✨｜sdxl

[molten gull]

yes

can you give me an example how you actually do it ?

[hoary saddle]

Edit the example to reflect the IP or localhost of your comfy installation then just copy paste your workflow into the prompt section with the API button. Then change the bottom part to the correct node numbers and their values

I use it for all my web app hooks and my sdxl blender addon

[molten gull]

i did save it in the API format

[hoary saddle]

Error in line 13? That's a comment line, same file?

[molten gull]

yep

i guess it means line 14 ?

but i still dont know whats wrong

[hoary saddle]

O line 708

Whats 708?

[hoary saddle]

Copy paste didn't work maybe, missing a {} or something

Replace between """

[molten gull]

i tried again, same error

can you send me the file you used that works ?

[hoary saddle]

Is prompt 6 your positive node?

I'm guessing no, with 700+ lines

[molten gull]

it's a pretty complicated workflow

[hoary saddle]

Change 6 and 3 to reflect positive node number and seed

[molten gull]

i commented those out

i dont want to change stuff there

but still the same error

do you have a working version?

[hoary saddle]

On mobile but I can help tomorrow

Here's a screenshot of it

Prompt load

Replace prompt load with save API output

Change nodes to reflect

[molten gull]

ah, maybe you can send me tomorrow then, would be really appreciated

[vivid tide]

Not just once.

[soft zealot]

eh?

[hoary saddle]

And you copy pasted entire API save results directly between:
prompt_text = """

"""

Still think your missing a bracket or something

[vivid tide]

You’ve been down the rabbit hole multiple times

[soft zealot]

ahhh gotcha 🙂

[molten gull]

i did that copy

[heady vale]

even though XL doesnt get the races correct, its a lot of fun generating them

[ionic gulch]

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

https://i.imgur.com/Li0m0IK.jpg

https://i.imgur.com/i4GdHdx.jpg

[ionic dragon]

did you use any lora's?

[dreamy sun]

Hey, quick question: How can i make it write a text like this?

[upbeat summit]

"a cat wearing a hat saying SDXL"

than you run a couple of seeds

[ionic dragon]

cool

[dreamy sun]

Is there more i can learn about text engineering?

[upbeat summit]

or you use img2img and inpaint

[upbeat summit]

mhh I only did a couple of quick studies. the syntax I posted works for quite well for 3-4 letter words.

"a man holding up a sign saying hello"

[ionic dragon]

does anyone know where to input and output this node?

[dreamy sun]

I know but it can't mostly do two letters together, i mean is there like engineer (Text) etc some people had ways to engineer like a year ago, there isn't any way to tell it is a text other than the plain input?

[bronze arch]

How much did sdxl cost to develop? Does anyone know

[ionic gulch]

here is a video that teaches lorastacker and also how to find out things like that by some trial&error:
https://www.youtube.com/watch?v=WFgf1I1aBxs
personally i prefer to just load them in a row with standard nodes:
https://i.imgur.com/STdAPeO.png

[molten gull]

any idea how/where the titles of the nodes is stored? doesnt seem to be in the json file ?

[molten gull]

i dont have it, maybe its not in the json file if you do that dev-mode-save-thing?

[soft zealot]

to be fair you didnt say "which JSON"

I found it in the JSON for my workflow, you appear to be looking somewhere lese.

There is more than one JSON file in the folder stcture you know!!

[molten gull]

i meant the dev-mode-save-json 🙂

[soft zealot]

pass, ask me another quesruion on sport

[molten gull]

another question: in dev-mode-save-json, how can i easily identify a certain node ?

uh, sports question: how would an icehockey goalkeeper easily identify a certain node in the dev-mode-save-json given that title aint there?

[soft zealot]

Are you Canadian by any chance?

[molten gull]

nah, german

[native knot]

One thing's for sure...he wouldn't puck around.

🥁 🥁 💥

[molten gull]

🙂

https://www.youtube.com/watch?v=2ic3xNfEP_o&t=32s

still didnt get an answer to my json question sadly though 🙂

[soft zealot]

why not save the node with a random name like "wibble" and then rather searching for "title" search for "wibble"

[molten gull]

it's all not in there

[soft zealot]

well there's your answer

[molten gull]

no, that's not the answer, it's the problem 🙂

[meager canopy]

If you're ever unsure about what it can connect to, drag a connector into an empty space and drop it. The menu will then open up, listing the possible nodes you can connect to it.

[soft zealot]

goof luck with that node though lol

https://www.youtube.com/watch?v=WFgf1I1aBxs

[ionic dragon]

ok

[ionic dragon]

will use this tq

[dense lichen]

is this a1111 or what?? .....I'm confused

[soft zealot]

It could be or one of its forks

[dense lichen]

how can iknow?

[soft zealot]

scroll to the bottom of the page

and read what it ssays or follow th elink tot he github page

[dense lichen]

it says a1111

thanks

[soft zealot]

shakes head

[dense lichen]

a1111 is the official from stable diffusion right??

[vast galleon]

Anyone have any experience using groundingdino? I'm thinking of trying to port it into comfyui but I'm having issues installing the modules.

https://github.com/IDEA-Research/Grounded-Segment-Anything

[dense lichen]

or from someone else?

[soft zealot]

no, there is no official (other than maybe a command line thingy)

this is belleding edge stufff, if you want the nice warm fuzzy embrace of a supported warp the STability AI (the people behind Stable DIffusion) off 2 web based services.

https://stability.ai/

[high skiff]

ok, just tore apart the cooling system on my GPU and put it all back together

it was so damn gross from the previous owner, took me over 2 hours to get it clean

Before (the top section of the GPU that looks cleaner was something I already started)

[uncut fiber]

smoker

[high skiff]

After

no, they didn't smoke

it's cooking oil vapor

[uncut fiber]

o.k. then.

[high skiff]

they live in a very very small like 350 square foot apartment in New York

he said that likely from cooking next to his PC, the oils gumed it up. It doesn't smell at all like cigarette smoke, and rather smelled almost like indian food when it got hot lmao

[uncut fiber]

i know similar thing is from el. cigarettes as well. Such oily dust. <- sorry not sure if anybody got it 🙂

[unreal plover]

Has anyone ever managed to use the ComfyUI API?

[arctic bloom]

took 4 generations for me to get correct wording on picture
with prompt "a cat wearing a hat saying SDXL"

[full idol]

What sampling methods are best for SDXL?

[dense chasm]

https://www.reddit.com/user/FrankChieng/comments/15p4qbb/sdxl10_base_model_lora_training_deploys_on/ just updates huggingface model card with the character maggieQ lora inference API,enjoy it

[ionic dragon]

@soft zealot @high skiff i need some serious help, i am working on a llm to generate sd prompts
if there was a tool to generate sd prompts on instructions, how would you ask it to generate?

everyone can answer

[slow arch]

how can i run models locally?

[strong field]

ComfyUI

[strong field]

Have you tried doing ChatGPT

[ionic dragon]

yea, not realible, it doesnt know what SD is

[vernal cloak]

Yeah. I also sugest running ComfyUI. It's a bit complecated. But it's the only UI that runs SDXL the fastest. As i'm currently aware of.
Automatic1111. (a big favorite of the community) Is currently not working as well with SDXL, right now.
So ComfyUI is the most stable choice to run SDXL. 👍

[strong field]

Sure thing! How about this SDXL prompt: "Compose a vivid and enchanting description of a serene sunset over a tranquil lake, where the sky is painted with hues of orange, pink, and gold, and the water reflects the stunning colors like a shimmering mirror?"

So you are looking for trigger words, formatting, and weights?

[dense chasm]

huggingface inference API is developed with Rust,a ML framework language,response speed is good

[strong field]

I would suggest using a local llm (langchain or otherwise) and take good existing SDXl prompts drop them in a database to start to train it

[abstract canyon]

any SD XL replicate users? wondering if there's a way to get weights/arguments {i.e. (((prompt here: 1.542)))} to work, or if there's a replicate specific syntax? (unless they do already work -- but for me they seem not to, or at least, to have notably different output than other methods 🤔 ). if there isn't a way to do so on replicate, are there are platforms that have api functionality which would work with prompt weights/arguments? my understanding is that stability.ai doesn't allow this, either

[soft zealot]

oh I dont know, chatgpt doe s quite well at knocking something up for use in the "lingustic positive prompt"

[coral tapir]

guys im very new to SD. just playing with different models learning the prompt and stuff. a big question I had in mind for a long time, imagine we generate a character with all its details. how can we get him in another pose without AI making total new character. It seems even with complete details, AI always generate a bit different look right? another example would be an environment. we create a castle in night, then we want the same castle in day.

[dense chasm]

You should learn how to do LoRA training

[strong field]

Lots of options in no particular order, image to image with appropriate denoise, control net, stacking, LoRA, textual inversions, hyper networks

[ionic dragon]

i didnt find chatgpt consistant at generating good prompts

[soft zealot]

This is AI, This is the way

[elfin flare]

This AI really likes clear, coherent sentences.

[dense lichen]

is it normal for a1111 to do nothing in the upscaling process?

[elfin flare]

Beautiful.

[ionic dragon]

so i am finetuning an llm, so it can generate reliable prompts, gonna use these to train. but i am looking for your answers on how you would ask the llm, so I can prepare the input accordingly
https://huggingface.co/datasets/Gustavosta/Stable-Diffusion-Prompts

[soft zealot]

If I was that good at desribing a scene to an LLM I wouldn't need an LLM to wrote the prompts.

Also define consistency?

Are you expecting short or long input s ?

Who is the target user demographic?

[dense lichen]

i dont see anything happening in cmd

is there anything i should know about when it comes to upscailing?

this is my very first time

haha..... it crashed

[soft zealot]

there should be some sort of progess inidcator IIRC but I havent used A1111 for a while

[ionic dragon]

i mean, if i wanted prompts of a cat, i would ask the following "generate a prompt of a cat based on cyberpunk style", so how would you phrase this and how you tweak it?
i am gonna traain it with prompts of 25-35 words

the user demographic could be everyone, will gonna launch a comfy node too, so that you can use the randomisation as the one you have in your workflow

[dense lichen]

the indicator is not moving

[elfin flare]

I just personally run a local LLM myself in order to generate cool prompts.

[soft zealot]

do you have the same issue if you use a normal SD1.5 model in A1111 rather than an SDXL model?

Personally II would dtart with non SDXL in A1111 rather than jumping tsraight in with SDXL as my understanding is SDXL support in A1111 is patchy and YMMV

and also what GPU/system specs

[dense lichen]

i was using 1.5 actually 👽

[dense lichen]

i have rtx 3050 laptop version

and R7 5800H

16gb ram

[coral tapir]

wow I thought it might get complicated )) any step by step guide?

[coral tapir]

ok... no idea how to start :/

[dense lichen]

@soft zealoti think that the problem was because i used a very high res for upscaling and rendering

it worked fine this time

[upbeat summit]

King of the Voxels!

[full idol]

Style Lora

[strong field]

Easiest to start with something like img2img, what are you using a1111 or comfy? Do you know what SeargeSDXL workflow is?

[coral tapir]

well im very new. just realized my PC is not good enough for sdXL. i have a1111. no i dont know about that

[soft zealot]

are you sure you're not working on a solution to a non existent problem?

NB there is a set of Nodes already out there that interacts with CHatGPT via API but you need a Paid for CHatGPT account

[ionic dragon]

i am basically working for myself, but i would like to share it with others too
chatgpt api is expensive af

[soft zealot]

it is if you want to integrate but opening a web browswer aint difficult

[strong field]

SD1.5 is still very good, I normally work in comfy but a1111 has decent workflow for beginners

[arctic bloom]

it's not a problem, but enhancing current level of prompting. need to look at this from positive persective or point of view

[strong field]

You can take an image to the img2img tab and then send to sketch to change some details, lower your denoise to stay close to the original below .5

There are examples out there but currently mobile I’ll look for some later

[soft zealot]

Im just taking it from the perspective of seeingg a growing number of people who cant even read readmes contained in workflows :p

[strong field]

Easiest way is to use Sunny’s promoting guide in database and start building from there. The bot throws out tons of prompts so the data is there.

[soft zealot]

you could di it in ewxcel with VBA if you really wanted, or access

[strong field]

But as for coaxing the llm, well if you need to learn to do that might as well do it with sdxl prompt like Winston said

[ionic dragon]

i didnt get desired results with chatgpt
and finetuning isnt too difficult just extracting prompts is timetaking, if i spend around 2hrs i can get it

[soft zealot]

takes a handful of seconds to ask CHatGPT and copy & paste back to Comfy and then refine as required

[strong field]

Honestly I’m intrigued, I might just throw a bunch of different prompts in a txt , then upsert to DB and see what I get. Project for later today

[soft zealot]

one shots are rare, dont expect them, you should always expect to discard more than you save

[strong field]

Yes agree with that

Ai generation is a numbers game

[soft zealot]

that was from the chatgpt prompt, I'm sure If I run 100 through I'll get something better lol

[strong field]

Ooo that’s a great start

[soft zealot]

and that was just using th eprompt exactly as supplied with no supporting terms and just some very basic negative (Text,watermark sort of thing) and no styles appled

[strong field]

Anybody got a prompt database already?

[ionic dragon]

but thats very tiring, over long term, i would save more time by just finetuning one rather than specifying and explaining chatgpt how to generate

[strong field]

Save me the trouble

[ionic dragon]

there are many on hf

[strong field]

I’ll have to look later

[soft zealot]

I didnt teach CHatGPT any rules I literally started a new sission adn aske dit to describe a picture meeting your ask

bear inm Mind that SDXL uses a more humanistic approch to Prompts than previous versions of stable diffusion

[ionic dragon]

i completely understand, but i had always had hard time getting the best outputs

[ionic dragon]

yeah true af

[strong field]

Lol was typing this out

[strong field]

Sd1.5 promoting is awful results in sdxl

[soft zealot]

Imean dont get me wroing, if someone wants to build something for their own amusement & education I have no issues with that. I'm not the target demographic lol

[soft zealot]

some of that is possibly down to the way Comfy has, according to WAS, implemented weighting compared to A1111 for example.

There's a handbags out thread , one moment

[strong field]

Lol yeah the weighting took a while to learn

[soft zealot]

https://github.com/comfyanonymous/ComfyUI/discussions/473

[trim orbit]

it does both. the G encoder is the one with better natural langauge processing. The L encoder is the exact same one we are used to from 1.5

[ionic dragon]

yeah yeah, i completely got your point, but my experience with those llms for this use case wasn't good
i got good results when i explained it how to generate and using multiple prompts, it would take around 7-10mins to setup for me

[soft zealot]

you're doing it worng then,

Once you ve done it once you save the instrictions in a .txt file, paste into chatgpt and you're ready to go in 30 seconds

[strong field]

Could you try custom instructions in chatgpt, write once and done?

[ionic dragon]

so using sentences for G and using the same prompt by splitting into keywords is better?
for example G: an aerial view of a woman in a forest wearing a floral dress looking at the sky L: woman, forest, dense, floral dress, aeriel view

[strong field]

Dude again with the same words out of my mouth lol

[soft zealot]

thats spoky

[trim orbit]

my point about using both is that you can paste chatgpt results into the G, but then add word salad to L just to guide results better

[ionic dragon]

i didnt use chatgpt, i used claude for prompts, i had to connect to the vpn, attach the files, and yet not all were great, but few were great

[strong field]

Oh sick going to save for later

[soft zealot]

im not saying its perfect , its one of seevral lol

[strong field]

I was wondering how Claude compares

[soft zealot]

this is another one Ive used

[ionic dragon]

i used claude only becuase it supported higher character count and you can attach files for free

[trim orbit]

why use vpn for claude? are they limiting it?

[ionic dragon]

no, its only available in US/UK

[trim orbit]

i hate when these AI companies limit releases. It speaks of unethical behavior to me

[strong field]

Lol would be hilarious if you taught Stablecode to write sdxl prompt fu

[coral tapir]

oh Im reading on comfy. its the nodebase right? should I start using it instead of a111?

[trim orbit]

i'll just refuse to use their product i guess

[soft zealot]

On the other hand wasnt it the Italian Goverment recently said it was going to bloack access to AI websites?

[strong field]

I 100% recommend comfy over a1111

[trim orbit]

google bard is another one. highly unethical moves by google to go the closed route

[trim orbit]

in the 21st century italy prosecuted and found guilty geologists because the earthquake happened. they're, well, they're italy

not exactly a beacon of democracy

[soft zealot]

Italy, where all the tanks have one forward geear and 6 reverse gears

[strong field]

Way to tell everyone they need a vpn

[trim orbit]

These people are host to the vatican. They know what a VPN is.

coughcpcough

[soft zealot]

ah it was temporary and it was specifically ChatGPT

https://www.euronews.com/next/2023/04/07/after-italy-blocked-access-to-openais-chatgpt-chatbot-will-the-rest-of-europe-follow

[coral tapir]

ok thanks. downloading 1.5gb now to start installing it

[soft zealot]

Ok so taking into accoung SDXL really really has to be forced (in my experience ) to get something a little naughty then this prompt generated by that isnt too bad 🙂

((page 3 style topless photo shoot taken with a Hasselblad 500C/M, cinematic shot)), ((best quality)), ((masterpiece)), ((realistic)), ((beautiful female model)), (Fiery auburn hair, tied in a loose bun), ((highly detailed face)) called Freya, (Leaves and flowers crown), ((flowing bohemian dress)) with earthy tones, (Golden hour in a sunlit forest), scenic, masterpiece, bokeh

[ionic dragon]

why EU organisation fucks all the countries in its territory, like for example, if Serbia has no prob with something and eu bans it, then even serbia has to follow it
duck eu

[trim orbit]

i love how that prompt crafting for sdxl is such a fresh new venue, but there's already people speaking about how it can't do certain things. Just last night i saw someone trying to go on about how "photographic" was a bad word to use in prompts because it doesn't describe a photograph it describes equipment. WEll, okay, that's a stretch, but then we have your use case here. Literally describing the technical qualities of the equipment used and getting satisfying results.

just love it

[ionic dragon]

@soft zealot forgot to add this, no offense
another main thing i observed is that, a few keywords are very well perceived by sd, like shot on fujifilm and so.
chatgpt isnt great at generate those keywords

[soft zealot]

Sadly I cant really post the same prompt run through the same workflow but Using Photon

[strong field]

Same in my experience

[trim orbit]

#🏞｜general-with-images message like, "photographic cameras", i had to pause there. More of the same can't culture that always annoys me. "You can't do this for all these vapid reasons"

[soft zealot]

suffice to say it it very accurately matches what I prompted

[strong field]

Woof

[trim orbit]

nice tidy node graph. i guess you hid the wires?

[soft zealot]

yup, its setup so this is my daily driver view , everything elsee that I dont need to touch on a daily basis is hidden off screen

thisis my helicopter view

[trim orbit]

love

only thing i find odd is that the readme for it, seems weird to put front and center in the middle of the workflow when it's only ever needed once or twice.

maybe you just love crediting people that much that you feel it should be the focal of the graph. that's wholesome to me

[glad dust]

As anyone got a simple basic workflow in ComfyUI for using Wildcards and Dynamic Prompts similar to Automatic 1111?

[trim orbit]

https://www.reddit.com/r/StableDiffusion/comments/15oeymz/depth_sdxl_controlnet_coming_soon_brace_yourselves/ they released safetensors for canny this morning. That's not how incomming brace yourself calls work!

imagine being on a navy ship an the captains like "BRACE FOR IMPACT! INCOMING!!!" and then a day later you're still braced like "ok...?"

captains like "ITS COMING!!!"

https://tenor.com/view/austin-powers-steamroller-imminent-run-over-take-your-time-gif-13595515

[soft zealot]

Its because I get pissed off with people who cant read

or say they dint see it

this way I can screen grab back and tell them theyre blithering cockwombles

[trim orbit]

oh. that is true motivation and now i understand.

it makes so much sense now

[soft zealot]

🙂

[trim orbit]

you know what they say though. Make something idiot proof and you'll meet a bigger idiot

[soft zealot]

yup, I have the same problem with designing stuff for work lol

[strong field]

the amount of people asking you silly questions is alarming

[soft zealot]

oh good its not just me noticed that then 🙂

[strong field]

interesting

[soft zealot]

depends on how well you tell it to do something. Its only as goods as the preson instructing it.

SISO!!

[strong field]

thats feeding custom database with prompt fu

not bad

Format is
Positive prompt:
negative prompt:

[soft zealot]

this really does produce some nicew ones

[ionic dragon]

but like using good prompts to train is way better than generating random prompts
lets close this convo

no offense

[soft zealot]

I win 🙂

[trim orbit]

my philosophy is that if a good image comes out, then it's a good prompt

hard to objectively quantify the quality of a prompt before firing it through the model. it's going to be a mostly subjective discussion and a lot of superstitions are going to be regarded

[soft zealot]

its like alll those people wqho insiste on using that NAI/ESPN thing or whatever it is . I'm sorry but its just a number that an odffest to the displayed seed, its nothing magical

[trim orbit]

the one prompt craft i find weird is when people put "extra fingers" and "extra digits" or "too many limbs" in their negatives.

[soft zealot]

oh ive used plenty of werider ones

[trim orbit]

OH NICE BUDDY JESUS THROUGH GEN2

[unreal plover]

Pikalabs is free for now, it is a lot of fun animating my SDXL images

[strong field]

I don’t even understand what this means or where it came from

[trim orbit]

i just find it weird and have a working theory about why it wouldn't do much, but my core prompting philosophy still applies. if a good image comes of it then do it

[soft zealot]

again with SDXL it does seem a case ofd less is more effective

[trim orbit]

people think if you describe broken generations then it won't do those. but it doesn't account that the model wasn't trained for broken generations or has knowledge about them

[ionic dragon]

i would say your points were spot on, but it doesnt really work for me, i find finetuning better

[viral grove]

Hey all, what do you think are the chances to get SDXL Lora training on an API service like stability.ai?

[strong field]

From my understanding, and per Scott Detweillers mouth, this does not work the way people think it works.

[trim orbit]

https://tenor.com/view/moumita-this-is-your-captain-captain-brace-for-impact-gif-15458126

imagine if captain tom gave that warning like, a day before the flight

[peak dove]

Anybody using this ComfyUI extension in A1111? https://youtu.be/H5103u5uRII

[gloomy barn]

Brace for the controlnet guy making a modified UI rather than controlnet for SDXL.

[trim orbit]

i think he's working with stability to create the new architecture

[gloomy barn]

That's good if he is

[trim orbit]

well, it's good no matter what he does because he published controlnet already and i'm grateful

[gloomy barn]

Yes, but we'd need something similar for SDXL

[trim orbit]

praise @illyasviel with that illbehavior

https://tenor.com/view/brace-for-impact-brace-lt-sylvia-tilly-star-trek-discovery-brace-yourselves-hold-tightly-gif-25252982

[vagrant wasp]

why does it pause sometimes and i need to hit enter in the cmd console

[gloomy barn]

I thank God each day for the meal I eat. Then the next day, I go get food and cook again - people can be thankful while still considering what comes next.

[trim orbit]

k

[gloomy barn]

What pauses? What UI are you using?

are you sure you have to hit enter? Could it be that it just takes time, and then you hit ENTER before it finishes, but it would finish without ENTER?

[trim orbit]

https://www.reddit.com/r/StableDiffusion/comments/15otedd/some_early_samples_of_a_sdxl_high_detail_pixel/ really loving these results

[gloomy barn]

Those are very good

[vagrant wasp]

it continues when i hit enter comfyUI .. a1111 also

[gloomy barn]

Ive not seen that

is there a message in the console before/after?

[trim orbit]

you installed a1111 wrong then. the terminal never needs interaction. go forth. delete a1111 and start over. this is your quest.

[gloomy barn]

he says also comfy

[trim orbit]

try to think of what custom steps you invented the first time and avoid those

seems like a core python problem then. probably has to reinstall all of python

it's just such far out behavior that my techsupport senses are tingling. this user haas fudged something up very deep in the pipeline from tweaking and experimenting where they don't understand.

[gloomy barn]

delete and reinstall solves many issues, saves much time

[trim orbit]

yeah instead of sorting out what was changed or how many changes got piled on top of it, nuke it from orbit

its the only way to be sure

[gloomy barn]

Send the Deathstar

[vagrant wasp]

im Comfy now

[trim orbit]

shrug don't know man. wish i could help but that's a messy problem.

[gloomy barn]

then delete and reinstall comfy.

[vagrant wasp]

i done so many times

sir

i think its good

[trim orbit]

if its in both comfy an a1 then it's further up chain

[gloomy barn]

Is something else running that uses the same resources?

[vagrant wasp]

sometimes cmd just is not responsive

[trim orbit]

laptop?

[vagrant wasp]

desktop

[gloomy barn]

when I run comfy and sdnext same time, they both hang a lot

[vagrant wasp]

running fine now

on 50 queue

comfy is so comfortable

[gloomy barn]

next it happens, look to see what messages there are in the cmd window

[vagrant wasp]

the progress bar especially stuck like an old game loading...

needed to spam enter to make it move

i think it's feeling better now

[gloomy barn]

once in awhile it craps out for me...I have to reboot...

[trim orbit]

amd motherboard?

[vagrant wasp]

gigabyte

[trim orbit]

amd cpu?

[vagrant wasp]

intel

[gloomy barn]

every time I generate, I get USB errors

[trim orbit]

hdd? ssd? lots of space?

[vagrant wasp]

yes

any training tutorial for comfyUI with 6 gb gpu ram

[trim orbit]

processes freezing up is often because the process is waiitng for it's turn on the mainbus, which is a low level irq situation managed by the windows kernel and drivers.

if your drive and chipset are good, maybe you've got a weird usb device plugged in. i once had a very old 2.0 hub plugged in acting like an extension cord, and it gummed up processes and made games stutter because it was acting all weird at the irq level

[vagrant wasp]

Kohya requires RTX and 12 Gb VRAM ?

[trim orbit]

for training sdxl on kohya, thats minimum i'd say

probably the most light weight training

[gloomy barn]

That could be my issue. I do have a hub plugged in.

[gloomy barn]

Get an A100.

[vagrant wasp]

wats that

the nvidia ai chip ?

[gloomy barn]

Higher end GPU with 80 VRAM

[vagrant wasp]

the ultimate dream gpu

[gloomy barn]

no

that would be, for now, the H100

[vagrant wasp]

a simple prompt and a beautiful image

[somber hill]

--network_train_unet_only uses 100 mb more vram very weird

[coral tapir]

ok i installed a worked a bit. seems very intuitive and powerful. where do I go next? I mainly want to try and produce base images so later i can make variations of it

[strong field]

Are you confident with tinkering?

If not start with the workflow examples provided on comfy git

Dreamshaper8 is a very good model to start with

And my personal fav tutorial

https://youtube.com/playlist?list=PLIF38owJLhR1EGDY4kOnsEnMyolZgza1x

[coral tapir]

IDK what is that... but i might be willing to spend time to learn anything. ok i will check the tutorial too. i have dreamshaper. BTW comfy was very easy to install and run. with SD I ran into 10 different issues when installing 😄 can I just remove SD and use comfy?

[strong field]

It’s allot of info but take it in bites and keep at it, once you get to the point where you say, hey that’s what I wanted it gets addicting

[strong field]

Comfy installs it’s open dependencies/files, if you are done with the others it’s ok to purge

[coral tapir]

it was just a portable version. nothing actually installed. i just clicked run

[strong field]

Oh the installer version, I think that’s still ok but I use the standalone

[coral tapir]

ok let me work with it for a few days to see what happens. thanks much for the help 🙂

[ashen oracle]

is it possible to generate typeface / interesting text that could be used for a title screen on a game, for example ? or still need to do this the old school way xd

[strong field]

Not sure sdxl does this flawlessly yet

Sd1.5 with control nets is great for text

[ashen oracle]

oh - i should look there then. i haven't tried that yet - will look for a tutorial thanks 🙂

[vagrant wasp]

do i need to use face fixer ?

[trim orbit]

glad to see that tutorial being held up. it's a video tutorial that feels more like an actual lesson with a teacher. relevant accurate info. straight to business. no "like and subscribe" shenanigans. I fel like it's not popular because it's not click baity.

youtube post thumbs down button pocolypse, has the worst culture of video tutorials.

[vagrant wasp]

how to save images

it goes into the temp folder

when comfy close it deletes it

[vocal rapids]

[trim orbit]

use save image node instead of image preveiw node. they show in /comfyui/outputs

[vagrant wasp]

thanks

[vagrant wasp]

how to save it as seed name

[trim orbit]

tricky comfyui tricks that i haven't figured out yet. the save node can convert the file name to an input if you right click it

[pliant sleet]

How do i turn SDXL into a coreML model?? For iOS

[strong field]

Lol is that possible sounds so crazy it just might work

[vagrant wasp]

the sde_gpu one is faster ?

sampler

[pliant sleet]

Lmk

[strong field]

Found this

https://huggingface.co/apple/coreml-stable-diffusion-xl-base

[pliant sleet]

Sick lol that’s what i need

[tender timber]

Does anyone have a good tutorial (but not a video one) that would explain pipes and how the inputs and outputs work?

[vagrant wasp]

fastest schedular and sampler ?

[upbeat summit]

good images 🙂

[molten gull]

doing those with my python script now

[hardy cipher]

guys, why is aiohttp such trash and why does it like to torment me?

[molten gull]

what is aiohttp and whats your problem ?

[hardy cipher]

blah blah, something about json, now your comfyui install doesn't work

[molten gull]

i m not sure how i can help with this information

[hardy cipher]

oh I know. it's been a hassle. I'm trying to just create a new venv environment

[hardy cipher]

basically some sort of dict_keys issue. and so none of the nodes were being recognized anymore. tried to google fixes on my own, but found absolutely nothing that appeared helpful.

not quite sure what to do with this

    text = dumps(data)
           ^^^^^^^^^^^
  File "C:\Python311\Lib\json\__init__.py", line 231, in dumps
    return _default_encoder.encode(obj)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Python311\Lib\json\encoder.py", line 200, in encode
    chunks = self.iterencode(o, _one_shot=True)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Python311\Lib\json\encoder.py", line 258, in iterencode
    return _iterencode(o, 0)
           ^^^^^^^^^^^^^^^^^
  File "C:\Python311\Lib\json\encoder.py", line 180, in default
    raise TypeError(f'Object of type {o.__class__.__name__} '
TypeError: Object of type dict_keys is not JSON serializable
INFO:aiohttp.access:127.0.0.1 [12/Aug/2023:18:21:13 +0000] "GET /object_info HTTP/1.1" 500 245 "http://127.0.0.1:8188/" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/115.0.0.0 Safari/537.36"```

tried to figure it out for a couple hours then just created a new venv environment which probably won't help anything, lol

[molten gull]

you're trying to get comfyUI to run ? or somethign else ?

[hardy cipher]

comfyui. I really have no idea what caused those errors to start

so basically it'd load fine, but no nodes

then when I would go to load a workflow I'd get bombarded with errors. and those are the end of them

and didn't want to come in here all needy trying to get tech support. still have the previous venv environment folder if it might turn out to be easily fixed. but at this point just starting anew

[molten gull]

did you try a clean install already ?

[hardy cipher]

I do like I installed pytorch with cuda then it tells me my pytorch install doesn't have cuda

clean as in the entire program? thus far just trying a new python environment

[molten gull]

you could try this here: https://github.com/comfyanonymous/ComfyUI/releases

a windows portable version

[hardy cipher]

yeah. if this doesn't workout I'll just wipe everything but my models and saves and start over with that

if I do start over I'm going to be much more mindful of these nodepacks. some of them seem to be trojan horses of pain

[trim orbit]

all trojan horses are malicious in nature

🤓

[hardy cipher]

man, I know about trojan horses. but that's another topic. one of these node installs though, broke aiohttp and so none of the nodes were being recognized

[molten gull]

what nodepacks are you talking about ?

[hardy cipher]

man, I don't know honestly

I realized pretty much all of them needed updated. thought it went fine. then restarted and all nodes red and undefined

[trim orbit]

unrelated. somethign i've noticed about kohya scripts lately is the example image generation still uses the vae that was first released for 1.0, but they've made the 0.9 the official 1.0 release now.

[hardy cipher]

alright, well at least it sees cuda now

[trim orbit]

oh i guess it's using the vae that's baked into the model i'm training it from. doiye.

[hardy cipher]

how can a trillion dollar company release such janky software? lol. I'm obviously not an expert. just seems like they could do a little better than this pytorch cuda nonsense

[trim orbit]

i've used trojan horses in my day. when i was aa young lad on the internet i used to pull the ol "hotboobs.jpg .exe" trick on friends

[hardy cipher]

I used to bind sub7 to .exe files that would play stupid flash videos

[trim orbit]

because i'ts bleeding edge frontiere research. pioneering aint ez

[trim orbit]

yeah thats the exact one i was into. such a fun time that tool was. the matrix chat!

[hardy cipher]

then I set the file to delete itself when it installed

[trim orbit]

okay i should stop admitting to malicious crimes

[hardy cipher]

hahaha

well I don't know if what I did was really too criminal, maybe on a petty level

found the master password on astalavista.box.sk

[trim orbit]

sub7 was mostly harmless and a fun exploration of weak security practices

[hardy cipher]

started scanning large blocks of ip addresses for installs, LOL

[trim orbit]

astalavista. what a time to be alive

[hardy cipher]

I found a bunch. then change port number, change password

[hardy cipher]

then got a letters from my isp telling me I should probably stop doing such things. probably hit the wrong ip addresses. I wasn't very smart about it

[trim orbit]

petty petty stuff that was made so easy by the protocols of early internet

[hardy cipher]

I just used it to freak out my friends mostly. have things print out on their printers at 4 am

[trim orbit]

war driving was a thing for a while. getting into random wifi networks

people would have their cars set up with all sorts of antennas

other people totally not me

[hardy cipher]

my buddy had a little setup to watch surveillance cameras

[trim orbit]

it's relatively secure now a days

[hardy cipher]

little crappy monitor and antenna

back then a lot of things literally had zero security

their security was not many people knew how to access them

[trim orbit]

encryption was breakable too

[strong field]

Hey I remember this

Lll

[molten gull]

stuff will change extremely soon extremely wild, once AGI is there

[trim orbit]

social engineering 2.0

[molten gull]

i fear it will probably be the end of humanity

[hardy cipher]

good news everyone, still a few minor node errors, but at least they are recognized again

humanity is ill equipped to deal with social media

[molten gull]

can i help you somehow pictures?

[strong field]

Well humanity + AGI will be something different altogether

[molten gull]

and no, AGI will produce that intelligence explosion

[trim orbit]

https://tenor.com/view/futurama-professor-farnsworth-hail-science-gif-5965670

[hardy cipher]

aig will create a dystopian nightmare

hoping utopian

[molten gull]

AGI will create nothing in that humans will have any place at all

[hardy cipher]

but I believe greed will lead to dystopian

[soft zealot]

PyTorch was originally developed by Meta aka Facebook.

"All your AI images belong to us"

Sorry sir , didn't you read the small print?

LOL

[trim orbit]

thats a greek pun that means both perfect place and no place. the word itself is a recognition that it's unattainable

[molten gull]

humans will not matter anymore, not at all. even timescale will be something completely different. stuff that happens for AGI will take place in milliseconds

[trim orbit]

point is, we need a better word

[hardy cipher]

well, some of the errors are things I just need to put the time in with. but then I don't really understand this at all. or if it's even an issue. I'm not well versed on python, trying to learn, but it's a lot

127.0.0.1 [12/Aug/2023:19:24:44 +0000] "GET /nested_node_defs HTTP/1.1" 200 159 "http://127.0.0.1:8188/" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/115.0.0.0 Safari/537.36"
127.0.0.1 [12/Aug/2023:19:24:44 +0000] "POST /mtb/debug HTTP/1.1" 200 187 "http://127.0.0.1:8188/" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/115.0.0.0 Safari/537.36"
127.0.0.1 [12/Aug/2023:19:24:44 +0000] "GET /pysssss/workflows HTTP/1.1" 200 184 "http://127.0.0.1:8188/" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/115.0.0.0 Safari/537.36"
127.0.0.1 [12/Aug/2023:19:24:44 +0000] "GET /extensions/pysssss/CustomScripts/assets/canvas2svg.js HTTP/1.1" 304 203 "http://127.0.0.1:8188/" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/115.0.0.0 Safari/537.36"
127.0.0.1 [12/Aug/2023:19:24:44 +0000] "GET /extensions/pysssss/CustomScripts/assets/favicon.ico HTTP/1.1" 304 177 "http://127.0.0.1:8188/" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/115.0.0.0 Safari/537.36"```

[strong field]

They will just pony up large scale missions to other galaxies cause what’s a couple thousand years to a hive mind

[molten gull]

my script-thingie is getting better

the stuff you posted there doesnt seem to be an error, just infos

[hardy cipher]

ahh, well that's good. thanks

[trim orbit]

a cool book i like is frank herbert's "Destination: Void" in that a team of engineers engineer an artificial conciousness outside of the solar system and the thing becomes jesus. since it's so aware and intelligent it has god like powers

and blinks them to a new galaxy where the second book starts . "The Jesus Incident"

[molten gull]

jfk, what will happen is that AGI with lets say IQ 100 will develop an AGI with IQ 101, which then will develop one with 102, which then will ...

[trim orbit]

Better than Dune imo. more digesteable

[molten gull]

and this will speed up uncontrollably and really fast

[hardy cipher]

the technological singularity

[molten gull]

read this here, it's long but really really really good:

https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-1.html

[trim orbit]

IQ doesn't matter. a real AGI will determine this in a blink

[hardy cipher]

back before gpt-4 was severely guardrailed and lobotomized I asked it to speculate on post singularity existence. then 2 singularities in the future, then 3, etc

people creating pocket universes before they eat breakfast

[trim orbit]

i wouldn't ever say gpt anything got lobotomized. it's entire architecture is limited and it's just a mechanical turk trick that it can do what ti does today

[hardy cipher]

well it's output has been. not the model itself

[trim orbit]

rote memorization is what it can do. thats about it

[molten gull]

people don't matter soon. it's like looking down at ants, we might find them interesting, and might like to study them for fun, but when they're in your living room you'll kill them without even thinking

but read that article, it explains most things really well

[hardy cipher]

problem I think is the transformer model structure right? it has no way to discern importance. so token in token out

[trim orbit]

read everything on waitbutwhy imo

it's such a great site

[hardy cipher]

what makes the human brain so efficient is it's ability to differentiate

short short term memory, short term memory, long term memory

most of what we take in is discarded within seconds

[soft zealot]

although that ability does vary greatly on the individual ;o)

[trim orbit]

needs a 2nd pass but the motif is money

[molten gull]

i dont have a good idea of how to do second pass stuff yet sadly 😦

[hardy cipher]

I actually question whether ai is going to surpass all the humans like that. it's something I've thought about for a while. but these days I wonder if humans are going to evolve with it. or the few elite at least. so they keep getting augmented as the AI increases in complexity and ability

[hardy cipher]

just run the same sampler on it again. adjust as needed. I like to sometimes darken the image a bit and/or add details before the second pass

upscale, add detail, downscale, darken, second pass

well multiply by itself

[trim orbit]

we'll install it in our brains as soon as it's possible. augmented will exist along side artificials. i see a future like her happening. except humans don't just fall in love with the ai. we blend and redevelop ourselves to become it

a new side to the anthropological story. the creator striving to become it's creation. legendary stuff really

"her" is a good movie how it's just so casual about way it unfolds

[molten gull]

just the same thing again ? with same seed, same pos/neg prompt just feed it the latent that i got in the end ?

[hardy cipher]

problem is the elite few will be able to get vastly more sophisticated augmentation. to the point that we could become effectively different species

[hardy cipher]

that's just how I do it, might not be the "right" way

[trim orbit]

i mean like a hires fix would do in automatic

sytan's workflow has a good example one set up

[molten gull]

can you show me a screenshot of a possible comfyUI setup ?

[trim orbit]

i geuss his is a 2x upscale though

[hardy cipher]

you certainly don't want to emulate mine, lol

well I could show you a barebones maybe

I like the searge node pack for simple sdxl stuff

[trim orbit]

i think if you drag a hires fixed workflow from auto in comfy, it knows how to make that a workflow

[molten gull]

so far i only made my own workflows in comfyui, never took anything from somewhere else yet

now that's gonna be some tea party 🙂

[trim orbit]

CHANGE PLACES

[native knot]

@soft zealot - On your workflow using the failfast custom node, how do you hide the links? All I can see in the drop-down is straight, linear, & spline.

[soft zealot]

Unless it’s been updated there should be a hide option

[native knot]

I've got nothing like that.

Looks like v1.6.0 removed links-render-mode stating that comfyUI integrated the feature into its core...unfortunately, they didn't include hidden as an option.

[hardy cipher]

I make weird things. there are lots of workflows online. I just experiment a lot. put something together, see how it works out

then knock it down like a sandcastle and build something else

[native knot]

Pulled down the linksRenderMode.js from the last version that had it. That did the trick.

[hardy cipher]

makes some cool stuff though.

[hasty smelt]

Hey everyone, something I haven't fully grasped yet. Apologies if the question seems simple. When someone adds the name of a random artist to their prompt, what does the model do? Does it access the web to fetch references of the requested style, or does the model need to have been trained with the selected artistic style? Again, I'm sorry for the question, I'm just trying to understand how SDXL performs so well with these prompts.

[hardy cipher]

it does not access the web

it has went over a ridiculous volume of images and data

[trim orbit]

no. everything the model knows was trained into it during a very expensive and long process with data center hardware

[molten gull]

so how would i best do that second pass thing ?

[trim orbit]

HUGE datasets of knowledge

its amazing that it all fits down into a disc sized file

[hardy cipher]

without changing the structure of the image to a degree I honestly don't know the best approach. my approaches do refine it, but at the cost expense of the original structure of the first image

[molten gull]

how much change ?

[hasty smelt]

thanks guys

[hardy cipher]

example

[molten gull]

wasnt it good before already ? 🙂

[hardy cipher]

yeah, first one was decent

[trim orbit]

sytans i think is best way to do it for starting out, but its a 2x upscale pass

[molten gull]

like what could you do with this ?

[molten gull]

what does 2x upscale pass mean ?

[trim orbit]

means the second pass increases width by 2 and height by 2

[molten gull]

and that would fix the images quality problem ?

[hardy cipher]

well not always

[hardy cipher]

you should try out the noise offset lora

[molten gull]

it does an upscale 2x, and then a downscale 0.5x again ?

[hardy cipher]

and experiment with skipping the last layers

what kind of an abomination of a random workflow did I load up?

[molten gull]

jesus, i can top that 🙂

[hardy cipher]

nice contrasts

[trim orbit]

https://tenor.com/view/teen-witch-top-that-rap-gif-6140710

[hardy cipher]

this image will load up one of my random workflows, but might have some nodes you don't have installed

[molten gull]

can you send a screenshot of the comfyUI with some second pass fix exmaples by chance?

[hardy cipher]

I'd suggest looking on civitai or github. their approaches are more better than mine

[molten gull]

got a link ? 🙂

[hardy cipher]

https://civitai.com/models/111463

[molten gull]

thanks

[hardy cipher]

also, not sure if you know, but you can drag a lot of images into comfy and it will load the workflow

[molten gull]

yeah, i know that, but its a good tip

[fiery knoll]

how are you getting auto1111 to v 1.5.x? Mine still shows 1.3.0 after "git pull"

[trim orbit]

git status

[molten gull]

has any of you managed to create an textual-embedding with a1111 and sdxl ? i get an error when i try to create one

[ashen oracle]

trying to generate a "game logo/title" with sdxl and some masking, not bad lol. if anyone has experience w this process on 1.5 or sdxl would love some pointers

[coral tapir]

ok watched those nice tutorials. 5 were there and 2 of them sdXL which I dont have. For the thing I asked I think the solution is that we can introduce new positive prompts in middle of the nodes so we change some details when we want. right?

[trim orbit]

dont think ti has been done for xl in auto. i used kohya when i made one but it only worked in comfyui

[hardy cipher]

might try prompt stylers. the one I use has typeset as well as a bunch of video game styles

basically just adds word to your prompt. you could technically do it yourself, but streamlines the process

*words

his right hand is something

fingers like tree branches

[soft zealot]

@clever verge Question if you're around.

Your revise dprompt stylers uses all the JSON files in the specified directory.

If possible would it be possible to add a slector so that each node you add could point at a different directory ? Ie give the user the ability ro split up the styles.

TIA

[molten gull]

is the auto guy still developing a1111 ? might TI come there ?

[hardy cipher]

isn't it multiple people? I don't know. just seems to have become a bogged down train wreck

[hardy cipher]

that looks good

[clever verge]

Yeah, I've thought about that but I'm not sure on how that should work redrawing the node. If anyone have an example it would be pretty easy to implement.

[hardy cipher]

if you're trying for more detail rather than going for a second pass you could experiment with noise offset and skipping layers. those od a lot to refine things

[soft zealot]

the nearest example I can thinkl of is the Load Image Node where yu can select an image via a browse function and the name is loaded in a widget.

Maybe that could be leveraged??

[fiery knoll]

thanks man. There was an issue I didn't see in the pull (some issue with the docker webui using symlinks for stuff, and some files the docker version had deleted.. I couldn't stash without re-checking them out)

# python -u webui.py --listen --port 7860 --allow-code --medvram --xformers --enable-insecure-extension-access --api # the docker version runs it all as root apparently

[soft zealot]

@clever verge well there is the dirty hack way of doing it lol

[fiery knoll]

apparently I'd need to put https://github.com/Stability-AI/generative-models somewhere, but it doesn't make sense (webui should be handling deps)

[clever verge]

Yeah, that's dirty but working! I want the functionality but since I found no simple yet effective way I did what I just did with multiple jsons.

[soft zealot]

which Im presuming you have figured out is this way lol
(and a quick edit) fine for personal usae but..................

[clever verge]

Really wish to have a multi-select that can filter by category

[hardy cipher]

it works pretty well though

using it now

[soft zealot]

@twri , is this slightly less dirty ??

and then have multiple nodes pointing a different json numbers ??

or maybe not, what do I know lol

[trim orbit]

wooooo updated comfy has a right click copy image

[soft zealot]

since when ?

I already have a Righ CLick Copy Image and have dione for errrrm, a while IIRC

[hardy cipher]

I like how civitai decided to default everyone to beta fast search that doesn't seem to give the option to restrict which version of stable diffusion you're searching for

right in time for sdxl

that's smart

whoa https://civitai.com/models/127485/azdwj

top notch

[soft zealot]

I just click on the models tab then select the filters to be whatever I want

[hardy cipher]

no, I did that.

but this beta quicksearch thing

it's a popup menu essentially

doesn't give that option

[soft zealot]

i disabled that weeks ago when they first launched it

[steel mist]

Oh, in the search box. Yeah that's actually super frustruating

[hardy cipher]

I opted out so now I'm back to normal

I actually downloaded a couple cool looking loras thinking they were sdxl since I have my settings to only search for sdxl atm

but nope, 2.5

1.5

I guess it's not too hard to figure out once you use it

[trim orbit]

some people i think are publishing sd 15 loras as xl loras because they've trained them with 1024 images and they think its a flex

[plush drift]

https://civitai.com/posts/492116

[trim orbit]

https://github.com/lllyasviel/Fooocus fooooocus

[soft zealot]

well this one stood out from th ebatch

[trim orbit]

looks like a neat ui

[soft zealot]

well its pissed me off already by forcing a download of specific models!!!!

Does this mean it will only work with one specific model?

that would be annoying

and this annoys me as well 🙂

[trim orbit]

at the very least you could just copy another xl model into it's place and it won't know the difference. but i think it's just downloading the base on first launch because it can't find any

yeah its for less advanced users to get into it i think . i'll recommend it to people after a test drive maybe

a nice economical vehical

[soft zealot]

no I suspect its downloading that specific file because its been hard coded too use that specific file in the initial install rather than going " ooooh look I have a directory fulll of files"

[hasty smelt]

Hey everyone, I know that SDXL uses 16GB of RAM. However, since I only have 16GB, is there a way to limit SDXL's usage to a maximum of 90% in ComfyUI or something similar? This way, I could reserve around 10% for my other tasks. It seems normal for SDXL to use up 99% of my RAM, causing stutters in applications like Spotify or others. I'm curious to know if I can reduce the memory consumption a bit, even if it means slightly longer rendering times. Thanks

[trim orbit]

its open source so i mean, nothing to suspect really

just look

[soft zealot]

Yes I did my symlink trick first befor running th einitial run.bat so all the folders were populated

[soft zealot]

its a n annpoyance

[trim orbit]

the launch with update script looks like it does a fetch from the git repo. must break your syms

[gloomy barn]

two things I dont like hearing right now: no controlnet for SDXL yet, can I meet Sunday with students

[glad grove]

yea we meeting with sensei Seagal soon

[gloomy barn]

hopefully not a lunch meeting...you'll get none

[trim orbit]

Joes holding the big boy cnets and could release them in the mean time. canny is really fun to toy with. wish i had normal, inpaint, openpose, the whole band

we gotta get the band back together

GO JOE

[soft zealot]

I killed it the dirt time, redid the symlinks and theyre still there as per the screenshot

[mellow tendon]

There is Canny for SDXL https://huggingface.co/diffusers/controlnet-canny-sdxl-1.0

Works great in ComfyUI with my brief testing of it.

[soft zealot]

And if this other chaps comment is true I will be really annoyed

https://github.com/lllyasviel/Fooocus/issues/31

[trim orbit]

they put out a safetensor this morning but last time i looked, the extension was hardcoded to not work with sdxl loaded

[trim orbit]

yeah. i don't think it's meant to replace any ui. i'ts made for people who want to dip their toes

think its for a different demographic

[soft zealot]

If its locked to one specific model it should be wirtten in big red ltters at the top of the bloddy page

[trim orbit]

"forced download" is a much too dramatic of a ticket headline. we're not trying to be tabloid journalists when making bug reports are we? cool

you seem mad about this. sorry for sharing smethiing new

i think i brought the eyes of sauron upon this controlnet creator. my bad

[soft zealot]

"Forced download" i san accurate description

[trim orbit]

it broke your symlink. be reasonable

[soft zealot]

no it hasnt broken the symlink, the symlink is still there and the files are still being downloaded

(seperate footnote ran out of space on that device so having to move some files around)

[trim orbit]

you don't have the base 1.0 model in your model folder? it wouldn't need to force dl if your symlinks weren't broken and it was still linked i imagine.

this isn't a conspiracy against you. you're just not the demographic. dial it back winston

[soft zealot]

I have the Base1.0 model WITHOUT the 0.9 VAE baked in which is the version that is being force downloaded

I chose not to download the baked in 0.9 vae and day to day I load the VAEs seperately

[trim orbit]

then you have the old base 1 wiht the old vae baked in. this one's configured to launch with the current base.

expected behavior is that it launches and looks for the standard model and downloads it

dial it back

[soft zealot]

and it would seem with no other SDXL model which is a bit sucky

[trim orbit]

easy to solve in that case. just symlink a different model into it's place instead

[soft zealot]

I ccould understand it downloading a model if it detected no models at all but that isnt the case

[trim orbit]

or change the open sourced code

it's really just manufacturing and lookign for problems. i'm not going to enable this behavior any more.

[soft zealot]

the point is I thought I would give this a whirl (in the sanme way that even though I had no real intention of using it ) I gave Stable Swarm a whirl but at the moment the OOTB install experience is not IMHO exact;ly srellar.

It has annoyances 🙂

and yes I suppose you could say it is looking for problems, its UAT which is part of my daily job

I would prefer it if an instaklller asked if I wanted to download something rather than just automatically downloading it

[trim orbit]

yeah. i loath the attitude. instead of seeing a new 1.0 release of a light weight project, you make a dramatic post about the first roadbump you found. you went into this looking for problems and irrationally found one with no effort on your part to find a work around.

expected behavior is that it downloads the base model. period. i brought contempt and hate down on his new release. i caused rain on his release day. makes me feel like shit. wonder how you feel bout it.

done for today

[soft zealot]

I mena it snot being binned offf, I'll still give it a dfair tryout once I've got the files

[keen nimbus]

My first gen in XL

[elfin flare]

What's currently the best SDXL checkpoint for prompt understanding?

I'm currently using base, but I'm curious about the others.

[upbeat summit]

Base SDXL is already insanely good imo. it's more about styles and concepts right now but we'll see what happens with fine-tunings in the near future.

[elfin flare]

Alrighty. Thank you for the response.

[upbeat summit]

I think the storytelling capabilities of SDXL are on a really high level. It really surprised me a couple of times

[autumn forum]

Agreed.

[elfin flare]

Indeed.

I think so far I've been able to do batch size 16 with sequential offload?

[upbeat summit]

You're welcome. Your artwork is really great 🙂 love the 3D pixel + voxel mash-ups

[elfin flare]

That one I like.

@upbeat summit Now combine it with Pikalabs

And...

[upbeat summit]

it's alive!

[heady vale]

my attempt without loras

[elfin flare]

Without a lora, and yet you got the style of that lora I use almost down.

Impressive.

[heady vale]

the voxels are more chunky but its similar

[elfin flare]

What I like though

is that it's almost as if the spear blade is sprite-like

Look how thin it looks.

[heady vale]

yeah combining pixel art with voxel really well

[elfin flare]

This is a really bad idea

I'm gonna try to do a batch size 32.

[heady vale]

really nice

[elfin flare]

Sequential offload will save me surely

[heady vale]

reminds me of lego world or something

[elfin flare]

It's like that combined with bionicles.

There is no way I'm going to survive a 32-size batch run.

I can do 16, but 32 might be impossible on an A770.

This one looks like a 3D print.

[heady vale]

ET phone home?

[elfin flare]

Allocation is out of device memory on current platform. As I expected.

I'll do 24.

[safe spear]

Is inpainting supported by SDXL? If so, is there a demo to use that function?

[heady vale]

XL has a good imagination

[upbeat summit]

you need SDXL-Sprite-Animation

[elfin flare]

I assume that's a sprite sheet LORA?

[upbeat summit]

it actually doesn't exist. it's something I would wish existed though so you could make game assets with your work 😄

[elfin flare]

Ah, I see.

[delicate kelp]

For the DynaVision fans out there - I've released the latest update which is the NSFW update people have been asking for 🙂 https://civitai.com/models/122606?modelVersionId=139617

[elfin flare]

My guy, I was just looking at the checkpoints and I saw that there.

Lmao

Now goddamn did the model get creative with this one's design

Holy hell, man.

[wicked dagger]

Have a look at this blog: https://ericri.medium.com/stable-diffusion-sdxl-wildcards-and-comfyui-e8483e00e1c

[fresh path]

I was like why is this guy writing so weirdly and then it says he’s likes the ex executive of windows

[elfin flare]

Time to spice it up. This is Crystal Clear XL

[upbeat summit]

these are awesome! great work

[hardy cipher]

I know you guys discuss it all the time in here, but what exactly are you using for lora training? was considering testing the waters. I'm sure I'd need to use colab or something though due to my inadequate hardware

[heady vale]

Awesome model dude, as always. I did some samples with and without refiner. Refiner messes the eyes. Without looks much better. Impressive

[delicate kelp]

oh yeah, don't use the refiner with my models, it'll screw all of em up. Someday we'll figure out how to train the refiners and maybe we can use them, but for now it's MUCH more preferable to just use high res fix (or the comfy equiv) and just run your image through upscale at 0.5 or so denoiser. should improve detail, fix any eye issues and make it look even more crisp

[shy kelp]

hi all, I'm running SD on my local machine and trying to use open pose with controlnet, I can get the pose preview working but my text to image generations don't seem to take into account the pose that was created, I have the controlnet enabled
any ideas why the generations are not taking the control net open pose information? thanks

[hardy cipher]

how much vram do you have?

[shy kelp]

if you're replying to me I have 8 GB vram

[hardy cipher]

might not be enough for controlnet with xl

[shy kelp]

ahh, thanks

[hardy cipher]

actually not sure the exact cutoff for that, hopefully someone figures out how to make it more efficient in time

[glad grove]

thers an openpose controlnet for xl?

[gloomy barn]

no

[hardy cipher]

I guess. I haven't even looked into it. I know my little 6gb 3060 would fail

only model I saw was canny. but who knows

[gloomy barn]

there is canny only, not working in a1111 or sdnext, but working in comfy

there is no openpose controlnet for SDXL

[glad grove]

then thats why it doesnt work for buddymooner,its broken

[hardy cipher]

hmm, maybe using 1.5 controlnet openpose with xl?

don't think that would work though

[shy kelp]

I just tried a smaller model and reduced resolution to 512x512 and its taking the pose information now

must be something with sdxl and only 8 gb vram

[glad grove]

nothing to do with vram its just broken,wait for the official release of controlnet xl

[shy kelp]

is there a release date for that?

[high skiff]

these gens look straight out of 1.5

Interesting results

[glad grove]

if it works with 512x512 u could create an img with that res then when its good upscale it on img2img

[heady vale]

yep. XL is catching up real fast

[hardy cipher]

all the goofs saying it wasn't as good as 1.5 literally the same day it was offiicially released, lol

[high skiff]

I am not sure I would say that, I just mean it has the same artifacts and general look as a lot of 1.5 models

it was more of a statement on rendering

[hardy cipher]

Sytan, a bit off topic, but any idea how to get this working with oobabooga? https://huggingface.co/stabilityai/stablecode-completion-alpha-3b-4k

I've got it to load, but not sure what settings it would need

[high skiff]

I have had no succes unfortunately

it seems as tho the non-instruct one is kinda just broke for ooba

I was trying to get it to work for about an hour when it launched, everything just returned incorrect context

[hardy cipher]

it'll respond in 0.2 seconds

with nothing

[high skiff]

yeah

check your command line, probably says wrong context or something

it'll say something like context between user and AI

[strong field]

have you tried the blokes version?

[hardy cipher]

I just barely understand how those language models work. it's such a crazy realm

[strong field]

he gives instruct

[hardy cipher]

I did not know thebloke had a version. I'll check that out

thanks!

[strong field]

https://huggingface.co/TheBloke/stablecode-completion-alpha-3b-4k-GPTQ

[hardy cipher]

there are so many weird intricacies to the open source LLMs, and they're all nonsensically named like weed strains. it's really bizarre

[hardy cipher]

lol

these are good

[heady vale]

very good

[hardy cipher]

have you tried it out? not exactly sure how to use it, but I guess I'll figure it out. good thing is a 3b model doesn't use too many resources

I have another coding model, but haven't messed with it much

[strong field]

ive been hearing wizard is better lol

ive messed with it here an there

write the prompt in the the left column of text gen, better results for me that way

i couldnt get it to respond unless i was in chat mode

honesly, i liked claude better for script writing

[hardy cipher]

only in chat mode? nice

which one of yu made this? https://civitai.com/models/125942/will-smith-sdxl-lora

those images

[heady vale]

loool

[glad grove]

no steven seagal lora yet for XL 😔

[heady vale]

wow...actually holding things too

[high skiff]

that is so crazy how much it picked up the 1.5 plastic rendering

Was it trained on 1.5 only?

[heady vale]

no idea

[high skiff]

Wanted to share this here. I spent 2.5 hours deep cleaning my recently purchased second hand 3090, and my god

[hardy cipher]

so apparently 8gb vram minimum for lora training? I wonder if it'd notice that I only have 6gb

[high skiff]

some parts of my card dropped by 15-25C

[hardy cipher]

nice. looks like a dirty air conditioner filter

[high skiff]

worst part was it was tar, not dry dust

so it was all sticky and greasy, so I had to use brushes, picks, and IPA to dissolve it

but my god did it help

[hardy cipher]

burning things next to the computer?

[upbeat summit]

smoker

[glad grove]

what checkpoint is that? it looks so real

[high skiff]

not a smoker, everybody keeps saying that lol

[hardy cipher]

this fan was an incredible hassle to get to

[high skiff]

its from bing close to a kitchen

[high skiff]

🫦

lmao

[glad grove]

[upbeat summit]

[hardy cipher]

had to take out like 50 screws to get all the cooling nonsense off

[upbeat summit]

well if you want your hardware to live longer, I remove the dust on a regular basis

[high skiff]

luckily, I didn't have to tear down the whole card

[hardy cipher]

it had actually seized up. figured it was done. but I cleaned it up and it started working again. it actually got better over time. bizarre

[high skiff]

its so crazy, I had my current PC for 6 years, and never had to clean it. There was barely any dust in my GPU and my CPU cooler after all that time

[hardy cipher]

it felt like it was about to fall apart so pushed it back together, lol

[high skiff]

like, its a little dusty, but it looks like some peoples PC's after a couple months

[upbeat summit]

not bad! I have to clean my PCs at least every 3 months

[high skiff]

Like this is my 3060ti after 2.5 years of being in my PC I used every day

I honestly wonder if its cause I live in a really dry place, maybe the dust doesn't stick with less humidity?

I never even cleaned that card once

[fresh path]

bro got the cleanest room in existence

bro running his pc in a clean room

[high skiff]

my 1080 I had for 3 years was the same

[hardy cipher]

have you ever had to put these little connectors on? pulled those things loose, thought I was supposed to solder them back on, lol

[high skiff]

my room is so dirty 💀

[hardy cipher]

in the end I realized that would have been a terrible idea

[high skiff]

my keyboard is absolutely foul lmao

[autumn forum]

i eat at my desk and have cats.. lmao safe to say my pc is duuuusty.

[high skiff]

I am really hairy, so soooo much hair builds up in my keyboard

its like a sweater, i swear

I have been considering thick foam modding my keyboard just to help with it

and I would also like a more thoccy keyboard as well

[hardy cipher]

https://tenor.com/view/star-wars-smiling-happy-hairy-gif-15918651

[high skiff]

basically

if I can find a proper foam mod for my keyboard, I'll do it

I can't seem to find any kits tho

[fresh path]

is it possible to use sdxl as an upscaler for images generated with 1.5?

[high skiff]

theorhetically, yeah

but not if its for like a finetuned model

[fresh path]

yeah I have some loras and stuff

[high skiff]

then no

[fresh path]

gotcha

I have a rly nice workflow right now but the hi res fix is kinda bad

also sdxl takes like 5 minutes per prompt for me

[high skiff]

what GPU?

[fresh path]

6900xt on windows

[high skiff]

damn

[fresh path]

I can go on ubuntu but like, sdxl perf is not much better

and it breaks alot

1.2s/it for a $1000 card

[hardy cipher]

sytan, what exactly do you use for lora training? I'm trying to figure out the best approach

I'm getting over 2 it/s with 6gb vram

[high skiff]

If you have only 6GB VRAM, you are not gonna be training any SDXL LoRA's unfortunately

[hardy cipher]

well yeah, trying to do it online

colab or something. that's where I trained 1.5 loras

simply for speed reasons

[high skiff]

ah, I just use kohya, or derrian works well as well

[hardy cipher]

cool, I'll see what I can figure out. wasn't sure if the same dreambooth trainer I used for 1.5 would work for xl or what

thanks

[high skiff]

from what i have seen, at least with how I train, 16GB VRAM is the minimum for SDXL LoRA's, and thats at like BS1-2, which is not as ideal cause of gradients

tho, it does seem like enabling additional steps of gradient accumulation does help tho

[hardy cipher]

well I'll just rent myself an a100. I tried training a model (not stable diffusion) with free colab once, and good lord. it would have taken days. but took less than an hour with the a100

I think they only let you use 40gb vram though. there's probably better options than colab

[high skiff]

I wouldn't recommend an A100 for LoRA's, they are overkill, and not much faster than a 3090/4090

And for the price of 1 A100 per hour, you could probably get an 8x 3090 system which would shred it

[hardy cipher]

do you have any suggestions for online training? I just use colab because it links to my google drive

[high skiff]

I train SDXL LoRA's on a 3090, and depending on the settings, I can get up to batch size 12

[high skiff]

I don't sorry man, i do everything local

[hardy cipher]

yeah, that's fair. I need to do some research. by default kohya on colab only has options for 1.5 and 2.1 models

any idea which one of these I should use? those descriptions are so ambiguous

I've been googling trying to figure it out. but everything seems to be moderate level clickbait

this looks better https://colab.research.google.com/github/camenduru/kohya_ss-colab/blob/main/kohya_ss_colab.ipynb

now I just need to figure out what 90 percent of these settings mean

[fresh path]

theres a runpod tutorial

make sure you have collab pro

[hardy cipher]

thanks. yeah, I have pro

[latent zealot]

Is there a way to swap faces, using a real image starting out? and swap it to either a face from another real image or a generated one ?

[queen adder]

when captioning your txt files do you not use any commas? because when i do blip captioning it will just put in the txt file something like man standing next to car

[hardy cipher]

yes. just look up "face swap" along with whatever UI you're using.

[latent zealot]

i get roop but that seems to be just generated to imported I kinda need the other way around but ill keep diggin

[hardy cipher]

ahh. well if you have gimp or photoshop you could do something rough there and then fix it up with inpainting

[soft zealot]

Apparently there is ReActor

https://youtu.be/FShlpMxbU0E?t=389

[latent zealot]

ill give it a shot too appreciated

[ionic dragon]

does she look like selena gomez?

[hardy cipher]

yeah

[crisp owl]

I can see the resemblance

[queen adder]

when training SDXL model, is 50s/it considered slow? im using RTX 3090 GPU

[ionic dragon]

hows the overall picture?

[hardy cipher]

I mean, it looks good. got some asymmetrical sleeves going though, lol

[hardy cipher]

full day of training, lol

I'm not an expert, but that seems pretty slow to me

[ionic dragon]

is face and neck coordinated?

[autumn forum]

yeah, are you using really big buckets

?

[ionic dragon]

which one?

[hardy cipher]

I'd say right. kind of a round face on the left

[vale eagle]

Your dim might be setting too high

[ionic dragon]

and the left one's face and neck arent coordinated?

[hardy cipher]

yeah, and a bit of a small head I think

[glad grove]

least cluttered italian bedroom

[arctic bloom]

comfyUI visualized?

[molten gull]

try this one @arctic bloom

it's unlabeled right out of the box

[arctic bloom]

nope, it doesn't.
also asked mate on windows to drag&drop it.
also -- no

[ionic dragon]

does she look like selena gomez?

[crisp owl]

I still see the resemblance

[ionic dragon]

on a scale of 1-10 how do you rate her resemblance to the face?

[crisp owl]

I mean, if you didn't say anything about who it was, I would have been able to pinpoint it to her face. so, between 7-10?

[upbeat summit]

it's close. ~96% there. if you would categorize this as fan art it's good. if it should be a very close representation of the real person the likeness it's not quite there yet

[heady vale]

cheeks need to be a little thicker

[crisp owl]

Depends if we're going for her look today, or younger her

[heady vale]

even in 2011 she was similar in the cheeks

[crisp owl]

Oh that wasn't in reference to your comment, just stating in general.

[upbeat summit]

but of course everyone would recognize at first glance who that should be - so it's really not bad 🙂

[ionic dragon]

the dataset was of recent pics of 2021, 2022 and 2023

[ionic dragon]

i trained with 16k steps, but its not still on point

[upbeat summit]

yeah. I can't help with training sadly but I've seen some amazing famous people Loras on Civitai that have an almost 100% likeness

depending on the seed some look absolutely on point

[ionic dragon]

i have generated v4 versions

and this is the best i got

[crisp owl]

Yeah I've made one lora, and one checkpoint, both on myself, just for kicks. THey were both...meh...and I knew nothing about what I was doing. So, take that as you will lol

[ionic dragon]

gotta try more

[hardy cipher]

how's the training compare to 1.5? any differences as far as your approach?

[ionic dragon]

i have never finetuned before

this is my 1st time

[meager canopy]

Are you training on Selene Gomez?

[hardy cipher]

no, that's actually steve buscemi

[thorny frost]

So we can conclude SD1 is finally history and SDXL has replaced it?

[hardy cipher]

nah

I don't think so

for some people maybe. but I don't think 1.5 is going away

[meager canopy]

Selena is already in base SDXL

[hardy cipher]

there's also a selena gomez lora on civitai

but it's all good if someone wants to do their own thing

[upbeat summit]

let him make his own Selena 😄

[thorny frost]

Why not, SDxl images are better and with the finetuned checkpoints not censored.

Maybe smalller requirements