✨｜sdxl

[soft zealot]

remeber SDXL1 is a firstrelease. You shouldnt be comparing it too fine tuned 1.5 models, compare it to vanilla 1.5 and it blows its socks off

[rich coral]

for me sdxl is a tool in the toolkit and an amazing one at that. sure life could have froze at 1.5 and I would be fine but sdxl just makes things better quicker for me

[fresh path]

its 10x slower tho

[arctic bloom]

there is no loader for refiner in workflow

[spark bear]

black cats = best cats

[indigo carbon]

yes there is, it's an A1111 extension

[woeful patio]

does it take significantly longer to generate like switching between models?

[ionic dragon]

i cant find lexica node

[soft zealot]

PMSL
first_nsfw_for_sdxl_v1

"use it like usual, read prompts and add the words ... i cant write here"

https://civitai.com/models/118794/firstnsfwforsdxlv1

[normal bridge]

with the supposed dreamshaperxl workflow i guess yeah

[arctic bloom]

then it should be explicity mentioned 🙂

[indigo carbon]

found a way to eliminate that issue by changing TOURCH arguments

[normal bridge]

usually you would feed latent into the refiner straight from base with noise leftover but dsXL seems to work different?

[soft zealot]

ah thats cios I renamed the title, WAS doesnt advertise what It does, its somehwere under Was Suite |Text IIRC

[indigo carbon]

it's not writing it in the metadata, it's a very new extension that keeps the refiner in the CPU's RAM and switches the UNET with the refiner UNET for the last few steps(you choose how many steps)

[arctic bloom]

ohh, nice

[glad flint]

I could never get 1.4 or 1.5 to do this kind of frame, I'm really tempted to clean a bunch of them up and add it to my personal clipart library but to do that I'd have to make a to-do list and open GIMP instead of generating more stuff

[indigo carbon]

this way, it doesn't effect performance, just takes ~3 seconds extra every gen

[glad flint]

LOVE art nouveau decorations, hate the time it takes to draft this stuff by hand

[woeful patio]

I'm suprised by how well sdxl handles pixel art, that's progress for sure

[mellow dome]

yea, sdxl does create higher quality image

[indigo carbon]

yeah, no noise with that extension whatsoever

[ionic dragon]

got it

[orchid cobalt]

what I need to do to use Swarm to create alterations of the image?

I can't connect image from Load Image with KSampler

different nodes

[normal bridge]

anybody care to explain CLIPTextEncodeSDXL or paste a resource for some information

[indigo carbon]

no refiner

refiner

[brazen patrol]

More than that. I used to try to make logos witha chess theme in 1.5 and gave up. Not only was it clueless on the idea, but its chess themes were just endless garbage. Not anymore

[fresh path]

try putting base on lower steps

[woeful patio]

Doesn't even need lora's to work either, I used to strugle with that in 1.5

[brazen patrol]

yes, it was one of the first things I tried, and I was quite amazed

[midnight shuttle]

Compare base model to base model. Compare SDXL to base 1.5 with no LoRA, no TI, no hypernet.

[glad flint]

this was something like "innovative web design with creative use of gradients, screenshot of a website"

[spark bear]

will take time for sdxl to take off, a lot of the popular 1.5 models are just mixes of the NAI leaked model and they don't have that for sdxl, very few people do actual finetunes, most just mix models

[radiant tartan]

hi, i'm using Comfyui SDXL to test out a few different things. is the official offset LoRa suppose to change the image fairly dramatically from the none LoRa image? i'm using the same fixed seed for both lora and no lora. I'm including my test images + an image card to drag into ComfyUI to show my workflow.
edit: oops the first image is like this Base -> with lora -> refiner

[glad flint]

BRUH I THINK IT WAS WILDCARD PROMPTS IN COMFY THAT WERE CAUSING THE SLOWDOWN!!!!
I'm back to 29 second generations 🥳

[arctic bloom]

download1 = w/o refiner
download2 = with refiner

[indigo carbon]

this is with more refiner steps

[glad flint]

wildcard prompts such as {option1|option2} were apparently doubling to tripling my generation time in comfyui, I've been trying to figure out what the problem is for a whole day now

[indigo carbon]

did I use the refiner correctly here? there is no significant difference- I used 17 refiner steps here

[arctic bloom]

i set 10 max steps for refiner to do it's job

[eternal fog]

Refiner doesn't do a lot for that sort of image. It seems like its more for faces/eyes

[normal bridge]

sdxl might be onto a new better helicopter design 🤔

[ionic gulch]

https://i.imgur.com/l9QFMq9.jpg
https://i.imgur.com/XppUkEr.jpg

[woeful patio]

the first one looks really good

[glad flint]

oof i accidentally replied to something, sorry

[woeful patio]

Damn I think it's on to something

[ionic gulch]

Here is a gallery with more in that style: https://imgur.com/a/4P8FNpj

[woeful patio]

damn are those done in comfy or a1111?

[normal bridge]

https://www.youtube.com/watch?v=P-DCgzFbo_o found this channel, with a unique artstyle think a lora might be nice for this one ^^

[ionic gulch]

Comfy - all with DPM++2M or Euler a and 80-100 total steps (80% base + 20% refiner, a few with 3 extra refiner steps in front of the base)

[glad flint]

"innovative web design, website about frogs, screenshot of a website"

[woeful patio]

Really cool, It might be because this is my favorite style but those are awesome

[young mantle]

Definitely

It just changed one corner of the cube

[glad flint]

art nouveau ornate web design

[woeful patio]

this combined with chatgpt would be great

[glad flint]

I've been trying to generate CSS with chatgpt and it hasn't been great 😂

[worldly terrace]

Tree textures and overall background vegetation textures are straight up bad. Anybody else?

Othervice getting damn pristine stuff out of that sdxl 1.0

[glad flint]

🤔 post example?

[worldly terrace]

But those textures make me grimace

[worldly terrace]

Sec

[brazen patrol]

It has trouble with simple geometric motifs a lot

[worldly terrace]

Simple bs prompt, but still exemplary of the general phenomena Ive encountered

[peak dove]

Make a primitive with required dimensions

[ionic gulch]

can i make a primitive with height and width combinations and use those values for two different inputs (height + weight)?

[peak dove]

You can use thesame primitive twice I think

Ask Scott Detweiler, or Comfy himself

[glad flint]

I get similar texture problems with realistic fur, sometimes adding to the prompt can help

[ionic gulch]

yes, but i need different values. i want to get a selection of the combinations in the blue box as dropdown or buttons or in some other way that allows me to pre-define multiple height/width combinations:
https://i.imgur.com/EURCa2C.png

[brazen patrol]

well, +1 for negative prompts. This next image is all identical except negative prompts: "distorted, bad geometry"

Prompt was: flat company logo of chess and motherboard, electronics, elegant, gorgeous design, minimalist

not quite logo, but I had not expected such a radical change from those two negs

[glad flint]

LOL, I love how the bottom right kind of stutters "MESS" while the other laptops say "CHESS"

[noble shoal]

I have coded myself a custom EmptyLatent node with a buidin Ratio/Dimesion selector. Quite handy

[brazen patrol]

Yes, I was caught offguard by its choice to add the words. Funny how the original tried to use 'motherboard' instead

[zealous horizon]

new color modes for comfyUI:

plain, color by type, rainbow and positive/negative

Positive negative will color nodes depending on the title

via: https://github.com/failfa-st/failfast-comfyui-extensions

[gloomy barn]

using serge cutom node 2.0 inpaint....after running it once, it doesn't run a second time. Seems the images have to be cleared from intermediate output, but how?
I have turned to saving it unrun, then reloading it to run a second time. It runs once, then runs very fast subsequent times - around .1 second - with no output

I take it back. Even if I save it before generating, if I reload the unrun saved workflow, it still runs in about ,1 with no output. I have to re-load the blank workflow...start from scratch

[boreal bough]

well someones making a statement XD

[soft zealot]

nothing to see here that was for somehwere else lol

[soft zealot]

Is this yours?

If so then a "Make selected nodes the same size" option would be nice please 🙂

[brazen patrol]

besides, it was misspelled. It is written "Seeks"

[boreal bough]

Mister MeSeeks

[wicked dagger]

The dynamic promtps built into ComfyUI is very barebones. A ComfyUI port of Auto1111's dynamic prompt extension was just released - https://github.com/adieyal/comfyui-dynamicprompts. It supports wildcards, combinatorial mode, I'm feeling lucky, magic prompt, jinja2 templates and the majority of the features available in Auto1111.

[zealous horizon]

feel free to pen a feature request on GitHub.
I'm still trying to find out how to do things in a clean way. A lot of reverse engineering.
But I think it should be possible to add such a feature.

[ionic dragon]

@soft zealot do you know anyone who trained an LLM before?

[soft zealot]

No

[soft zealot]

well its not an issue and I dont see a discussion page or a featre request page 🙂

ah ok use this

[unreal plover]

Name this

[zealous horizon]

in GitHub, everything is an issue. I added a new type "feature request": https://github.com/failfa-st/failfast-comfyui-extensions/issues/new/choose

[soft zealot]

god dang it, after I added some

[ionic gulch]

is you node available through comfy manager?

i found one for sdxl resolutions, but i don't know how it works, because it has some predefined resolutions, but still has seperate inputs for height and weight

[radiant tartan]

Did you get an answer to this? i'm looking for the same thing. mostly as how the "ascore" works

[soft zealot]

dang it, started playing witht e COlour Options and now I cant get it back to COmfy Default

[normal bridge]

nope sadly not

[soft zealot]

@zealous horizon grrr mutter mutter

So not only does it do nothing when I slect "Default" even after deleting yuor folder & restarting I stil have a fruit cocktail on the screen

[wise edge]

I've just done this. Those are Searge integer pair nodes. And just drag whichever width and height pair to those reroute pins.

[glad flint]

how it started VS how it's going
september 2022 stable diffusion 1.4 vs SDXL 1.0 today

[zealous horizon]

because the state has been saved in the nodes by Comfy. I'm still looking into supporting that. You can choose "plain" to get the default uncolored mode back

[glad flint]

I've been trying to recreate some 1.4 prompts

[polar epoch]

Damnit.. how do i tell it to make a tiny shark to bite him in the finger lol

[soft zealot]

Ive reset everything manually for now and reinstalled your extension but Im not toouching the colours agian lol

[bold osprey]

what are some crazy things you made with dreamshaperXL?

[woeful patio]

this has to be changed to 1024,1024 for lora training right?

[ionic gulch]

Thanks, nice idea. Maybe I've found a better solution:
https://i.imgur.com/d4rMHee.png
This is using a node I found in Comfy UI Manager - will have to test it - if it doesn't work I may use your solution.
Should be easy to add new resolutions to the code yourself, you want other / different values.

[wise edge]

Ah yes, well that looks like a more production ready solution 😄

[ionic gulch]

just don't know why it doesn't update width and height if you use the dropdown menu. but from a frist test it uses the values selected via dropdown.

@wise edge the name of the custom node is the same as shown in the picture - if you want to try it yourself

[wise edge]

Thanks! I`ll be checking that out

[worldly terrace]

You mean "realistic fur"? In my case would you suggest "realistic leaves" etc.?

[noble shoal]

Unfortunately not. At the moment it only lives on my PC

[ionic gulch]

i've found something similar in the comfyui manager (see screenshot above). just has some useless width/height fields that don't get updated by the dropdown menu

[boreal bough]

anyone have a way to unload comfy, without close/reopen the whole app?

[noble shoal]

If someone is interested in this node: https://drive.google.com/file/d/1iKMydI_uKv1k2PDswH3EBFfGm9hrPvd4/view?usp=sharing

[glad flint]

sometimes it's adding adjectives like "lush foliage", "majestic" to the entire prompt or adding interesting negative prompts
idk, you just have to experiment until you find what works
one time I added "itchy" to a negetive prompt and it helped something

[boreal bough]

thats a sampler issue

I have a pic for that. gimme moment

bit big, but it shows exactly what you have - and how to resolve it

either switch sampler, or go up in steps. both options work

[glad flint]

oh yeah, I forgot my texture problems were way worse when not using euler

[boreal bough]

@visual glade
is there currently a supported way to unload the vram? my lora training pushes it just over edge T.T
if I could just unload the models, then I could leave it open while training

[arctic bloom]

where do comfyUI's history is saved?

[visual glade]

unload the vram from comfyui?

[boreal bough]

output folder - if you save images
and in the ui, if your instance is still open

[visual glade]

the history isn't saved anywhere, it's in the software memory

but you can get the workflows by loading the images you generated

[zealous horizon]

Check for updates. both features you requested are implemented.

[boreal bough]

basically set it back to the state before I generate the first image - when it uses up 0 vram

[visual glade]

that should be the default right now

[boreal bough]

O_O

[visual glade]

how much is it taking after a generation?

in comfyui you can open a new tab and drag that image in

a new browser tab

[arctic bloom]

thanks,
recently reloading UI to test some stuff, so wanted to load previous queue query to compare with generation before reset/update

[boreal bough]

pre closing/after closing

[visual glade]

oh that's the amount pytorch takes by default

[boreal bough]

it works right now apparently. but sometimes it occupies 7 gb vram for hours

but clearly not always

[fierce hollow]

you're probably running with highvram and it's not unloading the unet to cpu after generation

try starting with --normalvram

[woeful patio]

RuntimeError: Sizes of tensors must match except in dimension 1. Expected size
56 but got size 55 for tensor number 1 in the list., Anyone know what this means?

[soft zealot]

Gracias

I see the Option and Ive selected it

And it works a treat 🙂

[boreal bough]

will try 👍

[worldly terrace]

Dang. No options for samplers in the service I'm relying to

[boreal bough]

can you adjust steps? issue solves itself around 50~60 steps

[fierce hollow]

also to add to this, if you have fewer than 8 (iirc) cpu threads it will try to move the text encoder to vram, no idea why it's the default

[worldly terrace]

Can't dabble with steps either 😦 but good to know that this ACTUALLY is not an issue. Thanks!

[spring fulcrum]

Is there any way to get comfyUI to use more vram??? I have a 4090 and it seems to be pegged at under 8gb vram since the update.. Would allowing it to use more vram get it to have more its/s right now I get about 6.5 its/s

[boreal bough]

batch size to 3, should 95% utilize it

[fierce hollow]

@visual glade Could the threads check here be lowered? I only have 6 but the text encoder still works much faster on cpu than moving it back and forth, I think 4 would be reasonable, or maybe it could just always stay on cpu with normalvram
https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/model_management.py#L367

[visual glade]

sure

I'll set it to less than 4

[fierce hollow]

thanks

[visual glade]

done

[primal vault]

most of the time you can just use the same prompt in main and secondary, making them different only helps with fine tuning the image in the last iterations of improving it

I always start with using the same prompts there until I find a nice seed, then fine tune the prompts if needed while locking in the seed

[ionic gulch]

is there a way to prevent nodes from changing size? especially primitive nodes always reduce their size if i save a workflow and reload it at a later time (pinning doesn't help for the size, only for the top left corner)

[primal vault]

also SDXL has been amazig so far and SAI definitely was right that lower weights on keywords that affect styles can improve the results quite a bit

[upbeat summit]

primitive nodes do currently not remember their size

[nimble heart]

I'd say it's up and down. Sometimes the reduced offset noise in 1.0 can wash out art if you don't cram the prompt full of tokens

[indigo carbon]

@visual glade remember when you claimed you managed to convert SDXL into TensorRT? how did you convert the safetensors to ONNX, as far as I know- ONNX doesn't support SDXL's shapes?

[visual glade]

why wouldn't it support it?

[indigo carbon]

ONNX isn't supporting SDXL's shapes, they are different from SD1 and SD2

[azure oxide]

i think ur referring to the a1111 extension

that specific one doesnt support it

nor did it support 2.1's unless you did something iirc

[visual glade]

yeah of course I didn't use any of the code from that extension

[indigo carbon]

when they built the latest version of ONNX they specified what architectures it's intended to support, I am certain SDXL wasn't one of them. did you change ONNX's code?

I have no idea how would you do it otherwise.

[visual glade]

I suggest you go read what onnx is and how it works

[civic sigil]

First finished version of my XL Animation workflow, would like to add reference_only and more controlnets as they roll out tho

[indigo carbon]

yes, I did. the conversion scripts on the HF diffusers repo mentions that SDXL isn't supported. I have no idea what kind of black magic script did you use for that conversion. the ONNX->TRT onversion is the simple part. however, I have no idea how would you convert SDXL to ONNX.

[eternal fog]

Reference only is here - https://gist.github.com/comfyanonymous/343e5675f9a2c8281fde0c440df2e2c6

However, from what I've tried, it's usefulness is limited

[civic sigil]

Doesnt support img2img

[visual glade]

by simply following the documentation

[azure oxide]

im more interested in how this is coming along tbh 😉

[indigo carbon]

which one? NVIDIA's?

[visual glade]

oh I started working on it, I'm trying to make the AIT and pytorch code the same

[visual glade]

the documentation that explains how to convert a pytorch model to onnx

[indigo carbon]

by the way, how's AIT performance wise in comparison to TRT?

[eternal fog]

Do TensorRT models just work with Comfy then? Or do you have to do special things.

[visual glade]

no and I don't plan on supporting them because they are not useful since the weights can't be swapped out

[eternal fog]

What's the actual speed difference like. I never bothered trying them before, because you have to convert every model you want to use.

[indigo carbon]

can AIT weights be swapped?

[visual glade]

if there's an official trt engine file that gets released for SDXL I might make a custom node to load it though

[indigo carbon]

TRT files are system specific

[visual glade]

of course and in case there's issues it's open source so I can fix it myself

not really

[indigo carbon]

nice! is it as fast as tensorrt?

[azure oxide]

ive seen reports of some crazy it/s's

[indigo carbon]

as crazy as TensorRT? I managed to get it to ~35it/s on 512x512 with SD1.5 (I have a 4070ti)

[azure oxide]

https://old.reddit.com/r/StableDiffusion/comments/xv7m89/metas_open_source_aitemplate_provides_a_24x/ just a quick search

tbh i wonder why it never took off

[indigo carbon]

oh jesus

[azure oxide]

i remember looking it up a while back, someone managed to get like 90+ it/s on their 4090 but i cant seem to find the thread in my history

[indigo carbon]

oh shit, also, the conversion is straight from safetensors to AIT, trying this out now.

[azure oxide]

oh i take it back, i confused it with this from yesterday https://old.reddit.com/r/StableDiffusion/comments/15clbwi/boom_99_its_batchsize1_with_torchcompile_on_a_4090/

[eternal fog]

The documentation for that says it requires Triton and I believe that only works on linux

[indigo carbon]

really? =[

[eternal fog]

Unless it's changed recently

The Kohya trainer can use Triton as well, but it won't install on Windows

[indigo carbon]

I'm getting MVS errors when using AIT hahaha

[eternal fog]

They claim for that AIT it uses less memory as well

[midnight shuttle]

Finally gettig ComfyUI going locally. If I move it to another location do I need to update any absolute paths other than the ones in the venv scripts?

[nimble heart]

the venv itself can't be moved easily

when it's created all its references are absolute

[eternal fog]

Yeah if you move the venv it will break

Even if you try change the config, it still seems to break

[midnight shuttle]

OK. I'll re-create it in the new location then. I was hoping I just needed to update the references in the venv scripts.

[nimble heart]

I tried that using a sed regex replace and it still broke

so I just remade the venv

maybe there's a shady python script for it on git or pip somewhere

[quartz sequoia]

@marble dew @hard fractal i propose everytime you get a new question of a upcoming feature, you tease us with a Picture or a Vid 👀 of "SOON™"......using that feature, like controlnet 👀 👀

just just a lil tease

of

SOON™

[tropic turret]

What am I doing wrong? 😦

[quartz sequoia]

so whens videoLDM for sdxl coming out 👀 ?

[quartz sequoia]

ur picture isnt supposed to be white.. thats what wrong

[upbeat summit]

have you checked the messages in the console?

[eternal fog]

And if you look in the outputs folder, is it actually saving a white image.

[tropic turret]

Yes, it is saving a white image.

[autumn forum]

i successfully managed to get my lora to do a reverse shot. albeit not consistantly but it did it

[indigo carbon]

@visual glade are you on linux? AIT is spitting OS compatibility issues at me

[quartz sequoia]

tbh you might have cracked the code in producing perfect white pictures

can anyone do that with SDXL?

[tropic turret]

😦

I really am just wanting to upscale the final image and save it lol

[eternal fog]

Do you get an image if you don't do the upscale step

[tropic turret]

ill try that

[eternal fog]

Also how the fuck do you get all your lines at right angles?

[tropic turret]

Ill find that for you

[hot rover]

Enabling the TAESD preview for each sampler might help you narrow things down.

[eternal fog]

I'd like to remove my spaghetti mess

[tropic turret]

This here

[quartz sequoia]

if you use ur Lora trigger word at the start of the prompt, it will almost always be consistant

[tropic turret]

I did that and it gives straight lines

[quartz sequoia]

token placement in prompts is important... especially for loras

a lora at the mid or end of a prompt is utter trash

but if its at the start it works almost always depending on the weighting sometimes

[autumn forum]

i know. lol. the lora is for moviestills. it has many features imbeded to it. weighting them all correctly is the hard part.

[soft zealot]

@eternal fog https://github.com/failfa-st/failfast-comfyui-extensions is probably better as also allows you to do other things include set the line width to 1

[quartz sequoia]

i found weighting the lora node.. to be a bit of you either get it right or its shit, weighting the trigger word and class is a bit more lenient

[tropic turret]

I got an image this time

Is it because the upscale model file is not in safetensors format?

[autumn forum]

no it doiesnt need to be safetensors. non of mine are

[tropic turret]

I was kinda going based off of this here: https://comfyanonymous.github.io/ComfyUI_examples/upscale_models/esrgan_example.png

[autumn forum]

mam you need to turn around.. 😂

[tropic turret]

But kept the default SDXL 1.0 workflow and trying to add upscale to it

[autumn forum]

that is the correct way to implement the upscale nodes

[upbeat summit]

yeah using this since yesterday. really nice QoL features

[indigo carbon]

@visual glade does AIT only work on linux?

[midnight shuttle]

I can see the different VRAM memory management strategy between A1111 and Comfy just by watching the VRAM graph. I can see where it loads and unloads models in real time whereas A1111 without --medvram just shoves everything in there until it overflows.

[quartz sequoia]

@visual glade is it possible to use like --share on my pc to connect to my laptop? or is there another method or still not implemented?

[autumn forum]

@quartz sequoia excuse the bad face but turns out it is pretty consistant

[soft zealot]

Look Ma, No Spaghetti ;o)

[hot rover]

I think you just type --listen 0.0.0.0 as part of your startup command

[autumn forum]

looks like a computer chip die haha

[soft zealot]

Lol

Even more so with straight lines at size 1

[quartz sequoia]

dope

[quartz sequoia]

i shall attempt thanks

[indigo carbon]

won't that only be relevant to Linux users though?

[mellow tendon]

Hmm I tried to install Roop with the ComfyUI package Manager and now ComfyUI will not launch, looks like it has broken the Python version

" RuntimeError: Failed to import transformers.models.clip.modeling_clip because of the following error (look up to see its traceback):
Descriptors cannot not be created directly.
Downgrade the protobuf package to 3.20.x or lower."

Anyone had this and know how to fix it, please?

[autumn forum]

did you install roop's dependencies?

[upbeat summit]

there could be a shortcut to just hide the link cables. but this goes a bit against what a node interface is. I understand having a simple interface with a compact overview makes sense, but I think UI projects like comfybox or mentaldiffusion that put a simplified UI on top of it could be overall a better solution.

but there's no right or wrong. if it makes sense to someone, it makes sense ;).

[mellow tendon]

hmm, I remember now, that I did try that the other night and it came up with a Windows file permission error.
There seemed to be 2 versions of Roop avalible so I just tried the 2nd one and that is when it broke ComfyUI.

[autumn forum]

hehehe this is getting me gitty with how well this is working

[midnight shuttle]

This is interesting. Now that I'm using Comfy I finally saw the VAE artifacts that I'd never seen before. Biggest difference is that I am using refiner in Comfy but not using refiner in A1111. I wonder if the VAE artifacts are mainly a refiner problem?

[upbeat summit]

is it a lora or prompting? I tried over the shoulder shots, but have not yet figured out a good prompt structure

[autumn forum]

my lora!

[upbeat summit]

nice! looks really good

[visual glade]

of course AIT works on windows

[indigo carbon]

it's spitting compatibility errors at me. am I supposed to use Docker?

[visual glade]

no, you can try: https://github.com/FizzleDorf/AIT

[wet nacelle]

Does anyone have a working XL 1.0 for control net workflow for comfy? If so can you send an image that uses it so that I can have the metadata attached to it?

[visual glade]

has anyone released controlnet weights for XL?

if they did you can just use it like a regular controlnet

[indigo carbon]

does this do the conversion? the errors I get are during the conversion process.

[wet nacelle]

I don't have a comfy control net workflow though. That's why I also need the workflow.

[visual glade]

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

[ionic gulch]

https://i.imgur.com/heAH7MC.jpg

[arctic bloom]

https://huggingface.co/SargeZT/controlnet-sd-xl-1.0-depth-zeed/tree/main

haven't tested it yet, tho (too few vram for it)

[eternal fog]

The main Repo suggests updates need to be made for it to be able to do SDXL https://github.com/facebookincubator/AITemplate/issues/851

Updates to the code in the example, so if you know how it works, then you can probably just update the conversion code.

[indigo carbon]

@visual glade claimed to have done it on SDXL

[eternal fog]

Probably because he knows how it works and can code it himself.

[visual glade]

yeah if you stick that .bin in the models/controlnet folder and use it like in my examples it should work

[pallid sail]

Is it my imagination or do the examples not match the depth input?

[eternal fog]

no they don't, so I'm not sure what that's actually doing

[indigo carbon]

oh wow, I'm getting 33it/s with https://github.com/FizzleDorf/AIT

no SDXL support though

[azure oxide]

what was your speed without it?

[tight fjord]

anyone can give me an idea of what I should be getting with a 3090 and SDXL in comfy? I think I hit a weird bug, used to be 3it/s, now 3s/it 😦

[indigo carbon]

~20

[midnight shuttle]

Should be more than that. Check console for any errors. Did you change anything?

[azure oxide]

nicenice

[indigo carbon]

I got to ~40 with TensorRT though

[tight fjord]

Nothing in console, but I might reinstall comfy just to be sure

[azure oxide]

did you change any generation settings?

[indigo carbon]

so it would be faster if the engine would be created on local machine, correct? @visual glade

[midnight shuttle]

Also watch Task Manager if on Windows to make sure it's using your GPU and that it's not using unexpected amounts of VRAM.

[azure oxide]

some samplers are slower than others and there might be some other generation settings that affect speed

[visual glade]

probably but those AIT files might not be the most optimized

[indigo carbon]

will the final version you're working on create the engine itself?

[visual glade]

no, the engine will be compiled separately and shipped with the standalone

[indigo carbon]

and the engine isn't model specific, correct?

so basically, everyone will get that insane boost

[visual glade]

it is model specific but not checkpoint specific

[indigo carbon]

that's amazing!

[quartz sequoia]

it didnt work

[quartz sequoia]

anyways to use my pc? and share the webui to run it on my laptop?

[gentle ridge]

when controlnet for sdxl realsinf

[visual glade]

on LAN use --listen, make sure it's unblocked in your firewall and access the LAN ip of that computer

[hardy cipher]

oh my, wha tdid I miss?

[quartz sequoia]

is it secure?

[visual glade]

as long as your LAN is secure

[quartz sequoia]

so no pass required?

ur webui isnt technically gradio

that makes sense

also i have no idea if my ur ui is updating for me

everytime i run the update / comfy manager update also

its always upto date

i never see it take in any pull requests when i start it up

[hardy cipher]

when I first started using comfyui recently I thought the tutorials were super lame and uninformative not realizing I could load the pictures to get the workflow, lol

[boreal bough]

as I just (unluckily) found out, some prompts in sdxl are out of the box insane disproportionally weighted - so if you add those words to your lora training, they make essentially zero impact unless you equally disproportionally add them to your dataset 🥲
guess you just have to be jedi and know ahead of time

[hardy cipher]

I need to figure out how to use the boolean commands to automate things. I haven't really seen much of that but I know there are lots of possibilities

[quartz sequoia]

how do i unblock it from my firewall? since --listen isnt working for me

[boreal bough]

'coffee shop' being one of them. trumps most other weights XD

[autumn forum]

that makes sense. that sucks. finding which keywords they are is probably pretty difficult

[spring fulcrum]

Idk what is going on with SDXL but I was making things like this.... the dark one with the golowing eyes, now they are coming out like this.... Can someone please help me???
.
.
Does the new SDXL 1.0 model need more steps to make good images?

[azure oxide]

poor comfy, probly has a million pings from various servers for tech support

[hardy cipher]

using a lighter latent image to start?

[azure oxide]

he the realest mvp for actually responding tho

[hardy cipher]

or whatever it's called

[boreal bough]

'view from behind' is one of them, and its one I often used for tagging - and also automated by vit-h. not sure what to do about that right now, as I have no good 'solution' for it

[autumn forum]

what are you trying to get and was is it giving you?

[candid walrus]

yeah i'm curious have been toying with the idea of trying to make some styled loras for sdxl and would love any time saving knowledge

[boreal bough]

did my proper captioning, and all views of rivet from behind, I tagged with 'view from behind'.
lora works great - but the moment I add 'view from behind' I just get photos of men from behind. as if my lora stopped existing. even if its the last prompt XD like wth

[hardy cipher]

"butt slap"

lol

[candid walrus]

i see completely what you mean now

[autumn forum]

maybe try giving it a different keyword? lol like facing away or something?

[candid walrus]

right, like some community 'hard to train' phases would be killer

[boreal bough]

view from behind vs same prompt, without view from behind

[upbeat summit]

a lombax!

[candid walrus]

right, like it almost got it, but then forgot about the subject

[hardy cipher]

hmm. maybe you could word it differently? not sure how exactly

[candid walrus]

'back'

[boreal bough]

there are definitely ways to get around it - but 'view from behind' shouldn't be incredibly biased towards men, to the point of ignoring 90% of the remaining prompt

its a core sdxl problem T.T

[autumn forum]

yeah you could try "back towards viewer" "back facing", "turned around". and yes i agree its like all the ones from behind were trained on men and it kinda stuck with that

[candid walrus]

@boreal bough what program is that first screenshot from

[boreal bough]

could be solved with TE training - but once I add that in, I can forget about making a 'newbie friendly tutorial', cause this is getting complicated

[boreal bough]

hydrus network. hard to learn the first time, but scales well from small till 5k datasets for manual tagging

[boreal bough]

I claim innocence!

its for sfw tutorial XD

app lets you organize tags real well

[hardy cipher]

lolol

[quartz sequoia]

can SDXL outpaint?

[hardy cipher]

anyone know why I would get a "header too large" error when trying to use a safetensors vae? switched to ckpt and it worked fine

[midnight shuttle]

Sounds like it's not really a VAE file but maybe a full model or something else? Or else it's a corrupted download.

[hardy cipher]

ahh. well it 's vae-ft-mse-840000-ema-pruned.safetensors. used the ckpt version and no problem.

[midnight shuttle]

Maybe try downloading it again in case it's corrupted.

[hardy cipher]

yeah, that could very well be the case. I'll do that

thanks. sometimes these errors are kind of hard to figure out without doing some digging or having someone clue me in.

[void skiff]

SDXL is not as the other AI kids, not long ago I did not like Euler and Euler A, but now SDXL can give okay but very random result. What have become your (any who read this) favorite sampler for SDXL?

[hardy cipher]

what's wrong with euler A?

[boreal bough]

#✨｜sdxl message

[autumn forum]

this is a cool question. my favorite was dpm++_2m_sde karras in 1.5 and now it seems to be a mix of euler and dpm++2m karras. sde seems to have kinda shit the bed lol

[boreal bough]

(base only)

but refiner doesn't change composition, so it makes for a good comparison

[midnight shuttle]

I haven't found the samplers to make that much of a difference. Maybe Euler is a little "softer" and some of the DPM are a bit more intricate detailed. But I could never guess what sampelr was used for an image.

[glad flint]

I love sde in 1.5, but in SDXL Euler is friend

[void skiff]

Eurler A was jumpy and changed result as fast as Sampling Steps was altered

[midnight shuttle]

That's because it's an ancestral sampler. It never converges.

[hardy cipher]

is it possible to copy an entire group of nodes in comfyui?

[urban fjord]

I guess you can try a seperate LoRA for just behind and then merge them...

Yes, control drag to select multiple nodes in ComfyUI, I think.

[spring fulcrum]

Does sdxl require more steps for a good image? I used to use 0-13 base then 13-20 refiner then 10 upscaler steps. Has this methodology changed since 1.0 released?

[hardy cipher]

ugh, so obvious now that you said

that

[glad flint]

I do the opposite, 13-20 base, 0-13 refiner

[boreal bough]

yep. I can definitely solve it - in multiple ways. But while writing a guide I'm doing my best to not turn it into an endless flowchart that takes multiple months of lora training to understand T.T

[brazen patrol]

depends on the sampler, no?

There are cases where the Refiner is actually counterproductive

[boreal bough]

and if it's too complex, it ends in "too long; didn't read" when a youtube video (looks like it) gets them 80% of the way there in 15 minutes. Even if that's not true at all

[nimble heart]

does that... work?

[midnight shuttle]

Oh, maybe they mean "13 to 20 steps on the base and then 0 to 13 steps on the refiner". Not actual step numbers.

[glad flint]

YEAH LOL
i thought you were talking about number of steps

somewhere my old high school teacher is taking a nap, woke up and screamed something about units, and fell back asleep confused

[boreal bough]

poor man had a revelation from an alternate universe self. tried to warn yall!

[gloomy barn]

I downloaded this SEARGE custom node. Run the inpaind. Once it's run one time, it runs in about .1 seconds every other time, outputs nothing. I have to reload it - either blank or saved with what I want to do before running it.

Is there a way to get it to run more than once?

[hoary saddle]

anyone here using comfyUI API? curious if anyone else experiences it freezing the PC after a couple dozen renders, after a bit the queue shows "Prompt Executed" but fails to render, then system locks up and have to hard reset

[hardy cipher]

that doesn't sound good

are you sure everything is up to date?

[gloomy barn]

mine doesn't freeze, but as I explained above, it won't run inpaint more than once

[hardy cipher]

every time I have issues like that it's because something isn't installed or updated

[trim orbit]

check mic is controlnet gtg yet mic check check 1 2

[hot rover]

As far as face fixes go, what works best with SDXL? Photo real faces all look weird in the eyes with SEARGE SDXL workflow and another that I tried. I've tried using a facedetailer node setup that I recreated from a 1.5 workflow but the results weren't great because I have limited knowledge on which model to feed the DetailedDebug node now that there are two different models to consider. It'd be nice to pass the image through codeformer or GFPGAN within the node network.

[hoary saddle]

yeah pretty clean fresh install

[gloomy barn]

we've taken such a huge step back with sdxl...

[hoary saddle]

only thing erroring is stage_model_clip, whatever that is

doesn't effect it via web ui

[broken ferry]

Sorry if this is a dumb question - are the settings for this discord's bot image generator(s) shown anywhere?

I finally downloaded the sdxl models for my local stablediffusion installation and the images it's generating are just really, really awful compared to what the bot is generating. I'm just using default settings from the automatic1111 web ui, and using exactly the same prompt as I am with the bot, and it's not even close (like child's drawing in mspaint vs a nice painting.

[gloomy barn]

in the past, people had answers and fixes. Now...nada. Faces look weird, workflows won't run more than once if at all

[midnight shuttle]

In what sense? It seems a lot better than base 1.5 or base 2.1. But you can't compare it with finetuned models. Only base.

[midnight shuttle]

That's not a model issue. That's a UI issue.

[trim orbit]

codeformer still works

[hardy cipher]

I don't even really know what you're asking? why inpaint only runs once? why things look weird?

[gloomy barn]

SDNext and Auto 1111 not generating clear images....look like at best refiner isn't really applying. Comfy running something once then never again unless it's reloaded

[trim orbit]

prompting is a new vernacular to learn too

[gloomy barn]

Clothing styles are lost...inconsistend or not there

[trim orbit]

"can't" is a bad attitude to approach these systems with

[midnight shuttle]

A1111 SDXL images are clear enough for me. Also ran many Comfy gens with no freezes.

[hardy cipher]

I mean, maybe go back to what brings you joy and wait until people figure xl out?

[midnight shuttle]

Compare with base 1.5, not finetuned.

But yeah, if SDXL doesn't work for you, use a model that does.

[gloomy barn]

base 1.5 if I prompted a certain clothing, it was somewhat consistent

[hardy cipher]

well 1.5 was so finely fine tuned

with all it's addons

[gloomy barn]

inpaint ran more than once, images weren't blurry in A1111 and SDNext

[midnight shuttle]

SDXL has more training data, so if you say "pretty dress" it will have a lot more different dresses.

[hardy cipher]

sdxl is like a new supercar that none of really know how to drive yet

[gloomy barn]

if I say shorts....it does long pants, baggy shorts, one leg long and one short, one leg missing of character,

[hardy cipher]

lol, those shorts

[gloomy barn]

When you try to make images of people playing tennis or basketball in the 70's, you need shorts

[midnight shuttle]

Use old model for waifus while learning new model. Problem solved.

[gloomy barn]

not the baggy shorts they wear today

I dont do waffles\

I tried to make an image of a tennis player with standard, not baggy, shorts

[glad flint]

SDXL is blowing my mind that I can have a whole cityscape + vast landscape with Maxfeild Parrish influences, and then add "and blue pineapple plants" and it doesn't tint random things blue or have a giant floating pineapple, it actually tries to incorperate pineapple. They aren't always blue, but sometimes they are, and it's amazing lol

[hardy cipher]

out of the box with no loras or anything else sdxl can do a lot of things as well or better than most of the stuff I could do with 1.5. and I'm not bad with 1.5. I just think it's one of those things that takes time to get used to

[gloomy barn]

I want to try to inpaint in Comfy,.,,but can only run inpaint one time...then it stops outputting anything

[midnight shuttle]

I think the SDXL base model is better but it will require time to learn how to use it and more time for people to finetune it.

[gloomy barn]

using the searge worfflow

can it be reset?

[midnight shuttle]

Build your own workflow so you know it's built correctly.

[gloomy barn]

instead of re-loading the flow?

I am not clear on how to build it for inpaint, so I am using searge

[midnight shuttle]

I think you need to use a load image node, then a VAE encode node.

[gloomy barn]

just as with 1.,5 I wouldn't have known how to do it, so I used a1111 and sdnext

[midnight shuttle]

There's no SDXL inpainting model yet so I haven't really done much with SDXL inpaint.

Might also need a mask encode node. I can't remember.

[gloomy barn]

but with sdxl, those give what look like unrefined images. Comfy has face issues...maybe there is a face fix somewhere for comfy

[hardy cipher]

I think there is a face fix

[midnight shuttle]

The model is the model. Comfy and A1111 use the same model. It's just more complicated to use Codeformer in Comfy.

[hardy cipher]

I like to have the freedom to make my own mistakes and mess everything up

[gloomy barn]

right. I have started using comfy because a1111 and sdnext arent giving good results.. not sure if they're actually applying the refiner

[hardy cipher]

bro, just use other people's workflows. and if they don't exist yet I guess just wait

[midnight shuttle]

A1111 is definitely not applying the refiner. Comfy and SD.Next are applying the refiner if and only if you configure them to do so.

[gloomy barn]

I am

Which is why I asked if there's a way to make it run over and over again

\rather than just one time...then it has to be reloaded

[hardy cipher]

I'm really trying to figure out some new things myself. lots of things that don't work, lol

but that's how I learn this kind of stuff

[gloomy barn]

I add the imagte, the inpaint image, inpaint, run it...it outputs., Hit QUEUE again...it goes for .1 second, claims to be done...outputs nothing. Then I have to reload the workflow

[hardy cipher]

if I'm using a 2.1 unclip model do the other models in the workflow need to be 2.1? I hope this isn't a dumb question

[upbeat summit]

that's how you do it 😉

[gloomy barn]

and you ask if it's already been done

[hardy cipher]

might be a problem with the node?

[spring fulcrum]

Don't you have to generate an image before it can be refined?

[hardy cipher]

lol. was not going for this

kind of cool though

[cedar tusk]

happy accidents, and all that

[midnight shuttle]

Crazy that just 1 year ago someone would have probably paid money to buy that and hang it in a gallery.

[hardy cipher]

they're my favorite

[midnight shuttle]

And someone would have spent many hours creating it.

[gloomy barn]

searge has some kind of discord alternative link to live chat about his workflow...but that won't create an account, so I can't get into it

[deft quiver]

people still will, lol

[cedar tusk]

got this happy accident while messing around with the clip sizes

[midnight shuttle]

For a little while longer until AI gen goes fully mainstream. Once it's as easy as Word or Excel no one will pay for anything other than highly specialized art.

[hardy cipher]

another recent happy accident for me

[gloomy barn]

Ah, if I change either the initial image or the image in the inpaint node, it runs again

[hardy cipher]

I like when they sort of render in liminal latent space. on the precipice of different concepts

[glad flint]

for me a bad generation used to be body horror where it looked like someone's face melted off and the subject of the portrait used to stare at you with dead blacked smeared eyes
This usually happened to me when my prompt was really complex in 1.4 and 1.5 and I wasn't using a LORA or anything for correction

In SDXL my bad generations are usually just too generic and rarely there's some floating object or something. Sometimes things are just "too perfect". It's hard to prompt back in that vintage look and the quirks that come with it.

These images look like the subject of my 1.4 and 1.5 prompts died of tuberculosis and are now in heaven

[cedar tusk]

Got themselves a case of the vapours

[glad flint]

I thought we were talking about number of steps LOL

[hardy cipher]

my bad renders are basic and dont pop. not sure how to describe it

[gloomy barn]

show one

[spring fulcrum]

ya but if you use the refiner at steps 0-13 wouldn't that mean you are starting the refiner before there is an image? then if you use the base at steps 13-20 thats when it starts to create the image itself? Maybe I don't get how this works.

[hardy cipher]

a bad one?

[gloomy barn]

yes

[glad flint]

nah I was confused and thought we were talking about how many steps we were using, not which order

[ionic gulch]

getting closer to a good workflow after adding the offset lora - using -0.08 for offset lora works great for that style of pictures:
https://i.imgur.com/CMatwx7.jpg

[gloomy barn]

anyone have a good workflow for inpaint?

[hardy cipher]

that one isn't very good

[gloomy barn]

like to fix handy pandies?

[gloomy barn]

looks like no vae applied

[hardy cipher]

well it was more than that

I moved a bunch of nodes when zoomed out

not realizing I was still changing parameters

[gloomy barn]

yes, but it's washed out

[midnight shuttle]

Some VAE had to be applied. That's the only way to get a human-readable image out of the latent space.

[hardy cipher]

so that took a while to undo

[gloomy barn]

we=ll...then the wrong vae

[hardy cipher]

I think it was

orangemix or something on sdxl

[gloomy barn]

in sdnext, I was getting washed out images until I changed to default VAE none

[hardy cipher]

but didn't realize because never meant to set it

[cerulean valley]

I ask the same thing
how do you get straigh lines in the workflow instead curves lines, thanks !

[midnight shuttle]

VAE "none" means to use the VAE built in to the model. A VAE is always used.

[gloomy barn]

right. well,. it's odd

when I used the sdxl vae in sdnext, everything was washed out

when I did none to use the baked in, it was fine

[hardy cipher]

if you drag a node and let go in empty space there will be a "reroute" option

[gloomy barn]

on comfy, I used the sdxl vae....looks fine

[hardy cipher]

or rather, from the input or output

[gloomy barn]

SDNext generations look blocky, unfinished. Comfy looks better, but issues in the eyes

[molten breach]

SDXL inpainting model wen? 😎

[hardy cipher]

cutting an unclip model's output off a few steps early produces some strange results

[gloomy barn]

I have a workflow for inpaint, but it only works once

[hardy cipher]

meant to put the cutoff on the sdxl base model, not the unclip, but either way

[cedar tusk]

So what's the skinny on refiner prompt versus base prompt? I usually keep it the same

[gloomy barn]

I have two prompts, one is for refiner, both go to the original image too

[hardy cipher]

how do you structure them?

I haven't figured out what my approach will be. still experimenting

[hardy cipher]

I've decided most of the online advice is pretty much. people saying all sorts of conflicting things

*pretty much nonsense

[winter raptor]

Any of yall played around much with SDXL on the StableStudio tauri build?

[hardy cipher]

I don't know what that is, but I wouldn't mind knowing

[midnight shuttle]

One person says it and 1000 copycats repeat it as gospel without understanding.

[hardy cipher]

I'll take note of what people say, but I don't just believe it. and a lot of it is subjective, so there really isn't a right or wrong way per se

[brazen patrol]

Are you talking about SDXL, Social Media, or what?

[hardy cipher]

well sdxl right now

try to not believe or read anything on social media

[brazen patrol]

Nonsense, It is the only source of truth

and knowledge

and wisdom

🤣

[midnight shuttle]

The best is if you learn enough about how it works so you understand it. And not just doing "magic spells" and incantations.

[hardy cipher]

yes

however, many things will start off as magic spells and incantations. we must put in the effort to demystify them

[midnight shuttle]

Following tutorials without understanding counts as incantations.

[hardy cipher]

understanding what?

the coding, machine learning? the model specifics?

[midnight shuttle]

Understanding at least on some basic and honest level what the steps are doing.

Some basic understanding of all those. Just the basics.

[hardy cipher]

well you have to be bad at something before being good at it. we were all born knowing basically nothing

[winged tree]

thats completely missing the point here

[hardy cipher]

what is the point?

[midnight shuttle]

Yes, but the information is out there to learn by researching. Not to say anyone knows it all. I don't. But some learning is needed for any new technology.

[hardy cipher]

that people shouldn't follow totorials if they don't have a deeper understanding of how it works?

[midnight shuttle]

Learn by doing will often result in wrong learning.

[winged tree]

This, you're not going to learn anything by not knowing how it actually works

[hardy cipher]

well okay

but how do you learn if you don't get experience in it?

should a person just read about it for a fdw months before looking at tutorilals?

[brazen patrol]

I disagree. I have no clue what each sampler does, and yet have experimented to geta feel for them and choose three that I rotate

[midnight shuttle]

No, but maybe a few hours.

[hardy cipher]

I know how things work because I used them when I didn't

[winged tree]

because you can learn 'cfg should be 4-6[or insert whatever]' and 'learn' that... but you're not learning anything by following a guide that doesn't explain why that is 4-6 vs, 8-10. for example

[hardy cipher]

and figured it out

I don't agree at all, but okay

this just seems like some strange quasi-gatekeeping to me. I don't care what someone else wants to understand, lol. it's their choice. doesn't impact me

[winged tree]

lol gate keeping XD

ight

didn't come here to discuss, just wanted to drop

[hardy cipher]

I like to learn everything I can about whatever I'm doing. but I realize not everyone has that want. and I don't see any reason to really be concerned with that

[vast galleon]

https://civitai.com/models/119038/sdxl-controlnetinpaint-workflow check this out. currently working on V2.0

[winged tree]

what was that made with?

[hardy cipher]

extremely long frankensteined prompt

sometimes I'll just put two completely different prompts together to see what happens

or 3

[winged tree]

guessing auto?

[hardy cipher]

ooh. made that on dreamstudio. mess with that sometimes on my phone when I'm out doing things

[winged tree]

ah, that's why i couldn't import anything lol

[hardy cipher]

yeah, I noticed that. kind of annoying. you'd think the people that make the model would be all about having the prompts in the metadata

[hardy cipher]

man, this unclip workflow is stuck in some weird nether region, but kind of making cool things

[winged tree]

Not necessarily, not like they intend most of them to be made on that, and with the pace of development it adds extra layers and more work

[hardy cipher]

that is true

I can't be mad at these, not what I'm going for though

[winged tree]

haha nice

some of my first ones for the prompt, and not too upset by that

[hardy cipher]

yeah, he looks manic, lol

nice. that's clean

[gloomy barn]

all searge's workflows are the same...they run once, then won't run a second time unless I remove the feeder image and add a new one, even the same one

[hardy cipher]

I don't know what to tell you, bud. they've worked for me so far

[winged tree]

does any seeds change or is it laid out to generate the same result each time? because comfyUI will skip redoing stuff

[hardy cipher]

true

[winged tree]

personally haven't used it but sounds like that's the issue if other people are fine

[gloomy barn]

No idea. it';s the searge inpaint and searge img2img

[hardy cipher]

so are you changing things int he flow before trying it again?

[gloomy barn]

no, when I inpaint, it has to generate lots of images to get one that's good generally

the first one isn't always fixed...in fact, almost never

[hardy cipher]

well what he's saying is if the inputs or whatever are identical it won't rerun them

[gloomy barn]

so it doesn't change it each time like a111 or sdnext

So it's one chance to inpaint...that's it

[winged tree]

Only if you set the seed to fixed or such

[hardy cipher]

seed number?

[gloomy barn]

I have seed set, and then random variance

[hardy cipher]

random variance?

[winged tree]

feel free to post the workflow, could solve a lot

[gloomy barn]

it is the searge inpaint and img2img workflow

[hardy cipher]

yeah, haven't used that one myself

[gloomy barn]

I dont know how to post it. An image is unseeable...the workflow is huge

[winged tree]

I understand that, but it very well could be a connection being wrong

post the image

it contains the workflow

[gloomy barn]

To get the image in one screen, the names disappear

[winged tree]

erm... post the generated image

[gloomy barn]

Here's the one searge uses

there's that, and then the json

[hardy cipher]

tilt shift?

[gloomy barn]

I have to load the json - the workflow doesn't always load from the image

That's for img2iimg...note there's nowhere to load the initial image

so I have to use the json

Here's one I made with it.

[hardy cipher]

I don't think mine willload it because it's rendering something. wanted to check it out. not sure I can help though. but curious

[winged tree]

you can, just gotta go into your history after or queue and select load from that and it'll be back

[hardy cipher]

that's fine. it'll just be a few seconds. are you trying to make peaky blinders there?

[gloomy barn]

The seed is set to random, but it doesn't generate a second time. If I change anything, then it will generate again

[winged tree]

this seed?

there's a lot so idk if there's another in there

[hardy cipher]

just change it manually then

[gloomy barn]

also, that seed is set. It doesnt change with the uploaded image

So I cant generate the same image over and over to try to get good hands, for example

[hardy cipher]

doesn't it say "fixed" right in that screenshot?

[gloomy barn]

yes, but I mean when I set that to random

[hardy cipher]

simple workaround it to just change it manually

until you figure it out

[gloomy barn]

well...now its working...generating over and over whether it's fixed or randomized

Ah., no

[winged tree]

no idea, a lot didn't load for me because i don't have those plugins/extensions but glad its worki--

[hardy cipher]

bruh

[gloomy barn]

it generated 3x, then wouldn't generate again

[fleet veldt]

what are folks using for easy training right now?

[gloomy barn]

in sdnext, sometimed generate over and over until it gives one with good hands...each generation is slightly different\

in comfy, cant do it

,,,,sometimes

Just did it again...it generated twice in a row, then wouldn't generate again

[hardy cipher]

no idea why the image won't load for me, but it's not

[urban fjord]

I'm using Kohya-ss gui for training, it works and I'm too lazt to look for something else.

[fleet veldt]

anyone try onetrainer?

[fleet veldt]

heard of this, cmdline or .. nvm you answered. ill check it out again. it's been months since i looked at that.

[half ivy]

i am here
wot did i miss '

[pure crystal]

this is a shitty sdxl gen. we have really leveled up

[tall lion]

that would be a good idea

[hardy cipher]

no one knows. I'm just over even trying to figure it out, lol

I've read completely opposite things about it both being stated as valid

[half ivy]

surely refiner is anything not forcing a change, anything positive..
so like "detailed" doesnt have a shape and wont affect a space

[hardy cipher]

read it's not needed. read it's vital. read it can be the same as base prompt, then not. then that it should just include base prompt and a few words for fine tuning, etc

[vast galleon]

I got mask upscale working now.

[hardy cipher]

that is a clusterific mess, lol

[midnight shuttle]

Eventually someone will really understand how it works, and then they can explain why it is or is not needed with evidence to back up their claims.

[vast galleon]

Yup. Trying to keep it tidy but i need to get rid of wires and find out how to name parameters

[midnight shuttle]

Right now you just have people randomly spouting things without understanding.

[hardy cipher]

believe me, I understand. I like to do lots of image modifications between renders and it can get twisted very quickly

[vast galleon]

If anyone wants to play with it. Nested Nodes, Masquerade Nodes, and WAS Suite is required.

[hardy cipher]

I think I have all those. thanks. I'll take a look

[tall lion]

for controlnet and SDXL?

[vast galleon]

No controlnet or inpainting model required. It has edge detection and a negative blend in the inpainting mask to help keep features

[hardy cipher]

sweet. I'll load it in a fw minutes when this thing is done

[pure crystal]

well done, nice

[hardy cipher]

trying to figure out the unclip models. or at least sort of grasp them

[pure crystal]

https://www.youtube.com/watch?v=lVSNJYfvO9o&t=2s Anyone know this channel? Comfy workflows and custom nodes

[hardy cipher]

I'm apparently one of the few people on the internet that does not really like watching tutorial videos

[half ivy]

looks slightly better with straight lines

[hardy cipher]

lol

[half ivy]

dont 😂

[hardy cipher]

you don't

[vast galleon]

To blur a mask, use radius

[hardy cipher]

excellent work!

man, I can't get over how wild these unclip images turn out

[vast galleon]

how do you get the straight lines?

[half ivy]

it dont

[hardy cipher]

like, not saying they're the most mind blowing thing ever

but very interesting

I think I'm zeroing in on a sweet spot of some kind.

[half ivy]

cant remember 😂
tryna find it

[dark jasper]

hey anyone have good settings for lora training?

[hardy cipher]

not really there yet, whereever there is

ooh, last image of that round got borked,lol. stupid edge filter

I erally like how it just runs what's needed, so when it's something at the end it's super quick

[vast galleon]

control mask contrast with black_level and white_level

[hardy cipher]

alright, I need to check this out

[vast galleon]

Working on mask by prompt, but it's getting late

[half ivy]

@vast galleon think it was this one
https://github.com/pythongosssss/ComfyUI-Custom-Scripts

[hardy cipher]

lol

When loading the graph, the following node types were not found:
Model_Inputs
Text_Encoders
VAE_Inpaint_Encoder
Inpainting
Edge_Detection
Blend_Node
Samplers
Mask_Upscale
Nodes that have failed to load will show as red on the graph.

I thought I had most of the nodes, but missed those I guess

[lilac raven]

Is there a way to know if a model knows a specific term?

[half ivy]

same same

[vast galleon]

looks greatt. im gonna pick it up

[vast galleon]

You will need masquerade nodes, nested nodes, and WAS suite

[midnight shuttle]

I don't think so unless you have access to the original training data. By the nature of the model, it "knows" all words. The words will have some impact on the generation even if they are nonsense.

[hardy cipher]

hmm. got another error. I'm sure it's resolvable

Loading aborted due to error reloading workflow data

TypeError: Cannot read properties of null (reading 'type')
TypeError: Cannot read properties of null (reading 'type')
at ComfyNode.inheritPrimitiveWidgets (http://127.0.0.1:8188/extensions/ComfyUI_NestedNodeBuilder/nestedNode.js:186:26)
at ComfyNode.nestedNodeSetup (http://127.0.0.1:8188/extensions/ComfyUI_NestedNodeBuilder/nestedNode.js:84:14)
at ComfyNode.onAdded (http://127.0.0.1:8188/extensions/ComfyUI_NestedNodeBuilder/nestedNode.js:124:18)
at nodeType.onAdded (http://127.0.0.1:8188/extensions/tinyterraNodes/ttNwidgets.js:392:49)
at LGraph.add (http://127.0.0.1:8188/lib/litegraph.core.js:1453:18)
at LGraph.configure (http://127.0.0.1:8188/lib/litegraph.core.js:2232:22)
at ComfyApp.loadGraphData (http://127.0.0.1:8188/scripts/app.js:1199:15)
at http://127.0.0.1:8188/extensions/ComfyUI_NestedNodeBuilder/nodeMenu.js:193:17
This may be due to the following script:
/extensions/ComfyUI_NestedNodeBuilder/nestedNode.js

[vast galleon]

trying to think if there are any requirements I might have forgotten

[hardy cipher]

well I have nested nodes and thought I had all that node packages

[half ivy]

a short story 😂

[hardy cipher]

I'll figure it out and let you know

[vast galleon]

maybe some of them have been updated?

[vast galleon]

Check out #✨｜sdxl message

install missing custom nodes

[hardy cipher]

I'll look when this thing finishes

I just odn't even know what to think of these results, lol

[vast galleon]

it takes some finangling

Boolean will invert the mask. But I just realized blur won't work if the mask is inverted, so I will have to fix that.

[hardy cipher]

and now this guy? what?

[half ivy]

did this but there were no options to install, got all the ones you mentoned, WAS etc

[vast galleon]

bare with me. gonna un-nest and try to see what's missing. everything is working for me

[hardy cipher]

I have all those node packages though, I use a lot of them

I'll just restart. that usually helps

well restart the comfy

[half ivy]

done that already

[hardy cipher]

once these apparent updates finish I'll restart

[half ivy]

also not the 1st time, might be a comfy thing

[vast galleon]

these are the only extensions I have

I wish it said which node the error was coming from

[hardy cipher]

it does

well for me

[vast galleon]

Oh, ok. those are all of the nested nodes I built. how do I make them available?

[hardy cipher]

"there is an extension that needs updated"

lol, thanks for the deets

[vast galleon]

put these into \ComfyUI\custom_nodes\ComfyUI_NestedNodeBuilder\nested_nodes

[hardy cipher]

oh yeah! that does make sense

I have mixed feelings about nested nodes, or maybe i'm not using them right. sometimes they're great, but other times they'll do things like swallow a parameter adjustment or output display or something

[vast galleon]

I am not fond of them either. I plan on going through and trying to build my own custom ones, but I'm trying to build it first

[hardy cipher]

yeah, how do you go about starting that process? I'm curious. I have some ideas floating around in my head, but pretty new to this particular interface so not really planning on diving in immediately

[vast galleon]

youtube tutorials lol. also I have experience with blender and houdini so I'm learning quickly

[hardy cipher]

only started learning about python because of stable diffusion, lol. but I'd taken a few random programming classes in the past so the logic is still in my head

and nodes satisfy my weird brain. and allow me to quantify some of the strange things I'll do with photoshop

not so much any particular single thing, but I've done so much as far as layers and filtering. but these nodes get out of hand quickly

[urban breach]

Thanks, is anyone else getting a weird primitive box underneath the "Image Source". It's already has a list of files that were in the JSON but not files i have

so it won't run because the "file doesn't exist" type thing

* LoadImage 131:
  - Custom validation failed for node: image - Invalid image file: NINTCHDBPICT000270813068 (1).jpg

[hardy cipher]

maybe modify the json?

[urban breach]

🤷‍♂️ just seems unintended, you upload a picture

[half ivy]

File "<frozen importlib._bootstrap_external>", line 1073, in get_data
FileNotFoundError: [Errno 2] No such file or directory: 'C:\Users\34\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_NestedNodeBuilder\init.py'

Cannot import C:\Users\34\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_NestedNodeBuilder module for custom nodes: [Errno 2] No such file or directory: 'C:\Users\34\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_NestedNodeBuilder\init.py'

[hardy cipher]

alright. I'm going to restart browser as well

[vast galleon]

That's a test image. You can load an image through the load image node in the inpainting group and then select the image with the primitive under image source so you have a view of the source image after the mask is painted on the one in the inpainting group

[hardy cipher]

it loaded fine but got some errors. scanning them to see what they are

[tribal jackal]

Trained my first lora on one of my dogs today! Pretty solid results, but two "issues". My lora is 1.7GB and the execution time jumps from about 20s to anywhere from 300-500s

[vast galleon]

try that one, maybe it changes when you make a nested node?

[hardy cipher]

ooh, zinger ModuleNotFoundError: No module named 'global_table'

[urban breach]

You upload a picture to the 2nd node and it should show up in the 1st node primitive?

[tribal jackal]

It works, but I was expecting a smaller file size from a lora, and not sure what common causes might lead to extremely long load times

[hardy cipher]

I wonder if there are any others

[rich coral]

yah i just made one too, HUGE! Wonder how the citivai ones are much smaller

[tribal jackal]

high five!

[urban breach]

(i saved it after adding the mask)

[vast galleon]

it should, but I don't know how without using the image primitive underneath

[vast galleon]

that makes sense

[hardy cipher]

ModuleNotFoundError: No module named 'comfy.ldm.models.diffusion.ddpm' tbh not exactly sure how to resolve that one

[autumn forum]

use network rank 8. the network being high will make huge file sizes

[tribal jackal]

Ah, I went with 256 same as the tutorial I think lol

What's the rank tradeoff again?

[autumn forum]

the tutorials are wayy wrong. lol. network 8 and alpha 1.

[tribal jackal]

Thanks bud. Yeah, I made some personal tweaks along the way as well.

But only when I was confident it wouldn't hurt

[autumn forum]

sdxl is very good with loras just dont go much heigher than network rank 8. cuz itll make huge files that you dont need. mess with learning rate and unet learn. i went from 0.0001 to 0.0004 on both learning rate and unet learn rate it made differences but not like a complete burn so try as you wish.

[tribal jackal]

I might be able to mitigate the lower rank training tradeoff if I increase the subject sample image data set size. I want more variation in blankets and background for him organically anyway.

[midnight shuttle]

LoRA is low-rank adaptation. Apparently 256 isn't low enough. 😆

[urban breach]

I converted image back to widget for the first one? seems to be working now. without that primitive. i'm not super familiar with all this yet

[vast galleon]

I took the primitive out. I like your method better

[tribal jackal]

I also normalized each of the learning rates to 0.0004

[finite moth]

whats the difference between a checkpoint and a Lora?

[urban breach]

i mean for a first generation with adjusting nothing 😀

[midnight shuttle]

A checkpoint is a whole model that can generate images by itself. A LoRA is extra data that changes the images a model generates.

[tribal jackal]

loras tend to be smaller, and can be used in tandem with any checkpoint

[finite moth]

dope

[vast galleon]

Lol face of hulk hogan

[urban breach]

i just threw a random image i saw on here in there lol

[vast galleon]

that's funny that it brought some of my test images into it

[urban breach]

yeah it must save the history for some nodes

[ionic gulch]

offset lora is helpful to hide problems with faces and hands:
https://i.imgur.com/Z3y448w.jpg
https://i.imgur.com/cFKHaLe.jpg

[urban breach]

that red sash thing looks great

[ionic gulch]

more red sash:
https://i.imgur.com/nqR5rXv.jpg

it would look even better with less offset, but you can see the broken faces without the heavy contrast.

[urban breach]

i have seen the offset lora in ComfyUI but with the weight set to 0, is that correct? not sure how to use it right

[hardy cipher]

lol RuntimeError: Added route will never be executed, method HEAD is already registered

ugh

[ionic gulch]

@urban breach to get the dark sillouette you have to use negative values.

the 3 images use -0.8 to -1.5

[urban breach]

is that for the model or clip strength or both?

[ionic gulch]

i've made a simple workflow that changes both values at the same time - so both for those pictures.

[urban breach]

ahh ok

really cool style

[ionic gulch]

maybe you can get better results if you try to change them separately.

very small negative values help in a lot of cases. even -0.05 can make a difference, if you just want a higher contrast in regular pictures.

and if you set it to 0 the lora isn't used. so you can keep it in the workflow all the time.

[urban breach]

cool where do you put that in the workflow, right after checkpoint?

[ionic gulch]

yes and only for the base model at the moment. haven't tried refiner because people said lora doesn't work well with refiner.

[urban breach]

ah interesting

[ionic gulch]

i still use refiner for the last 20%, but without a lora node.

with small negative values you get "glowing" effects like that:
https://i.imgur.com/JFzpTKG.jpg

[urban breach]

at last a good way to hide the 6 fingers on my snowy vampire girl 😂

actually it looks more like an assortment of random fingers under the gloves now lol

[autumn forum]

in 22 minutes we will see if my lora training of UnBlurTool will work! time to squash the pesky blur everyone has been annoyed with

[hoary saddle]

any good solid updates from @nimble heart or @high skiff with their node examples?

[high skiff]

Submitting 1.0 workflow to Comfy soon for approval for the wiki

have some full quality image examples of what my new upscale workflow can do

These are all 2048x res

This one is 4096x

[ionic gulch]

@high skiff do you use "target width" and "target height" for your upscaler?

[high skiff]

I do, but not in any particular way

[hoary saddle]

can't wait to test drive it!

[high skiff]

as in, I set the value to whatever

current node look

[polar epoch]

Tried making a ultrawide image.. well, made a meme instead lol

[high skiff]

oh me oh my

looks like

https://tenor.com/view/distorted-cat-meme-distortion-cat-black-cat-meme-gif-14670421

lmfao

same energy lol

[spring fulcrum]

@high skiff I really hope you drop that sooon... I am pretty much not using SDXL until I can work with a better workflow.

[polar epoch]

Here, have the workflow lol.

[high skiff]

Goal is for later today, early tomorrow

[polar epoch]

Don't mind the 80MB size, it's just native 1440p upscaled 4x lol

[boreal bough]

squiiiiiiiiiiish

[spring fulcrum]

You are awesome... thank you

[ionic gulch]

here is some information about correct values for target height/width: https://arxiv.org/pdf/2307.01952.pdf - just using 4096x4096 works but it should be better to use values caclulated by that information

[autumn forum]

most high res cursed image ive seen. i love it

[polar epoch]

Got a better squish for ya

[boreal bough]

oh god why XD

[autumn forum]

failed diffusions needs to be more active lmao

[polar epoch]

Have my first attempt at text. The pecsnort!

Or is it pecsnourt

[high skiff]

I and many people have messed with it, it doesn't make much of a reliable difference across different prompts and stuff

[boreal bough]

when you forget to tell your lora that you're describing a game character - and sdxl decides that it's a cat/labrador mix XD

#notEvenMad

[high skiff]

crash!

jk jk lol

Ratchet!

[high skiff]

that is too cute haha

[polar epoch]

Now to find the right CFG to not have face slaps all over lol

[tribal jackal]

Oooh I wonder if that's why it spits out so many errors but still produces something

[polar epoch]

Is there a benefit to have leftover noise with comfy?

[tribal jackal]

Did that with the base sdxl offset lora and my new one

I'm still getting good results, but if it speeds things up that'd be great

Granted, mine's a character lora not an offset lora

[polar epoch]

I just add lora for both base and refiner, somewhat works with sub SDXL loras

[ionic gulch]

for style/offset it shouldn't matter too much if you don't use lora for refiner. for faces/celebs it's more problematic, because the refiner without lora will remove some of that information.

[tribal jackal]

After removing from refiner

0.06s difference

[polar epoch]

When i attempted zaraki kenpachi, a 1.5 lora with SDXL. Didn't get it quite right at all, but close enough lol

[hoary saddle]

didn't realize you could use 1.5 lora's with it, i gotta try that

[polar epoch]

I just throw around different loras and see what sticks

Ariel for instance

[gloomy barn]

How did you get this. I am right now trying to get characters facing away from the viewer. Can't get it to happen.

I've tried walking away, walking away from camera, walking away from viewer, pitching to a batter, serving to an oppoent in the distance, - always facing the camera/viewer

[polar epoch]

that one was just random. Drop it into comfyUI and you can see it's workload and prompts. But generally just
Positive: "facing away", back towards camera
Negative: Looking at ccamera, facing camera, eye contact

And such

[charred beacon]

When I put facing away, it usually works but they don't look back

[gloomy barn]

I'll try the negatives

[urban breach]

i use "rear view" and "rear side view"

[boreal bough]

"view from behind" <- but biased towards men.
or "back view"

[polar epoch]

This is gonna take forever ultrawide native res with 80 steps

[tribal jackal]

This is incredible.

[polar epoch]

One last attempt, then bedtime at this damned 6am's..

[charred beacon]

I like making cursed images

[gloomy barn]

tried all of those...still facing the camera

[glad flint]

"back turned" but it can impart thematic coldness to some prompts
a weird hack can be adding prompts that focus on "hairstyle", especially for women characters

[gloomy barn]

I have (rear view:1.4) in positive, forward view in negative along with facing the camera, eyes to camera, looking at camera

[glad flint]

I accidentally found out (at least in 1.5) that trying to specify hairstyles ended up with a character with their back completely turned some of the time

[polar epoch]

Oh me too! Have a toothless kitten i tried to give shark teeth, well, chimera happen

[charred beacon]

Have a kid eating beans with his soul leaving his body

[polar epoch]

Have a cricket i made earlier today that i gave a hat, but also half turned into a slight cat lol

[gloomy barn]

tried each one and then all in the same prompt...still facing the camera

[boreal bough]

it's not ultra ultra wide until it's 512x2048 XD

goold old 1:4

[polar epoch]

Rookie numbers

3440x1440 x4

[charred beacon]

Well I'll be damned

[gloomy barn]

of course when I make a simple prompt "man with back to camera" that works

[ionic gulch]

offset lora?

[gloomy barn]

I'll be honest, it was less painful having a broken neck than trying to get this walking away from camera

[boreal bough]

dont forget that many prompts are biased towards front facing, so trying to overthrow front facing, by using more backfacing words usually just leads to worse quality. better try trimming, or rewording the original prompt

[gloomy barn]

I put "faciong away from camera" in the front of the prompt. It made the torse facing away, but the legs facing to the side

[boreal bough]

wait a minute XD

[polar epoch]

Finally. 1440p divided by 2, then upscale by 4x

[gloomy barn]

maybe when there's something like controlnet for sdxl 1 it will have the ability to rotate something - we'll all be pleasantly surprised

[livid basin]

So what do people think about 1.0 refiner?

[gloomy barn]

True...I see the quality is down when it has the torso facing away

[livid basin]

Good bad? Base only?

Or do peeps really think refiner helps

[gloomy barn]

for me, anything base looks blocky

refiner makes it not blocky

[charred beacon]

Is it bad to not do any upscaling? Biggest I make is 1920x1080

[polar epoch]

here we go. Wallpaper worthy

[boreal bough]

if you go the way of loras, then you're stuck with base only.
Ideally I wish I we could use refiner for everything, as it definitely is worth it when used right

[trail obsidian]

Hi guys having some trouble. Whenever i go less than 0.4 de-noise on high res fix i always get a weird result with the picture quality? Meanwhile others using it just fine. What am i doing wrong?

[gloomy barn]

you cant use refiner with lora?

[livid basin]

I guess you can, but the LoRA would apply to base only

[trail obsidian]

custom models like dreamshaper dont need refiner

high res fix is great if it works

does bascially the same job

[livid basin]

What solver do we use for the refiner? DDIM?

DPM?

[trail obsidian]

ive seen best results with Base: 20 Steps, CFG 7, DPM++ 2s a

Refiner: 15 steps, CFG 8, DPM++ 2M,

[livid basin]

I'm using dpmpp 2m karras for base

[trail obsidian]

issue is low denoise fucks the image quality for me for some reason

[livid basin]

But I've heard dpm is not good if you use <15 steps, apparently it can fuck it up with few steps

[trail obsidian]

try DPM++ 2s a most use it results are the best imho its euler a for sdxl

[boreal bough]

both ancestrals converge pretty fast, between 35~50 steps.
as for the steps before, their difference is genuinely minor (base only experience, where I generated enough comparison grids)
refiner may be a different experience

[trail obsidian]

Yeah was rather meaning euler a was the popular rockstar of SD 1.5 now i feel its DPMPP 2s ancestral

[livid basin]

ok

I've ended up with DPM++ 2M for both base & refiner

With lower_order_final=True

[trim orbit]

ii use whatever and it's always coming out good. hard to determine whats best. even old ddim is fast and great

[livid basin]

50 steps base, 15 steps refiner

[polar epoch]

Alrighty, time for a break for tonight. Gonna go wilder tomorrow lol

[livid basin]

without lower_order_final, 15 steps fails with DPM++ 2M on refiner

[hardy cipher]

orly?

[boreal bough]

if you ever have to go base only, then use this though
(steps 25 or 50, depending if the extra bit of quality is worth the double time for you)

[heady vale]

If only this page was finished https://blenderneko.github.io/ComfyUI-docs/Core Nodes/Sampling/samplers/

[hardy cipher]

is that so?

[boreal bough]

you can use the infinity grid from mcmonkey to make your own comparisons

or use this for reference
#✨｜sdxl message

[nocturne dove]

There is a lot of information here: https://stable-diffusion-art.com/samplers/ although not for sdxl.

[heady vale]

Yeah saw that but doesnt have dpmpp_sde_gpu, which Im curious about

[boreal bough]

oh yeah that ones new XD wasn't there when I generated the graph

[vale eagle]

Do we have something like xyz grid in comfyui?

[boreal bough]

StableSwarmUI - then give that the port of your comfyui. then you can gen all graphs

[boreal bough]

https://discord.com/channels/1002292111942635562/1134546054872842290
Has its own channel as well, if you get stuck on how to use it

pretty simple though

[spring fulcrum]

Does anyone know how to get ComfyUI to save images to a folder outside of the comfy folder? I would like to save my images to a folder on a server I have and it is not in the Comfy folder to say the least... Is that something that can be changed in the python file "folder_paths.py"?

[tribal jackal]

lora I trained on one of my dogs

[heady vale]

edit the bat file and add something like this at the end of the first line
--output-directory D:\SD_Output

[spring fulcrum]

Awesome, that's easy enough to do... Thank you.

[boreal bough]

just checked code. it calls for the output path variable... so it should be configurable somewhere

[spring fulcrum]

That worked perfectly... Thank you

[urban breach]

The new Marvel series, "Electric Dog". Coming to a streaming service near you!

[turbid garnet]

how much gpu mem is needed for sdxl base 1.0?

[urban breach]

Anyone have a pun for Electrocute + Dog

[zinc cargo]

something to do with davedevil comics

[boreal bough]

around 8

6 should be barely possible now, though I haven't tested it

[tribal jackal]

He's got a thunderous bark, but that's nothing compared to the bite.

[trim orbit]

hard to resist those eyes

[peak dove]

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

[tribal jackal]

You're telling me! I spent 6 hours deciding which ones to cut and what to crop out of the sample set!

Like having to say no to that face a thousand times

[trim orbit]

dataset building is a dogs life

[peak dove]

Scott Detweiler's Refiner-Base-Refiner Workflow - Triple-Proces R-B-R

[tribal jackal]

the payoff though

[dense chasm]

does anyone knows how to run sdxl lora training on colab?

[rich coral]

That is pretty impressive. My first go has been very satisfying too. Can't wait to make a bunch of these and start mixing them. strange amazing times.