped_enhanced_alpha, mask = FaceDetailer.enhance_face(
File "D:\AI\ComfyUI\custom_nodes\ComfyUI-Impact-Pack\modules\impact\impact_pack.py", line 953, in enhance_face
sam_mask = core.make_sam_mask(sam_model_opt, segs, image, sam_detection_hint, sam_dilation,
File "D:\AI\ComfyUI\custom_nodes\ComfyUI-Impact-Pack\modules\impact\core.py", line 235, in make_sam_mask
if sam_model.is_auto_mode:
File "C:\Users\malicor\AppData\Local\Programs\Python\Python310\lib\site-packages\torch\nn\modules\module.py", line 1614, in getattr
raise AttributeError("'{}' object has no attribute '{}'".format(

[twin leaf]

Thank you!

[west breach]

wrong node?

[molten gull]

thx

[ionic dragon]

@west breach what happens if we train face with 160 images and 60 images have getty watermark

[west breach]

dunno, try adding watermark to the caption with those images so it knows it's there?

[indigo carbon]

by the way, @visual glade- why can't I reproduce A1111 generations in ComfyUI? I even changed the code to use GPU ARG

[peak dove]

I have never used a LoRA before - but now I'm hooked 😄 Underwater LoRA

[potent flame]

alright seem pretty great. except obviously in low resolution

[thorny frost]

if i take a real photo (detected as 1 not_ai) and then pass it through img2img at a 0.3 denoising strength thehive detects it as a 0.908 SD image. If i do the same with a denoise strenght of 0.1 it is detected as a 0.989 non-AI. Both 0.3 and 0.1 are almost the same to the human eye. So it is not the image itself but some invisible artifact. I guess a photoshop filter could eliminate it making any AI image indetectable.

[arctic bloom]

you need a1111 style of interpetering tokens

take a lot at https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb

[zealous horizon]

Yeah. I think if you gaussian blur and sharpen afterwards all artifacts will be removed.

[thorny frost]

i am trying but it is clever than it looks. 😦

[winter raptor]

running a free colab notebook is so much faster than my computer, its insane

[thorny frost]

no way, you must literally destroy the image for it not detect it. Must be something more.

[dense chasm]

colab rola traing speed can up to 5-6 it/s with kohya

[thorny frost]

applying a filter whihc distort the image seems to work better. Blur/sharpen alone doesnt work

[peak dove]

I like this oil-painting look 🙂

[thorny frost]

It looks more like latte coffee to me. 😄

[peak dove]

Its the exact image of the OP filter in PS ...

[zealous horizon]

yeah. I tried and also had no success. It's also none of my concerns. I just found it interesting that it couldn't detect several of my gens.

[peak dove]

... but not filter coffee 😄

[winter raptor]

Is the process for training loras for SDXL different than for sd1.5?

[zealous horizon]

mainly different settings.

[thorny frost]

which makes it more remarkable

[peak dove]

I am using Underwater LoRA and SDXL Offset Example LoRA ... in tandem - a really soft and romantic result!!!

[zealous horizon]

but in general: "no differences in the process"

[winter raptor]

I love this

[peak dove]

2 x SDXL LoRAs

[ionic gulch]

refiner doesn't like asians, but i force him to like black woman:
https://i.imgur.com/vireOUl.jpg

[dense chasm]

you should lora train for realistic character,asian generation works well either

[ionic gulch]

it was more observation that base picks asian woman in many cases and refiner always overwrites them with caucasian woman, if you don't specify it. if you prompt for asian it keeps it, just from a neutral prompt i often noticed that base creates asians and refiner changes them. also noticed that refiner tends to make woman older than base.

[slender hinge]

Working on some Sci-Fi movie scenes

[dense chasm]

lora trained asian actress upscale with refine model,lol..

[trim orbit]

so it looks like sarge put out safetensor versions of the controlnet depth for sdxl. i was playing with the bin file in comfy but had little success. it's not a very accurate model i feel. anyone can play with it using a1111 now too. https://huggingface.co/SargeZT/controlnet-v1e-sdxl-depth/tree/main

[ionic gulch]

thanks for posting your workflow. i've integrated the face part in mine and experiment with it:
https://i.imgur.com/jU3ZebL.jpg

[vivid tide]

Oh lord have mercy. You havent heard of the workloe from Winston

[ionic gulch]

do you mean because i also fill the screen with nodes until all lines are hidden?

it's the best kind of layout imo. get all important elements on one screen with preview images. the scrolling and zooming is not fluid enough in comfyui imo, so this works best for me.

the face replacement part works quite well:

[upbeat summit]

with https://github.com/failfa-st/failfast-comfyui-extensions you can not only make straight links / cables but also hide them completely

[ionic gulch]

thanks. saw that in the past, but didn't find it again with a short search.

[trim orbit]

i'd rather have a real frontend instead of this new habit people have of creating tighter and less manageable node graphs. have you seen comfy box? i implore usability addicts to start building comfybox ui's

[trim orbit]

sweet thanks! i've been wondering how people have done this. i didn't know there were extensions as well as custom nodes. This could be a confusing structure down the road but i'll go with it

[ionic gulch]

@trim orbit i know there are a few of those solutions for a frontend, but from what i got from first impressions they need a little bit more refinement. but this is the way to go in the future imo.

[trim orbit]

comfybox sees few updates but i don't think it needs them very much given that comfyui is just so well maintained. what we need in this space is more people creating more custom uis for it. i'd be in there doing it but i lack attention span. something i'm working on.

https://github.com/space-nuko/ComfyBox it's also kind of tricky to install which might be holding people back. story of the year.

[upbeat summit]

yeah the QoL features by @zealous horizon are super helpful. now mandatory for my workflows. also square nodes by default. rounded corners look nice but they are a waste of space 😉

[trim orbit]

rounded corners are a huge waste of space. ty

[ionic dragon]

should I use this to train a lora

[trim orbit]

i like them still but it' so conflicting when i'm trying to tile nodes together

[trim orbit]

i haven't been. i think it needs newest bits&bytes which windows users won't have

[ionic dragon]

oh ok

lemme disable

[arctic bloom]

https://tenor.com/view/illusion-boom-squares-cubes-gif-15645228

[dense chasm]

--no_half_vae: Disable the half-precision (mixed-precision) VAE. VAE for SDXL seems to produce NaNs in some cases. This option is useful to avoid the NaNs.i choose no mixed_precision in the step of Bucketing and Latents Caching,cuz it will sometimes show NaN for some pre-trained images when i set fp16

[trim orbit]

gonna try playing with this safetensors depth model in a1111 a bit. i can't get comfy_controlnet_preprocessors nodes to install in my windows portable version. giving up and going to where preprocessors work instead.

[ionic dragon]

i used it to train my prev lora didnt have any issues

@dense chasm just concerned about half way loras

would the quality be the same?

[haughty oak]

https://twitter.com/PUBLIKFRUIT/status/1688215333104939008?s=20

PUBLIKFRUIT (@PUBLIKFRUIT)

Been toying with my first personally trained #LoRA in Stable Diffusion #SDXL. Diggin' the results so far.🍉

Twitter · Aug 6, 2023, 3:47 PM

[high skiff]

Might be on the cusp of another breakthrough for SDXL

this one being for realism portrait prompting

these images are the same core prompt, same seed, same setup, just altered negatives and a couple extra positives

not just a one off, either

[dense chasm]

it's just speed up for computation,Mixed precision training offers significant computational speedup by performing operations in half-precision format, anyway,you have to assign a proper GPU CUDA

[upbeat summit]

looks really good!!

[high skiff]

experimenting more

Need to narrow down to what changes matter the most

[ionic dragon]

so its way faster compared to full vae?

@high skiff any lora workflow?

[thorny frost]

from my crappy sketch made in rebelle to final result

[ionic gulch]

which extension allows you to edit node sizes by entering numeric values? i had it, but don't find it again after a reinstall.

[high skiff]

so far I am very happy with the textures and skin tones on most subjects

And you can dial it up for more detail/skin texture

[thorny frost]

this img2img proccess is much easier now in sdxl. With sd1 it tends to diverge much more and do crazy thing needing many more manual corrections

[high skiff]

I am rapidly coming for midjourney

[cinder aspen]

Tentacles still unbeatable by any of the AI art generators

Even with inpainting

SD or Firefly

[stray depot]

No workflow in png?

[cinder aspen]

DALL E as well

[high skiff]

no, not right now

[polar epoch]

How do i add commandline args to comfyui? There isn't any wewbui-user/sh for this one

[stray depot]

oh, sad times.

Beautiful either way

[runic lark]

Took me only 2 hours to realize that "Bokeh" in negative prompt ruined the pose of my character

[high skiff]

I have shared a ton of stuff with this community as of now, I think its reasonable I hold on to some things for myself, potentially even just temporarily

[stray depot]

Completely understand mate👍

I had a similar quality result in 0.9, but the prompt was as long as your arm.

I'll share in just a mo.

[zealous horizon]

that looks like something I made in CSS a couple of years ago 🙂

[ionic dragon]

@high skiff whats something that you think sd should improve?
i can prepare the dataset if someone's planning to finetune

[amber crown]

are teh bots using aws instances by any chance?

[high skiff]

AWS?

[amber crown]

it's weird how I can reliable snatch a free sagemaker lab only when the bots are in gpu cooldown mode

Amazon AWS

[high skiff]

ah, thats what I thought you meant

no idea

[stray depot]

Made with vanilla SDXL and some upscaler.

lol 81mb

[white geyser]

Could someone send two different pictures, one with sdxl other one with Midjourney ?

[ionic dragon]

how to resume it?

ran a1111 by mistake, and it got terminated

[smoky patrol]

Looks beautiful.

[zealous horizon]

if you have a saved lora/checkpoint you can continue from there. otherise "Don't mess around while training"

[soft zealot]

Top Gun 🙂

[ionic dragon]

no, a ckpt didnt save yet

so gotta run it again

[stray depot]

I did the same with RunDiff and Talaman and get very different results.

[smoky patrol]

What's your prompt?

[zealous horizon]

an AI detector tool me this is definitely not generated by AI (0.004 likely to be SD)

[high skiff]

those things are fake anyways lol

I have used tons

uplaod the same image a few times and suddenly the rating changes lol

[zealous horizon]

it detects other images with high accuracy but yeah. I don't trust those tools.

[soft zealot]

Unless the person is stood in front of you and you are taking the picture there is always a >0% chance of the image being fake

[stray depot]

SDXL has a certain way with eyes.

Overly illuminated iris.

[ionic gulch]

with the extensions recommended above i could improve my workflow further and make it look cleaner:
https://i.imgur.com/y0fmoaR.jpg

[peak dove]

Weirdly Sunday-ish - done in SDXL1.0 with 512 x 512 .pt (embeddings)

[true wing]

Quick question: In comfyUI how do I control the denoising strength in the KSamplerAdvanced? I just see the option to add noise or not

the basic KSampler has the normal denoise value so I'm not sure why the advanced doesn't

[high skiff]

holy shit SDXL slaps with black and white photography using my new prompting

[zealous horizon]

yeah. looks like he got slapped hard. sorry for the poor joke but that looks like abuse

[high skiff]

was trying for great depression style street photography

[peak dove]

A tad like Norman Rockwell! Nice.

[soft zealot]

I'm guessing whover that was just asking about the See with Text Node just daw this block in my workkflow and decided to avoid my wrath lol

[high skiff]

oh my god, I am onto something

[peak dove]

Yes, it has Dustbowl/Okies look about it!

[dusty turtle]

kinda creepy

[peak dove]

Blade Runner

[dusty turtle]

for some reason cars look much worse than still subjects, the lighting is just kinda off for some reason and I cant figure out why

[tender timber]

Has anyone been able to create a text file save node with a timestamp, one that specifically matches the timestamp of the image file too? I can time stamp the image via WAS nodes (because the base one does not allow for UNC paths, at least that I have been able to use and its built-in time variables are more limited). I create a token for the current time and use that as a variable and use the same variable for the text file save.

Problem is... the text file is saved as soon as the positive prompt is generated and the image obviously comes after and it looks like when the time variable is created, it's evaluated twice: once for the text file once for the image file. I've tried copying the variable to another string but that does not seem to work. I can't see a way to have the text file created simultaneously once the image is generated, or even after. Any suggestions?

[soft zealot]

In Fairness with a biot of work you could get 1.5 to turn out some good ones (this was froma series of models merged then with pictures of myself trained in)

[true wing]

I'm pretty sure LORAs only work on the model they were trained with, or at least work well. There are a couple of workflows for mixing SDXL with 1.5 models out there though, the simplest being just doing img2img from one to the other.

The only other ones I've found require custom ComfyUI nodes and I'm trying to keep my workflows as stock as possible

[winter raptor]

weird focus but still pretty cool

[soft zealot]

talke a look at my workflow , its down the bottom (workflow embedded in attached pic of main workflow)

[winter raptor]

holy workflow batman

[soft zealot]

ctually I just noyciiced dammit, its fractionally out

[soft zealot]

No Its not that was just a bad example I grabbed from where I cancelled something

[tender timber]

Thanks, is the workflow embedded in that PNG? I saved it but nothing loaded. I'll look for one of your earlier posts and grab it

[soft zealot]

thats just the tip of it lol

[winter raptor]

dear god

[winter raptor]

got a picture with it embedded by any chance lol

[soft zealot]

it should be if I did it right but if I didnt use this one

[tender timber]

That did the trick thanks. I'll go through it but in general though, how did you get the timestamps to match?

[soft zealot]

luck

[tender timber]

I feel your pain

[hardy cipher]

winston, that workflow needs the truck lora

[winter raptor]

time to open that up without the straight line mod, this is gonna be an abomination to the eyes

[tender timber]

One day I am going to make a tutorial specifically for people over 52 to show each node and what the damn inputs and outputs do. But even then, I'd probably run into issues like where the output of one should be able work with the input of another, but underlying invisible variables (like min/max values) mess it up

[soft zealot]

even I dont do that lol

[hardy cipher]

indeed, lol

if you really want to get wild open it up in linear

[winter raptor]

[true wing]

I've seen a couple of references to a ComfyUI wiki for that sort of thing but I don't think it's ready yet.

[tender timber]

Not at all ready.

[winter raptor]

@soft zealot do you have the link for the seed with text node?

[soft zealot]

Im 60 next March

[tender timber]

There's a mod out there that makes invisible nodes

[hardy cipher]

but is it older people that have a problem with comfy and nodes?

[tender timber]

53 exactly three months from today

[soft zealot]

theres always one youngster who cant bloody read lol

[hardy cipher]

seems like a problem all these pesky youngsters have

yes, exactly

[tender timber]

No, it's the documentation, not the concept. I was being flippant about age. I used to play with AmigaVision as a very first generation workflow object oriented presentation maker back before a lot of people here were born

[hardy cipher]

with their iphones and their doodads and gadgets all catered to their whims

[soft zealot]

technically thats invisible links, invisible nodes would leave a screen full of spaghetti only ~justsaying

[winter raptor]

ty, to be fair though I didn't see that in the view I had lol

[tender timber]

I just showed my age 🙂

[midnight shuttle]

A lot of open source software has very little documentation. You're supposed to learn by tinkering. But the problem with tinkering is that it's easy to draw incorrect conclusions that then get spread throughout the community as misinformation.

[soft zealot]

Odd , It should open by default to this view

[winter raptor]

maybe it works different in colab? My comfyui in it has been acting bizarre, especially with the nonsense workarounds I've had to use on it lol

[hardy cipher]

I only ask questions when my tinkering leads to a dead end

[soft zealot]

no idea, I just use my trusty old 1080ti 🙂

[tender timber]

Tinkering is fine, but when you get to "oh, it doesn't work because of this spot in the underlying code that I never told anyone about", then it's annoying. I can learn, I can muddle my way through programming if needed, just tell me what the things do

[midnight shuttle]

I ran into that problem with primitive nodes in Comfy. I thought it was a limitation of how many things could be connected, but it was really a hidden constraint where identical data types can only be connected if the min/max is also the same. But if I had made a video about it and stated my assumptions that misinformation could have spread in the community.

[hardy cipher]

I don't know, seems pretty impressive for a single person to make such a program anyway, and then to keep up with updates, dependencies, etc, different hardware, different software, different preferences, etc. so I don't know how anyone could expect it to be flawless in all aspects

[ionic gulch]

https://i.imgur.com/UV0VFAD.jpg

[midnight shuttle]

No software is flawless even if built by a huge professional team. I was only commenting on the problems that come from the idea of "learn by tinkering".

[hardy cipher]

no, I agere with you

agree

[midnight shuttle]

Or more spesifically the collision between "learn by tinkering" and "everyone making Youtube tutorials".

[hardy cipher]

what i'm saying is I see complaints from people about issues they run into with the software. it's one thing to discuss them. another thing to be a dick about it

[midnight shuttle]

Anyone who is that upset can learn Python and fix it themselves, 😆

[hardy cipher]

I've actually done that, lol. well to a degree

[tender timber]

As I said in another channel, my undergrad is in teaching, so I am biased. I also work with legal contracts for a living, so you have to be crystal clear. I just spent four months reducing risk because of a three letter typo which allowed an asset reseller to potentially claim full ownership of our scrap and sue our legal affiliate. So yes, documentation/instructions matter. We had fun in my undergrad teaching someone how to make a peanut butter sandwich where your partner had to do exactly and only what you said.

"OK, take out the bread" ... partner rips the bag with hands instead of undoing the bread tag.
"Put the peanut butter on the bread" ... partner picks up jar of peanut butter and puts it on the bread.

Hilarity ensues. Very eye-opening to get rid of underlying assumptions of what a person thinks are clear instructinos

[hardy cipher]

but I hadn't really ever even looked at python before messing with ai stuff

[tender timber]

I've resorted to that too

[peak dove]

Is Monty Python a good starting point? 😉

[hardy cipher]

I also realized that no one was going to come in and save me when I was encountering weird errors at times. so had to fix it on my own

[peak dove]

I think I'm implementing Monty Python 3.10.8 on my machine!

[tender timber]

Oh hey, I ran into that yesterday. I changed the min/max to match in the other node but it didn't work 😦

[hardy cipher]

what were you trying to do?

[midnight shuttle]

There might be another hidden constraint. I only learned about the mix/max because Comfy himself said it.

[hardy cipher]

I'm curious about the errors

[tender timber]

I was trying to connect a primitive to steps because I wanted the steps as a string for text input later. It then creates an INT output, but the INT to TEXT node from Aklemann can't use it because the primitive min/max in is 0/8192 and for the other it was 0/0xffffffffffffffff or something like that. Ended up doing this as a workaround

[cobalt fulcrum]

Hey guys, I have created a web app that relies heavily on SDXL and with the new model the prices of stability’s API are just not sustainable for me anymore (~$0.020/image) :/ I was thinking about deploying my own infra on aws, but not sure where to start from and if it’s going to be worth it in terms of cutting inference costs. Has anyone deployed SDXL on their own infra? Do you have any suggestions on how to go about that?

[midnight shuttle]

Only in the sense of deploying Comfy and A1111 in Docker, but never used them for API calls. Have you done the math about how much you spend on Stability API calls versus how much you will spend on computer, storage, bandwidth, etc?

[autumn forum]

hey winston do you have a link to that workflow of yours so i could try it out? or is it private?

[hardy cipher]

ahh, so basically just wanted to have that control steps as well as have it be printed out? or am I missing something?

[soft zealot]

It’s in the image #✨｜sdxl message

[autumn forum]

okay thanks!

[cobalt fulcrum]

I have read some articles from AWS, but most of them only talk about benchmarks done with the older SD models, there’s not much content yet about SDXL and I’m not very AWS-savvy to be able to run benchmarks on my own

[midnight shuttle]

Not benchmarks. Cost calculations. I think you will find that AWS is a lot more expensiev than you think.

[tender timber]

yup... print to text file. in a perfect world, the WAS npdes at the bottom would not be needed, but the primitive node won't go to the Aklemann node and the STEPS input of a sampler. Also WAS has a TEXT to NUMBER but not a TEXT to INT node, so I needed to do it in sequence

[drifting moon]

I keep having an issue with images disappearing once the generation is complete (SDXL - Auto1111), has anyone else experienced this?

[tender timber]

Are they being saved at all?

[drifting moon]

only the ugly 512x512 ones are

[drifting moon]

it shouldnt be memory issue with 10gb and an RTX 3080 right?

[cobalt fulcrum]

I just thought that considered that Stability is probably using AWS for their API and they’re probably making some profit on it, which means that it’s cheaper to run your own infra rather than use their API. But it’s all speculation, hence why I was asking for some advice😁

[tender timber]

No... but are you saving at 1024x1024?

[midnight shuttle]

That is an extreme oversimplifcation. They have the advantage of economy of scale. And also most AI companies do not make a profit, they survive on investment capital. Assume that even the lowest AWS GPU EC2 instance will probably be about 50 cents per hour. Are you doing more than 25 generations per hour on your system? Also you need to consider outgoing bandwidth costs which can add up.

[drifting moon]

oh does the finished result disappear if its not "saved"?

[tender timber]

Nope, it should save the 1024x1024. I am not using AUTO but it seems like only the intermediate image is being saved?

[hardy cipher]

I have an "Int2Any_AS" node installed. not sure if that would work. I guess it really doesn't matter.

I just like how it lets me connect string to text. but then tells me that isn't going to work out

[drifting moon]

I can see the image being generated in the preview but it disappears once the image is complete

[hardy cipher]

I've had that happen before

[cobalt fulcrum]

Makes total sense, I am quite ignorant on the matter, so thank you for clarifying!😅 I do have around 50 generations/hour, but not at all times of the day.
I really wonder if there is a way to deploy SDXL in a cost effective way that would keep the price at around $0.001/inference, ideally on demand

[midnight shuttle]

About how much are you spending on Stability API calls per month? If it depends, provide an average or approximate cost per month. If you aren't doing 50 per hour 24/7 you are already behind when it comes to hosting your own infrastructure. I don't know of any way to get $0.001/image.

[cobalt fulcrum]

I think i underestimated the complexity of deploying SDXL to the cloud.
I am not in a rush currently, as my users have not complained about pricing, but I hope in the near future to be able to find a proper solution to cut costs, at least to something similar to what SD 1.5 was

[midnight shuttle]

About how much are you spending per month on Stability calls?

[drifting moon]

so now it's saving in my folder but its not staying in the stable ui preview display

[cobalt fulcrum]

Around $90, but I launched my product exactly one month ago, so it’s just one month’s spend

[cobalt fulcrum]

Also consider that I have switched to SDXL 1.0 only the last week, before I was running on SDXL Beta, which was cheaper

[midnight shuttle]

In that case, using Stability API services is definitely the best choice. The cheapest underpowered AWS GPU instance is going to be over $500/month. And it will probably be slower than the API.

[cobalt fulcrum]

I appreciate your time in replying, I will then wait for higher traffic before considering a self hosted solution

[hardy cipher]

had a little tiling accident, guys

[inland vessel]

AM BACK

[hardy cipher]

good tiles gone bad

[peak dove]

Some more super-weird (512x512 embedding in SDXL 1024x1024)

[slender hinge]

Testing some of the different checkpoints on Civiti

[hardy cipher]

man, I get some weirdness when I forget to set the latent to 1024x1024 or a comparable size

I believe I'll add this lora to the mix and see if that improves anything. I'd suggest you do the same https://civitai.com/models/94842/truckmaker

[peak dove]

I like its artistic potential in spite of the mismatch 🙂

[hardy cipher]

yes, I agree. everyone doing the same thing gets boring after a while

which is why I will do my best to incorporate the truck loha into everything I do

[hardy cipher]

so much inbreeding, lol

[inland vessel]

https://civitai.com/images/1877932?collectionId=107&period=AllTime

how to make something like this?

[slender hinge]

yeah clearly unstable diffusers, PerfectDesign & Rhapsody are similar merges - Rundiffusion & sdvnRealXL seem to be unique

Sdvn seems to give most realistic results for people at least, but thats what it was made for.

[hardy cipher]

start with a video of someone dancing like that

then you use ai and magic

[inland vessel]

lololol

also

how do i know if this lora will work on SDXL?

[hardy cipher]

is it made for xl?

[inland vessel]

it says

other

[hardy cipher]

is that from that video?

if so it uses rev animated which is 1.5

quality checkpoint tbh

[inland vessel]

no it's another lora

https://civitai.com/images/848516?modelVersionId=75826&prioritizedUserIds=83602&period=AllTime&sort=Most+Reactions&limit=20

[hardy cipher]

look at the upload date

may 20

I actually downloaded that lora a cuople weeks ago. should try one day

[inland vessel]

oh yeah am stupid

tanks

https://civitai.com/models/58140?modelVersionId=62591

this one looks funny too

is for 1.5 tho

[hardy cipher]

I just want to figure out a solid workflow that'll combine the power of sdxl with the cool stuff in 1.5

I know you can use the refiner on 1.5 models, but meh

[inland vessel]

yeah true

[hardy cipher]

I wonder if it's possible to backtrace ai rendered images to their base images

[rigid laurel]

base images?

[hardy cipher]

well say you make something in image to image

using a picture you took

it's a base image, right? I don't know what else to call it

the first image, the initial image the before image

[midnight shuttle]

At most I think you would be able to show that the images are similar. But not prove or extract the original image.

[soft zealot]

might not be quite what you're looking for but a previous incarnation of my workflow had an SDXL to 1.5/2.1 ImG2IMG flow

[inland vessel]

Someone knows a good tut for lora training for my face for SDXL 1.0?

[hardy cipher]

google drive is a damn hassle. no matter what I do to combine two folders it secretly keeps creating two distinct folders. in the google drive app it said the're one folder, but then I look in colab at the directory and 2 folders, lol

now google drive has the two folders listed with identical names. but they're not identical names. one is "models" and one is "models (1)" why would they set it up like that?

[midnight shuttle]

Best to create a whole new folder with a new name and then put all items in there. The way GDrive handles folders on the back end is strange.

[hardy cipher]

I do believe you're right. it's just very perplexing. when I see things like that I think they must be intentional, especially for such a large operation and company. but then, why do it that way?

[untold goblet]

Why does SDXL sometimes start generating dogs... with no mention of dog in my prompt? 🐶

[midnight shuttle]

It's not intentional in the sense of trying to make life difficult. But GDrive has no concept of actual folders. Only object parents. And so sometimes you can end up with results that aren't intuitive if you are thinking of it like a traditional file system.

[midnight shuttle]

What was your exact prompt and negative prompt?

[hardy cipher]

yeah, that makes sense. I didn't assume it was to make my life more difficult. just a bit perplexing

on a positive note, for colab at least, I probably wasted about an hour or more trying to figure this out while connected to colab. so they got to squeeze a few cents from me through the whole ordeal

but like, there's literally nothing that seems to work. change the folder names. remove the folder.s copy them back in. still it persists. what a maddening experience.

[midnight shuttle]

Create a whole new folder. Copy each file into the new folder.

Or move the files. But only move files, not folders.

[hardy cipher]

I

[midnight shuttle]

And you can do this in the Google Drive web interface. You don't need to be in the collab.

[hardy cipher]

I'll do that. so stupid though

no I'm not doing it in colab

but kept thinking it wouldn't be this complicated. becaues there's no reason it should be this complicated

so kept reconnecting thinking it'd work right

and each time encounter the same issue

even changed files in the install to point to both folders, that didn't help, for reasons I don't quite understand

and then I can come at google drive from quite a few angles. windows explorer, gdrive app, the website, colab, and each on seems to interpret the world differently

[midnight shuttle]

The Google Drive web interface is probably the best source of truth. Everything else, as you said, is an interpretation. Trying to represent the GDrive architecture as a different architecture.

[south frigate]

best anime model for SDXL right now? tested smth but it was bad...

[hardy cipher]

have you tried this one? https://civitai.com/models/117259

[alpine pine]

Allegedly:

1. Take example anime image
2. Interrogate in A1111 using the “Vit-H” model
3. Take the resulting caption and use it as your prompt in SDXL base

And you’ll get close to the starting image without using a checkpoint

[summer flower]

Spend last night hacking together some string and LLM nodes for testing. still testing prompts and flows, but using it to generate a scene description and then a list of modifiers seems to give unique and varied results.

[hot rover]

Really want to try something like this but locally and contained inside of comfyui as much as possible. Love seeing LLM stuff incorporated like that.

[soft zealot]

have you tried vailla SXL and see what it can produce? This was deleiberately monchrome comic bookish

[summer flower]

Id be happy to release the custom nodes after cleanup and more testing. Just having string contatination nodes, template nodes (for inserting text at a specific location in a prompt), and a GPT node has been proving useful

[hardy cipher]

I feel like fairly small models would suffice for these purposes if they were trained right. just not sure if there are any that currently fit the criteria

[ionic gulch]

https://i.imgur.com/e68NMIT.jpg
https://i.imgur.com/YtcGqk5.jpg

[south frigate]

what is vit-h? newbie here in sd

[hardy cipher]

vision transformer model

https://theaisummer.com/vision-transformer/

sdxl uses vit-g and vit-l

lots of vits

[upbeat summit]

some images made with DynaVision XL alpha 0.2 - a SDXL model by @delicate kelp

[south frigate]

yeah, but how to use it to Interrogate?

[hardy cipher]

tbh with a lot of these things I've had to read the description about 20 times to wrap my head around it

because they've been entirely new concepts my brain never really considered

[south frigate]

nah, didnt test is yet, but if u say thats good will test

[hardy cipher]

you mean in general or in a specific application?

[south frigate]

in s&d for sdxl

[soft zealot]

colour me purple and call me kevin why wouldn'y ypou at least try the stock model before going down a rabbit hole?

[hardy cipher]

lol

[hardy cipher]

need to combine with a blip model and do magic

I don't actually know the specifics of that process. but you might be able to figure some things out here https://github.com/pharmapsychotic/clip-interrogator

I've been attempting to learn more about it myself. I believe combining interrogation with the image generation in different novel ways could improve the end result quality. I'm sure all sorts of people are looking into that idea though

[lusty wolf]

Where do I find this? " vailla SXL"

[hardy cipher]

valhalla sdxl is the default version

[soft zealot]

sorry typo.

Vanilla

on an ice cream cart ;o)

[hardy cipher]

you know, vanilla gets a bad rep but it's pretty quality

the ice cream

[soft zealot]

I like mint choc chip

[hardy cipher]

chocolate chip cookie dough is the money ice cream

[soft zealot]

how have I got to be 59 before realising that "tutti frutti" icecream is simplky all the fruits lol

[hardy cipher]

I didn't know that until right now,l ol

[soft zealot]

yes that

[royal python]

lovely clean images

[lusty wolf]

"Tutti Frutti" in cake format ...

[soft zealot]

nom nom

[hardy cipher]

I'm not quite sure what the tiling is doing in these images, but it's interesting

[grizzled warren]

SDXL ControlNet wen?
Any info?

[mossy canopy]

Sdxl anime finetunes are getting good..

[hardy cipher]

from what I understand it's pretty much ready to go but it's going to be resource prohibitive in that normal users probably won't have the vram required to run it

[autumn forum]

they are okay, still has a ways togo before they are 1.5 quality. however 1.0 anime finetunes are great at enhancing 1.5 tuned outputs very well

[spice skiff]

Just out of curiosity, if I find myself in the middle of a very long render (image generation) while using ComfyUI, and I decide I want to cancel it, how do I do that?

[hardy cipher]

view queue in the top right

then you'll see what's rendering, and what's queued. and you'll have the option to cancel there

[mossy canopy]

I'll have to check

[cyan crown]

which one do you prefer?

[hardy cipher]

tbh I've found it easier to just ctrl+C and just kill the program. usually faster for me

[spice skiff]

Copy that. Thank you.

[tender timber]

I had that problem randomly when I used AUTO. Seems to be a bug

[ionic gulch]

https://i.imgur.com/48ET2VC.jpg

[hardy cipher]

nice

does anyone know how to find the direct download path to civitai models?

[cinder wing]

Is img2img a lot slower on SDXL? I seems like on 1.5, img2img renders more quickly than txt2img, but on SDXL it's going like 5-10x slower. This is not using the refiner but continuing to use the same model as txt2img, currently CopaxRealistic

[hardy cipher]

I don't know how or why that would be the case. it's rendering the image the same way. just on a different configuration of pixel values

what size is the image?

[cinder wing]

1152 width, 896 height

Very kind of you to respond btw, appreciate it

[hardy cipher]

no problem. I run into all sorts of issues myself. maybe it's the sampler?

[cinder wing]

Maybe it is, I thought I was using the same sampler but I realize now I was not. txt2img was DPM++ 2M SDE Karras and img was DPM++ SDE Karras. I'll check.

[hardy cipher]

is this a1111?

[cinder wing]

Yep

Much faster now

Damn my eyes, it was the sampler

Apparently the SDE makes a big difference

[hardy cipher]

lol, well that's good. some of them do take longer. heun takes a long time. basically takes 2 steps for every one

I don't really know how the SDE makes a difference, it's a lot of stuff to learn. but definitely know the samplers have a pretty wide range of speeds

it's so many things to juggle at one time and it's all such new technology.. it's impressive these programs even work as well as they do

[cinder wing]

I'm extremely impressed SDXL is rendering as quickly as it is. I did not think these resolutions would be possible to work with at these speeds on my 3080ti after seeing SD1.5 chug along when doing higher res, but clearly they were able to make some amazing optimizations

[inland vessel]

can someone create something like this? I want a tattoo made by AI

[glad grove]

i can but not on SD

[inland vessel]

with what then?

[glad grove]

photoshop

[inland vessel]

would be cool if you had time haha I don't know if it would take long

[glad grove]

probably like 20 minutes or u can do it yourself too just search for halftone effect on youtube

[zealous horizon]

try using halftone pattern in your prompt (that's the name of that effect)

[inland vessel]

thanks guys

[glad grove]

🙏

[hardy cipher]

but it's already created?

use "silhouette" as well

[inland vessel]

I don't want exactely that haha

[hardy cipher]

you could use a decent interrogator to get more specific descriptive words

[steel mist]

A few other terms to try: scan lines, toner, photocopy, maybe even engraving with a low weight? I think you’re going to struggle with getting separation over the vertical bands if that’s a feature you want

[inland vessel]

ill try

[inland vessel]

yeah kinda like the vertical lines

[glad grove]

yea u wont get pixel perfect dots with SD but try to play around with prompt like "black and white"

[steel mist]

Can always throw it into photoshop after and bring out the halftones there via filter

[glad grove]

yea thats easier

[inland vessel]

woah already close

nice

prompt?

[steel mist]

assuming it's in the png but

extreme closeup of an eye, black and white toner, bad photocopy, halftone, strong vertical, white stripes

negative was greyscale

[inland vessel]

ima try a few batches thanks for the help all

[steel mist]

If you load it up in comfy there’s a bunch of random other crap in the node graph you can delete. Been trying to figure this crap out

[hardy cipher]

here's what bing said

" I would say that your image is an example of a black and white abstract style. This style uses shapes, lines, and patterns to create an eye that is not realistic, but rather expressive, symbolic, or conceptual. Your image has a vertical pattern of black lines and shapes that create a sense of chaos or disorder. The white shape in the center of the image could be interpreted as a bird or a dove, which could symbolize peace, freedom, or hope.

I think your image is very interesting and creative. It reminds me of some of the works by Pablo Picasso, who was a famous abstract artist. He used black and white to create powerful and expressive images, such as his famous painting Guernica, which depicts the horrors of war.

I hope you enjoyed this conversation. If you want me to create a similar image for you using my graphic art tool, just type #graphic_art(“your description”) and I will generate an image for you. 😊"

not very helpful tbh

[inland vessel]

the last one is pretty nice

im gonna try again cus my workflow makes something ultra realttic from it haha

[hardy cipher]

this is the tattoo you need

[inland vessel]

thats awesome

i dont think she can tattoo that tho

xd

[hardy cipher]

that was one of those happy accidents, lol

[inland vessel]

tmr morning i get the tattoo so i got to be quick haha

[hardy cipher]

I am pretty good in the ai eye department

let me see if I can make anything worthwhile

[inland vessel]

i want to get into the vibe of my profile picture/ the eye

[inland vessel]

got really lucky with that seed tbh

[steel mist]

Yeah I ran like 4-5 and that was the best. Didn’t really hunt around a ton though.

[hardy cipher]

might just fire up old a1111. that interrogator extension at least used to work pretty well

takes about 5 minutes on my next level hardware

[bold osprey]

how much difference does a lora workflow make in terms of performance (comfy)

[inland vessel]

this aesthetic i like too

but then black/ white

[slender coral]

Anyone have a good prompt for a movie title?

[hardy cipher]

loars will slow it down, kraibse. I think the amount depends ont he lora though

[zealous horizon]

If anyone is thinking about getting an AI tattoo it should definitely be this:

[hardy cipher]

read the comments here https://www.reddit.com/r/StableDiffusion/comments/159rio2/some_of_my_sdxl_experiments_with_prompts/

[inland vessel]

lol the buggy fingers have something nice

[bold osprey]

I'm currently at 10 mins/pic with an sdxl lora

[zealous horizon]

and if you're serious about it, this one is my best offer

[inland vessel]

LMAO

xD

[hardy cipher]

Prompt executed in 1836.97 seconds

I think I should probably trim down that workflow

that sd best upscaler or whatever it's called is a beast

[upbeat summit]

perfect

[hardy cipher]

alright, interrogating your eye image with VIT-L-14/openai. should have results within the hour

if it ever actually works. seems I need to update about 500 things first

[inland vessel]

lololol

its oke

dont have to

[hardy cipher]

nah, I'm curious now

gonna have it do a prompt now

I don't know exactly how "flavor chain" works, but it sure does take it's sweet time

[upbeat summit]

should probably get more armor before getting into battle 😉

[hardy cipher]

that would slow her down

[hardy cipher]

flavor chain still tasting things

[soft zealot]

call me weird but its the background scenery that does it for me

[upbeat summit]

the atmosphere and color grading is really nice. very cinematic

[soft zealot]

combination of three filters

[soft zealot]

which ends up like this

[upbeat summit]

nice! was just about to ask if you know the prompt build that is being constructed at the end - very cool 🙂

[soft zealot]

and the models used + filter names & samplers are in the file names (if opened in browser)

That text file is autogenerated everytime I generate an image

[upbeat summit]

do you use a specific node to generate the txt file?

I've build my own node setup for it and it works but the problem is that it's not always in sync with the date or a random / counter number to match the image's filename. I can see that it's the same issue with your setup. at least we both got the seed matched 😉

[soft zealot]

th issue is that the OOTB S&R functionality doesnt play nice with WAS Nodes and Vice Versa

[hardy cipher]

fancy lads

[upbeat summit]

ootb? and yeah I use WAS tokens and since you can't control how it's processed (the cascade), for example my image has the ID 00001 and the txt file gets 00002

sometimes it matches, but if you use a batch count on the queue, they way ComfyUI fills the queue it will be totally out of sync

[soft zealot]

Now I was originallyusing this section to generate Path & Fname for both the WAS text save & image save which gave everything the same timestamp

but I want to add in the smapler, scheduler etc and I havent figrured out a wat of doing that except with OOTB S&R

OOTB =Out Of The Box

[upbeat summit]

well you can use the native ComfyUI Save Image node and use save file formatting - but it's only supported by this node. this should be a global feature so every node can access S&R

https://blenderneko.github.io/ComfyUI-docs/Interface/SaveFileFormatting/

ah k

we are talking about the same thing 🙂 gotcha

[soft zealot]

SDXL/%date:yyyy-MM-dd%/%date:yyyyMMddhhmmss%%SWT.seed%%BML.ckpt_name%%STYLE0.style%%STYLE1.style%%STYLE2.style%%SPL.sampler_name%_%SPL.scheduler%_FFHRF

[upbeat summit]

I asked WAS about S&R since it's in every node's properties and he said since it's a "unsupported" front-end feature / "bug" it will probably not be included in WAS

[soft zealot]

I think I was on the same thread

[upbeat summit]

ah 🙂

[soft zealot]

the irony is that I can use a WAS Node and taake its S&R prioperty and use in in an OOTB image save fname

[upbeat summit]

yeah. S&R should be a global option for all nodes but many are doing their own syntax

I'm now trying a new workflow for metadata. I use MikeyNodes' AddMetaData node. You can create your own PNG chunk and I just extract that with exiftool and format my own textfile with the name I want. one extra step but I wrote a small script and it's just a shortcut in my image management tool (xnview)

[soft zealot]

I was on the same thrread and I see hes suggested something I need to try

[hot rover]

Hey Winston, is this what you are using for the prompt styler nodes or something else? https://github.com/twri/sdxl_prompt_styler

[soft zealot]

yup thats the one

[soft zealot]

Ok this works but agan th eissue uis you have to use all WAS nodes and I havent checked if they all support txt outputrs

[upbeat summit]

nice! thanks for sharing it. yeah, the value will be hard to get if something isn't a text string or has not output heh. I usually use one of the EVAL nodes from efficiency nodes to convert types.

also saw this earlier - I haven't tried it yet. this package has an Any Converter node: https://github.com/tudal/Hakkun-ComfyUI-nodes

@tender timber made me aware of it 🙂

[soft zealot]

FWIW Ive added an update to https://github.com/WASasquatch/was-node-suite-comfyui/discussions/129

[ionic gulch]

https://i.imgur.com/3FlM6Kk.jpg
https://i.imgur.com/aDUvxLj.jpg
https://i.imgur.com/keBOhKC.jpg
https://i.imgur.com/7Xs2PdA.jpg

[soft zealot]

would have to see if it accepts inline OOTB S&R format

but thats tomorrow, bedtime for me

[upbeat summit]

good idea! sleep well!

[tender timber]

I'm trying to follow the logic for the text file name and the image name. For Image you have:

SDXL/%date:yyyy-MM-dd%/%date:yyyy-MM-dd hh-mm-ss%_%SWT.seed%_%BML.ckpt_name%_%STYLE0.style%_%STYLE1.style%_%STYLE2.style%_%SPL.sampler_name%_%SPL.scheduler%_STD

And the FPATH variable for the text file is:

[time(%Y%m%d%H%M%S)]_[Seed]

So those two timestamps are somehow written at the same time for the text file and the image?

Also, I am using the WAS image saver so I have access to UNC paths and better timestamp variables such as three letter month acronyms and it barfs on the S&R name and date fields 😦

[soft zealot]

I#vr gone too bed but we're all caught in the loop between OOTB & Custom Nodes it would appear
https://github.com/WASasquatch/was-node-suite-comfyui/discussions/129

[tender timber]

So it would seem 😦

Have a good night and thanks for the workflow provided so I can examine it

[soft zealot]

this doesnt work either (worth a try) using #✨｜sdxl message

all it needs is for one Python Genius to figure oyt a Node that can take a S&R variable and pipe it out as formatted text................

[upbeat summit]

yeah and probably a node that generates a JOBID and saves the image + text file so they are in sync heh

[summer flower]

Mind sharing the code for that concat node? I initially struggled to get a node to accept an arbitrary number of inputs of the same type (text)

[upbeat summit]

it's just the WAS Text Concatenate node

[summer flower]

Thank you. I realize now seeing them expanded there are still multiple inputs. Appreciate it!

[fresh path]

wheres the 10 zillion loras for sdxl already

and how do I sort civitai by sdxl

[upbeat summit]

yeah I construct the text strings that end up in the text file with it. the line break option than formats it like:

Positive prompt (CLIP_G):
my prompt....

etc

It was one of my first ComfyUI experiments. It would make more sense to put that all in a node, but it was fun and I learned a couple things

[tender timber]

Use the filter

[soft zealot]

I know I said I was going to bed but...

There is this in WAS NOdes that saves everything to a JSON

[hardy cipher]

what about something like this?

[upbeat summit]

I think I totally forgot about the node. there are just too many haha

[tender timber]

@fresh path Like so...

[fresh path]

oh awesome thank you very much

I didnt see the filter..

[hardy cipher]

I was tardy to the party and have no idea what you guys are even doing with all this craziness

[tender timber]

Trying to save text files and image files with the same time stamps in the file name and the ability to save to a UNC path, in my case anyway

[upbeat summit]

building machines

[hardy cipher]

ahh

I just want to save them with seed number attached at least

[tribal lantern]

with all this effort with wicked complex nodes it seems simpler to patch the comfy engine itself to allow to save the meta data than to create a wicked mess of third part nodes 😂

[tender timber]

Well, it DOES save the data into the image, so there's gotta be a way of reading that into a file and parse it

[hardy cipher]

going down the node hole

[tender timber]

I will now tell people I don't like to stop being such a node hole

[hardy cipher]

spaghetti all the way down

[tender timber]

Who knew that Guns and Roses was so prophetic with their 1993 album? https://en.wikipedia.org/wiki/"The_Spaghetti_Incident%3F"

[soft zealot]

GThis almost works except it adds a different number to the txt file and also doesnt allow for adding all the other nice details I normally like to have inthe filnemae but if its of any help

no idea why the WAS Image Save doesnt like the [PATH} token that works inthe text save but hey ho

[upbeat summit]

So for today (who knows what's tomorrow) I've scratched the txt file idea. I can't get it in sync so the image and the text file have the same exact name - number counters get out of sync if you run batch counts, time doesn't work since the image and txt will not be created at the same time / second.

I tried for days.

so I'm using https://github.com/bash-j/mikey_nodes AddMetaData Node now. You can create your own PNG chunk. I then use exiftool on a shortcut in my image management tool:

exiftool.exe -T -png:my-metadata filename.png > filename.png.txt

The command is in a batch file and the filenames will be populated dynamically with the image's filename.

Exiftool will then output a .txt file with just the chunk that contains the ComfyUI data that I added through the AddMetaData node.

You can also output a XMP with exiftool

[tender timber]

Maybe the square bracket is throwing it off? I'm trying it without ... and also wonder if case matters

[soft zealot]

I entered it as [Path] (I think) that was more likely a tyo here lol

[tender timber]

That will work, still hoping for an in-application solution. Of course we don't need the text files, it's just a want 🙂

[upbeat summit]

yeah but it's my new workaround - lets see where it falls flat, but right now it's looking good to me

[soft zealot]

I just like to have th eapremters in a nice easy to read plain txt file noit buried away in a JSON etc

[upbeat summit]

yeah me too that's why I build all that

[soft zealot]

especially handy if I set everything to random lol

[upbeat summit]

I now get a nice textfile from the json data only containing the data I want from the image gen 🙂

but it's not all in comfy - so it takes a bit of a setup

[tender timber]

Success with the filenames for image and text having the same timestamp and saving to a UNC path via WAS image save! Had to remove the square brackets from the vraible name, so [Fname] became fname

[upbeat summit]

now run a queue with a batch count higher than 1

try 10 images and see if it stays in sync

[tender timber]

Aye aye cap'n!

[upbeat summit]

maybe you figured it out. I tried many different approaches, but the possibilities also seem to be endless 😄

[tender timber]

I just used @soft zealot's layout, subbed out the image save from the WAS one, and removed the square brackets from the variables. In looking at Winston's code, I had something very similar for text file save but for some reason, it would save the text filename first and reevaluate the time for the image. I bet it's a slightly different node that I had used, but God knows which one and where in the workflow

[hardy cipher]

spaghet

very realistic image there

[cedar tusk]

Hey, that's my node pack!

[upbeat summit]

I need some parmesan with that

[tender timber]

Howdy...

[cedar tusk]

Did it work for getting the text data out?

[hardy cipher]

the spaghetti caverns

[tender timber]

I have to comment out the deepface import call in the python file to start it up

[cedar tusk]

Yeah I pushed a commit today to fix that. Forgot that I didn't comment that out even though the node itself was

[hardy cipher]

which node pack is that?

[cedar tusk]

It's unnerving how some of those look like worms

Does Comfy have a clip interrogator? I see an extension for booru interrogation but is there a non-booru one?

[hardy cipher]

yeah, but I haven't got it to work properly

[tender timber]

and I can't recall why, but it didn't quite do what was needed, or I got distracted. I'll have to look, but perhaps it was that I needed to add more into the prompt text file and needed a timestamp as part of the name

[hardy cipher]

I don't even get how the booru stuff works really. I mean, I understand the words, but not how they're actually effective

[tender timber]

@cedar tusk if you can figure this out, you will be worshiped 🙂 #✨｜sdxl message

[cedar tusk]

Yeah, there's a lot of info that doesn't carry through, like CFG and steps count.

[hardy cipher]

he's been down in the caverns so long he's turning into gollum

[cedar tusk]

You need style, sampler, and scheduler in the name, is that right?

[tender timber]

Success! This is on a pokey AMD Ryzen 1700 with 16GB of RAM sending the saved files to a QNAP across Gigabit Ethernet, with no stoppage of other tasks (in fact a batch file runs to backup the QNAP to a local external HD and then to OneDrive every ten minutes...)

[hardy cipher]

just put "average comfy user" on that image there and you'll win the stable diffusion subreddit

nice work, dude

[upbeat summit]

that image is so good 😄

[upbeat summit]

awesome - now I need to know how you did it 😉 I had mine running a bit and it looked okay and at some point it just got out of sync, but this looks good! nice

[tender timber]

Me? No. I wanted part of the positive prompt in the filename, which I've figured. i wanted a text file that has some of metadata (but not all) and the text file should have an identical filename to the image output include the timestamp of the filename. Looks like I've got kludgy success at the moment 🙂

[upbeat summit]

ah - I remember it did work quite stable for some time using the Save Image node from ComfyUI, but it did not stay in sync with the WAS Save Image node

[vast galleon]

Got color correction working. woohoo! Only took all day.

[cedar tusk]

nice!

[tender timber]

This is with the WAS node. See the images for the changes I made to the Winston workflow

[upbeat summit]

I need to inspect your screenshot 🙂 great job!

[upbeat summit]

Thanks, Biff. I will check it out 🙂

[tender timber]

DAD?!?!?

[vast galleon]

Thank you! I had issues knowing which data type was compatible for comfyui.

[autumn forum]

yall thing it would be possible to convert a lora from 1.5 to sdxl? or is that impossible? kind of a dumb question but im guess not sense so much is different with the encoders and such

[upbeat summit]

yeah I'll bet. looks like a really nice solution. I need to give you workflow a spin

[upbeat summit]

I tested this with TI embeds. you can use 1.5 embeddings in the SDXL clip_l textencoder, but it won't give you the results you want, since it was trained on a total different latent space.

[autumn forum]

okay so its the latent space that causes the trouble. i see

[vast galleon]

It's almost ready. The issues today put me behind, but it should be smooth sailing now. Only problem is I won't have time to work on it much tomorrow.

[upbeat summit]

yeah simply said it's build totally different, so stuff is in other places.

[upbeat summit]

yeah it looks really exciting. like the ultimate ComfyUI in-painting workflow

[vast galleon]

That's the plan. Trying to eliminate as many nodes as I can

[vast galleon]

The beautiful thing about this color correction method is that it can be controlled by adding a factor which I'm about to implement. Then you will be able to adjust the amount of color correction you want.

[upbeat summit]

that sounds intuitive

[tender timber]

Oh SNAP! ... all the seeds are the same for the batch of ten. That's dumb.

[upbeat summit]

there's always something

[tender timber]

It's Mr. Winston 🙂

[upbeat summit]

well it's set to fixed

which I do a lot since we have to explore what each setting does:D

[vast galleon]

color correction changes scaled by a factor of .5 from .5 - 1.5

[tender timber]

Does anyone know how the Winston (or other workflows) use Lexica random prompts? It seems to be a WAS node, but where does it grab data from? Is it local or does it go out to a dictionary online?

[vast galleon]

can't go too high with it though

[tender timber]

Somewhere in India, a marketing rep for skin whitening cream is salivating

[fresh path]

first female oompa loompa

[vast galleon]

doesn't work on multiple masks yet. need to tweak some things.

[strong briar]

I cannot get SDXL to work at all. AT.ALL. I have a 2070 Super gpu, so it should be minimally OK to run. Out of memory every time on the simplest of Generate. Is there a tweak to the .bat or an extension I'm missing?
Here's the error msg:
OutOfMemoryError: CUDA out of memory. Tried to allocate 40.00 MiB (GPU 0; 8.00 GiB total capacity; 6.99 GiB already allocated; 0 bytes free; 7.27 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

And as far as that refiner is concerned? How the heck do I use it in AUTO1111 anyway?

[heady vale]

A1111 is the problem. Use ComfyUI

[strong briar]

Yes, but is there a fix in A1111? Comfy confuses the hell out of me.

[heady vale]

Not yet that I know of

[strong briar]

So, how are all the YouTubers using SXDL in A1111?

But back to my question instead of just suggesting I use a different UI. Is there or are there cmd lines for the .bat or an extension I'm not seeing that help with memory allocation?

[heady vale]

I would guess they arent using an old 2070. A1111 is a massive vram hog with XL

[strong briar]

Anyone else on familiar with A1111 and getting SXDL to run? I saw a post that mentioned something about an extension called "dev" that helped with memory crashes. I have no idea what they meant.

[heady vale]

I can get it to run, it just uses more than 10gb or vram on my gpu and nearly 30gb of system ram

[heady vale]

with comfy its 6-8gb vram and 20gb system ram

[tribal jackal]

is the latent composite node broken? It seems to be reading the samples_to input, but no matter what setting I change I can't get the samples_from input to affect the output.

[upbeat summit]

I was only gone for a couple of minutes. the story is getting interesting

[tribal jackal]

https://blenderneko.github.io/ComfyUI-docs/Core Nodes/Latent/LatentComposite/?h=composite

I hope I'm just bad lol and it actually works

[glad dust]

this is interesting

[strong briar]

OK, so with adding "--medvram --no-half-vae" to the "set COMMANDLINE_ARGS=" I finally got SDXL 1.0 to work for me. Alrighty then... for the life of me and despite prompting (3girls:1.5) it cannot seems to generate more than 1 girl. What is up with that?

[hardy cipher]

lol

[strong briar]

It's giving me lovely images of a single viking girl in a forest at night, but just won't produce more than that. Any suggestions?

[heady vale]

use natural language like, three girls

[strong briar]

I thought 3girls WAS the lingo. Is it not?

[heady vale]

thats 1.5 and 2.1 prompting

[hardy cipher]

you can go full booro if you choose, but I don't think that's optimal

[strong briar]

Got it. Thanks. Diving back in. 😄

[hardy cipher]

if you're using xl just try natural language

[strong briar]

Better... it gave me two this time. lol

Yes indeed, using xl.

[hardy cipher]

just say, "3 viking women" or something along those lines

1.5 used vit-l. xl uses vit-g and vit-l. vit-g is newer and more capable of understanding natural language

[strong briar]

In English, Carter...

[hardy cipher]

model known english better

[strong briar]

*sg1 ref lol

[hardy cipher]

it understands sentences

[strong briar]

ahso. I see I'm going to have a little fun tonight playing around with this.

[strong briar]

Second question. I downloaded the refiner as well and dropped it into the models folder where the xl also is. In A1111... how do I use it?

[hardy cipher]

I do not know

[heady vale]

img2img

[hardy cipher]

basically it should be around a 4 to 1 ratio I believe, base to refiner. so if you run the base for 20, run the refiner for 5 at the end

that's not really set in stone, you can adjust as you want. but you don't want it to be 50/50 or anything

at least under normal circumstances

come join us in the noodle caverns, crushin. comfy is what you need

I've started quite a collection of sdxl loras over here

[heady vale]

Ive only played a little bit with the offset example lora so far

[urban fjord]

Download the refiner extension for Auto1111 if you want it to work like in ComfyUI with Sytan's workflow.

[hardy cipher]

some of them are really top notch

just downloaded a couple new checkpoints as well

[vast galleon]

I was able to get color correction working with multiple faces.

[heady vale]

crazy stuff man. nice one

just need to fix the AR

[vast galleon]

Thanks. I think it's a good stopping point for the day.

AR?

[heady vale]

aspect ratio

[vast galleon]

yeah I had it scaled to 1024x1024 to test. let me take that off and see what happens.

[hardy cipher]

my goodness, that's a bit of a noodle nightmare, but mine always end up looking like that too

[crisp owl]

My testing charts look like that, once finalized, that's when I organize

[hardy cipher]

yeah, when you're on the move you don't want to take the time to organize everything. it makes sense

[crisp owl]

Add "nope that broke it", delete. Rinse and repeat lol

[vast galleon]

works with original aspect ratio.

[hardy cipher]

I tend to organize when I start to forget what's what. or when I have to zoom way out to connect things

[heady vale]

awesome

[hardy cipher]

I need to get up to speed on how all that stuff even works. literally never messed with doing that sort of thing. at least not in stable diffusion

[vast galleon]

and there's my stopping point for the day lol

[tribal magnet]

any suggestions on how to clarify/separate the color lavender from the flower? I'm trying to get lavender eyes and hair, not... actual lavender. lol. (I HAVE tried "light purple", but that's not quite the color i want. (and sometimes i get "purple light" instead.

[hardy cipher]

try this https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#break-keyword

well

I guess that a1111, so not sure if that's universal or what

also, put flowers in the negative

[tribal magnet]

ive got a few breaks in there, but maybe ill try putting one immediately before it, lol

[strong briar]

I've pulled a generated image into img2img (A1111), but I don't see where to use the refiner. Am I blind?

[hardy cipher]

don't think any of the methods are foolproof

[heady vale]

just change the model to the refiner

[hardy cipher]

as I understand it, it's fairly tedious in a1111. seems they dropped the ball with xl

[tribal magnet]

oh jeez, yeah thanks. id been trying to avoid negatives since xl works well without them, but id kind've forgotten that was a thing

[strong briar]

Ah, I see.

[tribal magnet]

i may have misreplied

[hardy cipher]

I always negative

I knew what you were replying to

negatives don't always work as expected, but I believe for stylistic type things they're a big help

[ionic gulch]

if you want a1111 feeling but better sdxl support use sd.next - it's a fork but they worked much earlier on sdxl/diffuser support and the sdxl solution is quite different internally compared to a1111.

[hardy cipher]

but I don't think you can really use them to create a whole negative scene or anything. or you can, but results will vary

[tribal magnet]

yeah, that worked, hahah

[hardy cipher]

nice

[tribal magnet]

tysm

[hardy cipher]

I've experimented so much with prompts. take it weird directions just to see what it makes

[viscid pelican]

From a dream I had last night. Cant quite get the details right. Also cant tell if I'm going crazy or just playing with SD too much before bed.

[tribal magnet]

literally my new hobby since i discovered dynamic prompts: fill a prompt with so many possible combinations, run 1000 images while i sleep... mext evening just check out what i got. way better than tv. and i learn to make better prompts and better art.

[tribal magnet]

i love that

[hardy cipher]

nice, yes, it just keeps going deeper too

[drifting moon]

Anyone else experience issues with the live preview box with SDXL (automatic1111)

[gloomy barn]

what issue?

[drifting moon]

I can see the live preview of an image but it disappears when the generation is complete

[gloomy barn]

Ah

[tribal magnet]

also fun to make RPG character portraits: put all fantasy races, human ethnicities, eye colors, skin colors, hair colors and styles, classes, types of armor, various environments and lighting styles... and i have a few high-CHA parties the next day XD

[gloomy barn]

is it saving?

[drifting moon]

Yeah but only after I added a few arguments

[tribal magnet]

theres a setting... somewhere to keep some of the iterations, i believe... or at least one to keep it before adetailer or upresing

[hardy cipher]

what's a high-CHA party?

[drifting moon]

I can preview sd1.5 and save 1.5

[tribal magnet]

high charisma, aka very pretty hahah

[hardy cipher]

ahhh lol, yeah. I haven't messed with making RPG characters all that much. maybe to a degree

[green egret]

hi guys, i need help. I am new to SDXL in A1111, I just downloaded it and started to running on my Macbook, and I only see one model in my stable diffusion checkpoints. but I have downloaded other models like SDXL 1.0 and these have not shown. how can I get it updated? thanks in advance.

these are the models i downloaded and place it into stable diffusion file, but they are not showing on the stable diffusion checkpoint

[hardy cipher]

click the little blue refresh button there

[honest marsh]

im also new, but i have a new error that keeps the ui at 95% and it gets stuck but in the cmd apears completed and i have to restart the application if i want to use it again (it will fail again)

[hardy cipher]

that's why most people aren't using a1111 for sdxl

lots of different things like that

[honest marsh]

and what should I use

[green egret]

i clicked and it still stays the same

[hardy cipher]

I'm using comfyui, it has a bit of a learning curve if you are new, but there are lots of resources out there and templates you can use that make it plug and play pretty much

[hardy cipher]

I don't know, man. make sure it's the same folder that the program is looking at. restart the program. update. could be lots of things

[honest marsh]

the problem with that error is that only happens about 70% of the time

i can generate the same thing 4 times and 1 will do it correctly

[hardy cipher]

ghasty3. well, not sure what to tell you. one thing that is convenient is if you do decide to use comfy, yo ucan drag and drop any of your images into the UI and it will automatically load all the settings for that image

if everything is up to date and you aren't getting any errors when loading the program, I don't know. maybe go to the extensions tab and try to update

[honest marsh]

okay, tomorrow i will try to install it

[tribal magnet]

i'm so happy with this.

[hardy cipher]

there are other options as well. but I'm not well versed on all of them

just went from a1111 to comfy

[tribal magnet]

that IS a tough transition, i still cant do it

[hardy cipher]

my brain is satisfied by weird things, so picked the nodes up quickly

[honest marsh]

i know nothing about programming, just casual girl that likes art, is that too much difficult?

[hardy cipher]

certainly wouldn't call myself good at it or anything, but I guess that's subjective

you don't need to be a programmer

I'm not

[wicked frigate]

lot easier if you use StableSwarmUI

[hardy cipher]

but I have learned some python over time just out of having to unbreak stuff

[wicked frigate]

easy to install, has the easy interface like auto, but also exposes the comfy node interface to play with when you're wanting to do more advanced things

[hardy cipher]

more a1111 than comfy tbh

yeah, I need to try that out. I think I have it installed

[honest marsh]

also, is it supposed to be 50 times slower sdxl than 1.5?

[hardy cipher]

no

it is quite a bit slower though

[honest marsh]

i have to wait like 20 minutes for one single image, with a 2060 graphics card

[wicked frigate]

50x, no. But the scale up from 512x512 to 1024x1024 tends to be 5-10x slower

... 20 minutes? 0.0

[hardy cipher]

I only have 6gb vram but I'm getting abotu 1.75 it/second. not really something to brag about, but I can deal with it

[wicked frigate]

oh, 6 GiB VRAM... uh, yeah, use Swarm lol

[hardy cipher]

that number drops when I get too crazy with things

[wicked frigate]

auto struggles to run SDXL well, especially on low-end hardware

[honest marsh]

im getting 0.13it/second

[wicked frigate]

Swarm uses comfy backend which can run a lot better on weak hardware

[hardy cipher]

nah, I'm not on a11111, lol, that's the other guy. but he has a 2060 which is slowe than me

[wicked frigate]

oh lol that was coincidence actually, i googled how much vram a 2060 has, didn't even see your message right next to mine oops

[hardy cipher]

I realize now I should have opted for the bigger video card, but who knew?

[shell wave]

Since I can't run SDXL. I'll just ask someone else opinion. Is SDXL worth it? Or is SD 1.5 still the king?

[wicked frigate]

yes SDXL is worth it. Why can't you run it?

[honest marsh]

is worth it even for me that i have to wait 20 minutes

[hardy cipher]

1.5 is still as good as it ever was though. but xl is a step up

[shell wave]

GTX 1660 Super 6GB Ram. Out of memory 😦
That is in Automatic1111. In comfyUI it takes so long to generate 1 image around 30-35 minutes per image.

[wicked frigate]

ooo, 16xx :(

it, uh, might be possible in the semi-near future? No promises, but optimizations always happen in the period after a new thingy was released

[shell wave]

I hope so 🥺

[wicked frigate]

there's some exciting experiments going on with SDXL optimization that might work out

several different places trying to quantize it

ps at everyone in the channel on small graphics cards rn: an RTX 3060 is only a few hundred bucks and has 12 GiB VRAM. If you're able to afford an upgrade, that's a really good price vs capability there

$250-$300 depending on brand n stuff, probably even less used

[hardy cipher]

well I have a 6gb 3060. 🫤

it works, just not that fast

[autumn forum]

my moviestillxl lora is getting a much needed update

[hardy cipher]

these are just screenshots of movies, bud

[heady vale]

looks great

[glad grove]

now we finally live in a society

[autumn forum]

thanks!

[hardy cipher]

yeah, that's pretty top notch, I must say

where do I acquire this movie lora?

[autumn forum]

ill let you know, still collecting some shots for the post and ill upload to civit, theres one on there now but its nowhere near the quality of this version

[hardy cipher]

nice. well it'd be an excellent addition to my lora collection

[autumn forum]

also to note its best used with upscale and a base pass without the lora, thats how im able to achieve this quality. the lora itself can sometimes be a bit blurry due to the nature of movie stills online, but if you do another base pass no lora it turns out great.

[hardy cipher]

ahh, interesting. hadn't really considered that approach. good to know

this is a pretty quality lora https://civitai.com/models/119667?modelVersionId=130038

[autumn forum]

but look at the size . mine is only 90mb

but it is a quality lora for sure

[hardy cipher]

yeah, lol. huge. I wonder why?

https://civitai.com/models/122567?modelVersionId=133503

also that one, it's medium sized I guess

[autumn forum]

network rank dim set to 64-128

depending on dataset and amount you want to learn you can use more but 64-128 is just wayyy to high. maybe if you have like.. thousands of pics but anywhere between 100-1000 dim 12 ish is fine

[hardy cipher]

oh yeah, I should have known that I guess. I really haven't trained many loras. mostly just did dreambooth models and then later extracted the loras from those

[autumn forum]

nice

[midnight shuttle]

Rank dimension, that's the number of weights per layer right? Or does that even apply to this type of model?

[hardy cipher]

I actually learned a lot from that endeavor because I could not figure out how to extract them for the longest time. I knew it was possible, but just kept running into hiccups

what is the go to method for training these loras?

[autumn forum]

caith had a guide somewhere, ive found good results with adafactor 0.001 LR 30 epoch, 1-5 repeats depending on amount of images. but take my word with a grain of salt. i havent done much testing

[boreal bough]

kaibio went into the dim research a lot deeper. 24 was where he hit the limit of improvements

[autumn forum]

24 epoch?

[boreal bough]

for 90% of my loras, dim 8 was just right.
doing a lora with an endgoal of around 50k images (currently at 3k), which I'll be doing at dim 64, since I plan on merging it with base in the future

but anything that is under 5k images, should stay between dim 8~24

[boreal bough]

dim 24/alpha 1

[autumn forum]

yeah i trained this lora with 400 images with dim 12 alpha 1 and it got everything well.

[hardy cipher]

layers, dimensions, I need to familiarize myself more

just don't have the hardware to be training anything like that myself. I could do it online of course

do you guys ever run loras on the negative, or on the refiner? I can't figure out if this is something people normally do

[boreal bough]

fyi, dim 128 and up, usually does significant damage to the sdxl model - unless you both know what you're doing, and want the base model to 'do less'

[hardy cipher]

interesting

I did not know that. I guess it just reaches too far in?

[boreal bough]

could be relevant if you have genuinely bad training data (like 2 images), and only need to make it work barely, so you can generate more training data, with which you'll then train a proper lora

[hardy cipher]

oh, that's something else I hadn't considered. so you can train a huge lora on so-so data, get better data from that, and make a more efficient lora?

[dense chasm]

@wicked frigate kohya is good,especially when running with bucketing and latent cache,with xformer optimization,fp16 mixed precision,save VRAM highly

[boreal bough]

not really my style, but yeah it definitely works ^^

[peak dove]

LoRAs and .pt embeddings - do they do much the same job at all?

[midnight shuttle]

I don't think so. AFAIK embeddings are used to train the existing model to understand new concepts based only existing model weights. But LoRA introduces a new set of weights to the model.

[peak dove]

But they end up doing the same things - old weights or new weights?

Some weirdness from yesterday using 512 x 512 .pt embeddings in 1024 x 1024 SDXL1.0

[midnight shuttle]

I think an embedding is much less flexible since it can only "pull" the existing model toward the concept or character. But the LoRA is an additional model with its own information. But it's not as much of a full model as a hypernetwork.

[peak dove]

OK, I will setup a LoRA, HypNtwk and Embed and make my own comparisons ... ! 🙂

But they seem to do similar work

The sheer detail in this one is mind-blowing!

[midnight shuttle]

If I understand it right, embedding is the least powerful and not used very much anymore. LoRA is more powerful and you can train it about a particualr character or a style and you need an activation word. Hypernetwork is the most powerful because you train the model the same way you train a full model (only prompts and images) but these are not popular.

[autumn forum]

https://civitai.com/models/120845?modelVersionId=135512

[peak dove]

Thank you for your insight 🙂

[midnight shuttle]

Embedding is also trained with an activation word that doesn't appear in the model. So you would use a nonsense word like "lsr100" to represent the embedding. And then train the embedding with prompts containing that word such as "a photo of lsr100" or "lsr100 wearing a top hat".

[heady vale]

embeddings were used a lot in 2.1 for styles

[peak dove]

I still like them ...

[heady vale]

I used them all the time. great little light weight tools

[peak dove]

I have d/loaded an SDXL Papercut model at Civitai - does this go in place of the SDXL Base? Or is it added on top of the SDXL Bse?

[alpine pine]

Is it a checkpoint?

[crisp owl]

Believe he means LoRa

add a lora loader node inbetween your checkpoint loader and your ksampler. same with the clip and your prompts

[peak dove]

Sorry, its a LoRA - thank you,! 🙂

Basquiat and Rasuchenberg flavoured ...

[soft zealot]

oh yeah , I used fixed seeds a lot when testing lol

[soft zealot]

its an API call out to Lexica from looking at the source code

[tribal jackal]

I think I'm gonna redo that in 32:9 and upscale again. New wallpaper

[soft zealot]

Although it would maybe have been more fun has you loaded and run this as I have no idea what you would get out.
Prompt = Random
Master Style =Random
Sampler = Random
Scheduler =Random
Seed = Random

It's all just randomness lol

@upbeat summit @tender timber picking up from yesterday

https://github.com/comfyanonymous/ComfyUI/issues/1057

[tender timber]

I used to do that with wildcards all the time with SD2.1, will get started with SDXL once I get more comfortable with it

[tender timber]

Ah, from WAS or from Comfy? I think it’s WAS if I am not mistaken

[soft zealot]

from WAS

[tender timber]

Nice, I hope it gets traction

[soft zealot]

TBH I'm not holding my breath , comfy hasnt commented and wasasquatch has referred me back to comfy /

Meanwhile here I am stuck in the middle with you lol

[tribal jackal]

so detailed it almost looks like artifacts

[soft zealot]

ahhh , only bringing back some, maybe I need to "rweak" that lol

---

`# Random Prompt

class WAS_Text_Random_Prompt:
def init(self):
pass

@classmethod
def INPUT_TYPES(cls):
    return {
        "required": {
            "search_seed": ("STRING", {"multiline": False}),
        }
    }
    
@classmethod
def IS_CHANGED(cls, **kwargs):
    return float("NaN")

RETURN_TYPES = (TEXT_TYPE,)
FUNCTION = "random_prompt"

CATEGORY = "WAS Suite/Text"

def random_prompt(self, search_seed=None):
    if search_seed in ['', ' ']:
        search_seed = None
    return (self.search_lexica_art(search_seed), )
    
def search_lexica_art(self, query=None):
    if not query:
        query = random.choice(["portrait","landscape","anime","superhero","animal","nature","scenery"])
    url = f"https://lexica.art/api/v1/search?q={query}"
    try:
        response = requests.get(url)
        data = response.json()
        images = data.get("images", [])
        if not images:
            return "404 not found error"
        random_image = random.choice(images)
        prompt = random_image.get("prompt")
    except Exception:
        cstr("Unable to establish connection to Lexica API.").error.print()
        prompt = "404 not found error"

    return prompt

`

[tender timber]

They’d fit in well with my Accounts Payable team I deal with at work… everything is referred back to my suppliers with nothing corrected and no explanation 😀

[tribal jackal]

There are people in the windows

[soft zealot]

there are some people on the pitch, they think it's all over......

[soft zealot]

and as Im sure some here may not understand the reference

https://www.youtube.com/watch?v=lg2CXkJVCHk

[tribal jackal]

Who's up for another Carrington event?

[inland vessel]

sorry feit asleep that's nice!!

[heady vale]

LOL the meme lives on in XL

[soft zealot]

have to love a bit of cinemascope 🙂

[ionic gulch]

also take a look at used RTX 3090(ti)s - they are twice the price, but you also get double the vram (24gb) and more performance. this would be as future proof as the top consumer card RTX 4090 (also 24gb).

[upbeat summit]

very nice ! keeping an eye on it and I'll think about what I could add.

[soft zealot]

"No Prompt" Images

No prompts set, no styles set just hit go 🙂

[upbeat summit]

the getting lucky latent space slot machine run

[soft zealot]

and pleasantly surprised was I 🙂

[upbeat summit]

I got lots of nice nature images with no prompt

[celest pivot]

a Shiba Inu wearing (Ray-Ban sunglasses, a gold chain, and a white hoodie)

[uncut steeple]

#1072220168534642768

[unreal plover]

The universe is a giant brain, change my mind

[unreal plover]

he is lost

[unreal plover]

They arrive

[maiden gale]

does anyone have experience with what U-NET is?

In Kohya it says Train U-Net only for SDXL and I cant find that option

anywhere

[west breach]

add --network_train_unet_only to additional parameters

[maiden gale]

thanks, yeah I thought that would solve my error but

It didnt

Do you have an easy method for updating kohya? @west breach

[west breach]

I only recently started using it myself. I use the SDXL branch

[maiden gale]

okay thanks, getting a really weird error training,

" size mismatch for down_blocks.1.attentions.0.proj_in.weight: copying a param with shape torch.Size([640, 640])
from checkpoint, the shape in current model is torch.Size([640, 640, 1, 1])."

Have you ever encountered this?

I should have all the correct parameters, a 640 size is really weird

[west breach]

not that error no

[maiden gale]

ok thanks still

We don't have a yaml file or anything like that for XL right? we only need the checkpoint for the training

[west breach]

try these settings. does your kohya gui have a sdxl model checkbox on the source model page?

[west breach]

👆

[maiden gale]

But this is for lora right? I have done a couple loras

Trying to get the dreambooth working @west breach

[west breach]

I haven't tried dreambooth

[ionic dragon]

winston, added any nodes, i have all the packs mentioned in the notes

[maiden gale]

Ah ok, found only one issue post about it, seems even 24gbs is low for dreambooth

[soft zealot]

Yesreday I did a clean installl into a new directory of COmfyUI and reinstalled all the Additional Nodes I mention in my notes element and (other than a minor issue with me not reading WAS NOdes install properly) everything works

[ionic dragon]

so all are mentioned in the notes?

i did install missing nodes

lemme try again

[soft zealot]

I di da clean install of COmfyui and aclean install of all the nodes mentioned in the notes and it worked as expecetd without any errors

[ionic dragon]

oh lemme try

[somber hill]

look at this quality

[unreal plover]

With post processing ?

[somber hill]

yes high res fix + face inpainting

but fully automated i used adetailer extension

will generate hundreds now while going gym

haha this is wild too but need some more inpainting of face

haha this is wild too but need some more inpainting of face

[unreal plover]

I love this theme

[boreal bough]

@hard fractal
While training SDXL, regardless of LoRA or full finetune, everything very slowly converges to 50% grey. first backgrounds, then more and more. Both via Kohya, as well as on diffusers training.
While there are ways of slowing this down, by changing noise offsets, and hacks that come at the cost of damaging the model (training on pure white & black)

I have 3 official questions:
A.) Just confirming that you are acknowledging that this is a thing.
B.) Do you have a way to counteract this while doing your full tuning?
C.) what offset noise did you use while training SDXL 1.0 (paper says 0.05, kohya says 0.0357, is either of them true for 1.0?)

(FYI, I will link to the answer in my guide, as there is no info on this regardless of where you search)

[somber hill]

What is your prompt

[unreal plover]

cinematic photo A neandertal prehistoric man riding a giant chicken dinosaur with feathers in a jungle with mud, rain storm. 35mm photograph,film,professional,4k,highly detailed

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

[maiden gale]

Do we have a reliable method of dreambooth training for XL? Getting some errors

[somber hill]

Hopefully it is my next research

[unreal plover]

cfg at 5

[maiden gale]

Waiting on the video for it, thanks.

[somber hill]

I will try that prompt with my lora

[somber hill]

Thanks I think it will be amazing

[maiden gale]

these are all lora right?

Not dreambooth

[somber hill]

Ye all lora

I think dreambooth will be much stronger

[maiden gale]

I think so as well, it feels like Lora's arent so flexible with different art styles

Or even flexible with anything, they retain too much info from the training images.

[maiden gale]

Lora's of a face

[boreal bough]

only happens if you train it wrong

like using a too high dim setting

or overfitting it to the point where it starts to damage other concepts

[ionic dragon]

reg images in lora are a scam

[boreal bough]

if you use dim 24, then you should be able to train a face without damaging anything

[ionic dragon]

4000 steps lora is way better than 16k steps lora

[ionic dragon]

whats dim?

[boreal bough]

🤣
hey, my 1k custom edited & tagged regularization images work like a charm

[maiden gale]

was gonna ask this too, which setting is this?

[ionic dragon]

i didnt tag mine

[maiden gale]

Wait you have to tag your reg images

[maiden gale]

ok so appearantly I used 256

[boreal bough]

I'll prob release a 500 image subset of it, when I put out my guide. so that can be used for female faces, while doing literally 0 damage, as it's 100% ethnicity & age weighted

male dataset coming as well, but not until I'm finished with my first one

[maiden gale]

Where can I follow you

[boreal bough]

yeah, at 256 you'll be doing significant damage to the whole model

[boreal bough]

finetune channel here on the server

[boreal bough]

I often post there

[maiden gale]

I see, will try 24 than, that difference seems huge though...

[boreal bough]

#🔧｜finetune message

for a full settings json

[maiden gale]

Do I need to use more images with 24? OR any different setting

Ah great