✨｜sdxl

[nimble heart]

brb googling parts list for the Luty smg

[vital ermine]

Then consider not getting the 600. You have fought the hard fight and it sucks but mega corps suck ass juice anddo jack shit so getting something is better than getting fuck and left for dead in a ditch by them.

[high skiff]

I don't have that as an option

its not my money to just forfeit

[vital ermine]

not up to you is my point and it sucks

get robbed and they then screw you some more. Seen it, been a victim to it, numerous times and it is getting worse.

[high skiff]

If they don't give me my money back tomorrow, I will reverse the transaction

then if they send collections after me, I have grounds to sue, so try me

[vital ermine]

good luck as that used to work though can take up to 120 days or some long ass time.

a sig to that one

Furry

[pure crystal]

even the watermark looks cool

[vital ermine]

yeah

I tried to stick an AIT node right after my checkpoint loader node and it had a ton of errors. I thought that is how you do it?

weird error

[shy kelp]

does it have different result, without and with ait

[vital ermine]

no idea I can't get ait into my workflow

[strong field]

Did you have control net in the loop? What was the error?

I’m troubleshooting weird AIT errors popping up recently

[vital ermine]

no controlnet in comfy at all for me yet

it happens when it hits the ksampler

works fine the minute I delete ait

[strong field]

Oof, did you update comfy or AIT recent?

[vital ermine]

the v2.json from the download site works as well

I think so

I know I did ait yesterday

[strong field]

This looks related to all the shenanigans popping up with memory

Unfortunately

Thanks for the info

[vital ermine]

I did roll back my driver to 531.79

did a clean ddu newest version

[shy kelp]

i want to do a model comparison, what sampler should i use

[strong field]

Yeah, smart memory management seems to part of the mix too

[vital ermine]

Nvidia did get on a thread I was on and admit they know the issue to be fixed in a future update. I lay odds the fix will be a toggle we turn it on or off in the control panel.

[strong field]

Lol

[vital ermine]

gamers gotta game

[ionic gulch]

https://i.imgur.com/hoFq7nY.jpg

[strong field]

Dpmpp 2m Karras is a favorite

[nimble heart]

if you're using the refiner, ddim/normal is what it was trained on I believe. karras scheduling wrecks the step swap

[strong field]

Oo care to explain this to a dumb dumb?

[nimble heart]

comfyui?

[strong field]

Yes

What do you mean by step swap

[nimble heart]

so when using the refiner, the intended way to get the best results is to stop the base model before it finishes denoising then switch to the refiner for the remaining steps

done using KSampler Advanced

so for base model: total steps 30, start 0, end 25

refiner: total 30, start 25, end 9999

it'll run 25 steps on base then 5 on refiner

[strong field]

Oh I see, and Karass screws with that left over noise?

[nimble heart]

yes

or something like that

not a scientist

I didn't go to highschool

[visual glade]

what I just pushed should reduce VAE memory usage by a bit on 3000 series and up

[nimble heart]

bf16 commits right?

[visual glade]

yes

just need more people to test: --bf16-vae to see if I can enable it on more cards

[nimble heart]

I may as well test it again on my xtx but I think the ROCM support for it is sketchy

[visual glade]

need someone with 2000 series to test if it lowers memory usage

and yes AMD too though I'll give it a test on mine tommorow

[nimble heart]

last time I tested it actually increased memory over fp32

that was rocm 5.5 though

[glad grove]

does sdxl still eats 24gb of ram or was that fixed?

[visual glade]

yeah on ComfyUI it runs on free colab which only has 12GB ram

[upbeat summit]

@strong field I also noticed since a recent update to ComfyUI when using AIT + LoRAs and you tweak the strength (not sure yet exactly how to reproduce it), the image suddenly looks like you are using the LoRA at 5.0 strength - so 100% overcooked.

after bypassing and re-enabling the AITemplate Loader node, the problem is fixed. this could be some memory management issue? maybe...

[nimble heart]

using 5.6 latest comfy it still trips the slower vae decode earlier than full precision vae does. Tiny bit faster on 1k but even 1.5k causes it to use the slow decoder.

[visual glade]

I guess AMD doesn't have real bf16 support then

[strong field]

Do I need to run with args on 3000 series?

[visual glade]

no if you update (git pull or update/update_comfyui.bat) it's enabled by default on 3000 series an up

[strong field]

That is bizarre

[nimble heart]

I saw commits for bf16 stuff on rdna 3 in the composable kernel repo a little while ago, but that might end up only on the pro models

[shy kelp]

3 model comparison
dpmpp2m karras
seed:473478811476380

[glad grove]

sensei 🙇‍♂️

[nimble heart]

it claims rdna 3 cards have bf16 compute https://gpuopen.com/learn/wmma_on_rdna3/

so maybe it's just another Eventually™️

[upbeat summit]

after restarting ComfyUI or as I just found out, bypassing and re-enabling the AITemplateLoader node, it generates correctly again

[strong field]

Sdxl is remarkable

[strong field]

getting late here, ill have more time to review tomorrow

[upbeat summit]

I think this could all be related somehow

[strong field]

it certainly seems that way

[shy kelp]

you guys are genius

[strong field]

wow what did you do, its NOTICEABLEy better (rhetorical)

[nimble heart]

apparently bf16 stands for "Brain Float" which is certainly one of the acronyms

[nimble heart]

--bf16-vae is the default

maybe one day it'll work on navi 3 cards too...

[strong field]

@upbeat summit can you try a SD1.5 lora in your workflow, testing purposes only

should load QUICk, but throw a bunch of warnings in console

[visual glade]

fp16 is unsuitable for the VAE because it overflows way too easily and gives black images so bf16 is better for it

[strong field]

which means this is isolated to SDXL

[nimble heart]

wonder if you could use bf16 on attention too to avoid issues like SD 2.X's black gens with autocast

[strong field]

isnt AIT using fp16??

[visual glade]

fp16 is good for the unet

and kind of works with the SD1.x VAE but it still gives black images sometimes

[nimble heart]

I found only the anime vae gives black on fp16. the 1.5 mse vae is always fine with it

[strong field]

and do we care about the precision issues on bf16?

[nimble heart]

nah

[shy kelp]

this is crazy

[visual glade]

I compared and there's no visible difference between bf16 and fp32 for the VAE

[strong field]

excellent news

[shy kelp]

which one is better

[glad grove]

the one with the detachable legs

[shy kelp]

HAHA

[strong field]

@upbeat summit last thing for me today, but digging through the commits, it looks like there is an issue with lora conditioning in AIT, since change in Comfy/lora.py

That hypothesis fits the behavior we are seeing

[upbeat summit]

yeah but seems it has changed for 3000 series cards. 4000 series cards already had that problem before the latest update

[strong field]

yeah that part I havent gotten to yet, lol

Maybe it was always screwed up for 3000 and i just didnt notice, that makes more sense to me, than the change in behavior

[upbeat summit]

well maybe we have some good data at some point to file an issue in one of the repos or even find a fix 😉

[strong field]

this is what im hoping

still gotta tackle more tomorrow

[upbeat summit]

it's really not fun to use AIT + LoRAs right now

[upbeat summit]

thanks for looking into it, too!

[visual glade]

it's probably just patching the model weights more than once

[turbid garnet]

When can we expect SDXL Tensort RT as an HF pipeline - https://huggingface.co/stabilityai/stable-diffusion-xl-1.0-tensorrt

[strong field]

i was just about to ask you about the weights, so it looks like im on the right track

Where do your model weights come from, AIT using CompVis

only ask cause you might save me time digging through the code

am i looking in the right place 😉

[visual glade]

no

that's to get the weight values when lowvram is applied

[strong field]

yeah i just read that 😦

[upbeat summit]

glad you got much more understanding of it than I do. I just tried running with --highvram and issue is gone here too just like with --gpu-only

[strong field]

is it stable?

[upbeat summit]

running a bigger batch now with --highvram

[strong field]

Just trying to understand how you handle weights different from AIT, if you got a quick answer I'd be grateful, but if not thats ok, Ill keep digging

[vital ermine]

Diabeetus

[visual glade]

ait is probably just mistakenly calling the .patch_model function when it should not be

[strong field]

youre not wrong

[visual glade]

it gets called in the load_models_gpu function now

[strong field]

THANK YOU

[mellow dome]

LMAO

[upbeat summit]

so far so good. running AIT + 4 loras now with --highvram. after 40 images no delays or freezes so far. will continue doing tests.

*edit: tried 7 loras now - still working. so looks stable so far. 7 loras only reduced the overall gen time by ~1.5s

[glad grove]

cats sequel lookin nice

[terse pewter]

Amazing..Thank you

[steady grove]

god emporer cat

[vital ermine]

Not sure what that one's lineage is but sheesh

[tepid surge]

makes me think of the desert from full metal alchemist for some reason

[vital ermine]

Interesting

I am playing around with nodes to see what all these can do

well

[tepid surge]

That's a mood, i'm between playing Homeworl: Deserts of Kharak (which is prob still free on epic atm if you or anyone else wants to nab it) or generating mech stuff casue i've been watching a playthrough of the armor core campaign and giant nimble robot noises cx (sry i slow on the typing today, scatter brained)

[vital ermine]

No worries

[uncut fiber]

i installed manual comfui, if any change. Getting this error, i think i got same in portable
missing {'cond_stage_model.clip_l.text_projection', 'cond_stage_model.clip_l.logit_scale'}

[unborn stirrup]

Hi guys, I am now finetuing sdxl model, I used kohya ss, i am still pullzed with V pred like loss this parametre. Do you know the meaning of this number

[indigo carbon]

just compiled AIT modules for Cnet. is there other types except Canny?

[uncut fiber]

i made some weird things but getting what i want. Must test without that nonsense node connection 🙂

[shy kelp]

what type of this art style called?

seems fun

[vital ermine]

I forget but it was done on old posters in the 1920-1940

[gusty canopy]

peaceful north american forest with little girl sleeping under a tree, very lush and beautiful

[indigo carbon]

is there a node for drawing an image similar to how the sketch feature in A1111 worked?

[uncut fiber]

i need bitten mushroom, cant get it.
@shy kelp dl it and put it in A1111 interrogation in IMG2IMG
To me it look familiar as well.

[indigo vine]

gave it a try, got that

[uncut fiber]

😄

[west breach]

what about 'half eaten mushroom'?

[uncut fiber]

will try thank you @west breach

[indigo carbon]

@west breach is there a node for making sketches like in A1111?

[west breach]

I haven't used one like that, I just used mspaint

[indigo carbon]

I feel like ComfyUI needs to add some basic features like that. it's somewhat weird for people moving from A1111 that are used to making a sketch in webui to have to open a separate program and do it there

[vale forum]

@junior grailtnSeraph you there?

@junior grailtnseraph you there?

[west breach]

have you edited mask with the load image node? I wonder if you can then convert mask to image

[vale forum]

Can we talk?

[austere ferry]

Hey, if I have an SDXL | ComfyUI tutorial, can I drop it here?

[indigo carbon]

didn't try that, will try

[indigo carbon]

there is no option to edit mask. but it is possible to convert mask to image

[west breach]

if you right click on load image node you can open mask editor

[indigo carbon]

is it possible to change colors like that?

it only does black

[west breach]

i think cause it's supposed to be a binary mask

I haven't played around with controlnet much yet, does this seem normal?

[indigo carbon]

controlnet isn't the way to do sketch

you do VAE encode to an image of a sketch, then connect to Ksampler

[ivory blaze]

has anyone ever zoomed out too far using deforum?

[west breach]

but the apply controlnet node takes an image as input

[ivory blaze]

cos, it said, screw the image, here's the artist's tools

guess i zoomed out too far :\

[indigo carbon]

it's one way to do it, but that's not how it was done on A1111

[upbeat summit]

btw. I tried running ComfyUI with --highvram and it fixed the issue for me for now. it's stable so far after a test run of 100 images with 7 loras in the mix.

[indigo carbon]

that's irrelevant for someone like me, I have a 4070ti

highvram isn't meant for it

[upbeat summit]

sure - but it's another piece of information

[shy kelp]

3 models comparison using same seeds and samplers

[upbeat summit]

@strong field is also taking a look at it since it now happens on 3000 series cards as well

[indigo carbon]

idk, currently working on CNET for AIT, I managed to compile some modules, but I'm unsure about some components of CNET

[nimble heart]

sketch was only just added to webui like two months ago

unless you mean masking which you can do in comfy with right click

[uncut fiber]

@upbeat summit where to put --highvram in Comfui? And how many VRAM do you have? Thank you!

[nimble heart]

probably wouldn't be super hard to make a colored version of the mask paint node

[upbeat summit]

24gb and you start ComfyUI with it as a command line argument. python main.py --highvram

[uncut fiber]

oh thats 3x more 🙂 Thank you very much just hoped it can help with my issue 🙂

[upbeat summit]

what is your issue?

[nimble heart]

If you only have 8 don't use high vram

[uncut fiber]

still same, AIT getting some errors, but reinstall to different place doesnt help. Some errors with importing but i think i get rid off it. will post what problem i have in cli

Not sure it is related
missing {'cond_stage_model.clip_l.logit_scale', 'cond_stage_model.clip_l.text_projection'}
yes i have normalvram when running it

[upbeat summit]

what GPU do you have?

[uncut fiber]

3070

it used to work perfectly

just now it running slower than without it

[upbeat summit]

I had this problem before, but I can't remember what it was. either I had an in-compatible model inserted like clipvision or it was something with the latent image

[upbeat summit]

I guess you will have updated ComfyUI to the latest version. Maybe delete the AIT custom_node directory and re-download it and try it with Tdg8uU's workflow https://civitai.com/models/133818/tdg8uu-aitworkflow?

[uncut fiber]

o.k. I will try it, thank you very much @upbeat summit !!!

[upbeat summit]

you're welcome. let me know how it goes!

[uncut steeple]

Testing a slightly different approach right now If anyone wanna try

[ivory blaze]

cant you also use

[proud lance]

1 girls tall big

[strange mist]

anyone else getting this error with latest comfyui update?

  img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8))```

[uncut fiber]

@upbeat summit unfortunately still same issue. But really thank you for your effort. I will live without it as performance could be worse. 👍

[uncut steeple]

#1072220168534642768

[uncut steeple]

Did you run the full update bat?

Did you update your extensions

[strange mist]

updated using git pull, same for extensions.
tried to run revision, that's when I encountered the error. it gave me blank images.
then tried to run the simplest workflow. error persisted.
now disabling custom nodes to see if it works.

[uncut steeple]

Non AIT vs AIT, 1024 gen is faster but the img2img with 2048 is slower

[uncut steeple]

Dont you run the bats?

[strange mist]

I usually just git pull

[uncut steeple]

id use the update bats since its suggested

[shy kelp]

which file

[native knot]

There's also a Depth module, don't forget that one.

[uncut steeple]

I dont think you have to do that, comfyUI does set up the settings itself according to your GPU afaik

[indigo carbon]

I was asking about architecture. AIT isn't checkpoint specific.

Do all SDXL CNETs share the same architecture?

[indigo carbon]

I actually made a specialized AIT workflow for SDXL. Currently trying to implement CNET

It's accepted as the official one for that node

Sadly, it's currently not possible to use AIT on IPAdapter or CLIP vision. But I was able to compile modules for canny. Not sure if I also need for depth because they seem to share the same architecture.

[idle zenith]

any idea what's wrong?

was trying to make this: https://youtu.be/sCOsZFH7xlU?si=wv5sP-uH2iu3zjd6

[uncut steeple]

The RGB mouse curser kills me

[uncut steeple]

Thats a 1.5 model and Lora? This here is the #✨｜sdxl channel

[idle zenith]

sorry but i couldn't share it to sd chat so tried here

[glad grove]

change vae

[idle zenith]

i didn't use any vae with it

gonna try with vae now

[glad grove]

either change vae or increase the high res fix denoising strenght from 0.3 to 0.45 or 0.5

[idle zenith]

tried both way.. it's the same

zombie look

[strange mist]

combined these two and got this:

no text prompt, 50/50 image weight

🤔

[uncut steeple]

You can use #🏞｜general-with-images for now If you want

[uncut fiber]

actor A is playing person B 🙂 i like it

[south frigate]

Is there any site that will convert a prompt from style 1.5 to SDXL?

[upbeat summit]

hmm that is strange

[upbeat summit]

I debugged this a lot. This fixes the 5-10sec delay that happens on every image gen when using AITemplate + LoRAs on a 4000 series card. Of course it only works right now for cards with lots of vram. but it might help track down the overall issue

[uncut steeple]

Interesting

[upbeat summit]

I don't know how you've set up ComfyUI, but if you're having the AITemplate + LoRA delay issue and have a GPU with 24gb vram, you can launch ComfyUI with python main.py --highvram or if you use the windows portable version you can edit the run_nvidia_gpu.bat and add the parameter --highvram there.

I'm not claiming this is a good fix for the issue and you might run into memory problems when working with multiple models or certain upscale workflows, but it currently fixed the 5 - 10 seconds delay issue for me when I use AITemplate and LoRAs at the same time.

[smoky patrol]

That's beautiful

[smoky patrol]

Ah, so familiar

[upbeat summit]

yeah. the performance gain of using AIT is almost reduced by 50% because of this delay. I'm not the only one with that issue, so I hope it can be fixed at some point without using the --highvram option

[smoky patrol]

Oo, very pretty. Love the water droplets, as always.

[upbeat summit]

how are you doing, @smoky patrol?

[smoky patrol]

I'm okay! Doing b e t t e r than the past week. How about yourself?

[upbeat summit]

same here. I'm glad temps went down a bit. the high humidity was very exhausting.

[indigo carbon]

I think a better approach to mix images together is to interrogate 2 images and combine tokens. heck, even chatgpt API can do that

[smoky patrol]

So over humidity. It's a gorgeous day out, and it's only going to be in the 80's, so I'm thinkin it's gonna be a mild, nice day.

[upbeat summit]

sounds great 🙂

[smoky patrol]

Which means...more time for running SDXL, hahaha!

Yeah, it's pretty cool rn!

[wooden pagoda]

Hi all,.. I'm having a ruddy nightmare trying to get my SD to install correctly,.. currently having issue with files in odd locations or not being where they ought to be,.. launched the webuser.ui.bat file and it stops, saying:

venv "H:\StableDiffusion\stable-diffusion-webui\venv\Scripts\Python.exe"
No Python at '"H:\PythonStableDiffusion\python.exe'
Press any key to continue . . .

When I click to continue, the box closes. So, why does Python appear to have to be in the Stable Diffusion folder,. ? I don't get it. Makes no sense.

My SD is installed to: H:\StableDiffusion\stable-diffusion-webui

[smoky patrol]

I love the whole woodworking thing. Super cool

[smoky patrol]

I recommend going over to #🤝｜tech-support

[uncut fiber]

it is combined with gustav klimt. Most noticeable on 3rd picture

[smoky patrol]

I can certainly see the influence in the colors/shapes

Would def be cool to have an actual piece like that, 3d printed or otherwise

[uncut fiber]

I just realized importance of word reminds 😄 So Klimt is everywhere 🙂
Now trying Picasso, reminds is nice word.

[soft zealot]

I use my ChatGPT assistant (as discussed yesterday bback up there somewhere) to do that

eg put in an old A1111 SD1.5 Prompt of

A glamour photoshoot in a dirty gritty New York Back Street for a fashion magazine cover .
(photo of a sexy mature woman in her 30's fully clothed in fashionble urban street wear),
(intricately detailed tattoos on arms:1.1),(large curvy body:1.1),(smoothed collar bone:1.1),(long neon pink colored hair),(glossy bismuth lips) a hint of a shy coquettish smile

Gets converted to

"A photo of a confident, mature woman in her 30s with neon pink hair, dressed in trendy urban streetwear, stands in a dirty gritty New York back street. Her curvy body contrasts the surroundings, accentuated by intricate tattoos on her arms. She exudes confidence with a shy coquettish smile, while her glossy bismuth lips catch the light. The image captures the juxtaposition of high fashion and the raw charm of the urban landscape. The high quality of the image ensures sharp details, vibrant colors, and an exquisite representation of the model's expression and style."

[glad grove]

when u write a 500 word prompt just for a girl's portrait

[soft zealot]

technically I didn't, my assistant did lol

and technically thats only 94 words :p

[glad grove]

that tattoo looks kinda suspicious 😬

[soft zealot]

its what got spat out at me 🙂

[uncut steeple]

cone head

[soft zealot]

I like that second on @uncut steeple

[strange mist]

what was familiar? the resulting art style? 🙂

[uncut fiber]

Aito portrait of Rembrandt 🙂

[wet nacelle]

@vital ermine and anyone else that can help.

Kohya Lora issues.

I am using this guide https://www.youtube.com/watch?v=AY6DMBCIZ3A&list=PLNccHdZquvdcE7yOmECuzRh2JmBhBs3ze&index=1&t=313s . I am also not using any reg folder because my Lora are quick poorly sketched out people creatures.

I have a 3070Ti with 8GBs of Vram. I also notice that when it runs it eats up 20+ Gigs of storage and then spits it back out at me when it fails. If you think I should suplly more information please let me know.

[fresh path]

you need more vram

minimum 12gb, I think

[smoky patrol]

Ah, pink and blue hair

[smoky patrol]

Colors, style, shapes

Clarity

[smoky patrol]

Painty ocean!!!

[strange mist]

I see 😁 I don't have a background in art.
but I'm slowly picking up bits and pieces about artists, art movements and styles.

[smoky patrol]

I had far too much fun making SDXL return to beta styles yesterday, but also using a lot of the new stuff

[smoky patrol]

That's so...ominous

[uncut steeple]

👀

[smoky patrol]

It's a lot of fun stuff to learn!! Def makes SD easier to work with, imho, cuz you can explore nuances you never knew about

[smoky patrol]

He's so....Handsome.

[uncut steeple]

Playing around with AI Self Portraits

[strange mist]

I agree. it's super fun and helpful to learn about art when it comes to prompting. though it does seem like cheating a bit at times 😄

[strange mist]

are you using an image of you as a base? 🤔

[uncut steeple]

Nope no base image, just txt2img

[strange mist]

I see. my pfp is a self-portrait of mine 😁
generated with the sdxl bot.

[uncut steeple]

My pfp is a warframe wearing a dress

[strange mist]

nice. I see it's also animated 👌

[smoky patrol]

I LIKE IT.

These are some of my abominations from yesterday

[smoky patrol]

Artists use references all the time

[strong field]

my initial changes didnt work, seems to be related to memory management still, but the sampling/weight issue is an additional thing i found

[willow bane]

Can I share my work here

[pure crystal]

yes!

[willow bane]

https://youtube.com/shorts/92CKpU6zq74?feature=share

Is it good?

[willow bane]

.

I am new to sd

[uncut steeple]

Hey in the future if you share your art please post it directly in discord and not as an external link for self advertisement

[willow bane]

Okk that's why I am asking earlier

Before sending

[uncut steeple]

No worries, just letting you know

[uncut fiber]

again that comics look, i somehow cant make it. Even with R. Crumb....

[wet nacelle]

@visual glade I'm getting the issue where the Vram usage won't go back down after a generation. I am up to date and this was solved before. Could it be that a node was made to solve this issue? My other workflow does not have this issue.

[wet nacelle]

I should mention that I'm on a Lora workflow.

[wet nacelle]

You should be able to drag drop this and see the nodes.

[wet nacelle]

When a gen is finished it goes down and then quickly ramps back up into a high usage flat line.

[visual glade]

try using the regular sampler node instead of that custom one

[wet nacelle]

Alright

[tepid surge]

how are you today?
response:

---

i found out something on start up was eatting up 2.5 gigs of vram

now i can run the AIT without any vram issues, yippe! ✨

[tepid surge]

bleh^ so 70s

but but but

gren

and and and ~~~

[civic saddle]

Has anyone trained on a p40? I'm a dev with an old clunker and I want to build a 1 liner training program for cli.

[tepid surge]

it's kinda stressing me out how fast it's making images rn cx

[pure crystal]

@tepid surge do you have a prompt generator script going?

[tepid surge]

fs fs

i craft a prompt when i need to but i'm pointing it in a direction and just letting it eat

[pure crystal]

got it, so much fun

[tepid surge]

i'm looking for landscape concept art kinda stuff like this so I can collect visual information for game stuff

world be building

[pure crystal]

I'm cranking out Morticias lol

[tepid surge]

what an intense face >u< v cool

love the like green house/plants and glass stuffs xwx

[tepid surge]

when Thaevil one generates: ^
when i generate: cx

ayee how are peeps organizing outputs, i need to figure out something so I can just gen and sort

[crisp owl]

organizing? 😬

[soft zealot]

auto date time stamped folders and then if Im doing something specific I may create a subfolder

In a standard image save node if you enter the value as

%date:yyyy-MM-dd%/COmfyUI

It will auto create the date stamped folder

[half ivy]

The mini eye in the middle of the forehead is so common 😂😭😅

[tepid surge]

i see i see, i have the file name and meta data sorted I just about to like be in 3 generic file explorer windows tossing stuff into diff folders (rip retinas) cx

[soft zealot]

for "zombies" ?

Itsan aiming point isnt it?

[crisp owl]

I did this zombie image last night showing my mom how all this AI stuff works

[tepid surge]

i'm looking forward to ai on like phone, that'll be a trip

that's how i feel a mom would feel but like in our time frame

[soft zealot]

My actual 3 image save nodes use 1 of

SDXL/%date:yyyy-MM-dd%/%date:yyyyMMddhhmmss%%SWT.seed%%M1.ckpt_name%%M2.ckpt_name%%STYLE0.style%%STYLE1.style%%STYLE2.style%_%STEPS.value%_STD

SDXL/%date:yyyy-MM-dd%/%date:yyyyMMddhhmmss%%SWT.seed%%M1.ckpt_name%%M2.ckpt_name%%STYLE0.style%%STYLE1.style%%STYLE2.style%_%STEPS.value%_HRF

SDXL/%date:yyyy-MM-dd%/%date:yyyyMMddhhmmss%%SWT.seed%%BML.ckpt_name%%RML.ckpt_name%%STYLE0.style%%STYLE1.style%%STYLE2.style%_%STEPS.value%_FFHRF

for either :

a "Standard" output (STD)
an "upscaled" output (HRF)
an "face detailed" output (FFHRF)

as well as having a datetimestamp, seed, models & styles used n the filename

[stone fossil]

a I-phone its called. 😉

Sorry I will leave myself out.

[tepid surge]

https://tenor.com/view/anime-girl-re-zero-starting-life-in-another-world-emilia-gif-17724709

cx

i almost forgot i couldnt post stuff with nudity

#freethenipple

[crisp owl]

/rstablediffusion server has a couple nsfw channels

not reddit

[soft zealot]

start ComfyUI on your PC with

--listen 0.0.0.0

and then access it from any device on the same network by the IP of the device you're on

[tepid surge]

yee, i'm talking personal assistant and 10-20 years down the line with tensor chips and stuff, and thnks btw

Okee, here my solution: if there are outputs i see i like on my main screen where the graph is, i'll pluck it from outputs into one of the other folders

time to gen x.x

https://tenor.com/view/speed-gif-5593963

[round walrus]

Anyone has idea of when 1.6 will be released, please?

[tepid surge]

1.6?

sd xl's going on 1.6? i havent even finsihed my slice of pizza

i like the terrain and water contrast

one of the things i think the The Cycle game did well was it's enviroment art direction

[tepid surge]

"Warning: Storm Imitate"

[tepid maple]

hmm

why vae broken for no reason ?

[tepid surge]

wow

[tepid maple]

what the hell happened on my webui

it is not working anymore O_O

[tepid surge]

i'd bet an update

[tepid maple]

didnt update it 😦

fuck

[upbeat summit]

I see. I've also send you a message.

[tepid maple]

ok fuk it time to use comfyui than

[wet nacelle]

I don't actually need to use the Lora keyword do I? Is it just the weight of the node?

[tepid surge]

Man FromSoftware did good with Fires of Rubicon cx

[tepid surge]

yee techncially

though the training might be specifc, hence the keyword to prompt those subjects/whatever, but yee tech non needed (least as in how I understand it atm x.x)

[wet nacelle]

I see what you mean.

[tepid surge]

Me coping after feeling like i just "um actually"-ed someone
Sorry if it came off weird Mack cx

[wet nacelle]

No I understood wjat you meant.

[tepid maple]

is there a vae problem on this iamge or am i trippin

prompt simple just "minions playing"

[tepid maple]

like this one

the lines on the floor felt like shadows from window

i wish it was 🙂

[tepid surge]

how dare you prompt minions cx

[tepid surge]

heh 💯

[tepid maple]

prompt it again add the shadows from windows

wanna know what wil lhappen 🙂

[tepid surge]

i kinda on a train that cant cahnge course rn cx

[tepid maple]

oh come ooon

lazy ass

😄

[tepid maple]

scheduled ?

[tepid surge]

do i look like a lazy ass to you cx
I am, generating for world building, watching armor core 6's campaign and talking to you all at the same time cx

[tepid maple]

left side of the explosion lookslike pritzel

lol

your model is hungry dreaming about food

[tepid surge]

mm pretzel buns

[wet nacelle]

Me when I! I say I say! Me when I wish I could make Loras!

[tepid surge]

goth mmy x.x

[wet nacelle]

@daring smelt *tricks

[daring smelt]

no way did you ping me for this lolllll

[wet nacelle]

ha

[daring smelt]

maybe there's only one trick

[wet nacelle]

HA

[upbeat summit]

hello fruit 🙂

[tepid surge]

"tricks for the trickless?"

[tepid surge]

these low poly so good

[half egret]

i wondering anyone help me with sdxl in automatic1111 i been having this strange effect where at 100% it remove all saturation in image

https://www.youtube.com/watch?v=Zxf-wekQA7s&ab_channel=SadSilly

i made a video showing it

i have tried disabling vae all together it does not seem make a difference

[daring smelt]

hiiiiiiiii

[tepid surge]

could be with how the preview's made

[half egret]

we thought that might be cause i changed all setting in live preview to match final image

[crisp owl]

Yeah I had this vae issue in A1111 before as well.
It was a weird one off issue, I changed the vae, changed it back, restarted the UI a couple times, generated a couple tests, changed the checkpoints, changed the vae again a couple times, and then it worked every single time after that. No updates, nothing else, just those.

[tepid surge]

man armor core 6 such a pretty game, i wanna make shit like this but that's soo many people etc cx So much creativity and skill

[half egret]

i just wondering anyone found a fix for it i really dont want to use comfyui lol

[half egret]

you cant prompt edit in comfy

[tepid surge]

come to the dark side

https://tenor.com/view/if-you-only-knew-the-power-dark-side-gif-10061896

[half egret]

i cant go back to cave man prompting style xD

[tepid surge]

https://tenor.com/view/joji-filthy-frank-welcome-to-the-rice-fields-rice-fields-gif-15724064

[wet nacelle]

You sound brain dead

[tepid surge]

okey here's my one weird trick

[half egret]

if comfy add prompt editing i would switch over xD

[hardy cipher]

prompt editing?

[wet nacelle]

yeah the fuck?

[hardy cipher]

what does that even mean? you don't think you can edit prompts?

[wet nacelle]

@half egretRead this you fool. It was replying to you.

[half egret]

in automatic you can control when token are turn on an off

[tepid surge]

token as in aspects of the prompt?

[wet nacelle]

Like the words maybe?

Tokens is a broad term.

[half egret]

@crisp owl did you mean reload ui butten ?

[wet nacelle]

does he mean |

[half egret]

you type [A:B:step] <-- a get replacaed with B at step

[wet nacelle]

You can make a prompt that allows it to step rate a portion of a token (word/words)

[crisp owl]

I did all the reloads I possibly could. Reload button, restarted entire UI, restarted PC, etc

[half egret]

🤔

[wet nacelle]

Auto is kool but his UI is awful now and I see no use anymore.

[hardy cipher]

probably talking about what this does

and there are all sorts of different things that can modify prompts in all sorts of different ways

thing is I don't think the naysayers are looking for solutions as much as looking for problems\

[crisp owl]

This was one of my messages regarding this very issue back in June 2nd in another server>

using XYZ plot to test several checkpoints. Determined what one I liked the best, then switched to that checkpoint and started by running a batch of 8 images, and noticed each of them was very muted. Thought it was just happen chance, tried adding keywords like "vivid" etc to my prompt, but didn't help. Decided to see if doing a couple with Hires. Fix would resolve it using a couple different upscalers, but still got muted colors. I then noticed that the pics I got from the hires fix using the 4x_NMKD-Siax_200k upscaler looked weird, not like the final images usually do, it was like extra grainy. That's when I suspected the VAE wasn't activating.
I then restarted my CLI, switched checkpoints, and got a normal generation. Switched back to my original, and started getting normal images once again with that also.

[half egret]

that might of solved it

i dont know why restarting would fix it

[crisp owl]

Something just hangs up the vae in certain situations, that's all I could think of.

Probably a conflict with an extension, code mashing together or something

[half egret]

thank you very much for the help ❤️

[tepid surge]

My honest reaction when a program hires something from the 10th dimension to make a bug

i need coffee, i'm crashing

[crisp owl]

catlice

[half egret]

getting nice saturation now

[crisp owl]

img2img wrong aspect ratio?

[vital ermine]

nah, I was screwing around

[crisp owl]

ah lol

[vital ermine]

still playing with nodes

[crisp owl]

the endless battle

[tepid surge]

this one makes me sad, feels real

[hardy cipher]

I like to believe that somewhere in the multiverse everything we create comes to be

[hardy cipher]

so somewhere there is a realm full of Jesus clones in mortal kombat style fights

thanks to me

[crisp owl]

Gonna be some gnarly hentei universes then from what I see on CivitAI

[tepid surge]

yeee, that one felt like a scene form an art film about a tragic early life from the kids perspective

[tepid surge]

so cuute!

[hardy cipher]

yes, they exist as a reminder of what can come to be if we follow the path of unfettered debauchery

[hardy cipher]

and then she became the female joker

[tepid surge]

insert gender being an inner thing

[tepid surge]

i so glad i figured out that vram thing for AIT, generations go zoom now

[tepid maple]

medvram ?

[tepid surge]

nah, thought of that route though
Something was eating my vram on start up, like 2 gigs but i tured it off (for start up)

[tepid surge]

words are hard

[tepid maple]

cant find what is breaking the failing replication

this is so annoying af

first vae problem now this

[tepid surge]

that's a mood fs fs cx

[tepid maple]

changing small setting updated soemthing

but no log for it

[tepid surge]

if i make a game i need to make sure it's multi cultural

[tepid maple]

one of this change fuked it up

holy fuk Ithink fonud it

what is this SHIT !

[tepid surge]

i still need to follow up on that stuff x.x

[vital ermine]

a koalapus

[pure crystal]

what are the composition words in this? or camera angle

[tepid surge]

Data dump:
Positive prompt:.painting, establishing shot of a Seductive ((game concept art)) , background is Townhouse, deep focus, Happy, Moonlit, Selective focus, Cinestill, vintage .Negative prompt:.blurry, old, line art.Model: .Seed: 964715408800005.Steps: 40.CFG scale: 10.00.Sampler: dpm_fast.Scheduler: normal.Start at step: 0.End at step: 10000.Denoising strength: 1.00

me: i hope no one takes advantage of me rn... xwx
I SAID!~

[wooden pagoda]

I have a few questions relating to the use of checkpoint files and then how to reference them (or parts thereof) in prompts,.. anyone feeling generous with their time to help me get to grips with getting some great results that aren't just made up of singular run-of-the-mill styles? I want to combine various styles but I just can't understand the YT tuts out there; it's massively technical and I'm stupid.

[pure crystal]

ask

[vital ermine]

Doom

[pure crystal]

what gave it away

[wooden pagoda]

Bob - Thanks. 👍 Ok,. so,. how do I combine two or three or more checkpoint files (is that styles?) and then reference them (or parts?) in my prompts?

[pure crystal]

merging checkpoints is possible

[pure crystal]

blending them with prompts works a few ways, remember 'style transfer' is free with this technology

[tepid surge]

they call it a ball park

[wooden pagoda]

So checkpoint files are styles?

[pure crystal]

say you trained a checkpoint on black and white cartoons, anything you prompt is going to have that look

but you can also do that by mixing concepts in prompts

[wooden pagoda]

But I have no idea about training,.. I know it's how MJ et al all get their learning,. but,. say I want to make a regency era (tasteful) nude and have it in an impressionist style,..

[pure crystal]

or weighting merges more towards one model or another. merging can sometimes wreck models a little because certain models have very specific weighting and the merging can mess it up

[wooden pagoda]

Oh, crap, ok. That's good to know (if disappointing).

[crisp owl]

easy enough to test it

[tepid surge]

kinda~, can be, def one the ways peep can embed styles

[crisp owl]

@wooden pagoda as I showed here, this a simple in-flow checkpoint merge. Can do as many as you want to test

[tepid surge]

man, it's like. expression done through diff technologies rather then any end-all be-all yet

[wooden pagoda]

Embed? Oh gawd, lol, this is so confusing. I'm not using ComfyUI,.. just the in-browser SDXL

[crisp owl]

ah gotcha, then you'd have to do a checkpoint merge through that UI and test the newly created checkpoint

[wooden pagoda]

I have merged a few already (in pairs),.. but,. how do I test it?

[crisp owl]

select that checkpoint

[tepid surge]

||like the tech is the + and () and x^y's and the data we use is the numbers||

[soft zealot]

or @wooden pagoda another option to mix the results from 2 different models/checkpoints is to use one Model as the base model and a totally diffferent model as the Refiner (if you're using a Base+Refiner style of workflow that is)

[tepid surge]

myorp
"Brother may i have some oats"

[wooden pagoda]

I merged a realistic one with a painterly one,.... *I THINK,.. might be wrong and even if I did,. I can't remember what I called it lol. (I know, I know, I'll have to remember what I called it and which two I had merged)

Refiner,.. I have heard of the term but not much else tbh,. it's soo confusing.

[crisp owl]

can just look in your models folder on your pc and see what the most recent checkpoint is

[pure crystal]

there are also LORAs, these are very small models that can be used and referenced as styles. will mostly be compatible among models

regency era, 3d cartoon style

[tepid surge]

@tepid maple

[wooden pagoda]

Earlier today, I tried to merge some but they didn't succeed,... Failed. I have absolutely no idea why though.

[soft zealot]

so using Sytanss Workflow as an example

There is a Base Model pass in one ksampler and a Refiner model pass in a second ksampler

This is/was the default method for SDXL0.9 & SDXL1/0

However you can also select say ProtoVision as the Base model & DynaVision as the refiner model

[pure crystal]

lol

[wooden pagoda]

I'm not using ComfyUI,.. it's a nightmare (nodes make NO sense to me whatsoever). I first saw nodes in Unreal Engine 4,... and HATED them. You need to understand coding terminology to get to grips with it. Instead, I'm using the SDXL in-browser method.

[pure crystal]

lol loved this movie

[tepid surge]

careful with nudity cx

[pure crystal]

I forgot where I was

[tepid surge]

that was me just a bit ago cx

[sinful falcon]

Taller. No, TALLER! (TALLER DANGIT:1.9)

[tepid surge]

#freethenipple

[soft zealot]

actuallly as its a viisual representation its actually easier for people without coding knowledge to use Nodes than the other way around..

All you're doing is joing little boxes with pieces of string.

Even toddlers can manage that ;o)

[pure crystal]

you can only prompt about 80% of a giraffe, this is a very good result

[wooden pagoda]

With nodes, you have to know what to look for,... what does what,.. why this connects with that and why not,.. and when, what order and when to use more than or less than 1, 2, 3 etc. It's not a thing that makes a micro-ounce of sense to me

[pure crystal]

regency era, grand fireplace, red velvet curtains, mahogany furniture, portrait, Georgian england, boudoir, (John Soane, Thomas Rowlandson, William Blake:1.3), (selective focus, cinestill,deep focus:1.3), highly detailed, painterly, inspired by William Brodie,

[wooden pagoda]

Bob - but,. which Checkpoint mdoel has that info in it? (Artists names)

[soft zealot]

there is no right or wroing way, there is just a preferred way 🙂

the best way to start is with a relativly simple workflow and then experiment from there, you're not going to tank your puter if you mess up.

Stop being a wuss ;o)

[pure crystal]

this is dynavision, a 3d animated oriented model

[thorny frost]

It is indeed pretty easy. It only allows the correct connections and tell you when you miss a connection.

[pure crystal]

feel free to use any of this stuff for any purpose, you will be surprised at how well it works

you MUST experiment and try stuff in this hobby

[wooden pagoda]

Tanking my PC isn't a concern (what on EARTH made you think that was a woerry for me?),... What concerns me, is the need to learn a shed-load of info why one box is connecting to another or two others etc etc etc etc,.. one word: Syntax. It makes no sense to me.

[sinful falcon]

(pats wall) "No one's gettin' in here."

[pure crystal]

ahahah :shake, shake, shake, Dad grimace: sturdy!

[wooden pagoda]

Look, no offence mugginses,.. 😄 but,. I'm here to figure out how to get some mixed styles into my renders,.. so far they've been utter tripe (since 11am today,. and it's now 9 1/2hrs later and I'ms till nonethewiser).

[pure crystal]

show your prompts and output

[wooden pagoda]

Output? I'm not using comfyui.

[pure crystal]

the images

[soft zealot]

it was an estoteric tamgential train of thought not meant to be taken literally.

Its only the same as an Otto Cycle engine

theres an input, something happens to it, it transforms and there's an output

[wooden pagoda]

Oh, sorry. 😄 Hm,.. they're mostly nudes. *Bad attempts,.. but nudeys.

[tepid surge]

https://tenor.com/view/heresy-detected-warhammer40k-warhammer40k-heresy-detected-warhammer40k-space-marines-gif-26543746

[pure crystal]

just blur them enough to post

or show your prompt

[wooden pagoda]

Sorry, kinda hard to know in just text-based form. 🙂

[tepid surge]

https://tenor.com/view/heresy-detected-no-heresy-no-nope-gif-25531410

[soft zealot]

I was only communicating in text , there were no images or anything attached :p

[tepid surge]

https://tenor.com/view/not-heresy-gif-24882220

[wooden pagoda]

Exactly why I misinterpreted it. 😉

[pure crystal]

@wooden pagoda try stuff, save seeds and workflows you're happy with, try more stuff

[wooden pagoda]

A natural redhead woman from the regency era (1820s) of the British isles with her back to the camera, scene is shoulder upwards, she is wearing nothing but a warm smile in a classic and elegant pose. She is graceful, mildly slender, with beautiful, healthy and natural skin with all of its pores and flaws. The lighting is soft, muted and simple. Her hair is styled of the era, her face is pale and she is relaxed and alluring, she is slightly fuller figured.

[wooden pagoda]

IIRC, I used the SDXL v1 Checkpoint (large 6.6Gb file) or possibly the Epicgasm one.

[pure crystal]

even though it's supposed to be natural language you want to keep the word count down

make this a little more terse and less like a romance novel lol

[pure crystal]

'style of the time' becomes 'regency style hair'

[wooden pagoda]

My main goal earlier today after that one came out, was to use ControlNet to fix her hand,. but since I couldn't grasp the concept to get the ControlNet to install, let alone work. I have no idea how to use the inpainting,.. or the img2img features,.. they're not intuitive.

[pure crystal]

yes it's enthusiast software, takes some time to learn

I also wanted to not learn this stuff and just inpaint my first week

[wooden pagoda]

I only installed it yesterday/today... because I'm SICK of paying to use MJ,.. it's not cheap for the pro version - anything less than that and everyone sees my prompts; so I've stopped using it.

[wooden pagoda]

"Oh, was it YOU whom stole my dress??! You rascal!"

[pure crystal]

eh share your prompts, there is too much experimentation required for everyone to learn everything themselves

[vital ermine]

LOL

[wooden pagoda]

When and when not - and how - to use a Lora file?

[pure crystal]

back tits, my fetish

[wooden pagoda]

Arnolda Schwarzenegger!!! "I have my back!!" 😄

[tepid surge]

"fetish you say?"

cx

[tepid surge]

hrrr

[pure crystal]

@wooden pagoda https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#extra-networks LORA stuff here

[pure crystal]

train stuff into a lora, use it across models. reference the lora using the prompt notation, weight the lora for how strongly you want it to blend into the output

[upbeat summit]

you're posting some great stuff 🙂 nice work

[pure crystal]

you can have style loras or concept (things, faces) loras

[vital ermine]

I made her a proper school marm type

[tepid surge]

you say that today cx
thank you though i appreciate it

[upbeat summit]

I did before!

[pure crystal]

just saw the koala lol

[tepid surge]

my posts today:

[vital ermine]

It was cute, lol

[wooden pagoda]

Looks like an area in RDR2.

[pure crystal]

british women woof

[wet nacelle]

https://civitai.com/user/Kappa_Neuro/models Is Kappa Neuro faking his stuff?

[wooden pagoda]

So imgh2img is akin to an image prompt in MJ? Am playing with it now, seems to be just that; but the rest of it; way too advanced for me.

[wooden pagoda]

How would you 'repair' her eyes and mouth?

[tepid surge]

cx

[wooden pagoda]

Midjurney 5.1 (IIRC). Took me HOURS to get this right.

[wooden pagoda]

Style is? *Painting, I mean

[tepid surge]

Pulls data dump lever:
Positive prompt:.(art by Graham Sutherland:1.3) , painting, Heartwarming Entrancing ((young women, hyper realism)) , plain background, shallow depth of field, Gloomcore, loop lighting, three colors .Negative prompt:.blurry, old, line art.Model: .Seed: 432108425717290.Steps: 20.CFG scale: 10.00.Sampler: euler_ancestral.Scheduler: normal.Start at step: 0.End at step: 10000.Denoising strength: 1.00

[wooden pagoda]

Thanks,.. TMI 😄 but,. how do I blend two or three artistic styles?

[pure crystal]

using the power of prompting, checkpoints and loras

[pure crystal]

you'll find "this lora, with this prompt always works across checkpoints"

[wooden pagoda]

I gathered that much 😄 hahaha 😄 - but what do I need to do in the SDXL browser to make it happen? Thanks. 🙂

[vital ermine]

I wouldn't as both of those befit the time period portrayed.

[pure crystal]

or "this prompt always works, even across checkpoints"

I dont know your setup

what's stopping you from trying?

[upbeat summit]

fusing some styles - cool 🙂

[vital ermine]

Inpainting in comfy sucks so I miss that the most, and outpaint just go lie in the road.

[tepid surge]

i'm gonna fuse something...

cx

[wooden pagoda]

You don't,.. I don't,. just that I only have the SDXL in my browser,,.. on PC,. w11.. 64Gb RAM, 10Gb on my RTX 3080 GPU and a cool Ryzen 9 5900

[tepid surge]

hope you're ready

[wooden pagoda]

Nicely Regency looking, not bad painting style too

[tepid surge]

pshhh whatever

cx

[wooden pagoda]

Susan Trump? 😄

[vital ermine]

LOL

[tepid maple]

nicee

[vital ermine]

Need lizzo in the middle

[tepid maple]

guess this get a million bucks

[vital ermine]

XL does know lizzo btw

[tepid maple]

you are generation pretty random shit 😄

I have no idea whats going on in your mind

[tepid surge]

---

"I'm busy"

"but there's always time for a good shag"

[glad grove]

you won the 4090

[wet nacelle]

ha

[tepid surge]

okay!

i'm out

zzzzzz

[wet nacelle]

night

[pure crystal]

oooooo that's the stuff

[pure crystal]

in his natural habitat

[pure crystal]

dont fall asleep

[crisp owl]

There was a bit discussion about that in here the other day

[terse pewter]

guys i was finally able to create that image, with garage with snow mobile, dead body, and a gas station

took me like 3 hours

[terse pewter]

❤️

[wet nacelle]

Any info then? I'm just lost on context.

[crisp owl]

There was a couple people discussing that exact thing, their suspicious that the Lora models are all fake, or at least mostly fake. Possible just trained on an extremely insignificant amount of data, then just uploaded to civitai. They noticed a large amount of them were nearly the same exact filesize (although they did have a different filehash).
But when attempting to use the Lora's, they were getting identical results when any weight on the lora, and then even removing the lora completely.

around here the conversation was happening
#✨｜sdxl message

[glad grove]

why tho? for the money?

[thorny frost]

donations probably

[crisp owl]

I suppose so. He's got a donation site

possible for competition also?
I don't follow what comps are active though, so not sure on that front

[thorny frost]

he said it was all an error and will upload the good ones soon and such

[crisp owl]

Also possible

I rarely use Lora's so I don't follow much of that side of things

[thorny frost]

in any case the very fac everybody was happy ´using´ the false loras show how innecesary they really were

[crisp owl]

Yeah, I think the only Lora's I've used was a couple character ones that simply didn't exist when prompting for

[visual glade]

there are fake loras now?

[glad grove]

they still needed when you need a specific character or a specific style

[crisp owl]

Some lora's were supposedly not doing anything when added/removed/weight increased/decreased.
But sound slike it may have been a mistake on the authors side

in the meantime, working on some of these

[thorny frost]

probably, or maybe civitai ha been testing all sistematically since then which i doubt

[visual glade]

by the way if anyone had some issues with memory on comfyui you should update because I fixed something

[crisp owl]

I'll do that and do more testing tonight.
I ran several tests a couple days ago and could not replicate my weird lengthy upscaling times. No samplers/schedulers would consistently cause the issue, it was just wonky for me

[autumn wind]

anyone have a clue on how to do those interpolation videos where they go from one image to another with SD?

[thorny frost]

btw dude, comfy is best thing since sliced bread dont know if anyone has said it already :p

[visual glade]

glad you like it

[autumn wind]

it doesn't seem hard to do

[thorny frost]

https://github.com/deforum-art/sd-webui-deforum

it is easy to do but hard to do it well

[autumn wind]

Have you ever tried it?

[thorny frost]

of course

https://youtu.be/aUJuwNxNUWQ?si=1Pnu7sWP_t4jWHqf

dont think you are going to get something like that the first day...

[crisp zinc]

how can i switch between pictures in the comfyui?

[thorny frost]

you want to load or save a picture?

[crisp owl]

If you processed batch_count, you can't.
If you processed batch_size, there will be an option to scroll through them in the output image on the UI

[autumn wind]

that's really cool

[dusty jewel]

trying to make a photo of a penguin being eaten alive by a polar bear, but it never seems to work despite the straight-forward prompt. is there some filter preventing this?

[crisp owl]

no filters

[glad grove]

its uncensored but its very hard to generate gore

[dusty jewel]

ah i see i see

RAW photo of a penguin being eaten alive by a polar bear, lots of blood and gore, 16mm, color graded portra 400 film, neutral colors, ultra realistic, shot with cinematic camera both the penguin and polar bear gets beaks, and they seem to be sharing some other meal

https://pbxt.replicate.delivery/hjHuxh7sALK4JtWMQIEJf6foYyfbclf7T66tz0vYzAtNJ16FB/out-0.png

it's for this article im writing about linux 😁

but what the hell is that ^^

😦

[glad grove]

are u using sdxl or 1.5?

[dusty jewel]

sdxl @ Replicate

[glad grove]

try with fighting each other,attacking each other,fighting

[dusty jewel]

hehe cool thanks

[dusty jewel]

ooooh that's good

[thorny frost]

The other pingüinos are pretty stoic

[dusty jewel]

haha

[crisp owl]

There's the full processed image of that one

[dusty jewel]

num_inference_steps, guidance_scale, prompt_strength?

Or can I just use yours?

I'd also need a photo of OpenBSD's pufferfish smiling with sunglasses, gold chain and a cap swimming in crystal clear water.... but i never get the cap or the gold chain :/

[crisp owl]

Yeah that flow isn't completed yet so I don't have reference notes in it yet.
Can use the image if you want

[dusty jewel]

OpenBSD vs Docker and Linux: Deploying Ruby on Rails in Production

cool thanks!

[cedar tusk]

Anybody else getting an error with the newest commit of IPAdapter (and newest ComfyUI commit)?

RuntimeError: Input type (torch.FloatTensor) and weight type (torch.HalfTensor) should be the same or input should be a MKLDNN tensor and weight is a dense tensor

[hardy cipher]

man, it's the server.py

[hardy cipher]

I started getting this error that popped up in execution.py, but seems to be triggered by server.py? I haven't really figured it out.

got prompt
Failed to validate prompt for output 3:
* AdaptiveCanny 2:
  - Required input is missing: images
Output will be ignored
invalid prompt: {'type': 'prompt_outputs_failed_validation', 'message': 'Prompt outputs failed validation', 'details': '', 'extra_info': {}}
got prompt
Output will be ignored
invalid prompt: {'type': 'prompt_outputs_failed_validation', 'message': 'Prompt outputs failed validation', 'details': '', 'extra_info': {}}

so reverted to some previous code I know for a fact worked. same error, so tripping this in server.py

    if len(good_outputs) == 0:
        errors_list = []
        for o, errors in errors:
            for error in errors:
                errors_list.append(f"{error['message']}: {error['details']}")
        errors_list = "\n".join(errors_list)

        error = {
            "type": "prompt_outputs_failed_validation",
            "message": "Prompt outputs failed validation",
            "details": errors_list,
            "extra_info": {}
        }

but wasn't before. so looked an it was modified 18 hours ago: "support on prompt event handler (#765)"

Allowing pre-prompt and pre-workflow checks before the prompt is executed enables pre-execution tasks that have a global impact, such as extension garbage collection, prompt editing, and other related actions.

anyway, found where some of the code was changed. turned this

        @routes.get("/extensions")
        async def get_extensions(request):
            files = glob.glob(os.path.join(self.web_root, 'extensions/**/*.js'), recursive=True)
            return web.json_response(list(map(lambda f: "/" + os.path.relpath(f, self.web_root).replace("\\", "/"), files)))

into this

        @routes.get("/extensions")
        async def get_extensions(request):
            files = glob.glob(os.path.join(
                self.web_root, 'extensions/**/*.js'), recursive=True)
            
            extensions = list(map(lambda f: "/" + os.path.relpath(f, self.web_root).replace("\\", "/"), files))
            
            for name, dir in nodes.EXTENSION_WEB_DIRS.items():
                files = glob.glob(os.path.join(dir, '**/*.js'), recursive=True)
                extensions.extend(list(map(lambda f: "/extensions/" + urllib.parse.quote(
                    name) + "/" + os.path.relpath(f, dir).replace("\\", "/"), files)))

            return web.json_response(extensions)

in server.py

so anyway, wall of nonsense, but it's quite a bummer because I can't even debug the nodes I was messing with without disabling the pre-check

it's just like "lol, nah, begone with you"

[cedar tusk]

@hardy cipher Weird. I noticed that two of the last six commits on Comfy were messing with clip vision, so I wondered if it was that

[visual glade]

I did mess with them but the output should be the same type as before the commits

weight type (torch.HalfTensor)
that's the problem

it should convert the input to fp16 if the weights are fp16

or turn on autocast

[cedar tusk]

ah, so the node has the toggle for f16 or f32, but the f16 isn't applying maybe?

[hardy cipher]

I'm so confused by my situation, lol. I finally got them working, then back to nada

maybe I'll just revert to the old code or something. but not even quite sure what the issue is

[cedar tusk]

Is enabling autocast something that should happen within the node or is that an external/launch setting? The node code implies that it is setting the dtype properly

Oh weird, I loaded up the SDXL workflow I had with adapter and that's giving the same error, even though it should be fp32...

[visual glade]

within the node

[soft bone]

@visual glade Is there a way that extra_model_paths.yaml can be refreshed without reloading the entire UI?

[eternal fog]

Are the AIT Nodes broken for anyone else with the newest Comfy update?