✨｜sdxl

[high skiff]

A member on our research team is working on their own more controllable open pose style model with control for body proportions as well

last I saw, it was actually going quite well. But controlnets are a pain in the ass to train lol

[radiant tartan]

ah very cool 😄 and thanks, can't wait

[high skiff]

Same here lol

I am nowhere smart enough to train something like that

his uses a floating point matrix instead of stick figures, so it should be easier to get more complex poses without tripping it up with overlapping "lines"

[steady grove]

Astronaut doing a side karate kick

[high skiff]

lol nice

[radiant tartan]

thanks 😄 yea thats about what i got using Depth but it looked more like a picasso. 🙂

[steady grove]

you usually should prompt a little closer to the image, as the controlnet won't have much context wiht just "astronaut". it's just another aid to the diffusion steps, and it's still a text to image system at it's core

[radiant tartan]

a simpler pose did work 🙂

[steady grove]

my astronaut in space reflected a desert and yours indoors reflected space. latent spaces are weird

same open pose, didn't work much but you can see it guiding

[radiant tartan]

normally i would use a more discriptive prompt, but i prefer to test with a basic prompt just to see how well the models work 🙂

[steady grove]

yeah you found the limit. weird poses.

[radiant tartan]

lol XD

[steady grove]

try a hand stand

[radiant tartan]

not bad,

[radiant tartan]

are these using openpose? or just prompting?

[steady grove]

open pose. astronaut doing a handstand on the surface of a comet.

[radiant tartan]

lol i used the same image

[steady grove]

great minds an all that

[heavy zinc]

Good evening everyone, can someone guide me with sdxl? I heard it needs powerful machine to run it.

[west breach]

You will get decent performance from the last few generations of nvidia cards with at least 6gb VRAM, preferably 8gb. It runs on my laptops 3060 6gb GPU too

[timber vapor]

hey all, does anyone know what the best SDXL inpainting model is? I saw one published in diffusers, but curious what y'all are using for inpainting/outpainting

[indigo carbon]

this is odd, I never asked for a heart emoji instead of the word "love"

[slender coral]

For training, I should use Auto11, and not compfy?

[upbeat summit]

yeah it does that with the word love on shirts sometimes

[delicate kelp]

Look what I've been cookin' up... a lil Dyna update in the works.

[heady vale]

nice glow effect

[heavy zinc]

I have RTX 3070 Ti, 16 GB RAM, i9-12900H

Are my specs powerful to run SDXL?

[half ivy]

i am here
what did i miss

[west breach]

The only issue you might have is running out of ram, your GPU and CPU look great

[half ivy]

yes
i have a 1660 and im fine, but, 32gb ram

overkill lets go!!!

can sdxl run multiple CN's?

[heavy zinc]

Oh, how can I adjust the VRAM or thread count to cause no issues

In stable diffusion

I have maximum of 8 VRAM and would like to use 3

[glad grove]

3 not enough

unless u want to wait 3 hours for a single image

[heavy zinc]

So 8?

[glad grove]

also i dont think theres a way for u to adjust vram

[heavy zinc]

Oh

Isn’t there a command or something

Like execute vram count

I am new to SD 🙂

[glad grove]

theres no such thing in auto1111

[heavy zinc]

Okay good to know

In the civitai web site, there are options for images I cant find in the web ui

[slender coral]

You can do this in python.

[heavy zinc]

Such as the vae, hires prompt etc

[heavy zinc]

Yes how can I do it

[slender coral]

This is done in the settings in auto11

[slender coral]

Google,

[heavy zinc]

Could tell me where to find them there?

[slender coral]

The settings tab.

[heavy zinc]

Yes but there are bunch of things

[slender coral]

You will need to read them, there's a Vae field on the first tab.

hires prompt is done in the actual prompt flags, there will be a checkbox for this.

This isn't rude, but I would watch a intruduction to save yourself some time.

They go over all this.

[heavy zinc]

I know have watched couple of videos and I am still learning

Is there a prompt to generate full body picture like from head to toe

[glad grove]

full body

[slender coral]

full body picture

[heavy zinc]

Oh as simple as that!

[steady grove]

mentiong both feet and eyes too

[heavy zinc]

I have successfully generate my first sdxl image in 7 minutes!

[steady grove]

shoes, legs, standing, lots of ways to describe

[heavy zinc]

Using Realism Engine

[heavy zinc]

I heard it doesn’t generate feet and hands properly. How can I fix that

[steady grove]

these are natural language systems. it turns your prompt into a latent target and denoises it

[heavy zinc]

Oh, is there a way to fix them?

[steady grove]

generate another image and see if it's closer. investigate prompt styles that target better hands, find refined models or loras that would get it better, inpainting , controlnet, lots of options

[heavy zinc]

Oh yeah styles, it’s a custom prompt, where do I find it

[steady grove]

there's an extension called adetailer for a1 webui. it'll automatically mask the hands and do an inpaint pass against them

[heavy zinc]

Refined models? Controlnet?

[heavy zinc]

What does it do? Why it mask the hands?

[steady grove]

yeah. anyone can take the base sdxl weights and refine them further on their own image sets

[steady grove]

inpainting is when you only do diffusion on a specific masked area of an image. changing only that area

[heavy zinc]

Oh

[steady grove]

that second pass with a prompt specifically for hands, helps a lot

controlnet is a system where you can give a visual control to the composition of the image. depthmaps and poses and other means of compositional control

https://github.com/Mikubill/sd-webui-controlnet
https://github.com/Bing-su/adetailer
two extensions you'll want

[heavy zinc]

There is also on model page on civitai on the right hand side a dropdown of files. Like full model fp16 and 32

What are those

[steady grove]

datatypes that the model uses. the amoutn of precision they have. floating point numbers (decimals) with 16bits of precision or 32bits of precision. super technical stuff. you'll only ever need half precision. fp16.

[heavy zinc]

I just pressed the download button at the top

I didn’t select 16 or 32

[steady grove]

more precision may be better for training but as far as i can tell, not needed for inferrence at all

default should download the half precision fp16 model

[slender coral]

I personally think bing create is one of the better generators out there, anyone know if there's a opensource model that's comparable?

[glad grove]

no

[steady grove]

sdxl

closest thing to any of the monolith models from corps

[slender coral]

Yeah I assumed, just wanted to check if I missed anything new.

[steady grove]

https://pixart-alpha.github.io/ theres this which uses t5 but we don't have weights yet

[slender coral]

Are they finding the NLP is the key to photoreal?

[steady grove]

it's a whole lot of things that get better results. a larger pretraining model than openclip helps. that's t5. also, a larger unet is important too.

https://twitter.com/EMostaque/status/1714982965212508228 emad keeps showing off stuff like this that has great prompt comprehension so i think they got some research up their sleeves

[supple knot]

@heavy zinc fyi, If you use ComfyUI it will cost less VRam and work without crashing

[heavy zinc]

What is ComfyUI

There is someting called badhandsv4

Does it work on SDXL?

[steady grove]

a node graph for affecting diffusion

[heavy zinc]

Just to be sure, SDXL is same as 1.6 right?

[steady grove]

1.6 is the version of automatic1111 you're using

[heavy zinc]

Then what is SDXL?

All I know it is the best version of Stable Diffusion

[glad grove]

i dont think u watched the videos u were talkin about

[steady grove]

sdxl is the 3rd generation of model released from stability ai. it's an architecture / base model

[heavy zinc]

Oh

I have Realism Engine SDXL model am using

Is it the same one?

[steady grove]

they release those "weights", a bunch of numbers that multiply together according to the inputs, for free. another program like automatic1111 or comfyui uses those weights to generate images with

realism engine is a version of sdxl that someone else refined with their own images. they train it for many hours so that it learns more.

[heavy zinc]

Oh now I get it

so Does bandhandsv4 works on SDXL?

[glad grove]

no but theres another neg embedding that does work for xl

[heavy zinc]

Yes i found it

I put it on embeddings folder?

[glad grove]

yes

[heavy zinc]

its not a pt file

its safetensor

and how do I use it ?

[glad grove]

some are pt and some are safetensors,just put the name of the embedding in the negative prompt box

[heavy zinc]

Is it in Textual Inversion tab?

[glad grove]

yes its also there just click it and it will add itself

[heavy zinc]

okay

So how do I adjust VRAM in python for SD?

[glad grove]

🤷‍♂️

[slender coral]

Vram is allocated automatically.

[glad grove]

one of the weirdest AI videos i have ever seen

[heavy zinc]

I cant fnd a vae which cis vae reverted

on the net

[heavy zinc]

What is the best way to enhance low res pic in SD?

[heavy zinc]

Why In Lora, It is says showing nothing? I have put my files in LORA folder and restarted sd

[steady grove]

are you trying to use a sd15 lora on sdxl?

things like loras, embeddings, hypernetworks, controlnets, they all need to be developed for the same architecture. sd1, sd2, sdxl

[heavy zinc]

oh

It says 1.5

it wont work?

[half cedar]

I am Legend ^

[supple knot]

@heavy zinc In general sd1.5 is trained to make 512 x 512 resolution ,sd2.0 is 768 x 768 and sdxl is 1024 x 1024 ,

[heavy zinc]

oh

when I load up SD , it is 512x512 by default

is there something wrong I have ?

[supple knot]

your using A1111?

[heavy zinc]

Yes

webui-user

[supple knot]

Im not to sure I only use ComfyUI, I do have A1111 that I started with but got into ComfyUI when they released SDXL and I havent went back, what are you trying to make "Text to image"?

[heavy zinc]

yes

[supple knot]

if you use a SD 1.5 model 512 x 512 is the right size

[heavy zinc]

What differences will it make when using ComfyUI?

I downloaded 1.6

but it says 512 512

[supple knot]

Comfy give you a better idea what's going on with the process and where its stuck at

[kindred shoal]

COmfyUI is also mroe complicated to use than A1111, still learning it myself

[heavy zinc]

oh

then its not for me

Can i Hires after generation is finished?

[crisp owl]

it "can" be more complicated to use

[heavy zinc]

yes its not for me

[kindred shoal]

By reusing the seed, yes

[crisp owl]

or you can use a premade build and just plugin the values much as any other ui

[heavy zinc]

how do I do that

[kindred shoal]

Under seed, right next to the box with the whole load of numbers, there's a green recycle button. CLick on the finished image that you'd like to add hires fix to, and then click on the green recycle button.

[high cove]

I saw this article from huggingface https://huggingface.co/blog/sdxl_jax, and was wondering if it would be possible to apply these same concepts to speed up SDXL model training, and if anyone has tried doing it already

[kindred shoal]

Then, with the same prompt, click generate.

[heavy zinc]

So first I click on the image

then the green icon?

[kindred shoal]

Mhm

It'll retrieve the seed, and reuse it.

To randomize it again, click on the dice icon.

but it wont randomize once you grab the seed.

[heavy zinc]

I have the seed number

Do I save it ?

or what

I just clicked the generated image on the righ side then opened Hires options with the same prompts

What do I do next?

[glad grove]

u should read the webui wiki

this one covers all the basics https://stable-diffusion-art.com/automatic1111/

[lofty helm]

always save every seed number. you'll need those later

[heavy zinc]

When do I need seed number for later?

[high skiff]

I am using @nimble heart's tinted latent node for SDXL, and man, it works better than offset noise when paired with my realism LoRA

Here are some extreme examples.

It seems to play much better with my first gen Realism LoRA than it does with finetuned realism models

get as bright or as dark as you would like

it can go brighter and darker, but thats a but much haha

and you can do COLORED tints as well

Prompt: A photograph of a warm forest at dusk

These are all the same seed/settings/prompt, just with different latent colors:

cyan latent, orange latent, dark grey latent, magenta Latent, yellowish green latent

its still really experimental, and I actually chose a pretty bad prompt for that haha

let me try some others in a more neutral prompt

[supple knot]

how can I find this node?

[high skiff]

just a sec

https://github.com/Beinsezii/bsz-cui-extras

I am using the RGB latent colorizer

[supple knot]

thanks ill check it out, always love your renders

[high skiff]

thanks :>

please stay tuned, gonna be testing a new setting validation and new image tagging for the V1.5 of my realism LoRA (trying 120 images instead of the 90 I have right now)

one step closer to the 500 image 2.0

And one step closer to the 2500 image 3.0

goal is to be able to control image framing, color grading, and realism subject effortlessly

while also having way better lighting than any of the realism finetunes (something I have already achieved)

"A cinematic portrait photograph of a pretty Japanese fashion model sitting in a room"

Normal Latent, dark grey latent, pink patent, blue latent, orange latent, yellow latent, white latent, blood red latent, purple latent:

Some prompts and secens do better. I was trying to get one pretty neutral. It also works well if you prompt for specific colors as a theme and match with the latent

"A cinematic portrait photograph of a white woman dressed in red in a red room with red lips"

Normal Latent, red 20%, red 50%, red 75%, red 100%

You can see it gets a bit much at the end lol

(please excuse the weird noise and slight image artifacts, I am benching my LoRA with no negatives to see what to correct in my dataset)

"A cinematic portrait photograph of a tiger in a forest at night"

normal, 25% black, 50% black, 75% black, 100% black

infinitely controllable offset noise

[noble shoal]

Yes, the node is awesome.

[high skiff]

Just be warned, it doesn't play with models as well as it plays with my LoRA, at least in dark values

[vital ermine]

https://civitai.com/models/173888?modelVersionId=195245

[nimble heart]

it's such a general purpose hack it should apply everywhere with a little tweaking of the values. Even works with the refiner if you really want

but yea what 100% black does will vary from model to model based on the contrast they cooked in

[indigo carbon]

I've found an anomaly in SDXL. why tf does it ALWAYS put the Apple logo on laptops

[nimble heart]

subliminal sponsorship

[indigo carbon]

also sometimes cars are often just Mercedes if not specified for some reason

[zinc cargo]

hey guys, how do i unbind the 8188 port or set it another way ?
PermissionError: [Errno 13] error while attempting to bind on address ('127.0.0.1', 8188): an attempt was made to access a socket in a way forbidden by its access permissions

[rustic shadow]

lil cuties

[indigo carbon]

there is a --port flag in ComfyUI

just add the flag to whatever your ComfyUI startup command is

and after you write the open/desired port you wish to use

[vital ermine]

tbh, though I lol'd, so many images to scrape for training are Apple laptops and Mercedes is big in cars too. A bit odd.

[indigo carbon]

yeah, I find it humorous that SDXL thinks the only cars are Mercedes and the only laptops are Apple

maybe a SAI x those brands collab is inevitable

[high skiff]

I am working my ass off to scrape images that are ideal for my dataset. it took me a week to get all 500

and its gonna take me forever to caption them as well, but I have faith it will lead to my LoRA kicking the asses of SDXL realism finetunes even more :p

[high skiff]

I was mostly referring to my model responding better to black offset specifically. The other models seem to kinda fight it and get muddier results

As seen in these:

The models go from left to right as follows: Mine, RealisticVisionXLV2, RealismEngine, RealStockPhoto

realistic vision plays really bad with this trick for black latents I have noticed

[indigo carbon]

SDXL and diffusion in general are meant to be general purpose, but LoRAs can be convenient sometimes

[high skiff]

when it comes to realistic photography, my LoRA is the GOAT for most things

At least from what I have seen and tested (around 700 prompts, from 9 different SDXL models)

Same order as above

mine seems to take much better to the black latents, while also having much better realism/lighting/background separation

thats 50% black offset

[indigo carbon]

idk man, realism doesn't seem to be an issue for my experience

[high skiff]

those look alright, but definitely not like my model

my model 9/10 times stops the competition from my tests, non cherry picked. Thats not to say it doesn't lose, cause it does sometimes

[kindred shoal]

they look real enough to me

i like the 1st one on the second row

Im a sucker for glowy bits

Green is real pretty too

[high skiff]

they look passable, yeah. Just not like they were taken on a real camera. Just kinda the standard flat depiction of real life with improper lighting, inconsistent background blur, and impossible dynamic range

Its a stylized realism look

[indigo carbon]

plus, the model I used there is actually a model I've calculatedly made by block merging. it's not meant to be realism or anything; just a good general model.

[high skiff]

In all of these, the model order goes:

Mine in top left, RealisticVisionXLV2 top right, RealismEngine bottom left, RealStockPhoto bottom right

[vital ermine]

Well, let's look at how much money they have and the circle they live in as well.

[high skiff]

mine consistently looks a lot more properly lit, with more realistic dynamic range, better foreground/background separation, and typically more coherent backgrounds with better fidelity

the whole point is it does really good realism with painfully basic prompts

same order here, Mine, RealVis, RealismEngine, RealPhoto

Entire prompt is "Dog"

"butterfly"

when unguided, realism engine is the closest to mine, and when guided, real stock photo gives me a run for my money occasionally on some prompts. Every once in a while real stock outputs very good images

[indigo carbon]

these are the parrots I usually get

the main way to improve SD is the text encoder itself, CLiP is kinda a bottleneck

[high skiff]

those aren't bad, though I assume you use negative prompting and such as well. I am not using negative prompts, just to see what they can do base.

All of these prompts are a single realatively short sentance

I do like the aesthetic of your images, but its still not "this was taken on a DSLR" like what I am going after specifically

which apparently takes a ton of work to do

evidently lol

[indigo carbon]

just promped "parrot in a jungle", that's it

SDXL has a SOLID UNET, but the text encoder is a stick in its wheel

[high skiff]

if thats really the case, then you have the best realism model by an absolute 10x landslide lol

[high skiff]

ah, the base style is helping some

also interesting you are below 1m res, which is something I have messed with too

[indigo carbon]

the base style is empty, it means nothing is being applied

[high skiff]

I have found SDXL looks a little better slightly above/below 1024x1024 eq

[high skiff]

I'd be interested to try this model you are talking about then, cause if this is really the output with no guidance, then your model is substantially better than all of the SDXL realism finetunes by a landslide

jesus christ, 67 steps? whats that about?

[indigo carbon]

uses AITemplate, sampling takes 10-15 seconds for batch 4 anyways

so I can go ham with the steps

[high skiff]

thats just wasteful regardless. Not sure why you would add all of those steps for no reason

[indigo carbon]

that's just how I like to set my workflow, my personal preference

[high skiff]

Fair enough I guess. I would prioritze speed/enegry savings over wasting energy for no gain 😅

is your model out? I would love to test it if so

[indigo carbon]

yeah, it is. it's on CivitAI and I explained exactly how I made it there

didn't train anything, just calculated model merging

[high skiff]

there has got to be something else going on here, cuase these results look impossibly good. I must test for myself

these results are so much better than any SDXL model I have ever used, I just don't see how its a mash up of all of them

[indigo carbon]

could also be my workflow =\

[high skiff]

I am willing to bet thats more likely

testing out in just a sec when its done

I am gonna drop realism engine, cause its consistently the least good on average

[upbeat summit]

greetings!

[high skiff]

for anybody interested, the absolute worst realism model I have used is juggernaut XL

it got absolutely trashed in every single image gen test I did

it was like 5 steps behind all the others I tested

and it breaks horrifically with any upscale diffusion

I am not sure what is up with that model

[noble shoal]

The parrot thing seems not to be that high of a challenge. Prompt: Parrot in jungle

[high skiff]

those look pretty bad compared to @indigo carbon

I am pretty nitpicky, and I see a lot of issues with those two

[noble shoal]

Yeah, for instance the burned out whites

[high skiff]

I see issues with TDG's as well, but there are a lot more positives than the ones you posted

[high skiff]

thats probably one of the most accurate things in the whole image, for it being from a camera at least

thats one of the reason why the SDXL realism finetunes always look so flat and muddy. They have impossible dynamic range that just makes images look artifical, cause real cameras don't capture things in perfect dynamic range

[indigo carbon]

for me it just makes these

[high skiff]

I had to train my model to revert that, and thats one of the things that makes it look much more realistic in lighting

[high skiff]

ok, these look significantly worse than the ones you posted above. Whats the difference?

[noble shoal]

He even missed the focus.

[high skiff]

oh shoot, i downloaded youre model 2 times lol

[indigo carbon]

I didn't change anything, Idk

only difference is seed

[high skiff]

hmm... weird

those look much much much less good/accurate

those look way more like base SDXL

way more scrambled, incoherent, messy, inaccurate lighting, no focus planes, plastic rendering on the feathers, wood texture looks bad, hallucinating other parrots and colors in the background

I would have never thought those were from the saem model had you not said something

[high skiff]

These get a 3/10 for me

[high skiff]

these get a 8/10 for me

that is crazy that those came from the same model

they went from pretty damn solid to pretty damn bad IMO

[high skiff]

the left ones have so much more fine details, better feather rendering, better light sources, proper rendering on the leaves with the textures, muchmore accurate wood, proper focus planes, better tonal contrast. I am having a hard time believing you didn't change something considerably in your workflow

[high skiff]

and these look as good as the first set

what are you changing, cause I do not believe you when you say its just seed lol

that is an unreasonable difference from seed to see if thats the case

[indigo carbon]

idk man, I'm just diffusin'

[high skiff]

I guess I will have my answers soon, gonna try it for myself

that second set of images looks way more like what I am expecting

[indigo carbon]

this conversation reminds me of Gus Fring talking with a scientist about the quality of their product

[high skiff]

how so? lol

[indigo carbon]

"this batch is different" "this batch is better"

anyways, idk man

[high skiff]

these look as good as 1/3

1,3, and 4 look the same

2 looks really bad by comparison

ok, yeah, just tested your model. Something is massively different with your workflow, and now you have my interest piqued, cause your model looks really bad for me lol

[indigo carbon]

you're talking about the batch itself or the batch(es)?

[high skiff]

I think all of the images in batch 1, 3, and 4 look 8/10 for me

While all of the images in batch 2 look really bad to me

and the results I just got from testing look even worse than 2 for me. So I really think your workflow is doing something really good

your batch 1/3/4 look solid is what I am saying

[noble shoal]

Wait, does a parrot has nose holes? 🤔

[indigo carbon]

it's just a basic workflow that uses AITemplate

[high skiff]

yes

interesting, if I add "photo" to the prompt, your model does much better

[indigo carbon]

literally not even close to the IPAdapter workflows I did

[high skiff]

dare I say, your model looks as good, if not slightly better than the realism models I am testing

wait, you are using unclip conditioning?

[indigo carbon]

in the normal workflow? nope

[high skiff]

@indigo carbonI am officially swapping out realism engine for your model in my testing going forward. I am very impressed with what I am seeing

[nimble heart]

"my workflow isnt that complicated"

[indigo carbon]

cool, don't credit me, all I did was the math for the block merging

[high skiff]

you may have ended up making the best realism merge out there that I have tested

ok nope, nevermind, it falls on its face like the others with portraits 😅

interesting

it did damn good with that parrot tho

[indigo carbon]

give me the portrait prompts, I kinda want to also see how it'll do them

[high skiff]

Portrait photograph of a pretty woman standing in a field of flowers wearing a suit and tie with braided hair at sunset

there is one

yeah, it looks like it struggles just as hard as the others for portraits without key word mashing

[high skiff]

yeah, super plastic

exactly what I am seeing

they also look stretched in what you sent. Did you not update your latent sizes?

[indigo carbon]

makes sense, like I mentioned earlier; that model isn't meant to be the best at anything, just a good general one

[high skiff]

Mine, real vision, yours, real stock photo

[noble shoal]

Very interesting. I've got basically the same person with the base model.

[indigo carbon]

LOAB?

this won't be the first time we see a woman reappearing in multiple generations

[high skiff]

"A portrait photograph of a tiger walking in the snow up in the mountains"

Mine, real vis, yours, realstockphoto

[indigo carbon]

yours is a LoRA, isn't it? the shapes are entirely different

[high skiff]

my LoRA is reallyyy overfit on a specific tiger image lmao

[high skiff]

yeah, I trained this LoRA really hard

and its super overfit on a specific tiger lmao

I only have one in the dataset

the new 500 image dataset has 7 different tigers

the tiger is the only thing that overfit in this LoRA lol

Oh, I also had the LoRA turned up too high. Thats why it was a little high on contrast

this image! lmao

it reallllyyyy grabbed onto that tiger lmao

[indigo carbon]

still, you are comparing 3 models that their main purpose is to be realism to a model that's meant to be generally decent

[high skiff]

I know, and yours is holding its won

*own

dare I say... I like it more than the other two

your generalized merge seems to be better at realism than the realism finetunes, albeit slightly. I am testing more

you can see again here how overfit my LoRA is on that tiger lmao

[indigo carbon]

it's similar shapes on the ears for some reason, odd.

[high skiff]

it just really overfit on that one image. A lot of subjects were represented with just one image, but the tiger held on SUPER HARD

so my new 5.2x bigger dataset has a few more tigers in it to fix that lol

and I am gonna remove that image

cause god damn does it love that tiger lmao

[indigo carbon]

this reminds me of when I tried to blend Markiplier and Shrek

[high skiff]

I think realvision is consistently the worse for animals

shrekiplier lol

[indigo carbon]

he's too strong, it's just markiplier dressed up like shrek

[high skiff]

A portrait photograph of a creepy ceramic garden gnome wearing glasses in a swimming pool

Mine, real vis, yours, real stock photo

mine looks the most like hes on the SO list lmao

but I like all of the results in different ways

yours looks the most like a terracotta/clay gnome in the mexican style

mine looks the creepies and the most like a real photo

real visions is super cute

and I like the tiles on the pool in stock photo lol

all of them get a pass for me

[indigo carbon]

that means it failed, you prompted using the word "creepy"

[high skiff]

yeah, true haha

mine wins in creepy, but I like them all!

they all get a pass in my heart lol

looks like the images I added for tree bark are working well too

god, real vision consistently looks ass for non people lmao

[indigo carbon]

also, a model (not SD) that uses T5 is about to be released, so we'll get a chance to see what other text encoder can do

[high skiff]

man, these all look bad lmao

[indigo carbon]

T5 is the same text encoder that OpenAI used on Dall-E 3 also

[high skiff]

Yeah, I know

Dalle3 is interesting

[indigo carbon]

I'm unsure if SAI will also adopt that encoder, because it'll need to be quantized to work on normal machines

[high skiff]

it listens really well (except for the times it flips you off and chooses to do literally nothing you ask for)

But its results usually look really bad IMO. Always flat, desaturated, no color grading, soft and lacking any real fine details and textures

I think, personally, that DE3 is the worst "looking" image generation model out right now

For visual aesthetics, I would take SD 1.5/SDXL/kandinsky/MJV4-5 over it ANY DAY

[indigo carbon]

yeah, the quality is bad. but the coherence is in its own league

[high skiff]

coherence = good
Aesthetics = ew

lmao

its like, every time I see an image from DE3 I think "Thats a cool concept, wonder what it would look like made by a visually pleasing model"

like "Wow, its so coherent! too bad it looks like ass and I don't wanna see it again"

[indigo carbon]

for sure, I'm unsure if the quality is bad due to T5 or if they fucked up something

because they did confirm it's latent diffusion

[high skiff]

I would rather have an aesthetic lump of crap than a flawlesly coherent boring blah that inspires nobody and nothing with the sheer impressive levels of drab-ness it exudes

I look at DE3 images, and I feel static in my brain

It feels the most lifeless of all of the models by a landslide, IMO

I would for sure take DE2 and its mountain over issues over DE3, just cause DE2 looks fun

[indigo carbon]

I think T5 might be a double edged sword, it's a tradeoff for coherence and quality

or it's just OpenAI being dumb.. we'll find out eventually when that experimental T5 model releases

[high skiff]

don't get me wrong, I hate OpenAI and all of the BS they have pulled, fucking over their investors, lying about what they are trying to do, all of that. And I am slightly biased against DE3, but I can recognize where its awesome, but god damn does it look like garbage aesthetically

[dense chasm]

run zoom-out and zoom in comfyUI with SD lol..

[high skiff]

oh thats kinda fun haha

that damn default comfy image is burned into the back of my retinas lol

[indigo carbon]

same, the "open" in OpenAI stands for "Open your wallet"

[nimble heart]

supposedly the quality improves if you dont use bing

but yea still not midjourney ✨ aesthetics 🌈

[high skiff]

I sure hope so, cause dalle looks great in comp, and nothing else. Zoom in, its static shit, look at the overall levels, its flat, uninspired and drab, look at the textures, they are dull and fuzzy

[nimble heart]

wonder if its pixel based like dfif

[indigo carbon]

if aesthetics is what you're talking about, SDXL is definately the winner there

[nimble heart]

XL is washed tf out

needs a black latent or post processing or something

[indigo carbon]

also DALL-E 3 is latent diffusion, so it could be T5 giving it that grainy effect

[nimble heart]

im mostly just wondering if they're running it at like the bare minimum steps for all the bing users

[wet glacier]

They all fail in one area: Skin. All models some to a higher degree than others still have the fake airbrushed look that has been all the rage on social media

[high skiff]

Yeah, i will agree thats probably the worst skin I have seen mine do

[nimble heart]

negative "model"?

or "instagram"

might help

"makeup" maybe

[high skiff]

my model can do skin texture, it just failed in that image

i had the LoRA set too high, and I also had the sampler set wrong

[wet glacier]

it doesn't matter, you need to train specifically for realistic (natural looking skin), not an easy feat

[high skiff]

I mean, I did, and it worked pretty good lol

just not on that image

[wet glacier]

please post some of the better looking skin 🙂

[nimble heart]

i think we're all missing the obvious solution here

reject human women and only diffuse attractive dragon women instead

[wet glacier]

hehe

[glad grove]

something like the skin on the right or more detailed?

[high skiff]

This is a crop

Please excuse the eyes. This LoRA is not done training at the moment. It didn't get a chance to fiully converge on high frequency details like eyes

[nimble heart]

i'd say eyes are mid freq

[high skiff]

also, I have tags specifically for more textured skin that I don't often use

[nimble heart]

high freq would be skin

[wet glacier]

the issue with 99% models is the lack of skin subsurface scattering

skin looks either plastic, dull, or artificially shiny

[high skiff]

uhhh

sub surface scattering is not a common thing in like 99% of photography, excluding ears lol

[wet glacier]

but in a real photo, you will see it

[high skiff]

[nimble heart]

maybe a black latent would show the pores better

[high skiff]

gonna have to x on that one

[high skiff]

maybe, I could try sometime

[wet glacier]

I use 100mp photos for work and I can assure you, that skin looks different

[high skiff]

mk

[nimble heart]

you ever see a dragon and just like think "would"?

[wet glacier]

it's the conundrum of the photoretoucher, how much skin texture is ok to show on a fashion shot

too liittle and it looks airbrushed

too much visible pores and it looks like you went too far with your high pass filtering

[indigo carbon]

yeah, like Donke from Shrek

[high skiff]

I know, lol

oh god

[wet glacier]

it's a fine line

[high skiff]

these realism models make people with "acne" look like they have skin deseases, oh my god lol

[wet glacier]

yep

[high skiff]

mine and real stock didn't, they look a lot better

[wet glacier]

or everyone looks like they came out from a Chanel makeup Ad

[high skiff]

I feel like you haven't messed around with SDXL enough if you think that TBH

[wet glacier]

really? 🙂

[high skiff]

and you're pretty passive aggressive as well, so I think I am gonna just move on. Have a good one lol

[wet glacier]

you don't have to be so sensitive, I was just trying to help with constructive criticism

[indigo carbon]

old people always have detailed skin with SDXL

[high skiff]

I am not sensitive, and I don't need your criticism, but thank you. I just don't wanna interact with you. So I won't

anyways

[wet glacier]

all good

[glad grove]

still waitin for that realistic skin pic

[high skiff]

I sent one above, though its not the greatest

[indigo carbon]

ew wtf

[high skiff]

makes old people really crusty lol

[glad grove]

no i mean to warlock,he should send one so we can see what kind of skin hes talkin about,show us an example 🙏

[high skiff]

ahhh, I get you

[indigo carbon]

the skin sure is more detailed when it makes old people

[kindred shoal]

Probably to exagerate age

[high skiff]

the problem is my LoRA has unrefined residual noise cause I didn't train it fully, so it adds some unwated patterns on skin when too close

[indigo carbon]

https://tenor.com/view/choco-choco-gif-26291245

[high skiff]

when I train it properly and get that to go away, it will look mucchhh better

[kindred shoal]

I would train a model but i heard that it's time consuming and requires some nice resources.

well I'd probably make a lora first, since that seems to probably be the easiest.

[high skiff]

my realism LoRA seems to work better than realism finetunes. LoRA's seem to be the way to go for dramatic changes

[wet glacier]

how about this

it's not perfect by any means

nothing is so far

[high skiff]

that... looks so crusty and artifical OMG

[nimble heart]

think it'd be a good idea to make a "latent gradient" node to go with the colored latents? for making simple patterns merging two latents

[high skiff]

I was wondering about that!

[upbeat summit]

I think the skin textures of older subjects is not necessarily more detailed. it has wrinkles and more stuff going on which gives it the impression of more detail but overall it's just as airbrushed missing pores and blemish.

[wet glacier]

btw, @nimble heart I have been using your workflow for some time and love it

[nimble heart]

kinda like this but without the math

[high skiff]

I think this skin looks WAY more realistic than that random non topology conforming skin

[nimble heart]

which one? Principled?

[high skiff]

but the residual noise makes it look a little too over-textured

grrrr

its an issue cause the LoRA didn't finish training

[wet glacier]

yes, principled

[nimble heart]

my non-principled workflows I haven't updated in like a few months at least

[nimble heart]

nice. try the colored latents in the newest version. pretty spicy

[wet glacier]

still trying to wrap my head around it though, trying to squeeze out the most out of it

[wet glacier]

I did 🙂

[nimble heart]

it's sorta like an auto1111 style pipeline

[nimble heart]

lmk if there's something in particular that's confusing so I can clarify.

[high skiff]

there is pores texture, but no fine lines/wrinkles, which makes it look more fake IMO

[glad grove]

looks like she needs some sleep

[wet glacier]

I like in particular the upscale and the latent color grading

love the skin pissing contest btw

[nimble heart]

latent color you kinda just have to try. Upscale was just reworked into it's own separate node so if you're talking about that, it's basically an auto1111 hiresfix. Just un-bypass the two purple nodes with ctrl-b and it'll do everything else for you more or less.

[nimble heart]

of you're still on the old all in one node then it's even closer to Auto.

ah maybe I misread I thought the upscale was confusing

I need sleep

toodles

[wet glacier]

I'm using this one

[high skiff]

what in the world is going on here

I am so confused

is that a sampler issue?

I am so sure it would look much better if I was able to get rid of that last bit of latent red noise on her face

[nimble heart]

yea that's basically it. the pixel buster node does math ops in this case to fix the black/white levels. Two sliders control scale and neutral.
The full Pixelbuster manual is it's own node called "Pixelbuster Help". pretty cool NGL

definitely not saying that cause I'm the one who wrote the library

[wet glacier]

That was another topic, I need to read the Help file

[wet glacier]

I can tell there's a lot under the hood

[high skiff]

Xiao, what is going on...

why the crunchies

are you using too few samples for the sampler you are on?

[nimble heart]

The principled node is mostly just existing ComfyUI nodes strung together automatically so it should behave as expected. Only change is the steps are dependant on denoising a la auto1111. I might revert that tbh, I haven't fully decided.

the other nodes aren't too bad either. I'd wager 95% of custom nodes are more complicated than mine.

[vale eagle]

Testing with my experimental fine tune. above image is like 35 steps, 5.5 cfg without refiner

[high skiff]

interesting

[nimble heart]

try it with a darker latent

that looks like a good candidate

[high skiff]

a slightly dark latent almost always makes images look a little better haha

[nimble heart]

make the tiger pop

[wet glacier]

That what I like about yours, it's not over bloated like a swiss army knife with 50 different options 😉

[indigo carbon]

I'm just messing around

[nimble heart]

at one point principled was longer with the scaling and some experimental stuff built in but I've sorta minimized it with as sane of defaults as I can. I used to even expose the XL clip sizes but now they're just auto calculated.

so far the only thing is principled won't work with conditioning nodes like IPA.

I could split the conditioning from the node but then it'd be super messy. Loras and whatnot already work so eh

[nimble heart]

@high skiff I'll try making that gradient node tomorrow probably till then night night

[high skiff]

alright, sounds good man

see you soon!

[west breach]

them apples

[indigo carbon]

also, what is realism engine? never heard that name before

[high skiff]

its a model for SDXL. Its not very good from my tests. It almost always comes in last

[west breach]

it makes everyone nude

[high skiff]

umm... I did see that happen a couple times lol

[indigo carbon]

sounds like typical CivitAI model, am I right?

[high skiff]

where girls were clearly wearing no bras under their dresses so it gave them some really prnounced nipples, or really scantily clad dresses lol

[indigo carbon]

yeah, sounds like something this community makes

[west breach]

I make pizza burgers

[high skiff]

what a crusty and prude old man >:C

lol

[indigo carbon]

he hates it

[indigo carbon]

how about now?

[high skiff]

I ❤️ Love SE X

lol

[indigo carbon]

I like how it made him depressed now

[indigo carbon]

they should join forces

[west breach]

yum

[noble shoal]

Now somebody needs to generate a newspaper article about the Pineapple pizza wars.

[west breach]

lol i like the pineapple head guy in this one, you can imagine they scrambled his voice in the interview

[glad grove]

why is there a guy with a pineapple head on there

[west breach]

a mushroom cloud made out of pineapple pizzas

[west breach]

creative use of pineapple rings as the wheels

[strange mist]

I'd eat the first dish as a taste-test
second one though, my taste buds can imagine how nasty that might taste

[vale eagle]

🤣

[west breach]

mcgruber!

[noble shoal]

Is that also served on a jeans? 🤔 Based on the beverages and the overall set-up I would say it was cooked and served by a drunk guy.

[vale eagle]

LOL

I tried several fruits and steak combination.

[half cedar]

Do "where are my farts" next

[west breach]

😄

[half cedar]

Hahaha gas pains, the worst

[west breach]

release

[uncut steeple]

[zinc cargo]

I used Dynavision with several Loras
Aether glitch and Aether bubbles from @icy brook
modern oil paint from @delicate kelp
and sktechbook from @sinful falcon

[lusty wolf]

Not eating crab again...

[copper spire]

Hello does anyone know if sdxl have an API

[icy brook]

Very nice!

[lusty wolf]

You mean ComfyUI run as an app. Yes, you can do this: Use MS Edge and install the Web IP as an API on your desktop.

[lusty wolf]

What's on the menu?

[cyan crown]

Pizza

[lusty wolf]

Without Pineapple!

[cyan crown]

of course

[noble shoal]

Everything else would break TOS Rule Nr. 4

[lusty wolf]

Smiling about it...

[cyan crown]

he's smiling too

and running for the victory

and after, having a Finnish wooden sauna

[copper spire]

Ah nice thank you

[heavy zinc]

Hi all, anyone have experience with tiled vae

[crisp owl]

Experience as in?
I use a tiled vae node in ComfyUI for my final upscaled vae decode

[south horizon]

Nice. I like the sketchy ones.

[heavy zinc]

Yes to use less vram

[crisp owl]

Yeah in ComfyUI, it will automatically switch to tiled vae if you don't specify it, but it will take a few extra seconds for it to detect that.
Alternatively you can set a tiled vae node manually if you know you always hit that threshold anyways, and thus save a few extra seconds.

[heavy zinc]

I heard ComfyUi is more complicated to set up, and I am new to SD

[crisp owl]

it can be, or you can just download a template someone already made and use it just like any other webui

Or use something like #🐝｜swarm-ui which has a more user friendly ui

but still uses Comfy on the backend

[icy brook]

John Bauer?

[glass notch]

Can someone ELI5 why DALL-E seems to "get" what the prompter is thinking of far more easily than SDXL? It's really a shame they put so many restrictions on its use

[crisp owl]

It was trained differently to have more focus on that, but it suffers in quality and graininess in the final images

[cyan crown]

yes

[icy brook]

On sdxl base only? I think it’s hard to make Nordic trolls anyway.

[hasty smelt]

Hi folks, does anyone know what prompt works well to get something similar to this style? I've already tried "vintage design," but nothing happens.

[glass notch]

Gravure, lithography

[cyan crown]

SDXL Base (I use more art Lora, but also without it's similar)

[hasty smelt]

thanks buddy

[hasty smelt]

THanks my buddy! I work now!

[glass notch]

maybe also linocut and woodcut. basically anything that has to do with etching the plate for printing in ye olden times

[hasty smelt]

I'll make a coffee and explore these valuable tips, thank you very much

[hasty smelt]

Thanks guys

This really works well!

[crystal gazelle]

Guys I am new to this, what command should I use to generate an image here?

[hasty smelt]

in bot use/dream

[crisp owl]

apparently prompting for an angry Thor, that means he has a light saber

[heady vale]

no, I am your father 😆

[crisp owl]

lol and zombie teeth with a shovel
🤣 gotta change up this prompt hahaha

[cyan crown]

Good night

[crisp owl]

popped eyes would hurt

[upbeat summit]

my poor unicorn

[noble shoal]

Horsetub

Horsehouse

[crisp owl]

Oh well now you're just horsing around

[pure crystal]

anyone know a video upscaler? preferably something I can run in a colab. I have a RIFE interpolator, that's working well

[noble shoal]

[wet nacelle]

https://tenor.com/view/jerma-sus-jerma985-jeremy-gif-19838735

HE GOING ALL THE FUCKING WAY!!!!

[hasty smelt]

Hey, is there any command for ComfyUI and SDXL for Low VRAM? I'd like to know if I can improve the performance of my 8GB 2070 Super.

[zinc cargo]

xenomorph it!

[crisp owl]

Comfy defaults to medvram.
I have a 2060S also with 8gb, I wouldn't consider going the full lowvram route

You probably should have --disable-smart-memory though put on

[hasty smelt]

I've read things more carefully, it seems like everything is fine with the use of the 2070S. I confess I was hoping for some miracle to reduce my rendering time.

[crisp owl]

Always a hope to reduce rendering times haha

[heavy zinc]

What does the refiner do, I forgot!

[crisp owl]

refines the image.

adds and refines minor minor details, is what it was designed to do.
Less needed as more finetunes have been released, but can still be used at your discretion

[steady grove]

have any of you 24 gb giants in here tried using this thing? apparantly the code downloads the weights

https://github.com/PixArt-alpha/PixArt-alpha

[heavy zinc]

So I can use the same model to refine?

[crisp owl]

This is the official refiner model
https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/blob/main/sd_xl_refiner_1.0_0.9vae.safetensors

[regal raptor]

hey guys, i read that it is recommended to have at least 32 go of ram for SDXL, would it work fine if i setup some pagefile from my ssd ?

[crisp owl]

Saying that, I often (mostly always) now use whatever model I want to do my refiner process.

Sometimes I want the look of Dynavision, but want it to finish out less animated, so will use a different realism model. And/or vice verse

make the process your own, experiment and see what you ultimately prefer

[vale eagle]

Their hf haven't upload the weight yet.

[steady grove]

ah. the code relies on the hugging face.

[hasty smelt]

Guys, do you have any Upscaler X2 for SDXL? Currently I only have X4.

[crisp owl]

for a straight 2x I just use the swift_srgan_2x

https://openmodeldb.info/models/2x-Swift-SRGAN

[heavy zinc]

Where do I put this file and how do I use it?

and What is the best Upscaler?

[crisp owl]

download it to your normal models folder where all your other models are, and then just select it for the refiner process.
Guessing you're using A1111, I haven't used A1111 since SDXL came out so I'm unsure where exactly that's all at

[heavy zinc]

so models/Stable Diffusion?

[crisp owl]

Yeah

[heavy zinc]

SDXL is on A1111? isn't it?

[hasty smelt]

Thanks, I'll try it, I'm working with images that don't need x4.

[crisp owl]

This is my favorite upscaler overall
https://openmodeldb.info/models/4x-NMKD-Siax-CX

[crisp owl]

sdxl is just a model, so yeah, if you use an sdxl model, then you're using sdxl

[heavy zinc]

I have two models

One called Realism Engine SDXL and RealVisXLV20

I think RealVisXL is based on the Realistic Vision model

[crisp owl]

I really like realvis personally

yes same guy that made realistic vision

[heavy zinc]

I just downloaded it now

any difference between the Xl and V5.1?

[crisp owl]

I tend to use ZavyChroma very often

SDXL is trained on these ratio's, 1MP sizes. Best to stick to these if you can

[heavy zinc]

I use 3:4 portrait mostly

its the best

So NMDK Is the best upscaler?

[glad grove]

its very good but theres not a "best upscaler"

[heavy zinc]

I have 4x-UltraSharp also

[crisp owl]

Yeah "best" is very subjective in all of this haha

[heavy zinc]

bBut its very slow

when I use Hires

[crisp owl]

It's why I said it's the one i prefer personally

[heavy zinc]

oh

There is Sampling Method, which one is most recommended?

[uncut fiber]

@crisp owl somebody said that DATs are very good upscalers. Have you try them?

[crisp owl]

Nope, I've meant to go on an upscaler journey, but just always have something else taking my time away lol

[uncut fiber]

🙂

[heavy zinc]

Anyone knows what is the recommended sampling method?

[crisp owl]

subjective like everything lol
I use either dpmpp_2m/Karras or dpmpp_3m_sde_gpu/exponential (don't think A1111 has that one yet)

[heavy zinc]

DPM++ SDE Karras

[uncut fiber]

i think it has it

[heavy zinc]

Yes iam using it

[crisp owl]

ah nice

[uncut fiber]

oh has not

probably on top

o.k. it has, i cant find sde 😄

[heavy zinc]

are there differences between them?

its on top of list

[crisp owl]

https://stable-diffusion-art.com/samplers/

missing a few, but a good read

[heavy zinc]

ok I will read

[hasty smelt]

Wow, my friend, what an insane speed this model is operating at! This was great! I dream of the day when the base model can do this with images.

[heavy zinc]

so the refiner file goes to Stable Diffusion folder and not the VAE?

[crisp owl]

correct

Unless A1111 is structured to read differently, but it's a model, so I'd figure where the other models are

[uncut fiber]

it is because other resize to 4x and then multiplied it 0,5 to get 2. This apparently just resize to 2x
You can as well try SwinIR 2x

[heavy zinc]

Okay

[hasty smelt]

Thanks

[heavy zinc]

What affects the processing speed/generation speed, the gpu, cpu, ram?

[crisp owl]

mostly gpu

[heavy zinc]

oh

how about ram

[glad grove]

only gpu matters if your gpu doesnt have enough memory then it will use shared video memory which would slow things even more

[crisp owl]

some but not a whole lot. Some things like models are just called from there, but doesn't do any processing

[heavy zinc]

I have RTX 3070 Ti

[uncut fiber]

if you have 16 you should be o.k. with 8 GB of RAM you will probably need higher page file

[heavy zinc]

what do you mean

[uncut fiber]

you asked about RAM, so i answered about RAM

oh sorry i wrotte VRAM 😦

[heavy zinc]

16 GB RAM is 8 VRAM=8 Threads

[uncut fiber]

???

if you have 16GB RAM, we are almost in same boat

[heavy zinc]

Yeah

it is equal to 8 VRAM, right?

[uncut fiber]

i think it is apples and oranges 🙂

[heavy zinc]

huh?

[glad grove]

idk what are u asking

[uncut fiber]

RAM and VRAM
Ram is way slower having not enough VRAM all process can be processed probably on RAM but will take ages. Therefore there are --medvram --lowvram switches.
BTW suggesting you using switch --medvram-sdxl @heavy zinc

[heavy zinc]

medvram sdxl?

what does it do?

[uncut fiber]

it is good for 8GB gpu cards. Ask if you should use it in #🤝｜tech-support

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimum-SDXL-Usage

[uncut fiber]

@heavy zinc depends on your sampler and number of samples

[modest pumice]

Is one able to run SDXL directly from the command line? I’m running An automated process from a GPU that I SSH into, so it’ll be ideal if I can run a generation with a single command. I’ve done this in previous versions but can’t find the txt2img.py file. Any help appreciated:)

[nimble heart]

Probably easiest to use the diffusers pipeline. You can make a really small python script that just forwards the args through argparse

works for 1.5, 2.1, and other diffusion models too

[indigo carbon]

Isn't diffusers pipeline not as efficient as Comfy though?

[nimble heart]

it's close. faster than auto for sure.

think comfy is still ahead on my card by like 10% from the subquad impl

on Nvidia with SDP it might be closer even

use torch autocast before you build the pipeline and it go zoom

[indigo carbon]

Also, Comfy will get more efficient very soon, AIT will be a built in thing in ComfyUI (including AMD)

[nimble heart]

but without autocast yea it crawls. like 5s/it vs 3it/s

I mean IDK how useful comfy is from a shell which was the question

if you really wanna minmax you can set up whatever speedhacks with diffusers. Pretty sure torch compile into tensorrt works for Nvidia

[nimble heart]

thought that was months ago and they gave up?

cause it wasn't seamless

[indigo carbon]

they gave up because the thing FizzleDorf and me(partially, just the compilation of the modules) was already pretty much seamless when it's deployed on supported hardware

[modest pumice]

thanks a lot. i'm experienced with py but much less so with sd. is this article a good place to start?

[indigo carbon]

but since new ComfyUI versions kill it, now it will be officially kept on life support

[nimble heart]

I've only used diffusers a few times but it's super easy. On the hugging face website they give you examples to get image gen working in like 3 lines of code.

basically just import pipeline, load model, prompt

[modest pumice]

https://huggingface.co/docs/diffusers/v0.21.0/en/using-diffusers/sdxl

[nimble heart]

yea and if you dive into it there's lots of functionality that most webuis have

[modest pumice]

perfect tysm man

[nimble heart]

yea I saw that new repo with a patch file

and the patch didn't apply cause I was a few commits ahead lol

[indigo carbon]

you were actually behind, it's not compatible yet because Comfy needs to do stuff on his end to make it work right away

[nimble heart]

thought the patch file was to bring it up?

[indigo carbon]

the AIT node will eventually become a built in node

[nimble heart]

that's literally what the description says

"apply this patch before it's integrated into ComfyUI"

[indigo carbon]

it won't work, I mean- you can try it..

[nimble heart]

regardless it's still missing a lot of shit in the new repo so I'm not going to try and make it work in amd anytime soon

[indigo carbon]

the only reason the repo will also stay as a custom node when built in is to also provide compilation scripts for AMD users and such

[nimble heart]

wonder what's stopping it from hot compiling

like Exllama does

does it take 10 minutes for a single kernel or something?

[indigo carbon]

it doesn't compile kernals

[nimble heart]

or whatever it does

[indigo carbon]

it compiles MSVC engines that are later used to load the models

[nimble heart]

msvc like Microsoft visual cpp?

sounds cursed as fuck

[indigo carbon]

and it can take even hours, the modules that have a range (the ones I uploaded to the repo that has the precompiled modules) took a few hours to compile

but if you are compiling for specific settings it should take a few minutes

[nimble heart]

so when you do a range it basically batch compiles for every possible combo of settings...

cursed 99

does tune affect it or only unet arch?

[indigo carbon]

that's the only way to make an optimization that'll work on any hardware and will be as fast as stuff like TRT and OneFlow

[nimble heart]

pretty sure one flow works with diffusers doesn't it

you should try it out

[indigo carbon]

OneFlow is not compatible with anything besides the most specific things it's meant for

[nimble heart]

damn. probably thinking something else

[indigo carbon]

the engines compiled with OneFlow aren't even architecture specific like the ones AIT can make and are even more painful to produce

[nimble heart]

what'd be extra baller is if AMD added their own backend to torch compile. then it'd be vendor agnostic

pretty sure they already have a graph compiling thing for that similar to cuda

[indigo carbon]

those engines include the model's weights and are checkpoint specific; not ideal. the engines made with AITemplate are architecture specific, so it's automatically a better optimization

also once the repo comes with prebuilt pip packages; the compilation scripts will just work ™️ on anything

[nimble heart]

™️

use :tm:

[indigo carbon]

heck, due to it being built in in future ComfyUI versions, it could just be included in ComfyUI itself

so then you would just use the node to make the engines, and you'll be able to use them right off the bat

[nimble heart]

maybe if you can multi thread the compilation it won't be too bad

[indigo carbon]

you'll just need to do it once for every single model architecture, it's not too bad

however if AITemplate would be as flexible as stuff like TRT, you'd need to compile for EVERY MODEL- and that's much worse

[modest pumice]

just got my first image there! thanks again

[half cedar]

(POV, close up portrait, cinematic lighting), (snobby anthropomorphic dolphin:1.3) wearing blue sweater and white sunglasses, pool lounge chair, (POV:1.3),1986, (extremely detailed depth, realistic color, saturated, deep blacks), style of Paul Kidby, highly detailed, style of comic book cover illustration, inspired by William Brodie,professional, clear, high contrast, high saturated, , vivid deep blacks, crystal clear

[crisp owl]

Would be cool to see how the AI see's reflections and draws them. Cause the reflections are often pretty spot on

[pallid path]

it guesses

[urban fjord]

I don't remember exactly, but doens't it just look at what it has done before and mirrors that.

[crisp owl]

No clue, but I'd be interested to know haha Just for the curiosity sake

[uncut fiber]

Here it failed a bit, columns reflections are perfect but... Probably not thinking about it as mirror?

[crisp owl]

Still a pretty decent job at capturing it. A bit wonky yeah, but the concept is there

[cobalt badge]

Hey guys, not sure if this is the right place for my question, but i'll give it a try: I want to recreate images with stable-diffusion-webui/SDXL, which were created using the BotChannels here. But it just produces junk / totally different images. I guess i am not using the correct model/settings or something else. Is there any information about which settings the Bots here use or any other way to reproduce the outputs locally?

[lusty wolf]

My white stage...

[celest egret]

is the refiner needed, or can I generate ok images without?

[hasty smelt]

Guys, the question might seem simplistic, but I'd like to know if I can delete the images in both the Output and Temp folders. I want to find out if deleting the contents can affect the generation of the next SDXL and ComfyUI content.

[indigo carbon]

I've blended a duck and a snowball.. LOL

[indigo carbon]

temp folder erases itself each time ComfyUI is started up

[steady grove]

https://huggingface.co/segmind/SSD-1B looks neat. the name confuses me though. distilled models in the past have been something different

[sinful falcon]

If anyone's looking for a good time, I nailed down a consistent comic book/graphic novel style I really love. Hope it's fun for someone else too! https://civitai.com/models/173569?modelVersionId=194889

[slender coral]

Hey lads, trying to get this to work in comfy https://civitai.com/models/23900/anylora-checkpoint?modelVersionId=95489 but unable to, I've read the civitai description but I don't see any info, am I missing something?

[nimble heart]

man even SD makes female fantasy races just different flavored sexy humans

worst trope

[hoary saddle]

agreed, 1 out of 10 got an actual female minotaur, rest just chicks

[nimble heart]

with those overfitted waifu models you can prompt animals and sprinkle in human anatomy words with a low weight and the overfitting will just bend the dragon or whatever into a human shape

might not work as much for pixel art tho

well i was gonna show an exampled but i guess i deleted it 🙃

[shy kelp]

how do i fix a lora doing this

[half cedar]

These are really great ^ nightmare fuel for me, would make good ecological propaganda

[slender coral]

propaganda

[upbeat summit]

the smell must be terrific!

[waxen galleon]

I enjoy the prompt generation on bing and would like to see that functionality added to SDXL.

[native knot]

I think I know what you're trying to convey, but can you expand on that? There's been quite a bit of discussion around that since its update.

I would imagine you mean how well it interprets the prompt.

[glad grove]

bing = good

[waxen galleon]

yes i do mean how well SDXL interprets the prompt. When using Bing it creates say a person sitting. I find that bing has a sentence structure that when I am using the same thing in SDXL i do not get the same results.

I feel very robotic adding alot of negative prompts using commas. Or for example with a seated subject having to put in the negative, kneeling, knelt, etc etc. i dont know how bing does it but somehow they pull it off.

I guess I like how on bing it feels like I am typing a story, and in SDXL it feels like I am just typing elements which feels more detached and makes me feel more like a programmer.

[native knot]

The reason that seems to make the most sense has to do with SDXL using CLIP and Bing's recent upgrade to DALL-E 3 using the GPT LLM; or at least that's what I've seen discussed, which makes sense to me. Now, implementing that locally? That's a whole other challenge. I don't think it'd be feasible locally. But there are some really knowledgeable people here who are deep in and know the underpinnings better than I do and I'm sure they could speak to it well.

[waxen galleon]

I think your right probably but im no expert. I do see smarter people than me saying that bing seems to send it to chat GPT or something to create the image. There is local chat GPT things right ? Maybe they just be an addon like the checkpoints? I would like to use the prompt style from bing in SDXL locally of course.

[native knot]

I think over time we'll get there; to me that seems inevitable. But it'll just be continued development both in software and hardware to get us there.

[waxen galleon]

I would also like to be able to get an image created in SDXL by another user load it into my pnginfo and have a list of their addons appear that it can download by checking a box.

[native knot]

If you move to ComfyUI and use Comfy Manager, you can do that with images created there.

[waxen galleon]

often when adding images to pnginfo with the metadata as a new user I am unable to tell what is a word and what is an addon.

[native knot]

It won't help with A1111 images, but it's a step in the right direction.

A lot of us here use Comfy.

[waxen galleon]

I am using A1111 so could this be better solved by building that into SDXL? that way we all can read it no matter who made it if it was made in sdxl?

I just followed a YT tutorial I had no reason to pick A1111 over anything else, i chose it arbitrarily not knowing the pros cons or even that other things exist.

[native knot]

It's not really about SDXL so much as it is the program wrapped around inserting the resulting image into whatever container it goes in.

[waxen galleon]

so what really is SDXL so i can better understand what I mean when i talk about it. Is it uhhh hmmmm i dont know what to compare it too, is it like windows, and A111 is a program?

[native knot]

You should check out ComfyUI. Don't get discouraged at first, just take it slow. Grab an SDXL workflow for ComfyUI, load it up, and begin experimenting. The other thing I recommend is start to read up on all the terminology so that you know what each setting really does...that helps a TON with tuning your image creation.

[glad grove]

theres a bunch of gpt models u can run locally but you need 💸 for it

[waxen galleon]

well I might do that next generation right now I cant get SDCL to do what I want either way. I can get bing too do it though which is aggravating since I have SDXL sitting here working just fine.

[waxen galleon]

I am willing to spend money once this whole hand thing gets fixed lol. then i fear maybe it will hit the normie sphere but maybe that isnt a bad thing.

[glad grove]

u could also hook gpt to SD so it "fixes" your prompt and sends it to SD but its useless since their bing integration is better

[native knot]

Stable Diffusion in general is image inference from models. The common model versions now are 1.5 and SDXL. XL can infer generally reliably for images sized about 1 megapixel, while 1.5 models infer at one quarter of that size.

[waxen galleon]

i am happy, very happy with the quality its amazing really! even with my junk 2070 to me it looks incredible.

[native knot]

It'll only get better as time goes on. Enjoy the ride. 🙂

[waxen galleon]

wait what is this? link it to SD? that might help solve my problem maybe I have been having in SDXL

[native knot]

It's not the same as what Bing does.

It's just using GPT to "enhance" the prompt.

[glad grove]

a small gpt model you could run it locally on 24gb vram so a 3090 or a 4090 would be good enough,if u want gpt 3 or 4 u would probably need a cluster of A100's thats like prob 50k$

[native knot]

But in the end, it's still just passing the verbiage to SDXL, which is interpreting it via CLIP.

[waxen galleon]

so am i using it right that the hands look wonky in SDXL because I use real low settings on purpose. if I dialed them up it would fix the hands? if it can id be interested in checking out combinining it with gpt.

[glad grove]

deformed hands are normal

doesnt matter if u prompt with gpt

[native knot]

Hands are always a challenge...eventually, add-ons will get better for SDXL with that, solving the issue.

[waxen galleon]

Ive seen some horrors on Bing as well, hands, extra legs, humans bent up like pretzels to fit in the frame, its bizarre. so it seems everyone is in the same boat for deformities.

so with SDXL will there be new versions, or does it just update the current version? for example would there be a SDXL 1.0 1.1 etc?

This DALLE on bing it uses its own code for the whole image? asking since when I have multiple subjects sometimes the front one is crystal clear, but the other ones looks like poorly rendered SD peoples faces.

[native knot]

It's possible. I don't know if anyone here actually knows if there are plans for that aside from those at Stable.

[glad grove]

you can fix faces with adetailer extension

[native knot]

(re: new versions)

[waxen galleon]

from what I have seen as a new user and outsider it seems to me that SDXL currently is superior to the other SD versions. That is my personal opinion only, speaking from seeing photorealistic images.

[glad grove]

both versions can create beautiful photrealistic images but sdxl is more coherent

[waxen galleon]

I better go thank you! just wanted to leave that feedback about the prompting method which I enjoy on bing and would like to see implemented. Computers will catch up if they need more computational power.

[native knot]

SD 1.5 has the advantage of being around longer and having lower requirements, therefore there are a lot of add-ons/extensions to give it better capabilities than it had out of the box. But give SDXL time and that will occur more and more. SDXL definitely has a better starting position.

[nimble heart]

👀

[vital ermine]

is that pytorch rocm or rocm repo?

[nimble heart]

yea

[vital ermine]

which one he asked?

[nimble heart]

https://github.com/ROCmSoftwarePlatform/pytorch/commit/ca5fbbc75d2f959d9d30012a7703294f1dbbcf5e

[vital ermine]

thank you

[half ivy]

Damn tattoos are getting good 👍

https://www.reddit.com/r/StableDiffusion/s/SVMb2R7t6x

[heavy zinc]

Hey guys is there any prompt generator which help you construct a prompt to generate proper images?

[lament karma]

Hello everyone,

I'm excited to share Fooocus-Control.
Fooocus-Control is a ⭐free⭐ image generating software (based on Fooocus , ControlNet ,👉SDXL , IP-Adapter , etc.).
Fooocus-Control adds more control to the original Fooocus software.

[nimble heart]

what's the word for those fucking michael bay black movie bar things so I can add it to negative

letterbox maybe

[indigo vine]

cinematic bars?

[noble shoal]

Letterbox is correct. Michael bay is incorrect. They would be exploding if they were from Michael Bay.

[nimble heart]

beyond the explosions his movies are super tonemapped and widescreen

or at least transformers was

[noble shoal]

Could it be that your latent offset node forces them if you have cinematic in the prompt and have it turned down to a negative value? Just assumptions here.

[nimble heart]

yea im using a 100% black latent which paired with underwater monster seems to pull from movie stills and results in black bars

turning it down to like 90% fixes but it's not as contrasty I guess

[noble shoal]

I love the nodes btw

[nimble heart]

memaid 💗

[nimble heart]

when it's not making IMAX letterboxing, 100% black can do some dope things

[hasty smelt]

Hey folks, are there any prompts besides 'brush strokes' that work effectively for emphasizing brushwork in painting?

[uncut fiber]

Try "thick paint", not sure it will work but probably will.