✨｜sdxl

[uncut fiber]

"Encaustic TEXTURE" adds plastics parts, but not like brush thing

[peak dove]

Try pentimento

Ebru

Impasto

[nimble heart]

think it's starting to have a bit too much going on for XL to handle

[noble shoal]

There can never be too much going on in a picture

[peak dove]

CLIP-ViT-H-14-laion2B-s32B-b79k.bin - must I d/load ALL those files ( https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/tree/main ) ? And where do I put them in comfyUI at all?

[noble shoal]

Aaaaaand...back to my old standards

[upbeat summit]

after making some myself I can attest to that burger sizes are really XL in this version of SD

[nimble heart]

if it's still not making backgrounds totally dark with a black latent a simple levels fix would help. Pixelbuster node is good for that, it's set up for levels already in the demo workflow.

[noble shoal]

The background was once black, but I did a noise injection before the upscale.

[west breach]

mmm burgers

[indigo carbon]

what do you need that for? I'm pretty sure this is one of the text encoders SDXL use

[uncut fiber]

Science fiction punk basicaly 🙂

[peak dove]

I found it thanks. Its a part of Scott Detweiler's IPAdapter setup. Anybody know where I can get an Image Batch Node for ComfyUI?

[uncut fiber]

i think it is called load image batch???

WAS Node Suite <- should be here

[peak dove]

My first results using ComfyUI IPAdapter - these 3 photos were input - no text prompt

And these were the SDXL results

I will get onto ControlNet later ... 🙂

[hoary saddle]

what's the method called when combining 2 images? can't recall but someone had a newer workflow for it using t2i, @indigo carbon maye?

[peak dove]

IPAdapter

[hoary saddle]

do you have an image with WF in it for this? i overwrote my previous one i think

[nimble heart]

if you're going for the most cursed workflow Olympics, my node pack has a custom ksampler node that can inject arbitrary latents partway through sampling

also Latentbuster

[noble shoal]

Fine, you got me at "cursed". I will check it out immediately.

[hasty smelt]

What upscaler you using? The image sound good.

[lapis gale]

just a fish

hippie family album

[uncut gull]

I'm actually just out back taking pics of my new holographic frosting. It looks like AI, so I'm posting it here.

[uncut gull]

This model's texture, promptfollowing, and text are awesome though.

[delicate kelp]

Hi everybody! I'm excited to announce the release of DynaVision XL 0.5.5.7. I've put a lot of work into this update, as the last update (0.5.4.3) kind of strayed from that magic that is DynaVision and pushed a little too hard towards realism. With this version, we've come back around to the Disney'esque charm that made DVXL so great to begin with! DVXL is now on it's own branch entirely as it has diverged 100% away from NVXL. This version I used multiple forward and back-merges between 0.5.4.3 and 0.3.7.1, working in a few new loras to help try to bring back the more exaggerated 3D cartoon look, while upping detail level and quality bolstered with some of my own training. Expect more detail, better hands, better coherence!

View the full changelog and download DynaVisionXL 0.5.5.7 on Civit!
https://civitai.com/models/122606?modelVersionId=198962

[bright valley]

Lookin' good

I just released my first LoRA today and I decided to try and tackle the hardest task in AI art, TEXT
So this is my logo LoRA. Pop in your text, colors, and style and it can pretty consistently do things.
I started small to even see if it worked, which it does, so I will be going back and stepping it up WAY farther, but for now:
https://civitai.com/models/176555

[noble shoal]

Why are you screaming at me?

Works fine

Awesome. I keep DynaVision in my repertoire.

[vast narwhal]

My favorite SDXL model is an old one, one of the firsts, it's called CristalClearXL but i can't find it anymore, i don't remember where i got, anyone knows?

[bright valley]

Ah a fellow jellyfisher in the wild, nice to meet you

[native knot]

How about "CrystalClearXL"?
https://civitai.com/models/122822/crystal-clear-xl

[vast narwhal]

Yes it's the one. THANKS. I really don't know why it doesn't appear when I search

[native knot]

Well, if you're looking on Civit, make sure A) you're spelling it correctly, and B) you don't have filters turned on that would hide it.

[nimble heart]

@vital ermine
https://github.com/ROCmSoftwarePlatform/pytorch/commit/306cb1ddabcbaa1b987c083fd466e964823e7ca4
using the new hip flash attention fns in pytorch

[hasty smelt]

Guys, is there any lora for SDXL that adds brush textures? I'm using a good prompt, but I don't think I can go any further than that. I would like to be able to have more control over the brush strokes like in digital paintings. Thanks

[supple knot]

@hasty smelt there's some on civitai site https://civitai.com/search/models?sortBy=models_v2&query=oil paint sdxl

[steady grove]

back to the future: Gundam

[remote jay]

A dick image on the moon as nazca lines.

[hasty smelt]

thanks buddy!

[glad grove]

i love Dick Cheney 🌚

[indigo carbon]

I've been holding that workflow back, it is seamless; but requires plenty of stuff

which is better?

[heady vale]

top one

[glad grove]

this one

[heady vale]

the silly lights on the chest and shoulders are almost unavoidable unfortunately with this style

[peak dove]

I love 543!!! 🙂

[vital ermine]

TY

[peak dove]

Dios de los Muertos #SDXL in ComfyUI IPAdapter

[noble shoal]

Prompt: BRO FREE GIFT DISCORD NITRO . Don't know what i have expected.

[sweet bane]

warning spider

[vale eagle]

https://civitai.com/models/177820

released my simple drawing SDXL lora

[noble shoal]

Since when does the Gigabot react to images? 🤔 Guess he liked it.

[turbid vigil]

anyone heard about this?

[vale eagle]

https://arxiv.org/abs/2302.04222

This paper

"pisons" bascially is an low donise img2img with other style that mixed up the image features. While training with that mixed up features, it would lead to model collapse.

[turbid vigil]

I'm guessing there's no way to detect those sort of images now ?

[vale eagle]

It might be detected by AI dector like https://hivemoderation.com/ai-generated-content-detection

Just tested. It can't

with very low percent of the "cloak"

[lusty wolf]

Light things up..

[dusk cape]

Anyone have suggestions for settings?

Trying to set up a workflow so I can render at various qualities

[inner garnet]

Hi, I have a couple of use cases for pixel art generation that I'm looking to fill- one is 8bit art for a gba game (FFTA) with these requirements:

• Palette control super important. Doesn't need to be perfect since I expect to need manual adjustments, but needs to be guided pretty heavily.
• Specific poses super important. The goal is generating characters that can fit with the pre-existing content.
• Img2img support is appreciated. Even if it has to be done 1 sprite at a time, that's ok.

I used SD a little at the start of the craze but haven't touched all the new tools since, I'd appreciate any alternatives that might work since the only options I see at the moment are:

• Train a lora for the style and use SDXL, I do have tons of character assets so it might be doable. I think I can pay to train one online?
• Without loras, use SDXL and hope that low noise img2img is enough?
• Buy Retro Diffussion and hope it happens to fit for the use cases. I don't mind the pricetag, but the gens I'm seeing in their discord look really samey- it however has 100% palette control and resolution isn't a problem, which is valuable.

I don't mind learning either Comfy or A1111, I'd just appreciate any guidance on how best to handle something like this where the targets are super-low resolution compared to the material the models are trained on.

[dusk cape]

use SDXL and hope that low noise img2img is enough?
I'm going to guess this won't work

or at least, if you put the noise low enough to work, you won't get very "different" looking arts

[indigo vine]

@zinc cargo nice lora 😉

[cyan crown]

Like these?

[noble shoal]

Oof, 5 minutes of inference

[upbeat summit]

these look great!

[indigo vine]

I saw yours on civitai and they are 👌 !

[zinc cargo]

thank you!
it turned out so well for you.

can i please ask you to post these on civit?

[peak dove]

I am blown away by SDXL, ComfyUI and IPAdapter!!!

[indigo vine]

done

[delicate kelp]

After Mandalorian ended, Grogu turned to street performances to make ends meet, though the lifestyle took a massive toll on him

[noble shoal]

Am i doing it right?

[shy kelp]

does anyone know a bit about how image models get trained

i have an idea

[indigo vine]

👀 👀 it would be way more expensive than gold

Copper Price
7,974.87+29.22+0.37%

Gold Price
1,985.60+5.63+0.28%

[uncut fiber]

You must have bad data. gold is 4x expensive on Nasdaq

[noble shoal]

Pffff, cheap ass ceramic 😅

[indigo vine]

found it here: https://markets.businessinsider.com/commodities/copper-price?op=1. What are those value then?

[uncut fiber]

it seems to me more like some sort of demand, that copper is much popular for some reason.

[noble shoal]

Demand will be through the roof when i have my copper bathroom business going.

[indigo vine]

ok, I got it. it's USD per Troy Ounce price

[indigo vine]

Gold Price Per 1 Kilogram 63745.29 USD
Copper Price Per 1 Kilogram 7.95 USD

so copper bathrooms are for losers

[uncut fiber]

@cyan crown @noble shoal it will be your Waterloo 🙂

[noble shoal]

Ok, so my Civitai Review will be: Cheap ass material. Do Gold next time.

Jeah, i put a few diamonds on this crappy material 😜

[uncut fiber]

got it its green oxide version as well?

[noble shoal]

Can't get it oxidized

[bright valley]

One of my friends made one of the coolest images I've ever seen with my new XL lora last night

[lusty wolf]

Don't keep us in suspense, show it... 🤦🏻‍♂️

[bright valley]

Sry it's taking forever to upload

I'm on mobile

Behold

[noble shoal]

Ohhhh, so you meant the coolest image made with your lora. And i thought with SDXL in general.

[bright valley]

Ohhh no I'm sorry yeah I meant with the use of my LoRA

I mean, still pretty awesome though right? 😁

Although I will be the first to admit, I'm noticeably biased, for obvious reasons.

[delicate kelp]

This is from my new CineVisionXL model I've been training. I'm reeeeeeally starting to dig the output.

[lusty wolf]

Copper is good for my hair...

[delicate kelp]

I recognize a quality @zinc cargo LoRA when I see one 😄

some movie posters

[static prawn]

which res to use for 1920x1080 wallpapers?

im not on a perfect pc

[cyan crown]

1344*768

and then upscale

[slender coral]

I don't really know where else to ask this, as no one answers these types of questions, but what exactly would be the purpose of a dataset?

https://ai.meta.com/datasets/segment-anything-downloads/

[pure crystal]

https://www.instagram.com/p/Cy3eZhkof94/?utm_source=ig_web_copy_link

immortal_gothic

:
● Reposted from @vampireoftheshire

● Tag us or use hashtag #immortalgothic or #immortal_gothic for a chance to be featured.

#immortalgothiconline #gothicwoman #gothicgirl #gothiclady #gothic #gotico #gothique #哥特 gothicsoul #gothicmood #gothicaesthetic #gothicvibes #gothicstyle #gothicbeauty gothicworld #gothicromance ilovegothic darkbeauty...

Likes

1284

Instagram

[dusk cape]

Datasets are used to train the AI

You show it millions of examples, and it learns to imitate them

[slender coral]

For the example I gave, what's your thoughts on training yourself? Should I even bother? Or just accept what's done already for me?

[dusk cape]

Just accept what's already done

If you want to get into training, start with something smaller, like a Lora

That will expose you to the concepts that are relevant, and you'll be able to decide for yourself whether you need to train a whole new model

[wet nacelle]

Does anyone know a way of prompting a low definition cctv footage?

[jolly brook]

not sure about prompting but I find this lora to be very useful for that

[wet nacelle]

Thanks for this. I'll mess with it.

[wet nacelle]

When you were using it what model did you use?

I'm trying to get it to work with base SDXL 1.0

[glad grove]

that lora is for 1.5 only

[wet nacelle]

ah

[zinc cargo]

when you give your wizard 18 STR

[indigo carbon]

idk why I decided to create a cubical duck

[delicate kelp]

New model I'm training

[shy kelp]

got my 4090

i am blown away how fast this is

[bright valley]

dude I wonder what this would look like with my LoRA, could probably make some dope movie title images

[delicate kelp]

it's with internal testers now. Will probably release it within the next week or so, depending on if I need to do any more training on it

[bright valley]

Sounds dope!

[verbal mulch]

Gave chatgpt ability to vision its thoughts in Stable Diffusion, asked it to represent itself

Redid again, i'm dying

[steady grove]

cause this is THILILERR

[verbal mulch]

asked chatgpt who is the most beautiful cat he saw in his life

looking for ideas for fun to know more about chatgpt lol

(questions to ask chatgpt)

will try this lol!

[noble shoal]

Little did he know

[heady vale]

hold still, this will only take a minute

[noble shoal]

Just going to drop this here. Thanks for the idea. https://civitai.com/models/178910/sdxl-cctv

[heady vale]

smile, youre on camera

[zinc cargo]

hi everyone 🙂
Installing comfy on a fresh windows here and being at it for too long, i dont even remember where to start.

is there a reccomended php version for windows 11?

[noble shoal]

Why don't you use the embedded version?

[heady vale]

I just use the portable version. Make a folder and unzip

[noble shoal]

Sorry, meant portable

[zinc cargo]

i'm goingn to have more SD tools so i better start at something, and i made it a habbit to keep each in its own venv

but honest to god, i dont remember the begining

[noble shoal]

Like the portable version?

[heady vale]

portable has everything in the folder you need

[zinc cargo]

is that what it is? a venv in a zip?

[indigo vine]

git clone the repo and launch run.bat file?

[noble shoal]

https://github.com/comfyanonymous/ComfyUI/releases/download/latest/ComfyUI_windows_portable_nvidia_cu121_or_cpu.7z

Unless you are on linux

[zinc cargo]

thanks

[noble shoal]

Just unpack and run. Done 🙂

[upbeat summit]

very nice! thanks for sharing it

[verbal mulch]

SDXL is incredible, I can't see any other way

[noble shoal]

Yeah, it's truly majestic

[rugged crow]

Was this you by any chance ?

[noble shoal]

One man? No way.

[rugged crow]

Ah! Oh well, thought it might have been you. Great stuff all the same.

[upbeat summit]

checking this out shortly. looks fun 🙂

[noble shoal]

It will absolutely smash your quality. Can't wait for the first complaints on civitai.

Output looks like shit. Doesn't work with my model...bla bla

[upbeat summit]

I love post processing effects and it is what I would expect from this. a reduction in quality for authenticity.

[west breach]

https://civitai.com/models/141592/pixelwave

Just posted new version of Pixelwave 04! Please check it out 🙏

[high escarp]

How are ppl generating a Videos in the bot?

[hoary saddle]

LMAO filmed on a potato. Will check this out later today. Good stuff

[turbid vigil]

are there major differences in sdxl prompt creation or is it more or less the same as SD 1.5 and such?

[pure crystal]

bananateer (banana skin person, banana rocketeer, bananacraft, bananacore, enzymatic browning,goggles:1.4)

[high skiff]

Wanted to share some of my favorite results from my Realism LoRA v1.5 hybrid compared to 3 other models

The following are all non-cherry picked.

The order of the results goes: top left (mine), top right (RealisticVisionXLV2), bottom left (RealismEngineXL), bottom right TDG general mix)

I feel my realism LoRA consistently provides better lighting, rendering, composition, colors, background fidelity, more accurate dynamic range, and better foreground/background separation

Please let me know what you guys think. 2.0 is in the works, and should HOPEFULLY be a huge improvement over what I have already shown here

Trained off a combined total of 210 images (90 images for first half, 120 images for the second)

Next version is gonna hopefully be the full 500

Some areas I have yet to even compare that should have some huge increases in quality should be drone shots and underwater photos, of which I purposely represented in my dataset

Also, please note that this version is not fully trained, and does take some hits on fine details consistency in some cases. That will be ironed out by the official release of 2.0 to the masses

(also, all of the prompts are just a single sentence describing the image, no key word/tag mashing for days)

Example prompt: a portrait photograph of a teddy bear sitting on a child's bed

[indigo vine]

👍

[noble shoal]

I like the Images from your LoRA. Nice work. 💪 But "wanted to share some of my favourite results" and "non-cherry picked" is a bit contradictory.

[high skiff]

as in, I have about 80 comparisons, and I wanted to share ones that I myself liked to show a range of different subjects

not cherry picked meaning, I gen it one time and whatever I gen is what I send (unless there is a fatal flaw across multiple like a cursed seed)

[noble shoal]

I am just screwing around. Will test the LoRA when it is released. Looks very promising.

[high skiff]

if I can check off the improvements I want with my team (I have some people writing whole new feature implementations for LoRA training to see if a dataset as high quality as mine can benefit), then I will likely have the best realism LoRA/finetune for SDXL there is

I am hoping for an improvement over my current one even bigger than my current one improves over the others

also, this LoRA upscales extremely well, and reliably as well

thats something I found that some SDXL models just can't do (biggest example is JuggernautXL. That model is trash with upscaling)

[noble shoal]

You might be onto something. You mentioned that you already train on very high resolution images, so you should be able to upscale it to the input image resolution without risking double things. Btw, I still use your upscale method. Haven't you been working on a new method?

[high skiff]

the new meothod is unfortunately unreliable and requires a lot of tinkering to get it working properly. It also likes to nuke certain models that are not as resistant to doubles

[noble shoal]

Noise injection?

[high skiff]

and as for the training, I am training at standard 1024x, but the images I use are like 4k-8k+

So the plans we have for future training are really promising

[high skiff]

kinda sorta, but more than that

its also unreliable in dark images

likes to always hallucinate things, but when it works on a golden sample image, its insanely good

[noble shoal]

Yeah, I know the struggles. You think you have it and then you prompt something else and everything falls apart.

[high skiff]

just a sec

[true geyser]

Hi everyone, I just started to use the baseXL model but when I generate an image it's having this color issue. What's the problem and how to fix it? I installed the refiner and VAE btw

After trying SDXL, I'm having the same issue with 1.5 too. I switched the VAE

[high skiff]

@noble shoalOk, I think comfy update and broke something for me, as my 2.0 high res fix is suddenly not working to 4x upscale without absolutely destroying my GPU. Even 2048x is showing 31GB VRAM usage which is not right. I was able to do 4096x on a 10GB GPU before, so IDK what is up here

but I was able to pull up a less ideal example of what my 2.0 high res fix can do when things work realtively correctly

these are just screenshots, but here is a base image I made with my 1.5 realism loRA

here is what that looks like zoomed in

and this is what my upscale at 4x will do (16x the pixel count)

this is by no means a fantastic example of what it can do at 4x, but its all I can seem to do right now until I fix this issue

ss of the 4x upscale

it really does help foliage massively

[high skiff]

ok yeah, this VRAM bug I am experiencing in comfy right now is making this all unusable

i can't work around generation times this horrible

I don't even know what would be the issue here

its like it doesn't realize I have 24GB VRAM. It hard caps everything at 16GB VRAM, making my 4096x upscaling I use to do in like 20 seconds take nearly 3 minutes to complete

and if I try to use GPU only like I normally do, it tries to use 31GB VRAM to upscale to 2048x, which is something I regulary did on a 10GB GPU, and I have friends who do the same fine on an 8GB GPU, so IDK what the hell is going on

[noble shoal]

Ok, who of you guys uploaded pictures of pregnant naked women in a changing room, captured on cctv onto my Civitai LoRA page? 🤣

[noble shoal]

Did you tried downgrading somehow?

[high skiff]

last time I did this was months ago, so I doubt I would easily be able to get my hands on something that old

[noble shoal]

You mean a comfy version this old?

[high skiff]

yeah

its not something I wanna mess with anyways

comfy UI is extremely volatile with anything VRAM related

[noble shoal]

oof...i need to check if i still have an old 7z on a stick, then i could upload it. Wasn't there a high vram flag?

[high skiff]

comfy told me that basically adding in any one node thats not stock has a high chance of nuking all of his VRAM optimizations and making them worse

[high skiff]

pretty sure that was appended in favor of just gpu-only, which is what I run daily

looks like all high VRAM does is keep the models in VRAM, which is not at all what I need

[noble shoal]

Hey, just saw that every release is still here: https://github.com/comfyanonymous/ComfyUI/releases

[high skiff]

mayeb at some time I will go back, but not at the moment. Just hoping to get this working without having to do a bunch of stuff

side note, anybody else been bleeding SDXL performance? Like has it been getting slower and slower for you lot as well?

When I first got this GPU, I was doing 4.5it/s @ 1024x1024 with DDIM at 100% power, and 3.3 it/s at 50% power

Now I get only 3.7it/s at 100%, and 1.9 at 50%

its definitely slowed down dramatically for me, that I know for sure

let me check my 1.5 speed while I am at it

1.5 speeds are down too

wonder if new NVIDIA drivers have been slowing our GPU's down for diffusion

[noble shoal]

Wasn't 531.79 the last unicorn?

[high skiff]

I think so, that was the one I was using for a while

that was also the one before the 3060 memory leak

[noble shoal]

I am still on it

[clever verge]

More recent works too

There were some issues in between though

[high skiff]

I am starting to think NVIDIA is slowing out GPU's down, cause this is a big loss of performance for me for no reason, and its been in the down as of late

[clever verge]

Last versions they've introduced functionally that should do the opposite but hey it needs software support for it to work.

[high skiff]

that tensor RT bullshit, yeah, I saw it

tensor RT sucks, even if it is even faster now

its a pain in the ass to use, very limited, its just worse AIT honestly

pretty sure AIT is a modified version of TRT

[noble shoal]

And even thats already a low bar

[high skiff]

I am pretty damn sure my diffusion slow down is NVIDIA driver related

[noble shoal]

Make the test with 531.79 and you will find out

[clever verge]

I have seen improvments but then I'm on the latest comfy, pytorch, kohya-ss and nvidia driver.

[high skiff]

I guess I can roll back and check

[noble shoal]

Would be quicker then roll back the comfy version

[high skiff]

yeah, and I doubt comfy would be the culprit

NVIDIA has a lot more reason to be slowing down their last gen GPU's

[clever verge]

As you said earlier there's a good risk some funky nodes messes things up

[high skiff]

for memory attention, sure, but I don't see how some installed nodes could eat over 40% of my perf in some cases

[clever verge]

you maybe have a more complex workflow now? 😄

[high skiff]

nope

I test them in the most bare bones workflows possible

tho I will say, I have seen no perf diff between simple and my normal workflow

[eternal fog]

😮 been waiting for this

https://huggingface.co/h94/IP-Adapter/blob/main/sdxl_models/ip-adapter-plus-face_sdxl_vit-h.bin

SDXL IPAdapter Face

[high skiff]

@clever verge@noble shoalrolled back, no difference

I am gonna restart my PC and bosot my GPU's VBIOS, tho I don't see that having a huge impact on perf

brb

[clever verge]

What were you running before downgrading?

[indigo carbon]

no, it's two different things. TRT keeps the weights and is checkpoint specific while AIT is architecture specific and doesn't keep the weights and is compatible with both AMD and NVIDIA

also TRT and AIT can theoretically work together at the same time. I know a guy called FizzleDorf that managed to pull that off; they got actual instant speeds. I have no idea how they did it, but they even sent screenshots

[indigo carbon]

I've just been using my old ComfyUI installation with PyTorch and Xformers built from source for a while and I didn't see any difference

[high skiff]

the absolute newest one

or well, I guess its a a week old

545.84

[indigo carbon]

I have two installations of ComfyUI; one is an installation from around a month ago and I use it for pure txt2img with AITemplate. the other is just for the new experimental stuff

[high skiff]

Ummm

My PC just blue screened

That's not fun

Something about the video scheduler

My PC may not appreciate going that far back in graphics drivers

[high skiff]

I will always be the king of the white tigers :p

[steady grove]

i got a few bluescreens this past month. not sure what started causing it. i went through and killed all overlays. nvidia overlay always causes it. videos in browsers seem to use GPU even when browser hardware acceleration is turned off too. So i gone into windows 11 gpu settings and turned browsers onto power savings plan and they'll use the integrated GPU

[indigo carbon]

literally prompted "white tiger in a jungle"

[steady grove]

i still get these warden blue screens though now and then, but doing those measures has lessened them a lot. python 3.10.11 seems to work most stable

[high skiff]

whatever your workflow is in insane, cause it makes your images look so much better than the model actually is haha

[kind horizon]

python-3.10.6

[high skiff]

wish i could see what sorcery you are doing, cause your mix 100% does not replicate results like that for me lol

[indigo carbon]

maybe you're using CLiP skip?

[high skiff]

nope

also, have one of my high res tigers

my high res fix 2.0 does really shine sometimes

[steady grove]

oops i had that wrong. DPC_WATCHDOG_VIOLATION was the bsod i keep getting

[indigo carbon]

my workflow is just text->style(typically base which is none)->text encode->Ksampler(using model loaded with AIT)->decode->rebatch->upscale

[steady grove]

it's always that one and only happens when i'm using python. never with games or 3dmark or other stress tests

[high skiff]

not sure man, your model doesn't give me results anywhere near as good as the ones you show

[steady grove]

whenever i flip AIT on my results go all to hell. anytime i try to diagnose i get a lot of gaslighting. i've stopped caring months ago

[high skiff]

my high res fix 2.0 can also fix some considerable facial deformations, tho the eyes still need a little work

[indigo carbon]

this is how my setup looks like when I'm cookin'. I usually do use a negative and sometimes CLiP_skip, but for the ones from a day or two ago I didn't use a negative prompt

[peak dove]

I am so impressed by ComfyUI's IPAdapter!!!

[high skiff]

didn't use a negative? the image shows a negative

[indigo carbon]

when I'm casually making stuff I use a negative, like in the screenshot

[high skiff]

ahhh, I get you

[indigo carbon]

I'm unsure what's so off-putting about that girl for me, the eyes are uncanny

[high skiff]

yeah, unfortunately, eyes are one of the things my LoRA has yet to full pick up on

[indigo carbon]

I think a ((bad eyes)) negative prompt would fix that, no?

[high skiff]

not likely

oh but that does remind me, I wanted to do a test

I wanna see what happens when I gen my negative prompt lmao

oh god

oh no lmao

ohnohnohnohnohnohno

[indigo carbon]

I also just tried that, it just makes ugly boobs

[high skiff]

I have a pretty hefty negative I use from time to time

omg, my realism LoRA made it WAY more palletable lol

@indigo carbonyour model

[indigo carbon]

ew wtf

[high skiff]

realistic vision

[indigo carbon]

what was your negative prompt?

[high skiff]

realism engine

(Ironically, thats a damn good hand lmao)

vs my LoRA lol

mine is for sure the most palletable lol

[indigo carbon]

it's abstract

my usual negative prompt is nsfw, painting, drawing, sketch, cartoon, manga, watermark, signature, label

so it just decided to make boobs with lables on them

[winged thorn]

I don’t think

[indigo carbon]

also, I heard if your negative prompt has things like ((bad quality)), worse quality, bad image in it; it will make the results worse

I usually just use nsfw, painting, drawing, sketch, cartoon, manga, watermark, signature, label and it easily makes stuff like this without a fat prompt or a style template

[nimble heart]

🤔

[indigo carbon]

tf is SSD1B?

[nimble heart]

exactly

[indigo carbon]

could be some unreleased SAI model Comfy didn't know was private or something

so the UI supports something which wasn't even released

[eternal fog]

https://fxtwitter.com/camenduru/status/1716817970255831414?s=20

ffs

That has a link to the SSD1B model

Looks like a distilled SD model

[indigo carbon]

SDXL is already fast though

well, when used with something like AITemplate

[nimble heart]

big "when"

might be nice to do the 1st pass in XL then hiresfix in ssd so it doesn't run at 7s/it

[indigo carbon]

Eventually diffusion will definitely be as fast as exLLaMa is for LLMs though. I saw FizzleDorf already achieve that by somehow using AITemplate and TensorRT at the same time

It's just that preparing it all in a simple environment and the flexibility of complex optimizations like TRT with more optimization layers or OneFlow is non existent. You would have to compile each model separately

[nimble heart]

yea the nice thing with the llm loaders is they basically just work™️

just select it from the dropdown in ooba

[eternal fog]

I need to figure out the best way to use that SDXL face IPAdapter, but it seems to work pretty well.

It picks up more than just the face if there's other stuff so I think I might need to mask off the images.

[indigo carbon]

I think a YoLo face detection model may come in handy for that

[eternal fog]

Yeah the impact pack can do it with face detailer, just trying to see if I can just do it standalone

[indigo carbon]

You would detect a face, then it would automatically crop around the face, then send to the IPAdapter node

[eternal fog]

Yeah

[indigo carbon]

Also, does this IPAdapter model also have image blending capabilities?

I'd assume it would be able to blend faces possibly

[eternal fog]

I mean you can use it like any other, but I'm not sure how you would blend 2 faces well.

I've tried running the FACE model into the Main Plus Model and it sort of works.

I've not tried blending 2 faces yet.

[indigo carbon]

What do you mean by "running the face model into the main plus model"?

Like chaining them together?

[eternal fog]

Yeah

I've tried 3 different ways so far.

1. Run the Main IPAdapter for X amount of Steps with Y left over, then run the face IPAdpater for the remaining Y steps.

2. Make an image and then img2img that with the Face IPAdapter

3. Run the Face IPAdapter chained with the Main IPAdapter

1 sort of works, but it either doesn't get enough of the face, or starts picking up other stuff in the face image and makes it go weird.

2. works well for the face, but it doesn't fit the face well with img2img so it starts to look a bit weird.

3. I'm not sure how to balance the weights properly, so it's doing strange things I'm not expecting.

[indigo carbon]

I think using FreeU with the chained method could be interesting

[eternal fog]

I'm using FreeU as well

[indigo carbon]

For me that resolved blending quite easily

[eternal fog]

Just messing around to see what gives the best affect and how I need to use the source images

You can see it in the previews concentrating on the face

But it picks up background things near the face as well and it can cause.... issues

[indigo carbon]

Why would it do that? Isn't it meant to just be zero-shot with the faces?

[eternal fog]

Not sure but it 100% picks up objects near the face. Maybe my weights are too high at the moment

[indigo carbon]

Well, nothing Photoshop nodes can't fix..

[eternal fog]

But for example I had a face that had christmas lights behind it

And it covered the person in lights

Even though they were only in the image fed into the Face IPAdapter

[indigo carbon]

You could have some kind of image photoshopping node to remove background from the face-focused image

[eternal fog]

It gets background colours as well. So I absoultely need to mask out the face.

I just need to find the right node for it.

As I say the FaceDetailer in the Impact Pack does it, but it also does other stuff, I just want the face. I might just grab a SAM node pack

It doesn't copy the input faces exactly, but it seems to be able to keep them reasonably consistent over multiple images

Very quick example

Face Input

Normal IPAdapter Input

And if I change the seed, the face is basically the same

So it's keeping the face, but the quality isn't fantastic and as you can see it's also pulling information like the bricks behind her.

Which then makes the image look really weird

Lowering the weight helps though, I need to play with it a bit more

As you were asking about mixing faces

With 1 face in the Face IPAdapter and 1 in the normal you get this

Which I think has worked pretty well to be fair, it's put the face into the source image

If I mix 2 faces into the face IPAdapter, it does very on the weight, but it does seem to try and blend the faces together

These two with some lower image weights on the photo did this

It seems to mess up backgrounds quite badly, the way I've got it setup now.

Hmm maybe this way works better.

First 10 steps you use the Main IPAdapter to build the image, and then remaining, in my case 25 steps, you run the Face IPAdapter

[eternal fog]

Yeah that seems to work better

Image without the IPAdapter face

Add in this face in the 2nd stage

Still looks a bit odd, but getting closer

[shy kelp]

dall-e 3 can create that Thanos snap effect that turns everything into ashes. SDXL can't, is this because SDXL wasn't trained with it?

[agile quarry]

dall-e directly or through a LLM that elaborates on your prompt for you? ChatGPT writes 4 prompts for each request unless you ask it to run your prompt unaltered

[uncut gull]

I haven't been able to upload to Discord from my PC in ages. Trying to get it working.

Success!

Who needs privacy when you have trendy architecture?

Giving text to ip-adapter gets you remarkably coherent text. (Not remotely what was on the ip-adapter image, but I gave it a long poem.)

[nimble heart]

adjusted Principled to be able to use arbitrary models so you can "refine" with whatever the fuck you want. Diffuse half the model with a photoreal and the other half with a toon model if you want.

[upbeat summit]

/r/lastimages/

[rose ether]

Travel Poster

Miami Beach poster

[wet nacelle]

I'd with her

[crisp owl]

Anything new in the past few days?
Been out of commission medically and finally coming to

[delicate kelp]

Hi Everybody, TGIF! I'm happy to announce the dual release of NightVisionXL v770 and ProtoVisionXL v630!

Both models have had a few rounds of fine tuning applied which boosts detail and coherence and makes the models more responsive to longer prompts than some of the other models out there. Full patch notes and sample images are all available on the listings, as always! Thanks and have a great weekend folks, happy prompting!!!

https://civitai.com/models/128607?modelVersionId=201643

https://civitai.com/models/125703/protovision-xl-high-fidelity-3d-photorealism-anime-hyperrealism-no-refiner-needed?modelVersionId=201514

[native knot]

The Transform...ish:

[native knot]

G.I. Guess So?

[native knot]

Masters of the Whoniverse?

Blundercats:

[native knot]

InspectYour Gadget

Teenage Mutant Knockoff Turtles:

[native knot]

Talespam:

[vale eagle]

yes, it is what dalle3 designed with

[agile quarry]

prompting Bing image creator lets you prompt directly. ChatGPT and Bing the chat bot prompt for you and alter your prompt

dalle3 was designed with better care taken to caption the dataset. "Improving Image Generation with Better Captions" is the name of Dall-e 3's research paper. the ChatGPT model is just GPT tuned to write prompts and access the dall-e model for you

[native knot]

She-Ra...ther be somewhere else.

Bratman:

[agile quarry]

" utilizing a LLM to "upsample" captions can be used to not only add missing
details, but also to disambiguate complex relationships which would be hard for a (relatively) small image
generation model to learn. The end result is that the model will often correctly render images that it would
have otherwise gotten wrong."
-- dall-3 paper

[supple knot]

@native knot can you make like a modern bad guy in the 1960s batman ,like bane or 2face

[native knot]

1960s version of Bane

[crisp owl]

Damn the 60's were bad ass

[vale eagle]

I don't think Bing image creator would be directly put your prompt into Dalle3 T5 text encoder. Using short prompt would decrease the capibility while the model trained with descriptive prompt. If Bing image creator didn't go through gpt, it would go through with Bing to transform your short prompt into descriptive prompt

[half ivy]

i am here!!
What did i miss? o0

[native knot]

1970s Gameshows

[agile quarry]

i havent used bing allot since subbing to Cgpt+ but i think you wrong the image creator doesnt alter your prompts at least not the diversity the LLM does . and Cgpt im pretty sure every image is made using the same seed either 500 or 5000 helps with consistency i think tough, they are different and bing is probably better 🙂 im not paying for Cgpt+ next month

" You asked me to create an image of a pig dressed as a hero, and I confirmed that I would try to create that. I did not provide any further elaboration on the prompt."
does bing even alter the prompt?

[vale eagle]

The upsample process could be different. But I still stand for my view. I have fine tuned SDXL with small descriptive caption dataset. While using longer training epoch, it decrease the capibility to handle short prompt.

upsample doesn't mean chat.

[agile quarry]

upsampling ....
me: Only Hope was left within her unbreakable house,
she remained under the lip of the jar, and did not
fly away

1 Photo of a detailed ancient Greek jar with intricate designs, slightly opened. The essence of evils like dark shadows and mist are escaping from the jar, while a luminous figure representing Hope remains inside, looking up towards the opening.

2 Illustration of a hand holding an ancient Greek jar. Dark wisps representing the evils of the world are being released into the air. At the bottom of the jar, a radiant figure of Hope shines brightly, contrasting with the darkness around her.

3 Watercolor painting of an ornate Greek jar on a stone pedestal in a dimly lit room. Dark ethereal forms, symbolizing the world's evils, rise and dissipate from the jar, but a glowing embodiment of Hope remains seated at the jar's base, casting a light.

4 Drawing of a cracked ancient Greek jar on a classical pedestal. The surroundings are shadowy, with tendrils of darkness emanating from the jar's opening. Deep inside, the silhouette of Hope is illuminated, offering a beacon of light amidst the gloom.

theres a link to the paper here https://openai.com/research/dall-e-3-system-card

[native knot]

Blacksploitation film shots:

[agile quarry]

bing just does what its told. also Cgpt's limiting images now 😦 BingFTW

[glad grove]

not if u tell it to create a somewhat lewd image

[nimble heart]

using a gradient of colored latents to encourage composition

[half cedar]

1980s office, high fashion toy designer, (POV:1.3),1986,

[native knot]

1970s Kung Fu films. Tossed a Bruce Lee in there.

[vale eagle]

I think you are right. Bing image creator seems wasn't do much upsampling compared with gpt version.

[nimble heart]

ringed blend ⛅

[shy kelp]

bing image creator

[native knot]

70s unusual Sports Illustrated covers:

[vale eagle]

Using lcm or ssd1b for the first pass to construct the coherent and using orthers lora, model or controlnet to get the fine details. It should able to improve speed with quality and control using comfyui

[shy kelp]

These are awesome! What a throwback for me. What model did you use? Or are you training?

[zinc cargo]

please change your nick to All_I_Can_eat 😛

[noble shoal]

Me at the all you can eat buffet

[zinc cargo]

stack`em

[peak dove]

Having fun with ComfyUI and IPAdapter SDXL model = SDXL_Base1.0_0.9VAE

[peak dove]

Wow!!!

[peak dove]

IPAdapter and ComfyUI SDXL - Something Surreal!

[native knot]

I believe everything I made for the 80s cartoon posts were done with Juggernaut V6. Basically, "80s <insert show name here> tv cartoon screenshot", and then sometimes I'll add to the prompt if I'm trying to grab something more specific. I'm also making sure to use a 4:3 ratio so that it's accurate to the screen size of the timeframe.

Juggernaut is, in my opinion, the most versatile checkpoint for XL by far. The work they did recently for V6 has pushed it to another level.

[peak dove]

I'm more of a Dynavision, Counterfeit, Unstable Diffusers kind of user ... not convinced by Juggernaut (yet!)

[native knot]

It's all art and subjective, and those checkpoints are also great. I use them at times, too.

[high skiff]

I managed to have one of the worst training runs for a LoRA I have ever seen lmao

[high skiff]

JuggernautXL is barred from any of my future realism model research as its completely non functional with upscaling, and I have also found some other weird behaviors I have not seen from other SDXL models in regards to latent spaces

[noble shoal]

Impossible, i had a 30 min training and got basically a NaN machine.

[high skiff]

I trained a combined total of 28,800 steps across 118 images for this new Realism LoRA trained for nearly 6 hours, and it's worse than base SDXL lmao

[noble shoal]

Ok, you won

[high skiff]

I have managed to reduct SDXL lmfao

[noble shoal]

Oh, i made that my hobby

[high skiff]

base SDXL vs my realism V2 candidate vs my realism 1.0 LoRA

its not actually worse in aesthetics, but it is worse in the sense that it doesn't let me have any control over anything

and somehow, 90% of the images it generates are of people sitting on the ground, when I have literally 0 images of anybody doing that in my dataset

me asking for my standard albino woman in a white room prompt, and it makes her sit always

I cannot generate her standing

[noble shoal]

SDXL_Sitting.safetensors confirmed. When release?

[west breach]

ubu

[high skiff]

man outsider a church

Also sitting for no reason

base vs 2.0 candidate vs 1.0

Mind you, this isn't even prompted for my 1.0 LoRA

base vs 1.0 vs 1.1 (hybrid mix of 1.0 and 2.0 candidate 1), vs 1.2 (hybrid of 1.0, 2.0 candidate 1, and 2.0 candidate 2)

its worse in every way lol

and if I take off the training wheels, lets see

[noble shoal]

I am feeling more and more lucky that i took the believable shit quality route with my trainings. I have no struggle with ultra sharp details or fancy composition.

[high skiff]

interesting

I would try going that way, but the whole point of my LoRA is to go for professional photographs lol

[noble shoal]

Looks like we have no competition. a selfie photo of a man in front of a church

[high skiff]

oh jesus you weren't kidding when you said shit quality lmao

[noble shoal]

At least he is not sitting

[high skiff]

ring doorbell LoRA haha

honestly, I could see a market for it lmao

[noble shoal]

It's on Civitai. It's made with the CCTV loRA

[high skiff]

ahhh lmao

uhh, you promoting?

[noble shoal]

Did he got hacked?

[high skiff]

not sure

@peak doveyou good?

[peak dove]

Hi, I'm good! You?

Should I not tell all about that freebie vector site? I shall remove if its wrong

[peak dove]

Hello?

[high skiff]

I don't know how the server would view that, potentially as an advertisement. We just weren't sure if you got hacked or not, so we wanted to know because there are a lot of links on discord that lead two people getting their accounts hacked

[peak dove]

OK, so its only Civit.ai which can be mentioned here?

I just thought that everybody's happy at a good quality freebie site 🙂

[noble shoal]

I think you can mention what you want, but it came so out of the blue, that it looked sus.

[peak dove]

OK, I will be more careful 🙂 Thanks for the heads up!

Vectorizer.ai - an excellent site which will - and for free - make all your photos/ai art into high quailty vector files. (® Certified account not hacked ®)

[shy kelp]

@high skiff

https://media.discordapp.net/attachments/100633018493837312/1081358928912330863/41231.gif

finally got one

[peak dove]

IPAdapter with prompt "chaos dustbowl okies dystopia nuclear winter"

[shy kelp]

the middle one in the center is really good

[peak dove]

Thank you 🙂

This setup,using Base 1.0 model

3 photo input in IPA uses a lot of memory

[high skiff]

uhhh, cool

I am not sure I even know who you are 😅

but nice

[shy kelp]

we talked about furry stuff

my old gpu broke so i haven't really been making much lately

[high skiff]

man, these two realism LoRA's I made hybridized together make some great realism shots

especially compared to base

maybe don't say that lmao

I'd prefer if you didn't

[shy kelp]

https://tenor.com/view/mads-mikkelsen-finger-heart-heart-gif-26006306

ok

[high skiff]

thanks haha

but yeah, I remember you

nice man!

[high skiff]

very nice

[high skiff]

alright, gonna setup another new realism 2.0 LoRA. I have some high homes for what this one can be like. Hoping to have some good style transfer with this

[velvet shore]

Sdxl is 🥶

[clever verge]

Too few images with people standing maybe?

[clever verge]

That are some outstanding images! Well done!

[zinc cargo]

who's coming to flex with me?
|https://civitai.com/models/177578/anthroid-pump-you-up
new lora that is

[high skiff]

oh no, it's for sure bad training, cause my previous ones with the same dataset was fine

I am redoing it and only 1/10th done its already working much better

[high skiff]

NEW LORA TEST IS A BIG SUCCESS ALREADY

only 20% done

less than 1/30th the time of the previous one that came out terrible

30% just saved, testing

hmm

[ionic dragon]

@high skiff i trained an llm lora to generate sd prompts, wanna try it?

[high skiff]

I have messed with that a lot, and I also have a specific prompting style that probably won't translate well, but maybe some other time

[ionic dragon]

messed with??

[high skiff]

using LLM's to prompt

[ionic dragon]

i dont think there's one which does what mine does atm

it works something like this

[hoary saddle]

Cool. I set up SD llm on one of my 4090 24gb systems and the responses are slow to generate. How did you set yours up and how responsive is it?

[ionic dragon]

idk ask on ooba server, i just load them, ig your ooba is broken

[hoary saddle]

Ooba?

[ionic dragon]

oobabooga

[vital ermine]

https://civitai.com/models/180637?modelVersionId=202725

[ionic dragon]

@uncut steeple @smoky patrol

[steady grove]

@wicked frigate only cause it's been there for half hour and you look online. virus links shouldn't hang aorund this long and i think these could potentially be bot removed

roles don't seem pingable and theres no way to report stuff i can find anymore. used to be a script i thought

[wicked frigate]

cleaned up most of it and poked in mod chat about it

guess most the team ain't around on a saturday morning

[uncut steeple]

It's Saturday evening

[peak dove]

IPAdapter in SDXL (ComfyUI) using MidJourney originals - model = SDXLNiji_V51

[uncut steeple]

Sick dude, love that style

[mellow dome]

have anyone try this?
https://huggingface.co/segmind/SSD-1B

https://www.reddit.com/r/StableDiffusion/comments/17ig80u/comfy_ui_now_supports_ssd1b_upto_70_speed_up_on/

[indigo carbon]

you can easily get faster speeds than that using AITemplate if you have a 4090

[mellow dome]

i dont have 4090

[mellow dome]

can i get better speed if i use that model

[indigo carbon]

this is the speed I'm used to with small models like that

that's 31.5it/s on SD1.5

that optimization also works on SDXL, and its usually faster than TRT also (not to mention there is no degradation)

[mellow dome]

thanks I'll search it up

https://github.com/FizzleDorf/ComfyUI-AIT 🤓

[indigo carbon]

yep, that's the one

I actually know the guy who made that

also I'm the one that provided the modules for SDXL

in total it takes a little over ~30s to do batch 4 like this

[half cedar]

I would have to see sun conure to be sure 🤔

[maiden gale]

does anyone know how to reproduce "my prompt is more importan" "controlnet is more important" options in comfy?

[hoary saddle]

did you use the model from here?https://huggingface.co/h94/IP-Adapter/tree/main/sdxl_models

getting errors with that ipadapter model

[sand pier]

are you 100% sure the clip vision you have is the sdxl one?

wait, that vit-h is supposed to be used with the 1.5 encoder.

[hoary saddle]

hmm, says sdxl in the name, will look for a different adapter model then

[wet nacelle]

HELP HIM!!!

[stone fossil]

He lost his remote? 😄

[wet nacelle]

YES!!!

https://tenor.com/view/scary-encounters-stairs-scre-funny-meme-gif-14577420

[hoary saddle]

that was it, thanks Kaku

[high skiff]

@clever verge@noble shoalnew LoRA with much better settings is done now

and its looking quite promising

[clever verge]

It's take a while to build confidence in the results.

[high skiff]

hell, all three of my Realism LoRA's right now are just looking good

[clever verge]

I took my one of my LoRA's out for something it was not intended for.

[high skiff]

realism 1.0, 1.5 hybrid, 2.0 candidate

[clever verge]

They all look good yes

Nice analogue look

[high skiff]

I didn't even include analog photos in the dataset, but it can infer

[clever verge]

Yeah, if there are enough other connections to analog photos it can happen

[hasty smelt]

Guys, does anyone have a good negative prompt for flowers?

[bright valley]

Speaking of new LoRA, I've made huge strides in generating text with SDXL with my new LoRA, Harrlogos. Here's a couple I've made today with it:

Here's the link if anybody wants to check it out https://civitai.com/models/176555/harrlogos?modelVersionId=201745

[mellow tendon]

Juggernaut XL v6 is knocking my socks off! especially for portraits (you still have to use Adetailer for the eyes though)
https://civitai.com/models/133005/juggernaut-xl

[glad grove]

looking alright

[wet nacelle]

Do you know if I can just install the model and stick it into ComfyUI? Do I need new nodes for this to work in regards to the speed? Do I need any specific files alongside the model file?

I found the details to be a bit confusing.

[glad grove]

it works on a1111 just drop it on model folder and done,so i assume its the same for comfy since its just some sort of disstiled model

[mellow dome]

yes, looks like it

[wet nacelle]

@glad grove@mellow domehave you both used it to what you think is the correct way?

[mellow dome]

man does anyone know how to use ai template: https://github.com/FizzleDorf/ComfyUI-AIT
the patch looks outdated 😭

[glad grove]

i just dropped it on models folder and selected it like a normal model on a1111,tbh i saw no difference in quality

[wet nacelle]

In speed though?

[glad grove]

for me it was at least 50% faster

[wet nacelle]

alright kool

I'll give it a try

[mellow dome]

did u try the ai template thing

[glad grove]

no cause i dont have an RTX

[wet nacelle]

I'll try to take a look at it

[mellow dome]

I just tried the SSD01B, it did increase the generation from 22 sec to 13 sec for me, which is insane

[wet nacelle]

@glad groveHey I keep getting this error.

@mellow dome

[glad grove]

update comfy i think they added support for it today

[wet nacelle]

FUCK!

k

I got it working

Thank you for the help @glad grove

[wet nacelle]

@glad groveWhat is your sampler and scheduler when using it?

[glad grove]

dpm 2m karras

[wet nacelle]

Thanks

[vital ermine]

https://civitai.com/models/181012?modelVersionId=203145

[upbeat summit]

congrats to the release. thanks for sharing it. looks great!

[vital ermine]

Yeah, side by side on base there is no comparison so I can only image on a FT model.

[nimble heart]

UIs still need to be patched for it because it's got some layout changes to unet

but beyond that it's drop-in yes

not worth the huge quality loss imo

[glad grove]

yea and its only base model so meh i deleted it

[nimble heart]

i'd say if someone's not using XL at all because of the perf, ssd1b is a lot better than 1.5

but if you're on 8GB+ and dont have issues running XL it's a bit too crunchy

[glad grove]

i mostly do anime so i stay on 1.5,only switch when i want a very coherent realistic img

[glad grove]

elmo in guantanamo

[indigo carbon]

I personally use an old version of ComfyUI and an old version of the node due to the latest version of ComfyUI often being broken; but I'm pretty sure the latest version of the node may be compatible (I haven't tested that)

[mellow dome]

I follow its readme and try to apply the patch but it does not work since comfy ui code is constantly changing

could be useful if we know which version the extension targeted

[indigo carbon]

huh, I just have 2 ComfyUI installations; one is an old commit with the old AITemplate node, and the other one I use for stuff like IPAdapter

[mellow dome]

this is the error i getting

[indigo carbon]

yeah, probably Comfy not keeping the node on life support like they said they would eventually

so you'd be better off having a separate installation for AITemplate at the moment

[verbal mulch]

is chatgpt providing good advices on how to prompt well in stable diffusion?

[nimble heart]

git log in comfyui and look for the commit that came right before the patch file is dated then git checkout so you can apply it

you dont really need to keep a separate install since git tracks everything, so to switch back you can just git checkout master

save the patch commits to a branch will make switching faster too

[wet glacier]

@nimble heart what did change in the latest version of your principled sampler? the outputs I generated with the previous version aren't compatible. It seems that the results are quite different if I use all the same settings.

[nimble heart]

if you know the commit/date you're coming from I could say more clearly but beyond that it's a sort of "it depends"

[wet glacier]

on top of my head it's the one just before your latest

[nimble heart]

if you're using the full workflow I fixed a bug in Pixelbuster where I guess comfy updated at some point so it had the wrong gamma for a few months till just now

just today I started a refactor of the conditioning for future work. There was a few janky commits with the wrong values I think so depending on what time of day you pulled it might be off? I think I fixed it now tho

[wet glacier]

it's no big deal though, I was just curious as to why my outputs from the day before weren't reproduceable anymore

[nimble heart]

comfy does that sometimes too if you generate 500 images it starts to drift almost compared to if you restart

might just be a Pytorch thing tbh

pull latest where I'm pretty sure the conditioning is at least to the point of working as before and if it's still wayyy off I can look into it more tomorrow

[wet glacier]

I was getting this when dragging the outputs from the day before

[nimble heart]

some breakage was necessary but directly here I should be able to add some basic parsing for things like BREAK/AND from your prompt

[wet glacier]

will do

[nimble heart]

ohhhhh

with the conditioning refactor

it works on non-XL models

so I renamed it principled sampler

instead of principled XL

[wet glacier]

ah!

[nimble heart]

now you can use any model as a base or refiner

can have real vision 1.5 base with anything v3 refiner

just cause why not

[wet glacier]

so in essence I'd have to rewire the samplers and it should work

[nimble heart]

don't even have to. the params are the same

so if you open the json in a text editor

change the PrincipledSDXL to PrincipledSampler

and it should just work™️

[wet glacier]

perfect, thank you so much

[nimble heart]

you can do that for changed params too fyi. if a param gets added/deleted you can just edit the widgets_values field to add/remove the new param value

works on other addon packs too

[wet glacier]

is that on your principled.py?

[nimble heart]

but yea I thought you meant the images looked different in which case besides the Pixelbuster change and the hour of time when conditioning was wonk I'm not sure

[nimble heart]

widgets_values is a field in the .json

[wet glacier]

got it

[nimble heart]

so if say I remove the "refiner_amount: 0.15" field

you can go to the list and delete the 0.15 out of the widgets_values list in the json

and it should just work™️

what I do 90% of the time to keep my example workflow updated

[wet glacier]

sounds good

[nimble heart]

you might be able to ctrl-shift-v to paste a node while retaining connections and have it update it's params but I haven't tried

anyway imma sleep @ me if something else comes up so I see it tomorrow

[wet glacier]

thank you gn

[half cedar]

Try a karate riot

[cyan crown]

can you share your prompt?

[noble shoal]

Uhm...here you go

[cyan crown]

so that is CCTV Lora?

[noble shoal]

Yeah. I have one on Civitai

[mellow dome]

@indigo carbon took some time to set it up lol, but no idea how to use the node lol

[indigo carbon]

oh, that's.. different. the old version of the node selects the right node on its own, but it seems as if now you have to select the right module.

I've provided modules that have a generation range, you can try those

if you're on windows, give these a try: https://huggingface.co/Fizzledorf/AITemplateXL/tree/main/modules/windows/sm80/bs4/1024

those modules are 64-2048, bs1-bs4

use the module with xlr in the name for the refiner, and the module with xl for base

I have no idea how efficient this version of the node is compared to the old ones; but it should still be miles faster than pure PyTorch

[cyan crown]

Thanks will try it

[mellow dome]

is this how I suppose to use the node ?

[indigo carbon]

what's the error message?

[indigo carbon]

I recognize that error, usually version mismatch

I can provide the versions I personally use; but keep in mind it must use an old ComfyUI commit and an old version of the node.

[mellow dome]

by old version of the node , which node you mean ? the ait loader ?

[indigo carbon]

I use this specific version: https://github.com/FizzleDorf/AIT/tree/acd1d80c52bc0713f8cc8e2f59fd50e1adb47ec0

with this specific ComfyUI commit: https://github.com/comfyanonymous/ComfyUI/tree/bc76b3829f5fbba7c5a439c7833d313a3ca87398

[mellow dome]

whoa lemme try, thanks for sharing !

[indigo carbon]

those two probably can't work without each other, so I personally have a separate ComfyUI installation with this commit that I use for AIT

[wet nacelle]

Great work on the stuff that you're doing.

My only goal in stable is to make the best hyper realism that I can. I'm always shocked by the anything but realism outputs that you make.

I can actually see a goal being excecated within your outputs. Almost as if it's telling a short story or saying a message.

[stone fossil]

I am doing a finetune so standby. 🙂

And it is just one of the things it can do ofc just playing with this now as it seems ok.

[lusty wolf]

Nodes to watermark my images directly.

[static prawn]

someone know why some of my images tend to get overexposed, too sharp to be realistic?

like here it feels like there is tooo much contrast

it still looks ok thought, but i miss that photography aspect

[crystal flame]

Avril Lavigne showing her curvy body, shapely ass, wide hips

[peak dove]

I d/loaded via ComfyUI Manager. If I rememher, I will check for the Huggingface page when I get home later

[wet nacelle]

https://tenor.com/view/skull-issues-gif-13031152103567454559

you fucking bastard

[turbid vigil]

this is really sick

nice one

what was the prompt for that?

[lusty wolf]

Spidey never gets old, for me anyway...

[fringe barn]

Are you guys able to get A1111 to do upscaling in Img2Img with SDXL?

[half cedar]

@wet nacelle 1993 game show sound stage, nickelodeon, production still

[cyan crown]

yes

[wet nacelle]

overall the image is awful, but at least the headline is fantastic.

[bright valley]

blown away by the stuff people are making with my LoRA

and not just with my name 😅

[south horizon]

text workin' pretty well

[bright valley]

Finally!

Achievement unlocked

[mellow tendon]

I have been working on some scary Vampires without needing a Lora (not that I have found a good SDXL Lora anyway).

[mellow tendon]

Oh someone did release one a few days ok https://civitai.com/models/178275/vampirexl I will have to check it out, but prefer not to use Loras ATM as I am loving the speed of TensorRT.

[heady vale]

I updated comfy and quickly reverted. Every image I previously generated was outputting a different image. Something wrong with seed gen