✨｜sdxl

[fierce hollow]

https://huggingface.co/TheBloke/MythoMax-L2-13B-GPTQ

this works well too:
https://huggingface.co/TheBloke/MythoMax-L2-Kimiko-v2-13B-GPTQ

[rustic garnet]

my poor harddrive ^^ this AI stuff is taking so much space

[fierce hollow]

keep in mind 13b models only work with 13b loras and so on

also you need like 12gb vram with full context

(that's before any sd models)

[ionic dragon]

i run it in 4bit

[rustic garnet]

yeah, it seems to wirk

[ionic dragon]

https://github.com/altoiddealer/history_cleaner

use this

so that it cleans the history and generates better prompts

[rustic garnet]

oh, I didn't used the chat Tab

the prompts from the lora are a bit random, though xD

sometimes it works really good, sometimes it feels as it would just add random style and artist names to the prompt

(contradicting stuff like "black and white, movie by james cameron, unreal engine"

[ionic dragon]

yeah, its not too good, its just a testing one

the prompts were from gustavo's dataset

so they arent really great

i am avoiding his dataset in the bigger training

[fierce hollow]

are there any public datasets?

I kinda want to train one for 7b

[rustic garnet]

maybe use less training data? I heard you can train a Lora with only a few hundred instruction-response pairs

[ionic dragon]

@rustic garnet where can we find some good prompts with context included with tags?

[ionic dragon]

yea, just searchn on hf

[rustic garnet]

hm, you could crawl the "pantheon" channel here in the sd discord. Not perfect, but might be not that bad either

[fierce hollow]

ah okay found the gustavosta one

[ionic dragon]

the one is sent you only had 80k, and 40% are repeated, but the instructions were kinda different like generate a sd prompt of a dog & give me a stable diffusion prompt which can generate a dog

[ionic dragon]

dont use that, its bad

[fierce hollow]

well, I don't see any others, also one option is just using danbooru dumps

it's how nai was trained after all

[ionic dragon]

https://huggingface.co/datasets/trevordark/sd-prompts

check this out

[fierce hollow]

also yeah, 2 million prompts is probably an overkill, just going to fry the llm

in the above kek

[ionic dragon]

filter them

[rustic garnet]

yeah, probably you subsample 200 prompts

or 500

[ionic dragon]

by length, and remove duplicates

[fierce hollow]

I feel like these should have instructions, otherwise it's just a meaningless string of words

[rustic garnet]

and then manually filter them

you can add instructions just via script

[ionic dragon]

use any llm to give them an input

[fierce hollow]

I mean, as a dataset, it should contain something like "Generate XYZ, make it look ABC." as instruction and then the prompt as response.

Sure you can generate the instructions afterwards

but that's less than ideal imo

[uncut fiber]

crumb/bloom-560m-RLHF-SD2-prompter-aesthetic
do you think it is realy 560 milions?

[fierce hollow]

isn't that a 560m parameter model?

[uncut fiber]

maybe i am noob 😦

will try something different than Gavota

[fierce hollow]

yeah seems to be a model not dataset, in any case parameter number doesn't correspond to dataset size

[uncut fiber]

max size of model is i know 1.4 GB it can say something probably.

[analog roost]

Holy shit, dynavisionxl is incredible. The image has NOTHING to do with the prompt ({A determined runner, muscles tensed, pushing hard to climb a steep hill, sweat dripping down their forehead, juxtaposed against a speaker backstage, taking a deep breath, their eyes closed momentarily as they mentally prepare to face their audience.}, digital photograph, realism, Dribble, side view, short telephoto, 70mm, studio lighting, high resolution, (challenge:1.25), (overcoming:1.2), (determination:1.15)) but the images themselves are insane

And this one is scary af

[sturdy sinew]

Hey sorry for my ignorance but what service offers that node interface?

[lusty wolf]

Boo!

[sturdy sinew]

Hey guys. So I just went back to image generation and reinstalled automatic1111 on my googlecollab (unless there are better options now that I am not aware of). I would like to make that kind of image with hidden words in a visual. I understand (but i may be wrong) that it's achieved with those QR code controlnets. I'm only seeing them for version 1.5. Is it indeed the way to go, and if so, does it exist for SDXL? Thanks in advance.

[glad grove]

no qr controlnet for sdxl

[shy kelp]

i was messing with one earlier that was available for sdxl but man, results were meh

[glad grove]

soon (tm)

[shy kelp]

30 minutes for that i guess idk

[shy kelp]

i like dynavision

[analog roost]

Sure, I just wish it followed the prompt...

[uncut gull]

You can check the status of your clegs at the top of the suatllee menu. 8K game ui.

OpenAI's Dall-E3 is coming out soon, and it can do text much better than SDXL, at least for standard fonts like Arial and TNR. Will it be able to write swear words? I bet not.

[fierce hollow]

it's comfyui

[zinc cargo]

anyone knows how to load pti files in comfy?

[heady vale]

it wont generate violent, adult or hateful content. Also wont generate public figures

[uncut fiber]

yes only for sd, which shouldnt be problem, just different models. It is controlnet model called QR monster. @sturdy sinew

[native knot]

You may have a value in a box for a file that doesn't exist or may have a different name on your system.

[fierce hollow]

does it matter that dall-e3 is coming out if you can't run it locally though

[indigo carbon]

dalle3 would be useless once oobabooga get's ComfyUI backend integration

[uncut gull]

They can say that, but they said that about GPT-4 too, right?

[indigo carbon]

also: no freeU

freeU

[fierce hollow]

I mean maybe if llms could generate and tweak workflows on their own, not sure just having them generate a prompt can produce anything similar to dall-e

freeU is pretty nice though, I think...

[heady vale]

we will find out soon I guess

[indigo carbon]

nah, I think all it needs is access to the unfinished latent and it could handle the writing

[fierce hollow]

sounds complicated

[uncut gull]

The LLM needs to understand vision, not just read a CLIP description and then right a prompt. It just doesn't come out right.

[indigo carbon]

I bet they pulled off something like that on dalle3, the quality of the images aren't as good as SDXL, but they have better text- so that's my theory

[indigo carbon]

LLaVa

[uncut gull]

My theory as of 10 minutes ago is they have something about as good as SDXL for text, then they're detecting text boxes after generation, and using AI to paste in text written in Arial or Times New Roman. Because in the background of some of their cherry-picked images, they have garbled text exactly like SDXL.

[fierce hollow]

llava/r is just clip-whatever-it-was-called + ocr, it's not really that great imo

it's not like it can actually see the image

oh right, clip interrogator

[indigo carbon]

yeah, dalle3 on it's own without LLM is just like closed source SDXL imo

[uncut gull]

It could have more parameters and be trained on way more images though. Because they have more money than Stability for fighting lawsuits, and they're not targeting consumer hardware.

[indigo carbon]

their results say otherwise. the quality of the images they showcased wasn't really as good as SDXL imo, they just had better text

it's certainly not pixel diffusion anymore

[uncut gull]

But those were prompted by OpenAI's engineers, right? I bet I could make it do better things. (But I'm just going to test it for prompt comprehension TBH. I have zero use-cases for a watermarked, censored, corporate-owned workflow.)

[indigo carbon]

idk, I tested some of the prompts they did on my SDXL workflow, SDXL did better on all of the images that didn't have text

[uncut gull]

I find that hard to believe. You know what that means??? I'll just have to test it. 🙂

[fierce hollow]

can you recreate this with sdxl though?
https://images.openai.com/blob/54facbbb-c94c-4884-8c94-5b984b19749c/dalle-image-map.png

with the same level of coherency

[indigo carbon]

probably, hold up

[uncut gull]

Right off the bat, the dunking nebula explosion is following the prompt much closer for Dall-E3 than for SDXL. Sure, I could use prompt engineering to try to get a nebula explosion to turn into a basketball player, but that's not the same as saying the model understands what I'm asking.

[indigo carbon]

that's actually one of the prompts I tested and SDXL did better lol

[uncut gull]

Pics or it didn't happen! SDXL cannot render the human body's shape as an explosion. It would mess up the edges.

You can immediately tell which of the two models actually understood what this prompt was saying, and which model was just picking a few random words as inspiration.
A middle-aged woman of Asian descent, her dark hair streaked with silver, appears fractured and splintered, intricately embedded within a sea of broken porcelain. The porcelain glistens with splatter paint patterns in a harmonious blend of glossy and matte blues, greens, oranges, and reds, capturing her dance in a surreal juxtaposition of movement and stillness. Her skin tone, a light hue like the porcelain, adds an almost mystical quality to her form.

(And, to be clear, I am 100% on open source's side, and I will continue using SDXL. But I'm also super obsessed with truth for personal reasons, so I probably come across as a jerk about facts.)

[uncut gull]

It's a dragon fruit. (The fruit is a banana, to be precise.)

[indigo carbon]

@visual glade I just updated to the latest ComfyUI and speed is still a fraction of what it was on commits from a month ago.. any idea?

[uncut fiber]

Weird thing as those porcelain thing, there isnt single piece have common fractured color, or is it there?

i have difficulty express myself in native language so i cant in english

[indigo carbon]

with commits from a while ago, I'm getting 3.2it/s on batch 4 with AIT, and with the latest commit I'm getting 4s/it also with AIT

[uncut gull]

I'm getting 1.1s/it with A1111, so I don't know.

[uncut fiber]

@uncut gull what is your gpu? 3060 or similar?

[uncut gull]

It just can't do it. SDXL cannot mess up the edges of a human's form.

[uncut gull]

LOL no, I'm maxed out. I have the 4090. Bought a new PC a few months ago.

[uncut fiber]

ok 😄

[indigo carbon]

DRAGON fruit

[uncut gull]

That is beautiful! What did you do?

[indigo carbon]

first attempt on my SDXL workflow. again, I'm thinking the reason why dalle3 is performing better than SDXL for you is likely your inference

[uncut gull]

No, like, what prompt did you use?

[indigo carbon]

"dragon made out of fruit"

that's it really

[indigo carbon]

that's a skibidi-looking fruit

[uncut gull]

I tried "dragon made of fruit" and it failed??? Has SDXL been able to recast materials this whole time but I was missing the word "out"? (Testing now!)

[indigo carbon]

no, again, it would also do that without the word "out". it's your inference, not the model.

[uncut gull]

No, it's definitely working 100% with "made out of" and 0% with "made of". It's the prompt.

[hardy cipher]

I must have set these to render then forgot about them, because I did not seem them until now

[indigo carbon]

"dragon made OF fruit"

[hardy cipher]

that thing might be your friend or devour you in your sleep. never really know

[indigo carbon]

it did just as good, idk man

[uncut gull]

OK, hang on. I must be remembering something wrong then. Let me check my prompt history and see what I actually tried.

Human with messed up edges!!!! It's real! 😄

[hardy cipher]

dalle-3 might have gpt-4, but it doesn't have my workflow or fruit skills

[indigo carbon]

yeah, GPT 4 is overkill for controlling diffusion, I bet we could do better by using Oobabooga x ComfyUI

you would just give it access to ComfyUI with a workflow loaded, and it will do the rest

[hardy cipher]

lol, you can already use them in conjunction with each other to a degree can't you?

[uncut gull]

Oh LOL the mistake was human error in my case. I didn't zoom in on the picture, and I thought it had failed. These are actually little dragon heads. 😐

[indigo carbon]

yeah, but it will be much better if Oobabooga was talking directly to ComfyUI

[uncut fiber]

so left is SDXL, i thought right one

[indigo carbon]

the bottom right one reminds me of harry potter 1 Voldemort when loosing the stone

[uncut gull]

Come on though, break the face into pieces somehow.

Still not broken.

Okay, I give up. It's an interesting challenge, but I want to try to get more interesting game UI designs.

[uncut fiber]

i like challenges

[uncut fiber]

@uncut gull i think you can try some nsfw mode. What i saw makes me sick, and it is only dynovision

[drowsy dust]

What settings are you using for FreeU?

[uncut gull]

?
NSFW for explosions?

[uncut fiber]

nsfw isnt just nudity no? Something hardly bare mature or adult, not sure what is right. Explosion of head can be pretty ugly... I think

[fierce hollow]

nsfl

[indigo carbon]

I'm currently liking JuggernautV2 for everything, some of the SDXL finetunes are mind boggling

[uncut gull]

True. I was turning the head into another material first (mercury or tile) and then trying to fracture that into exploding pieces. TBH my game UIs are much more beautiful than exploding heads, and I will stick with those for now.

[uncut fiber]

@uncut gull 👍

[uncut fiber]

this seems like proper explosion but in part we dont see it

[noble shoal]

@uncut fiber

[uncut fiber]

nice, but nobody know what exploded 😄

i see, its head!

[noble shoal]

Exactly

[uncut fiber]

but i think is impossible for example made headless robot, or statue. Probably with word torso? I know i tried robot, and cant.

torso is working

rarely

[stone fossil]

Blueprintify SD XL 1.0; https://civitai.com/user/Thaevilone/models

Blueprintify - Transforming Images into Architectural Blueprints

Introducing Blueprintify, an innovative LoRa (Text to Image) model designed to morph ordinary images into intricate architectural blueprints. With Blueprintify, you can effortlessly breathe new life into your visuals, creating captivating representations that showcase the beauty of structural design and precision.

Use trigger word: bl3uprint.

Trained on 3500 steps from a highly detailed by hand captioned large dataset.

• Special thanks the people who help me and who I care a lot about:
  @dapper dragon
  @upbeat summit l
  @osiworx
  @uncut fiber
  @abstract depot
  @lusty wolf

Some images do not seem synced well yet lol, give it a spin let me know what ya think. 🙂

https://civitai.com/models/150986/blueprintify-sd-xl-10

[uncut fiber]

scary

[stone fossil]

If it not while he nicely explains it step by step.

[noble shoal]

Ah, the north Korean space program.

[stone fossil]

Sorry dint want to bring politics into this channel it just felt like a funny thing.

🙂

[uncut fiber]

i would call it rocket program

[stone fossil]

But if u do it good lol.

Sometimes you dont even need the front side of the person lol.

And it is just funny like heck. 🙂

[noble shoal]

Oh, one of my ancestors 😅

[stone fossil]

😛

[noble shoal]

Wait, i am not Austrian. Nevermind

[stone fossil]

I like this one myself a lot as it is almost as hes trying to play c&c.

[uncut fiber]

its well made nice picture of ugly person

[stone fossil]

xD

So next subject.

[uncut fiber]

there is radioactive symbol?

[stone fossil]

Draws in the high hat.

[stone fossil]

There better be. 😄

[uncut fiber]

when i was there on civitai, i cant download your previous loras, i hope issue is on my side.

[stone fossil]

Should be civit is a bit of a cake tho.

[uncut fiber]

how would you describe this image?

[stone fossil]

mona lisa space dish silo bokeh backdrop.

[uncut fiber]

imagine how powerful lighthouse 😄
@stone fossil 1:0

[stone fossil]

Me using SD.

[uncut fiber]

i dont know her

i cant nor see gifs... When posted i see them, but cant post them, because not see them

[uncut gull]

If aliens who used magic instead of science took hand-written notes about nature, it would look like this.

[uncut fiber]

as you speaks about UI i know two games that changed UI and both game devs it took over 1 year. No small indie companies, quite big players

[uncut gull]

Not sure I know what you mean, but UI design is definitely challenging.

[uncut fiber]

https://tenor.com/view/the-voynich-manuscript-logo-the-voynich-manuscript-mystery-book-trending-rare-book-gif-10416031475225917112

maybe this is what they left here! 🙂

simply developers decided to change GUI of game. And it took them over 1 year of working

do you know this Manuscript? @stone fossil @uncut gull ?

[stone fossil]

It seems like a meat flip phone to me.

Samsung might like it. 🙂

[uncut gull]

I can't read the title. The watermark is covering it.

[uncut fiber]

it is old manuscript, nobody cant translate it nobody knows it origin. There is one small clue to italy. But who knows. It has as well unknown flowers, i think scans HQ can be found on net

Voynich manuscript
covered is probably THE

[uncut gull]

Tired of just looking at pictures of meat on your out-of-date phone? Introducing!

[uncut gull]

SDXL is great at darkmode UIs, but apparently it thinks light-mode UIs are just for pretty princess games. (Probably just a problem with my prompts though.)

[hardy cipher]

have to give it better directions maybe

[uncut gull]

The challenge is finding a way to keep all the colors and high-frequency details in there. That's what most of the prompt is doing.

[hardy cipher]

you mean just keeping it bright and all that

just tell it to make the opposite of this

[steady grove]

those squiggles are warped

[uncut fiber]

scream! wasnt here football derby

[tender timber]

Hey... was this note about "our" Sarge? 😦

#1020355799945580585 message

[steady grove]

Thats the sarge i've seen here. Very sad to hear. 😦

heavy

oh no. sargezt was on this server but never posted here

still, feels bad

[hardy cipher]

how are the details so on point?

I guess the zippers look a bit goofed

[lusty wolf]

Is this for ComfyUI. How is it implemented?

[hardy cipher]

these strawberries playing games

[ionic gulch]

Workflow V15.0 - optional image caption, new menu layout
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[steady grove]

why do all these custom workflows use gan upscalers in the mix? they always make gan scaling artifacts

sdganl

[hardy cipher]

you have a memory leak somewhere

hard to say what it is

something isn't clearing out as it should

[steady grove]

make sure to use nvidia studio drivers. the game ready have been crash with the new dlss3.5 introductions

[hardy cipher]

but I don't know about the vlad fork so I can't tell you what's going on really

JPS, how well versed are you on the different prompting tools and models? not like models devoted to only prompts, but at least used for them. I'm wondering if there are any options for enhancing a prompt rather than changing it's style

or maybe one of you fine folk know the answer for this

[steady grove]

studio drivers are the safe bet. i dont know where all the community tips came from about downgrading to older game ready drivers.

[hardy cipher]

lol

sdxl is stable diffusion

but user interface

[steady grove]

the generation speeds between xl and comfy on my system are negligable. i use auto more lately because comfy i always have to fudge around with the ui connections to do anything

[ionic gulch]

what do you mean with "without changing its style"? sai-enhance is something in that direction. it adds keywords like "masterpiece", without a specific artist or movie

[hardy cipher]

is a matter of preference, I personally like comfy, but it's kind of hardcore for some. a1111 just seems like one of those too many chefs not enough cooks situations

[steady grove]

the work between prompting is so much longer in comfy

stable swarm is a good ui but i'm still confused by it and don't feel as streamlined in my creative process as i do when using a1

[hardy cipher]

well I mean, I'm using some stylers in my workflow, but some of them drastically change the output, which is fine, but not always into doing that. and it doesn't have many options to just make it look cleaner, or mildly something. I don't know. I'd like to keep the essence of the base and pull other things out with it

stable swarm is pushing forward

implementing new things

not sure about the others really

I've only used a1111, comfy, and then stable swarm some

a1111 was what I began on. it was great at the time because not really anything else, but they sort of got stagnant. seems they're catching up now though

[ionic gulch]

my styler has artist, movie and general style - the general style has the sai prompts (the ones they use on their own website). if you pick a matching general style it shouldn't change too much.

[hardy cipher]

just seems like if several people were building a house and none of them communicated with one another

[steady grove]

stableswarm needs to get a ux designer consulting. this is what it looks like when i load in. .... why is the negative prompt and actual prompt on opposite sides of the ui? everything is just mishmashed in there and there's no reasoning to it at all

[hardy cipher]

well the designer is in this channel so you could probably talk to him about it

[steady grove]

yeah. i dont like to engage developers directly. if i'm not being paid to be somene's boss i feel really really weird talking to them about their project.

what i can say for sure is the UI is super confusing

[hardy cipher]

at this point in the evolution of these things I don't expect perfection from any of the options, just open minded developers willing to listen and push forward

well I'm not saying you need to say anything

I understand where you're coming from. and I do find it outlandish when people have started talking trash to someone that's created something they cna use for free. that's silly to me. but giving feedback seems fine to me

[ionic gulch]

@hardy cipher you could give your own part of the prompt or input image a higher weight to reduce influence of a prompt styler.

[hardy cipher]

that's a great idea and not sure why I didn't think of it. thanks

this ipamonster machine makes such clean things

[idle kelp]

how do i install comfyui manager to an existing comfyui

i want to download extensions but not sure how

[hardy cipher]

open command prompt or powershell

navigate to your custom nodes folder

[idle kelp]

mhm

[hardy cipher]

https://github.com/ltdrdata/ComfyUI-Manager

[ionic gulch]

same prompt and ip adapter image - styler examples (with the new caption option, so i can easier compare results)

[hardy cipher]

and then you need git

git pull into the custom node folder and it'll install itself

discusses it in that link

[idle kelp]

okay

thanks

im horrible at this stuff

[hardy cipher]

well you know, I actually don't even want my own prompt to hold much weight

[spring fulcrum]

The funny thing is by the time everything gets to a great spot in ux/ui developement we are all going to have to start from scratch with SDXL 2.0 etc... LMAO

[hardy cipher]

I'm sort of combining several things. the blip output from 3 images, then some words I'll add to and remove from here and there, and I'll concatenate those. then I run them through a styler.

[idle kelp]

how do i do the git clone with powershell into the custom_nodes folder

[tender timber]

Yeah, a big loss for us, his control net models were really good

[twilit nexus]

is it possible to make animated door or similar stuffusing stable diffusion?

[tender timber]

UPDATED Endless Nodes 🌌 are now pushed to GitHub (added aesthetic scoring and others)

I've updated the small set of nodes I am working on:

Took the node from https://github.com/strimmlarn that does aesthetic scoring and re-purposed it
Added an eight input number switch because I needed it
Added the Endless Nodes Parameterizer with Text_G and Text_L prompt box
Added the Parameterizer with a_score for both pos/neg
Added the Parameterizer with a_score for both pos/neg and Text_G and Text_L prompt box
Added the Endless Nodes Parameterizer

You can grab them from here:

https://github.com/tusharbhutt/Endless-Nodes

[hardy cipher]

copy paste what's on that page

I'm curious about the aesthetic scoring tools and what criteria they follow. because in my experience I've consistently disagreed with their ratints

we might just be going for different things

[steady grove]

i'm not sure that placing the negative and positive prompts together in the same UI needs to be "UX progress made". It's just programmer ui and no UX effort at all. That's why they need to get a ux designer, someone who thinks about stuff like creative flow mindsets and making the ux facilitate that flow, sooner than later

https://en.wikipedia.org/wiki/Flow_(psychology) bad ux is the enemy of this

[ionic gulch]

did a test with different ip-adapter weight: 0.7, 1.0 , 1.5, 2.0, 2.5, 0.5, 0.3, 0.15 (same seed and ip-adapter source image)

going above 0.5 doesn't really care much about the prompt anymore

[idle kelp]

i had to install git, ive cloned it in now

and it works

thanks

[hardy cipher]

I've been messing with model block weights. well groups of blocks

[steady grove]

as these systems come into play too, i notice ux problems. People are used to image to image denoise setting, higher meaning less of the original image. now these new systems have the setting higher means more like the oriignal image. Technically we can understand why these decisions were made. Realistically, it is bad UX and is contradicting paradigms

[hardy cipher]

well they can't all be winners. yowzers

[steady grove]

that's one angry nun

[hardy cipher]

overcooked all of them, but bottom right looks a bit unhealthy

[idle kelp]

with the comfyui manager do i still have to manually install loras?

[uncut fiber]

she just stay on head quite longer

[hardy cipher]

you put them in a folder. wouldn't really call it installing

but yes, you'd still have to put them in the folder yourself

[idle kelp]

okay

im not sure if ive even done that right as its hard to tell if the prompts are working half the time

[uncut fiber]

are you using default workflow?

[hardy cipher]

well if you put them in the right folder they show up in your lora node as an option

if they're in that menu they should work unless something is broken somewhere

[uncut fiber]

also lora must be for same model, 1,5 2,0+ SDXL as you are using

[hardy cipher]

well if it's not you'll figure it out pretty quickly

it's happened to me, civitai trickry

[idle kelp]

yeah i think i did mess up somewhere

[uncut fiber]

what workflow are you using? @idle kelp

[hardy cipher]

told it only sdxl stuff and downloaded a cool looking lora without thinking. turns out that they inserted 1.5 models as things I might like

so that made things goofy

[idle kelp]

im not sure i even know what that even means

i barely managed to get sdxl to work as is

[uncut fiber]

i am used lora maybe twice. SW uniforms to test and some another one

[hardy cipher]

"oh, you're looking exclusively for one type of model version? let's give you something incompatible and keep it a secret"

lol

[idle kelp]

i think i figured the loras out though

i just have to create nodes

instead of the text based loras i see a lot

[uncut fiber]

yes nodes is basic no nodes nothing in comfui 😄

[hardy cipher]

I really can't fathom why would compel the designer of civitai to have that happen

[steady grove]

thats more because uplaoders mislabel their stuff because 1.5 is more popular and they want to be popular

[hardy cipher]

it was labeled 1.5

and came out in lik may

[steady grove]

contests like giving away 4090s for popularity can have that community effect

[hardy cipher]

it's on me for not double checking

but really, I just didn't think I needed to since I restricted my search options

after that I checked the settings and turned off all the new junk that I could

did you see the one guy everyone was saying put out fake loras got unbanned?

[idle kelp]

where do i put the load lora node

as in what do i connect to it and connect it to

[hardy cipher]

do you see the inputs and outputs and how they have names and are colored?

[idle kelp]

yes

model and clip

[hardy cipher]

as a general rule you can use those as a guid

e

so your checkpoint output has model and clip

and lora loader has those as input

so connect those

[idle kelp]

okay

[hardy cipher]

now just remember that any output name/color can connect to any input of the same name/color. it's not always the best idea in some situations. but they are going to be sending/receiving the same type of data

[uncut fiber]

right to left

[hardy cipher]

right to left?

[uncut fiber]

right points to left points output to input

you mentioned it already

[ionic gulch]

same seed

[hardy cipher]

here's a basic flow. you can see how everything should connect. ftom there it's just adding more pieces

you'll notice all the lines are straight. that's because it's objectively the cleanest looking approach

[idle kelp]

that looks much cleaner than mine

i need to redo my lines

theyre all curved

it works fine though

[hardy cipher]

lol. well don't worry about that now. I'm just joking

but yeah, that's good

the one in the top left with the black blotch that replaced her face is moderately terrifying to me

[idle kelp]

that is a little unsettling

what does the cfg change?

ive just been messing with mine really

[hardy cipher]

cfg is essentially how much you want the image to conform to your prompt

[idle kelp]

ohh

[hardy cipher]

and there are drawbacks to both too high and too low

too low won't necessarily output a bad image, but it'll be pretty random and not very representative of your input

[hardy cipher]

but too high it becomes overly cooked and looks terrible

[hardy cipher]

super meat boy

[idle kelp]

yeah

[indigo carbon]

what?

[idle kelp]

it sort of got rid of the colour

20 cfg

i was using 7.5 - 12 before

[hardy cipher]

now some would say these are cooked. and they'd be right. but I like the aesthetic in some situations

[idle kelp]

the bottom right is actually pretty good

[hardy cipher]

messing with model merging and block weights along with some other tricky things

so tricky tricky

[idle kelp]

thats so good

[indigo carbon]

I have so many of those haha

[idle kelp]

oh my god

i dont even know how to

form how perfect those are

[indigo carbon]

I can crank batch images like that in under 30 seconds (per batch)

[uncut fiber]

20 is good for cartoon

[hardy cipher]

it started making backpacks yesterday

look at the details on that though

[uncut fiber]

sorry for this, but nice solved i "I" all should be mouse 🙂

[hardy cipher]

lol

these backpack images just look like photographs

[uncut fiber]

nice

[hardy cipher]

I was genuinely impressed by them when I realized what they were. I did not tell it to make backpacks

I think this is the same seed number as above. or they're sisters

[uncut fiber]

@idle kelp it is in settings if you want spagheti or straight lines. There are 3 posibilities, you cant redo it manualy. It is i think under small cogwheel in manager window.

3rd eye

[hardy cipher]

I need to amp some of these up and upscale

[idle kelp]

oh i see

thoughts on this?

[uncut fiber]

good, red eyes he must be angry 🙂

[idle kelp]

yeah

[hardy cipher]

same input, different model block weights

[uncut fiber]

physical meltdown 🙂

i spend too much time today in A1111 with CN

S/HE has heart on bad place 😄

[uncut fiber]

Cuba?

[cyan crown]

yust a car

[uncut fiber]

Those cars are on Cuba very much. Nice car!

@idle kelp do you know, somebody told you if you drag drop image containing workflow (nodes in some order) it is default to keep this workflow in picture.

[jagged scaffold]

Just to be sure: Has the Ultimate SD Upscaler been examined for any odd code?

[fierce hollow]

yes, I scrolled through its single python script available on github for everybody to look at just now, and can tell you it doesn't connect to the internet, send your prompts to the fbi or have a built-in keylogger, hope that helps

[uncut fiber]

he cant bite 😄

[uncut gull]

I just saw some epic 3D VR180 videos on DeoVR. I can't find anything VR related on Civitai. Does anyone know how these were made??? They do say Stable Diffusion in the title, and they obviously are. But I really want to make one! https://deovr.com/videos/180-vr-ai-overclocked-stable-diffusion-deforum-part-33-12461

[supple knot]

In a T2i node system, my clip is "clip g" in both clip loaders, is this correct ?or do I need one call "clip vision I"

[uncut fiber]

@uncut gull The official port of Deforum, an extensive script for 2D and 3D animations, supporting keyframable sequences, dynamic math parameters (even inside the prompts), dynamic masking, depth estimation and warping.

Supposing 3d animation based on depth estimation is it

[supple knot]

Is DeForum coming to SDXL ComfyUI?

[uncut gull]

I know how to use Deforum. 3D VR180 needs side-by-side images showing the same scene from two different angles, and those images need to be panoramic projections on a spherical lens. You can't just use depth maps to "slide" the image or something, because you get blank spots. (I know, I've tried.)
His VR180 3D is perfect though. No missing spots, no artifacts, no stereo glitches. There was even a brief moment when part of a landscape morphed into an astronaut, and the face inside the helmet had perfect stereo rendering.

[uncut fiber]

yes and this can be done with depth estimation bad not best yes

i have VR quite long time but wasnt in it as well long time

[supple knot]

A flatten 3d 360 video looks like this flatten , like a flatten globe map,so if you can get it to produce images like this then convert to 360 video

[uncut gull]

No it can't. You get blind spots. Imagine you're staring at a wall with a window. Move to the side, and the view out the window shifts. You can't do that in post. You can't change your view out the window.

[uncut gull]

360 panoramas have nothing to do with 3d VR180. You need two different camera angles on the same scene, with zero glitches in the stereo data.

[uncut fiber]

Depth has great quality also notice they used some special model. Also bet it is cherry picked. Yes and those two cameras are done throught estimated depth

save animation's depth maps as extra files
Save 3D depth maps

[uncut gull]

Okay, I'm gonna try it. I have Deforum zoom and I learned how to set up a stereo camera in Blender last week, and I can get depth maps and put them on a plane... Or something. We'll see how it goes!

But this example https://civitai.com/models/125100/stereodiffusion is awesome, it just has too many stereo glitches to be usable, and there's no sphere projection, so it doesn't work for VR.

[uncut fiber]

i realy dont want arguing. it is pointles. 1st my english, 2nd no reason
i have for vr great player, paid but cheap. Have almost unlimited posibilities. All sorts of projection and much more. I forgot its name 😄

here they upload video, and then get depth from it. Why? because stereo image only comes to my mind

[steady grove]

it doesn't look like those have stereo renders on them. i think they're just flat images projected into a vr space like 360 videos on youtube are

[uncut fiber]

Whirligig media player

[urban fjord]

dynamic math parameters
So deforum is still a disaster when it comes to userinterface. I like the idea but the execution is really botched.

[uncut fiber]

it is description, but yes it is quite messy UI

depth model has 1,28GB

[uncut gull]

Now that is one beautiful 3D! Not VR180 though. But we'll see if I can get that working.

[uncut fiber]

it is quite good will try short video from it

[steady grove]

deforum is a toolset targeted at technical directors and animators. people who plan stuff out meticulously and love doing it that way. i wouldn't say they botched their goals at all. it works so well in the realms they made it for

maybe what you want isn't something they've been aiming for

[analog roost]

Nothing to do with the prompt but damn, this looks like Cyberpunk 2077!

[urban fjord]

If you want to use it you kind of need a proper tool to create camera paths and export that into deforum. But if you have that then it might be useful.

I'm testing out using my controlnet for more consistent text in SDXL and while it's not perfect yet, it is showing promise.
Left one is without controlnet and right one with controlnet.

photo of a cake with the words "behind you"

[uncut fiber]

@uncut gull dont forget it is video, not real time head motion capture. As i understand it.

[steady grove]

you made a controlnet model? 😮

[uncut gull]

Yes, video.

It is 3D VR180, but the render takes up only so much of the field of view.

[urban fjord]

Controlnet-lllite, quite easy to train and quite small. 44 MB files if I'm not mistaken.

[uncut fiber]

my pruned models for 1,5 are 800MB 🙂

not my as made by me, just i have on pc. <- clarification

Sorry i forgot it cycles short videos. so about 8 second, also glitches, but i believe when some texture not so bad and probably can be somehow corrected.

[uncut gull]

Maybe you do the VR180 projection mapping in SDXL? It can definitely handle something like that.

[uncut fiber]

@uncut gull if whirliwig has demo, i suggesting to you try it. i think it has its homepage on xyz.

[uncut gull]

Looked up whirliwig, found nothing. But I'm not using a proprietary 3rd party tool I can't control, if that's what you mean.

[uncut fiber]

it is on steam so should be clean, it is VR player.

[uncut gull]

VR player? Hmm. I have a Quest 2 headset, and I have some player that I use on it. Day something? Daybox? Skybox?

[uncut fiber]

https://store.steampowered.com/app/451650/Whirligig_VR_Media_Player/

Steam

Whirligig VR Media Player

Whirligig is a VR media player for the playback and viewing of video files and images. It support a wide range of projection types, render paths and has many other features to make your experience as enjoyable as possible.Tutorials highlighting some of the features and how to use them here:http://steamcommunity.com/app/451650/discussions/0/15464...

Price

$3.99

Recommendations

453

▶ Play video

[upbeat summit]

Skybox is pretty great

[uncut fiber]

skybox, imho 0,1% of features whirligig has 🙂 Or used to have

[uncut gull]

I've never run into a problem though... So I don't know what else I could want.

[upbeat summit]

yeah it doesn't have the same features, but usability is also a bit better 😉

[uncut fiber]

i think there is free demo, used to be. So nothing to loose

[uncut gull]

ZoeDepth is still running on a single image, maxing out my GPU, and it's been 6 minutes???

[uncut fiber]

you may want different projection, different cubes, different anything you never dream about.

something wrong

it downloaded model probably 1,28 GB

[uncut gull]

But I definitely don't even know where I would find videos like that. I already know my Blender camera setup works with whatever projection I'm using now. (It's a spherical lens and a panoramic projection. Not a cube, whatever difference that makes.)

ZoeDepth will work, it did this for the last image too. It is just taking forever. I will switch to Midas after this one finishes.

[uncut fiber]

o.k. i found some mars surface scanning in stereo by japanesse and similar

if last image, it is over i bet

they said it finnish before 100% frequently

[uncut gull]

I mean previous image. I did one image before, and it worked. Now I am doing another image, so I know it will work. But it takes a very long time.
AAAAA! I am going to stop it.

[uncut fiber]

weird

[uncut gull]

Restarted, and downloading Midas 3.1 now. Hopefully it will be faster.

[uncut fiber]

i had only superlow res but it did 12,5 fps and 151 images was made in few second
it downloading its own

[uncut gull]

This is 1024x1024 res? Maybe that is why?

[uncut fiber]

but it is say 8 images in my case.

o.k. understand

[uncut gull]

Finished!

I just have to test in my VR headset now somehow. Hmm.
But I can already tell from crossing my eyes that it is very glitchy.

[uncut fiber]

for example can you in skybox watch single pictures? SBS Top and bottom?

[uncut gull]

Maybe. I just rerendered with Res101, and it's amazing!!! I will try copy-pasting the image to Quest 2, and try opening in Skybox. It might need video though.

[uncut fiber]

@uncut gull if you are unbiased you can at least watch video from steam about it, it cover some aspects of program. And i am not related to it at all, just fanboy of it.

@cyan crown reminds godfather.

[uncut gull]

It worked! 😄 Sort of. The projection wasn't spherical enough, but it was still really good. Also the image was too low-resolution. I need a higher resolution image somehow? But... Hmm. Well, I'll try it.

[uncut fiber]

in sw i talked about you can change curvature and barrelenes of screen 🙂

[uncut gull]

Yes, but I don't think this was the software. The image isn't "round" enough when I compare it to other VR180 images. I'm changing my prompt to make it rounder.

[uncut fiber]

i think it can make image rounder, watch steam video, should be mentioned

[uncut gull]

That is too much roundness. But wow. SDXL has a really good imagination about how to make a lens even more round than it can be in physical reality.

[uncut gull]

Okay, trying this one in my headset now...

[uncut gull]

The VR was absolutely perfect!!! 😁

Just needs higher resolution / more pixels now.

[uncut fiber]

just check how many custom things on right

[uncut gull]

LOL OK! Okay. I will try it out. Might be more convenient than copy-pasting onto the headset. 🙂

[uncut fiber]

good night! 😄 LOL

[upbeat summit]

you're posting some really great stuff. nice work!

[cyan crown]

thanks

I'm testing various artists and illustrators

[uncut gull]

8K VR180 SBS 3D scene. About to test in headset. 🙂

Wow! That is amazing in VR. It's really beautiful.

[uncut gull]

Okay. Deforum video, here I come. Not sure I can scale that up though...

[slender coral]

How's it looking?

[uncut gull]

It's phenomenal. Other than the smoke-like texture wisps you expect from SDXL, it's like standing inside an actual landscape photo. I'm working on making a Deform video now.

[uncut gull]

Urgh. I have no idea how to make video. It's all flickers and weird transitions.

[zinc cargo]

need help?

[uncut gull]

I feel like I'm hunting blindly through parameters here in Deforum. I also downloaded TemporalKit, but it doesn't seem related.

[zinc cargo]

it's not related, it's for when you wish to do vid2vid mostly

[uncut gull]

I understand more or less what these numbers mean, I just don't know how to get that smooth zoom with very little change from frame to frame.

[zinc cargo]

reduce the zoom to minimum and increase cadence

while making sure nither of them is too high or too low

dont forget high steps for sdxl

[uncut gull]

I have 50 steps for SDXL, and cadence is at 8. I just increased zoom speed, so maybe I should've decreased it instead.

[zinc cargo]

there's a whole discord server for deforum, if you're into diving into it.

[uncut gull]

I can believe it.

[zinc cargo]

deforum is what really hooked me on GenAI

[uncut gull]

My image keeps exploding instead of zooming.

[zinc cargo]

increase strength

[uncut gull]

I have strength at 0.85. Pretty sure that's why it's exploding, actually. But lowering it was leading to changes...

[zinc cargo]

you need to zoom slower, and then at some point reduce str so it can create something new

[uncut gull]

It turns into cloud nebulas at the end LOL.

[uncut gull]

Oh? Hmm. That makes sense. There is a noise schedule. But maybe the scenery should accommodate that instead. I need a scene that inherently contains zoomable information maybe?

[vast zenith]

Hey I’ve been kinda out of the loop since before xl came out, have the optimizations brought it down to be able to run on 8gb vram? Ty :)

[uncut gull]

Yes. There's a comfyui workflow for 4GB VRAM on Civitai.
https://civitai.com/models/147057?modelVersionId=163935

[vast zenith]

Oh great, tyvm

[zinc cargo]

no, it can keep creating visuals for the same prompt or entirely different ones, just need to give it "air to breath" by reducing str for a few frames (you either check out math for deforum to give it a function or try to give it a soundtrack first)

[uncut gull]

Well, the zoom view works too. But I am zooming too fast for sure. Still, it's looking better! It's not exploding anymore. I'll try giving it more to work with and slowing down.

[zinc cargo]

first, it's beautiful.
reduce noise to 0.02 and it will look nicer

i shuold try to do some for VR, my kids now are in their gorilla tag phase and that might get them into something else

is 12 too early to introduce kids to flying ai waifu?

[uncut gull]

😱 Depends on what you mean by waifu? But stick to beautiful landscapes and it should be fine?

[rustic garnet]

do you share the prompt? 🙈

[cyan crown]

positive :: impactful paint by Makoto Shinkai, woman at the train station, highly detailed, 8k, sharp, professional, clear, high contrast, vivid deep blacks, crystal clear

negative :: ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, blurry, bad anatomy, blurred, watermark, grainy, signature, cut off, draft, multiple hands, low contrast, low resolution, out of focus

[zinc cargo]

thanks for sharing @cyan crown !

[cyan crown]

sometimes I use a Lora called xl_more_art-full at 0..8 weight

only for artistic (not photo)

[rustic garnet]

thanks! I guessed it will be Makoto shinkai, I love using this author tag for anime images

I only know "Your name" from him, though

[hardy cipher]

That is a quality lora

[thin edge]

hello. I had a question

how are they creating those encrypted text in images?

or like encrypting one image in another?

[glad grove]

controlnet

[thin edge]

like this?

[glad grove]

controlnet qr

[thin edge]

cool cool

got ya

thanks

[urban fjord]

Oh, it's that simple. I do wonder how they made the training data...

I guess they don't actually have to put text bitmaps into the training data

[indigo carbon]

BLiP diffusion is giving me a hard time atm.. I'm not sure if it's possible to implement AIT to it

[noble shoal]

You convert an image into a threshold black and white image. Now you can train the threshold image against the original image. Do that a few thousand times and you have your qr controlnet. Easy. At least i think thats how it works

[uncut fiber]

@thin edge it is qrmonster, just for SD only now.

[indigo carbon]

I DIDN'T NOTICE

I just went into my room and saw that from afar and I'm dying rn

@mods?

this is not okay

[glad grove]

@halcyon tusk

[indigo carbon]

this is actually making me want to kill myself

I need to go to church for the rest of my life now

[uncut fiber]

i think ppl with good eyes see it bit worse. Hidden texts.
block is only way

[indigo carbon]

I actually just threw up

[noble shoal]

The gif's keep playing

[indigo carbon]

nah, a ban to this guy isn't enough

[uncut fiber]

oh no. those hardly call them people should have banned IP on discord

[indigo carbon]

maybe even lawsuit. this actually made me throw up

[glad grove]

its ok chibi seagal is here

[uncut fiber]

how you can lawsuit stupid around 10 years old kids? Whose IQ = their age

[halcyon tusk]

taken care of, thanks!

[indigo carbon]

jesus, thank god it's over

[upbeat summit]

thank you

[uncut fiber]

thank you!

[noble shoal]

Thank you

[halcyon tusk]

thank you guys! i was deep inside comfyUI hahaha

[indigo carbon]

I feel like I gotta go to church for the rest of my life now

[heady vale]

how does stuff like that bypass the bot?

[glad grove]

easiest way to avoid that is just removing embed permissions from new ppl

[upbeat summit]

let's have some nice images to forget about this

[noble shoal]

Yeah!!! Eyebleach

[halcyon tusk]

nothing is perfect, but will keep improving

[heady vale]

understandable. just that some of the worst possible images are gif embeds. happens too often

[uncut fiber]

probably bot cant check videos

[indigo carbon]

I need to put a gallon of this on my eyes now

[halcyon tusk]

it you go too hard on the filters, half of the images on this channel and other channels will get blurred, flagged, etc.

[indigo carbon]

my eyes are used to seeing thousands of lines of code, not the worst things humanity has to offer

[uncut fiber]

nice both generated? i mean text2img?

[urban fjord]

More examples from my Controlnet-lllite training to support better text generation. Still far from perfect but it is a lot more consistent on certain prompt than base SDLX.
Prompt:

movie poster with the words "cats & dogs"

And the input image, it doesn't need to be placed correctly for the Controlnet to use it.

[ionic gulch]

yes, all based on "detailed closeup portrait of a female in a beautiful japanese red and gold temple" with different style settings (styles are just additional text prompts). position of the girl is done with low weight control net depth. seed is different.

[uncut fiber]

nice consistent look

[snow eagle]

/

HhHhHhHhHhHh

Im,hh

[spring fulcrum]

Making some stickers for Halloween... I could use some input...

[upbeat summit]

those are great!

[vital ermine]

I especially like the one on the left

[uncut fiber]

culturist ?

[supple knot]

Here is a list of reaction emojis if you ideas for sticker Thumbs Up 👍
Thumbs Down 👎
Heart ❤️
Laughing Face 😂
Sad Face 😢
Angry Face 😠
Wow Face 😮
Clapping Hands 👏
Fire 🔥
OK Hand 👌
Peace Sign ✌️
100 Points 💯
Eyes 👀
Thinking Face 🤔
Heart Eyes 😍
Facepalm 🤦‍♂️
Shrug 🤷‍♀️
Hugging Face 🤗
Poop 💩
Thumbs Up and Down (Like/Dislike) 👍👎
Praying Hands 🙏
Party Popper 🎉
Face with Tears of Joy 😅
Rolling on the Floor Laughing 🤣

[hardy cipher]

Write prompts using emojis exclusively

[native knot]

Ok. So the prompt for the first one was 👍. But since I got an eyeball, I thought I'd try 👀. That gave me the other eye image, which I think makes more sense. But now I also think that it's mocking me in response to the 👍 .

[hardy cipher]

It most certainly is

[native knot]

This was for ❤️

[hardy cipher]

I'm fine with that one

[supple knot]

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

[vital ermine]

Zoom in on that, and just WOW

[shy kelp]

hand

[pure crystal]

Goonies - Fratelli's hideout

[crisp owl]

V2 out of my workflow.
https://civitai.com/models/142535?modelVersionId=169206

[hardy cipher]

man, if I could ever find a stabilized point with what I'm doing I might just put my workflow up, but I just keep pushing forward with other weird things, lol. might just start it over and try to be orderly this time around

[crisp owl]

It's hard to not just keep expanding it.
Have to call it done eventually lol

Especially when new features keep being released

[hardy cipher]

should probably streamline it a bit. I just keep trying new things, but then, some of them aren't really the kind of thing that most people would probably want to mess with. dynamic thresholding, model block merging. and not sure there are one size fits all optimal settings for those.

[crisp owl]

Yeah when I have stuff that's variable in that regard, I just put it near middle variable and note to play around with the weights as they wish

[hardy cipher]

could also just make it optional to use some of the stuff. or have a couple versions. maybe a low vram less indepth flow. but for me personally I can't stop myself from going deeper if there's the option to do so

the weights aren't really the thing that I think would get people, but dynamic thresholding is a tad tricky and not sure how I'd effectively explain it to anyone since I don't even know yet what all the settings actually do

on another note, found the longest node you can use. lists hundreds of weights that you can modify individually. pretty practical

[crisp owl]

Yeah I get that, some concepts are just out there for many.
But, then again, that isn't who would generally be looking for a workflow like you have

[hardy cipher]

true enough. well the flow without either of those two things worked great in my opinion. so I could certainly put that one up for people. not sure where though. I guess civitai. could put it on my github but not sure anyone would ever see it. need to update my listing in the manager too

ugh, or I could do none of these things

[hardy cipher]

made this from 3 random images my friend decided on earlier

[hardy cipher]

he went with macaroni salad, vintage godzilla, and ferrofluid. not seeing a lot of ferrofluid in the mix, but close enough

also these for some reason

[supple knot]

Where Do I find "BNK clip text encoder"?

[hardy cipher]

do you have the manager?

[prime bison]

salad monsters from the bot yesterday

Spaghetti monster looks good, I'll try that next

[hardy cipher]

nice. these ones are fun because not even describing the images anywhere. just put them in and see what it gives me

[supple knot]

you had a spaghetti monster long time ago also

[hardy cipher]

like I could put in 2 of those with that truck picture above, and something coherent will come out. sometimes it doesn't really even make sense to me, but always works, lol

[glad grove]

speakin of salads

[hardy cipher]

yes, but that took typing words. these are just automated. so certainly takes away some aspects of control. but I do still positive and negative prompt boxes if I want to manually add anything

[prime bison]

Spaghetti monster didn't work well. Didn't even give me fsm, oddly. Pasta monster worked better. Not that I tried either very hard...

[prime bison]

Try for scout badges. Might be able to get something better than this? This was the best arson badge I got. None of the stitched patches even came close 😦

[hardy cipher]

hmm

[prime bison]

I suppose I could anyways fix that one with sd15 img2img but cbf

I like the tomato

It definitely took the colors from the salad at least

[supple knot]

https://www.youtube.com/watch?v=fokOgbJAmu0&list=PLVEvxlmz1V-Y9-pOlf1z78Xp45pJL3-pC

[full idol]

why does sketch controlNET in comfyUI give weird results?

[hardy cipher]

well different aspects dominate depending on input and settings. I like to keep the same seed number and move things around to see how it changes the scene. this was definitely more cartoony and animated than I expected. next ones seem to be shaping up better though

[prime bison]

I did some killer tomato stuff with sd15 iirc. Dunno where it is though

[prime bison]

Also, such a corny movie

What's with the banana?

[supple knot]

better memory then a movie

[hardy cipher]

changed some weights. same seed numbers

[prime bison]

Tomato bun

[hardy cipher]

that guys face is cursed and terrible. oof

[glad grove]

is that miles morales burger 🧐

[hardy cipher]

maybe. I don't know miles first hand. but it keeps making me burgers for some reason. which is okay with me I guess

[glad grove]

speakin of food

[hardy cipher]

pasta and salad with a beautiful view of the sunset

[prime bison]

Better pasta monsters

[hardy cipher]

not bad

[prime bison]

Spaghetti monster didn't work. Pasta salad monster worked 2 times out of 6, I think

[hardy cipher]

lovely

really captured mr. seagal's expression

[glad grove]

tasty

[hardy cipher]

cfg might be just a tad bit too high there

[hardy cipher]

wow

[glad grove]

this a weird elephant

[prime bison]

Jack Nicholson

[hardy cipher]

that's his younger brother that decided to pursue a career in competitive eating

[simple vessel]

lora,korean girl,pretty face,on bed,materpiece,best quality,photography,

[hardy cipher]

materpiece

why'd you add lora to the prompt though?

those boys are cooked

[glad grove]

done

[hardy cipher]

your damn mammoth dominated the photograph

[prime bison]

Try putting these together? 🤔

Idk

Pizza monsters are harder

Was hoping for olives or pepperoni eyes, not googly eyes.

Neg googly eyes and I just get holes

[hardy cipher]

it has to be 3 pictures

[prime bison]

🤔

[hardy cipher]

I set it to kirigami

[prime bison]

I've got some fat Yoda?

[heady vale]

lol

[hardy cipher]

nice

that style really not getting the pizza

[glad grove]

pizza time

[prime bison]

I should do pumpkin spice pizza. It would go well with my other pumpkin spice stuff

[glad grove]

tasty

[hardy cipher]

alright, this style might work

[prime bison]

Pizza anyone?

[glad grove]

he loves pizza

[noble shoal]

That's fine as long as you don't put pineapple on it.

[prime bison]

I have so many pineapple pizza pictures

😏

[hardy cipher]

that's like me with strawberry pictures, but that was an accident

[glad grove]

he hates pineapple pizza

[noble shoal]

TOS Rule Nr. 4 doesn't allow pineapple pizza pictures

[glad grove]

now that u mentioned pineapple he destroyed it

[prime bison]

Mostly not pineapple pizza, but worse https://sd.efreakbnc.net/gallery/Discord-Bot-Pizza.html

[hardy cipher]

is that a stack of pancakes pizza in the bottom right?

[glad grove]

tasty

u want a bite?

[half ivy]

I am here!!

what did i miss?!

i saw that

o0

[glad grove]

steven seagal joined the server and livestreamed a full stack of 12 H100's giveaway

[prime bison]

Only 12?

[hardy cipher]

that's not even a terabyte of vram is it?

[half ivy]

wheres 34N3's demmit?!!

[glad grove]

hes too cheap 😔

[prime bison]

Man, I should be asleep

[uncut steeple]

Cant get the QR code monster to work in Comfy

[noble shoal]

1.5 or SDXL?

[uncut steeple]

Oh yeah its 1.5
gotta ask somewhere else

But yeah tried 1.5

[noble shoal]

I have tried the SDXL super early one, but the results are questionable.

[hardy cipher]

somehow turned the aesthetic score up to 999 earlier

that was something

[uncut fiber]

what is aesthetic score ? and is posible to use on already created images?

[hardy cipher]

I really don't know what exactly it does, but you don't want it above around 6 if you are going for photorealistic

[noble shoal]

Oh, coincidentally i had an eye pizza a moment ago

[uncut fiber]

i think it has something with beauty of image 🙂

[hardy cipher]

yes, but too much beauty can be harmful

[uncut fiber]

yes 🙂

[fierce hollow]

this is a pretty dope idea, salami, mozzarella, olives or something

[hardy cipher]

maybe some eyeballs as well

[uncut fiber]

pitty using QRmonster i cant be with you all here 😦

[woeful patio]

Weird question, I want to create ai voice dataset, does anyone know where I should start with this?

[noble shoal]

Do you mean Text to Speech?

[woeful patio]

yes sir

I was reading on gradio is that a good start?

[noble shoal]

You could look into https://github.com/neonbjb/tortoise-tts

[woeful patio]

cool, can you train models with that too?

[noble shoal]

I think that's possible, but i don't know for sure. Let me check whats state of the art right now

[noble shoal]

This also looks very promising, there is a demo available on Hugging Face: https://github.com/coqui-ai/TTS

[woeful patio]

that does look pretty good, I'll take a look. Thanks

[uncut fiber]

i think there were one author painting heads as green apples 🙂

[hardy cipher]

those other 3 images. very nice

[noble shoal]

They are staring into my soul 😱

[uncut fiber]

can anybody do bitten pizza? Oh i probably got what i have wrong...

[hardy cipher]

meaty boy

oh my

[noble shoal]

Better than Chicken Man

[hardy cipher]

why not both?

[noble shoal]

Send you a DM, cant show that here 🤣

[prime bison]

That's super meat boy to you

[hardy cipher]

well these should be interesting

[vital ermine]

check out its headlamp/headlight

[uncut fiber]

any english classical dish?

[shy kelp]

reminds me of land before time

[zinc cargo]

someone should shout on them to do it for SDXL

[zinc cargo]

seeing this and imaging her sayign I dont know in an angry voice

[hardy cipher]

pushing the limits

[hardy cipher]

khajiit

[winged scroll]

Hi everyone! I have created "game cards" thanks to SDXL. You can check out the whole project here if you are interested in the creation process. I would appreciate if you would leave a feedback.
https://www.behance.net/gallery/180764931/Ancient-Egypt-CSGO-Key-dropcom

SDXL 1.0 + Photoshop

and After Effect for animations

[ionic dragon]

@rustic garnet @fierce hollow do you guys know if anyone used pickascore to validate multiple images?

[vagrant inlet]

a cat with long hair sitting on a table, a photorealistic painting by Natasha Tan, cgsociety, photorealism, realistic anime cat, realistic anime 3 d style, realistic 3 d style

dream /a cat with long hair sitting on a table, a photorealistic painting by Natasha Tan, cgsociety, photorealism, realistic anime cat, realistic anime 3 d style, realistic 3 d style

[fierce hollow]

I have some classifier nodes but afaik there aren't any pickascore ones, looked into the code at some point but it wasn't very obvious of an implementation

[ionic dragon]

its not really good?

[fierce hollow]

it's quite good, just requires some effort to implement

[upbeat summit]

Bill Hawkins

[hardy cipher]

pushing this cfg hard. hitting unprecedented levels

[fierce hollow]

I have something like this setup, where it basically takes a bunch of images, picks the best one based on some classifier and if the score is higher than threshold it continues with the gen, otherwise stops

[hardy cipher]

what do you use to rate the images? because I've found myself disagreeing with the assessments of the models I tried out

[fierce hollow]

yeah, it's an issue for me too

tried a bunch of different ones but none are perfect

I often just pick all 4 and average the score, that tends to help, currently using... one sec

[hardy cipher]

up to 60 now

[fierce hollow]

cafe_aesthetic/style/waifu and vit-age-classifier

pickascore is nice since it also takes the clip embedding (is my understanding anyway) and checks how close the image follows the prompt

gotta implement it one of these days

[hardy cipher]

right on. I'd like to learn more about that stuff. and that seems to be a pretty unbiased approach

[ionic dragon]

oh ok

[hardy cipher]

alright, going for cfg 70. no one has ever tried this before now

alright, something is cooking

[full idol]

Has anyone got the "control-lora-sketch-rank256" working and is using it??

[hardy cipher]

no

[fierce hollow]

vit-age-classifier actually works pretty well by itself too in place of the aesthetic ones, if you give it a bit of a ramp to follow where the highest scores are assigned to [your-preferred-age-range-no-judgy] it tends to pick pretty good images, though it mostly works for realistic art

[hardy cipher]

see, I just make pictures of weird stuff

cfg 70.5

[upbeat summit]

have you tried https://github.com/mcmonkeyprojects/sd-dynamic-thresholding? there's a comfyui version now

[fierce hollow]

I kinda like low cfg now, like 3-5, works well with loras

[hardy cipher]

had to bump up that threshold percentile

3-5 is great and all. but have you ever tried 70?

[fierce hollow]

I mean yeah, I imagine most people cranked the sliders up to the max at some point and noped the hell away after one gen

[hardy cipher]

well it's not something everyone can handle, that's true

[fierce hollow]

dynamic cfg is nice too I guess, personally I never really found much of a use for it

[hardy cipher]

also doing this because I can

well I like to make things that are different from everything else, and these are definitely that

[vale eagle]

freeU, freeU depth map, ori

freeU is good for structure (usually) buy bad for color. The idea is to use depth map to keep the structure from freeU and generate with ori prompt

[hardy cipher]

that's pretty cool actually. nice work

everyone doing this stuff so quickly too

[fierce hollow]

if you ask me the colors from freeU look better than original even without the depth map

[vale eagle]

It usually add too much red and blue for me

[fierce hollow]

in your examples above the original feels kinda washed out but could be the case in general, not sure

[vale eagle]

ideally should use the same color tone like ori

still experimenting

[hardy cipher]

I'd say some kind of merging of the two would be ideal

or at least the option to do so. maybe some kind of sliding scale

cfg 70.592 looking pretty clean. not gonna lie

[vale eagle]

I think I should use recolor instead

[peak dove]

Dudley Doberman

[peak dove]

Bill Potter/Harry Gates!! 🥳

[full idol]

anyone ever found good use of the SDXL sketch control lora?

[vagrant inlet]

https://giphy.com/gifs/pup-parker-parkerpup-gfeYTTL8dTqIv9ErhX

[uncut fiber]

@foggy garnet qrmonster?

[foggy garnet]

sdxl depth

[uncut fiber]

oh great!

[uncut fiber]

very nice

reminded me atari 🙂

[uncut fiber]

i made those to myself in inkscape 🙂

those with alpha not sure how they would work with depth

this is bit more dense then i posted before

[foggy garnet]

i hope they train qrmonster for sdxl

[uncut fiber]

we talked about it yesterday and author claimed somewhere where repo is, he will. Just skip 2.1 models

[urban fjord]

I think there is a QR model for SDXL on Civitai and it seems kind of simple to train if you don't want to wait.

[vale eagle]

freeU, freeU depth map+ori ip-adapter+image blend, ori

[strange mist]

what is freeU?

[uncut fiber]

@strange mist
https://arxiv.org/abs/2309.11497
i am bad in english

[strange mist]

is there a node / extension for it? 😅
I don't have enough functioning brain cells left to read and understand research papers.

[vale eagle]

just update to latest comfyui

[strange mist]

oh so it's already implemented? 😮
cool beans

[vale eagle]

It seems to became my workflow with freeU

[uncut fiber]

can anybody share wf with animdiff?

[ionic dragon]

hey guys I need your help
can you convert this prompt into a good prompt antique photograph of a beautiful woman, sad eyes, sad smile, facial symmetry, frontal view, cracked and faded photo paper, staring at the camera, headshot, dark background, 1850, low light, dark, 4k by adding a bit more context and adding/remove tags

im training an llm to generate prompts, so your help by converting this prompt into a good prompt would be valuable for us

[noble shoal]

In this time range I would prompt for Daguerreotype. Try to mention old photo techniques instead of the damage and fading.

[ionic dragon]

it would be great, if you can type out the entire prompt

it is appreciated

[noble shoal]

I would like to test the prompt first before I write something, but unfortunately I am currently at work. But if I would prompt in my style I would probably go for: "Daguerreotype photo of a sad beautiful woman, ca. 1850" thats it.

[uncut fiber]

looks like facial asymmetry move their head on side 🙂

[ionic dragon]

the prompt is too short, i want to add a bit more context so that we get precise details as possible

[uncut steeple]

In the course of this project, you are challenged to replicate an antique photograph reminiscent of the mid-19th century ambiance, employing modern 4K resolution to capture haunting emotion and detailed aesthetics. The focus is on a beautiful woman, whose melancholic eyes and subtle, sad smile speak volumes. The attributes of facial symmetry, a defining element of classical beauty, should be well articulated in the image. Your model will pose for a headshot with a direct frontal view, engaging the camera lens with a soulful gaze that traverses time.

The photograph's aesthetics should mimic the age-old picture—cracked and faded photo paper that tells tales of bygone eras, accentuated by the stark contrast of a dark, enigmatic background. Low light conditions will engulf the scene, adding to the solemn yet profoundly evocative atmosphere.

The frame will capture more than just a melancholy visage; it's a recreation of a moment frozen in time, rekindled by modern technology. Your objective is to seamlessly marry the essence of 1850 with the precision of a 4K resolution, concocting a visual narrative that is as compelling as it is nostalgic.

Upon completion, the project should invite the audience to reflect on the timeless beauty and transient nature of human emotions, preserved through the evocative power of photography. Through this lens, allow the viewers to travel back in time, feeling the tangible sorrow and tender beauty radiating from the antique visage now reborn in high definition.

[ionic dragon]

is this gpt/llm generated?

[uncut steeple]

Sure

[ionic dragon]

the misc tags are missing

[noble shoal]

How many tokens? 🤣

[uncut steeple]

Enough

[noble shoal]

Flawless prompt

[ionic dragon]

like 8k, uhd, charcoal, unreal engine, etc, these tags can really boost an output

[uncut steeple]

Recreate an antique photograph, 1850s setting, beautiful woman with sad eyes and a melancholic smile, frontal view, facial symmetry emphasized, headshot against a dark background, low-light ambiance, image on cracked and faded photo paper effect, modern 4K resolution, emotion-transcending time, blending historic aesthetics with today's technology

[ionic dragon]

damn, its too good
did you type it out or gpt?

[zinc cargo]

that prompt is not so bad 🙂

[ionic dragon]

its too good

[uncut fiber]

i added yellow tint. Actualy wasnt those old photo manualy colorized?

[uncut steeple]

GPT

Im very lazy

[ionic dragon]

damn, gpt4?

[uncut steeple]

Yup