✨｜sdxl

[native knot]

Definitely keep everyone updated if you do find a way that works. Funny thing is that the page shows pictures as if there's a way for it to work with more than 1 image.

[soft zealot]

and for the next seed it rewote the initial as "A photo of a translucent, hexagonal dome tent with holographic patterns, set up amidst lush greenery and towering trees, showcasing its advanced design and eco-friendly materials."

[cedar tusk]

ah wait, can't combine clip visions, just conditioning

hmm

[eternal fog]

Although, the guy said he just implemented it in the node - https://github.com/laksjdjf/IPAdapter-ComfyUI/issues/11

[native knot]

Does anyone know of a node that displays the seed for the current image instead of the one that's used for the next seed?

[cedar tusk]

ha, but what does that mean

[soft zealot]

I sue Seed to Text to add ro my prompts.txt and also use in thr filename

[eternal fog]

translated from the readme

Multiple conditions.

Multiple images can be input by connecting nodes naturally. By combining with Mask, you can do something like separate conditioning for left and right.

The problem is that the background is also divided ^ ^;

[cedar tusk]

ok, so just chain them

[native knot]

Yeah, I would just like a text box node that has it in a spot I can display it in my workflow.

[cedar tusk]

neat

[eternal fog]

Yeah that didn't work before. Not sure what it will do. I'll give it a try again later.

[soft zealot]

@native knot so my image save filename is

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

[eternal fog]

Previously it just ignored every adapter but the last one you used.

[native knot]

Oh man...so if IPA has been updated to work with multiple images, I know what I'm playing with next.

[soft zealot]

I wonder if Tiny Terra Nodes textdebug node would do that

[soft zealot]

@native knot looks like it will

But.... you will also have to switch to SSeed With Text unles.....................

[native knot]

I'll poke at that in a bit. Thanks for looking. 🙂

[cedar tusk]

Well, chained IP-adapters seems to be too much for my poor VRAM

Getting black images only

[sand pier]

I wish Comfy had a bit better component for loras, stacking 4-5 loras together to create a image is a bit painful

[native knot]

Someone here has a LORA loader node that makes it way easier...let me see if I can find it if nobody beats me to it.

[sand pier]

that'd be great

[soft zealot]

ComfyRoll

[eternal fog]

How much VRAM do you have, seems to be working for me.

[cedar tusk]

8GB (2070 super)

[native knot]

The CR one is one of them...there was another that had an even bigger stack.

[eternal fog]

Ah I've got 10GB, and it was using around 8GB

[cedar tusk]

Lol just missed it

Works great with SD 1.5 though

[eternal fog]

I wonder the best way to use it

[noble shoal]

https://github.com/picturesonpictures/comfy_PoP

[eternal fog]

As I've been feeding the IPAdapter through Revision

[sand pier]

Looks way cleaner at least

[native knot]

That's the one...does like 10 LORAs

[soft zealot]

I just have 4 of these setup in line

[sand pier]

oh lord, thanks hahaha, ill have a look

[sand pier]

yeah same, it's just annoying having to mute, bypass etc.

[soft zealot]

just set to Off, they have switches (ComfyRoll again)

no muting/bypassing needed

[sand pier]

ah comfyroll, not regular one

yeah that's more useful

[soft zealot]

I raely use 1 lora let alone multiple of them but its there just in case

[cedar tusk]

neat!

[sand pier]

i love loras hahaha

they be handy

still use 1.5 a lot thanks to them

[thorny frost]

One of the things I like more about a1111 is the model card panels and Civitai helper. Would be nice to have something like that in comfy

[native knot]

SDXL has pretty much cut my everyday need for LORAs down to almost zero.

[cedar tusk]

Adapter has been to me as big a hit as LORAs were

[soft zealot]

build one ;o)

[cedar tusk]

Like any image is a LORA now, it feels

[sand pier]

ive not played around with adapters yet

probably should

[eternal fog]

I've got way too many things to feed in now lol, I've got 2 IPAdpaters, and also 2 Revision images.

[cedar tusk]

HA, I know how you feel

[cedar tusk]

What do you mean by Revision images?

[native knot]

My 3 biggest things with SDXL are AIT for performance, ControlNet, and IPAdapter. If they can get around to making ControlNet and AIT work together...

[cedar tusk]

Like second pass?

[eternal fog]

Revision/Clip vision G

[cedar tusk]

Got a link?

[soft zealot]

Using 21 or more images that get interrogasted to produce a clip encioded output instead of using a text iput

[cedar tusk]

controlnet color + adapter has been great to retain some of the info that gets lost

[cedar tusk]

waaaaaaaat

[eternal fog]

https://huggingface.co/stabilityai/control-lora

Has info on this page.

[soft zealot]

See https://huggingface.co/stabilityai/control-lora & https://www.youtube.com/watch?v=1SwDCqgXZ-M for an explanation

[cedar tusk]

oh right

for some reason my breain TOTALLY blanked on that

[eternal fog]

It's kinda similar to IPAdapter

But it's much more loose

It's "conceptual" rather than "copy" of the image style

[cedar tusk]

Is revision avaialble for 1.5 or is that just in the new SDXL group?

[thorny frost]

It is like mixing two prompts.

[eternal fog]

I believe it uses the SDXL Text Encoder, so SDXL only

I might be wrong though

[cedar tusk]

The number of times that stable diffusion has made me almost buy a new card

not a healthy number of times

[ionic gulch]

new version 5.7 of my workflow - this time ip adapter nodes got changed by its author (he combined two nodes into one) and broke my workflow - also did a lot of cleanup
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[strong field]

Just need to compile … lol

[eternal fog]

oh no lmao, You can now feed the clip vision output through the IPAdapters into Revision, which gets you different images again.

[cedar tusk]

you'll never be free haha

[eternal fog]

I'm not sure exactly what the technical difference is though

[native knot]

Worked like a charm. Thanks again!

[eternal fog]

I guess it's mixing the images differently

[strong field]

You can composite latents, so do the background separate and composite the characters on top of the background

[cedar tusk]

Revision outputs conditioning, right? So you can feed one revision into the _g and another into the _l to affect style versus content (sort of) right?

[eternal fog]

I think it only gets put into G as that's what the clip vision model is based on

Hmm feeding 2 image through IPAdapter isn't deterministic. You get different results every time, even with the exact same settings.

These were all the exact same settings

[strong field]

Deterministic when you blend the images first, use blend image node to set the balance

[sand pier]

is this the comfy extention you guys use for IPAdapter or? https://github.com/laksjdjf/IPAdapter-ComfyUI

[strong field]

I have had very good results with both IPA and Revision like that

[eternal fog]

why would you blend the images when you are wanting stuff out of both of them, just randomly blending images together and then feeding them I don't see how that would do what I wanted.

[strong field]

What are you trying to achieve?

[cedar tusk]

IPA has been updated to allow you to chain them together

thus allowing two image inputs

[eternal fog]

Just trying to work out how this is supposed to work really.

The Idea I have is get an image of say a person and than another image of a style and put them both through their own IPAdapters. It seems to work, but it gives different results every time.

[strong field]

That’s neat that they implemented native support for it

[cedar tusk]

It looks like the adapter is using cpu. Maybe it's getting internal randomness from that?

ip_state_dict is doing a torch.load() with map_location='cpu', but elsehwere in the code you see device as cuda. Might be related to non deterministic behavior 🤷‍♂️

[eternal fog]

I might post it as an issue on the Github page and see what the guy says

[cedar tusk]

Good idea, because that's above my pay grade. I can make neat little nodes but the actual diffusion logic flies way above my head

[strong field]

I’m not sure you can achieve this with just IPA, you will need a way to constrain the latent space (after encode) or the image space (harder from my experience)

[eternal fog]

It seems to work fine doing it, it's just not the same every time.

[strong field]

Ah ok

[cedar tusk]

Which sampler do you use? Some of the samplers themselves are non-deterministic

Though probably not to that extent

[eternal fog]

They aren't usually that bad, but I'll have a check in a bit

[cedar tusk]

DDIM might be worth checking, as that's a solution-based sampler right? No noise addition each iteration

[eternal fog]

ok yeah DDIM it stays pretty much the same

Right ok, looks like it might be because I was using the "GPU" dpmpp_3m_sde sampler, using the none GPU one it's still different but a lot closer.

[soft zealot]

Jumping back a bit @native knot

side by side comparison
L to R

No ChatGPT assist no Supplementary
No CHatGPT assist with supplementary
CHatGPT Assist + Supplementary
(all same seed)

[eternal fog]

@cedar tusk I think I'm getting a bug I was getting yesterday when sometimes it randomly decides it's not going to listen to the IPAdapter at all

Yeah I think it's doing some really strange stuff with how it's pulling the data through

[cedar tusk]

I just noticed that without changing any settings and keeping a fixed seed, Comfy dosn't recognize that it's already been calculated and re-runs the workflow

[eternal fog]

Mine isn't doing that

[cedar tusk]

Might be my image load node

[tired moon]

hello how can i use the sdlx patch on nocrypt colab? just load the models i want use?

[strong field]

you got an image of your latest workflow, been wanting to check out your ChatGPT assist

[soft zealot]

Probably not very useful like this as you cant see what connects where lol, would need to pull it out to see clearly and trun spaghetti back on

Or you could just load this young ladyt and try & work it out lol

[strong field]

ill pul lit apart hahah no owrries

thats the fun of it, "what the hell was winston thinking?!"

[soft zealot]

hehehehehehe 🙂

If it helps this is a simpler overview of just the switching elements with the final output onthe right being what you would feed into yuor conditioner Positive

[strong field]

Many thanks!

[soft zealot]

thing is after I get it working I rearrange everything to make it useable rather than flow

and even then I have the Style Warapping missed out from that lol

[strong field]

haha, i like to build from scratch ALOT, mosty testing, my usual workflows have tons of different tried and true methods, no organization

[soft zealot]

this helicopter view makes sense to me

Green are the daily driver zones (which get loaded in taht view when I open COnfyUI
Purple Are supplementary areas I may use on a semi daily basis
Blue are workzones that I dont need to touch once built
Yellow is old shit left in because..............

This is my "extended" daily view although some of the text can be difficult to read

[strong field]

Very nice, i see you have improved it substantially hahaha

[soft zealot]

and this has everything in view for daily use laid out in (what I think) is a logical arrangement

top right is my virtual "progress bar" where I can watch the various green boxes lighting up

[strong field]

ah that is slick i thought thats what this bar was

[soft zealot]

Image saves I can quickly break or make the jumpers depending on what I want to save

[strong field]

and i thought to myself, that is one of the coolest ideas i have seen

[soft zealot]

aye the purple bitts I justwanted a home for lol

If I was being Picky theres an argument that he LORA section should move to Lower Left closer to the Prompt ZOne

actually thats a good idea

[steady grove]

offtopic kindof. just having a nostalgia moment. i used to use this software called "Reason" which i've been thinking is the first node editor i've ever used. it's a skeurmorphic interface based around rack mounted audio hardware. very cool software!

[strong field]

hahahahhaha

love my reason rack

[soft zealot]

Oh what does that remind me off, I whink it was aWinamp plugin

[strong field]

just built this nonsense

[steady grove]

it's reason. got popular around the age of winamp for a bit. went huge when matrix came out and the propellorheads became maintstream for a minute

[strong field]

comparing new IPAdapter chaining behavior to existing image blend techniques for shits and grins

Blending is far more preditable and offers much more control imo

[steady grove]

ohh the one i saw that looked new is actually new. i figured it was just a manual instlal of what i was using already

[strong field]

so yeah Reason was my training ground, noodle pasta been on the menu for a long time

[steady grove]

mfw noobs acting like node editors are something new and foreign

[soft zealot]

1914 noodles

[strong field]

ok heres another comparison, fine tuning ipa settings to get something mroe realistic

PATCH BAy

used to do a ton of patching back in my younger days for work

[strong field]

ah that reminds me of Moog Model 15 lol

[steady grove]

patch cords are kickass

[eternal fog]

For this one are you not just basically overlaying the images on top of each other faintly. So it's not actually getting the images seperately, you just have an eye image with the other image faintly on top of it?

[steady grove]

when do we get skeumorphic patch cords in comfy? i dont even see a branch for it. bollocks

[soft zealot]

when you create some lol

[strong field]

Change your blend mode to whatever you want. Heck convert to latent space. Possibilities are endless

Open Source = "DIY im lazy over here"

[eternal fog]

Well chaining 2 together seems to make it go absolutely haywire for me anyway, it's making images that make absolutely no sense for the provided images.

[soft zealot]

meanwhile.
All 4 of these used the same basic Linguistic, Supporting & Negative Prompts and the same fixed seed but because I had my ChatGOT Prompt Assistant enabled each time I Generated I got a slightly different description extrapolated from my base prompt

[strong field]

reduce your strengths for starters

full strenght chaining on IPA is asking for monsters haha

[willow bane]

Hlo

[steady grove]

okay, either i need to find a foss library for rendering node connections as patch cords then get a coding LLM to implement it for me, or go back to school and learn. i'll get right on one of those options

[strong field]

I am loving this idea, its nuts

[willow bane]

Can we share a link here?

[eternal fog]

Oh I had.

But I had an image of a person and then a 3D model and it just starts making images of books, makes zero sense. But then after a few it makes something that makes sense and then goes back to making books. There's something wrong with it. I think sometimes it's just ignoring the values.

[strong field]

what was your prompt ?

[willow bane]

Yt link I want some information

[strong field]

and unclip conditioning set at?

[strong field]

sure

[eternal fog]

No prompt because I wanted it to just use the IPAdapter and it works perfectly fine when you have 1 image.

[strong field]

yeah two chaining is a weird one for sure, still learning how it works

[eternal fog]

You don't use unclip conditioning for IPAdpater. Or you didn't before.

[soft zealot]

Its a nice way (I think) of getting variation from a base prompt with a fixed seed

[willow bane]

I want to create something like this
https://youtube.com/shorts/CRY6C05QAPM?si=hVFo90Gsu1hbxBsf

[strong field]

I am not sure i see the use case, to be honest

[soft zealot]

who was that at ;o)

[strong field]

looks like Deforum

[willow bane]

Yes but I don't know prompts

[strong field]

aaron

[willow bane]

Is there any site where I can get these types of prompts

[upbeat summit]

@vital ermine the meme images posted here will finally look more like him. thanks for doing this 😉 https://civitai.com/models/135544/segalsdxlv2

[strong field]

I dont remember if it was in the standard workflow, but this is an example of coherency with it turned off....big difference

Same exact settings

[soft zealot]

ah iok.

If it was me the answer would have been.

My Base Linguistic Positve was

one white woman hold flowers on top of roof, cyberpunk world, lots of neon lights

With it being translated by my assitant as:

A photo of a confident woman holding a vibrant bouquet atop a rooftop surrounded by a cyberpunk cityscape. Neon lights illuminate the futuristic city, casting a mesmerizing glow on her as she embraces nature's beauty.

OR

A photo of a white woman standing on a rooftop, vibrant neon lights illuminating the cyberpunk cityscape as she holds a bouquet of flowers.

OR

a photo of a stylish cyberpunk rooftop with a glowing white woman holding flowers, vibrant neon lights

OR

A photo of a white woman on a rooftop holding flowers at dusk, against the backdrop of a futuristic cyberpunk cityscape illuminated by vibrant neon lights. (No deep-fried barbie dolls, with no text or watermarks)

[strong field]

This right here is gold

Just the concept

AI using AI to AI

[willow bane]

.

[soft zealot]

Its blending manual prompt crafting with ChatGPT assistance

[eternal fog]

How exactly are you setting up the conditioning, because the IP Adapter doesn't actually supply conditioning it just supplies the model.

So are you taking the Clip_Vision_Output and blending it with text conditioning?

On the readme it says to basically ignore the output on the IP Adapter, but it might be a miss translation

[strong field]

I dont have much experience with deforum, sorry

clip vision output to unclip, and clip text encode (prompt) to unclip, I do have a prompt

[strong field]

general idea

but its super messy since i just dragged into view, lol sorry

[eternal fog]

Ok, I think we are approaching this in different ways, so it's not cross compatible.

I'll stick with what I was doing before, that works fine. Looks like it just doesn't like having to images in the IPAdapter.

[strong field]

see thats interesing, i dont think that provides you with near enough control

[soft zealot]

@strong field and If I really get stuck:

[eternal fog]

I guess because you are using text prompts, it's taking the image conditioning and the IPAdapter. I guess essentially it's running both IPAdapter and Revision on the same input image.

[soft zealot]

why does that say 15 terms or lesss, thought I had told it 100 or less lol

[eternal fog]

So you have Image with IPAdapter + Image with Revision + Text Prompts

[strong field]

IPA very sensitive to weights and conditioning, so clearly it makes a difference. If your workflow works, just do you.

[strong field]

Yeah i guess you could think of it like IPA + Revision hybrid.

[eternal fog]

Yeah I think I understand how it's working now. Both for you and me.

[soft zealot]

with styles applied as well

[eternal fog]

I'm using revision but i'm using different images, because I'd rather use that than text prompt.

So I guess I could try putting in an extra unclip conditioning for the IPAdapter and see what it does.

[strong field]

Looking forward to implementing this, I trained an LLM to give me prompt ideas with just a "give me a cool <description> prompt, and it would generate, but i had to manual ask and manually copy paste and revise

i really like the idea of having it all built in

[strong field]

Yeah im definitely not using IPAdapter standard issue hahaa

Like i said, lots of experimenting

[soft zealot]

aye I used to use chatgpt in a browser to do similar then copy & paste, this is so much easier & slicker

[eternal fog]

Hmm running the IPAdapter through the unClip does let you make small adjustments, but it's adding an extra knob to my never ending list of knobs lol

[strong field]

yeah it really do be like that

[eternal fog]

I guess you could use it to push the output to be even more like the IPAdapter image, because you're also using the text encoding of that image as well as the adapter.

[strong field]

Yup that was my initial goal, take input image and refine, improve, add detail but stay very close to original

I found out you could make SMALL adjustments very well and still mainting OG integrity

[eternal fog]

Just a quick test left to right is, nothing via unclip conditioning, 0.5 strength, 1.0 strength.

[strong field]

Except one caveat, i didnt have image prompt

[soft zealot]

OG Prompt from a Random Pull from Lexica:
a cheetah and a bird in a buddy cop movie, movie still,

ChatGPT Assistant version:
A photo of a cheetah and a bird in a buddy cop movie would show the two main characters standing side by side, with the cheetah displaying its sleek and powerful build while the bird adds a touch of charm and wit. The high-quality image would capture the intensity of their partnership and the beautiful natural surroundings, immersing the viewer in the thrilling world of the movie.

OK doesnt always work lol

(although the CHeetah looks good)

Then again you could refine it manually from there

[eternal fog]

It's certainly an interesting option, I guess I'll leave it in to mess with

Next on my list to try out is - https://github.com/Ttl/ComfyUi_NNLatentUpscale

Looks interesting

[strong field]

Great gens btw

[eternal fog]

They'd look better after an upscale 🙂

[strong field]

16k?!

[eternal fog]

Lol I accidentally made a 350MB image yesterday, it tried to crash my web browser when I previewed it.

[strong field]

Omg hahahahahah

[soft zealot]

@strong field my other usage for CHatGPT assist is to take an old prompt originally written in A111 and then "transalte" it

eg

OG A1111 Prompt:
*(art style by (Utagawa Hiroshige:0.9) and (Reka Nyari:1.2) and (Jennifer Massaux:1.2) and (frank frazetta:0.8) and (salvador dali:0.9)),
(erotic boutique photography),
(best quality, 4k, 8k, ultra hires), (realistic, photo realistic, RAW photo), (hdr, sharp focus:1.2), (intricate texture),
(soothing tones, insane details, intricate details, hyper detailed,photorealistic, 8k, ultra realistic, volumetric lighting,(film grain:1.4)),
(Taken using a 35mm Hasselblad 500C/M camera using Lomography 400 color film at f/1.8),(bokeh:1.1),

(realistic skin detail),(skin pores),(hair fuzz),(skin wrinkles),(skin blemishes),

glamour photoshoot on an riverbank, early morning, a submissive elegant lady,wearing a kimono,standing,

(background of a river,cherry blossom)*

Which if I drop it in direct (with some styles applied) gives me the attched (which is srill nice)

[ionic gulch]

2x ipa + txt promt is working fine for me:
https://i.imgur.com/I2XeXlq.jpg

[soft zealot]

@strong field CHatGPT translated that as :

A photo of a glamour photoshoot on a riverbank at early morning. The image features a submissive elegant lady wearing a kimono, standing gracefully against the background of a river adorned with cherry blossoms. The photo is of the best quality, shot in 4k or 8k ultra hi-res. It showcases the art styles of Utagawa Hiroshige, Reka Nyari, Jennifer Massaux, Frank Frazetta, and Salvador Dali, combining their unique aesthetics into one captivating image. The lighting creates soothing tones and adds volumetric lighting effects. The photo is incredibly detailed and photorealistic, with intricate textures and sharp focus. Taken with a 35mm Hasselblad 500C/M camera and Lomography 400 color film at f/1.8, adding a touch of bokeh to beautifully blur the background. The image showcases realistic skin details, including pores, hair fuzz, wrinkles, and blemishes, adding a touch of authenticity to the subject. Overall, it is a stunning, high-quality image that captures the essence of elegance and beauty.

[strong field]

I see your weights 😉

[ionic gulch]

3x ipa + txt prompt works but is a little bit to inconsistent to be very useful:
https://i.imgur.com/z3DRHQl.jpg

[strong field]

Definitely like the sdxl version better

[soft zealot]

well technically theyre both SDXL model wise lol but yeah I get you actually mean the prompt itself ;o)

the functional vs the linguistic

[strong field]

Ah yes the revised prompt

I also wonder if summarizing the prompt would perform any different

[ionic gulch]

workflow 6.0 with up to 3x ipa (each one has an on/off switch) now on github:
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[native knot]

Closest I've been able to get so far to a workflow that says otherwise:

Without cheating by using a cnet, of course.

[strong field]

Looks like you ready for those wedding invitations

[native knot]

Yeah, but that workflow says otherwise.

[soft zealot]

Probably. I could ask the assistant to limit to say less than 40 words

[native knot]

🥁 🥁 💥

[tepid surge]

warhammer creature

[strong field]

Custom instruction and reuse the prompt to summarize? Might not be useful lol more token usage

[south frigate]

still bad, even with ur prompt, bad quality - look at face... idk what to do with it

[native knot]

I gave you a suggestion that should help. Change the resolution to something closer to 1 megapixel. You'll generally have better results.

[south frigate]

1024/768 is bad?

[native knot]

1344x768 instead of 1024x768.

[south frigate]

hmm, okay will try

[zinc cargo]

@icy brook

[south frigate]

tbh looks very similar @native knot

face is giga bad wtih it

somehow

[native knot]

SDXL generally does best with ratios that are divisible by 64 and are closest to 1 megapixel.

You can see a table I posted here:
#✨｜sdxl message

[native knot]

Sometimes you're just going to get bad results...that's part of how all this works. You can then tweak little things like the CFG up/down, the steps up/down.

[zinc cargo]

why there are 2 text prompts as well?

[south frigate]

yeah I know, I used a lot of SD 1.5 on A1111 but in Confy the images seem so grainy and fuzzy especially on faces in general (ofc with SDXL), I think I just have something set wrong

[ionic gulch]

sdxl supports two text inputs (text_g and text_l) - so the two prompts are for that concept. don't really understand in detail how it works. one should be used for the content and one more for general style afaik, but as i haven't researched enough i just use the same most of the time. guess in the future the txt prompt will get further improvements, after i've done more research. but for now people who know how it works can use both positive prompts supported by sdxl.

[native knot]

That's just it...there's no such thing as "wrong" or "right" in this case. You could keep generating the image with new seeds and stumble upon one that works perfectly. There's not really much wrong with the recent image that you posted other than the face, so simply finding a seed that works better for this image would likely work fine. There is a tremendous amount of randomness with seed values and any given value can provide something totally different than the next.

[icy brook]

Fun! 😀 I tried some flying before also actually 🙂 Not as good as yours though.

[south frigate]

it seems to me that for now it is much easier to get better quality on SD 1.5 and hires fix, am I wrong?

[icy brook]

Can actually play with it online here for those who want to do some transparent ghosts without local setup.

https://huggingface.co/spaces/multimodalart/LoraTheExplorer

[native knot]

I don't think that's true in a general sense. Maybe it's easier to get a very specific look in some cases, but SDXL is far more capable in its "basic" state early on than 1.5 was in its early stages. The only reason 1.5 is capable of what it is now is due to the massive amount of development around it and the tons of models, LORAs, etc. Give it a bit more time and SDXL will easily pass that up.

[south frigate]

OK, I understand. What about addons like facedetailer? I see that it's downloadable in the manager, I'll say that I want to install it, can I somehow add it and have the machine connect this "noodle" by itself? As a newbie I don't know how to do it right.

[native knot]

I don't know if facedtailer is compatible with XL as I don't use it.

[soft zealot]

Remover this is all being done pre conditioning

[soft zealot]

What checkpoint/model are you using? I could go look in the image but can’t be bothered lol.

[south frigate]

this one

@soft zealot why?

[soft zealot]

I cant run your workflow as its AIT enabled but what I would do is try a non AIT enabled workflow with the same model & prompt as a comparison

[soft zealot]

why what >

[south frigate]

why u asked for my model but now i know

[tepid surge]

https://tenor.com/view/ariel-little-mermaid-contract-sign-gif-8147904

[south frigate]

currently a large proportion of hentai models are suitable for more than just "hentai" , you may be glad you don't see the rest

[tepid surge]

he/she/they idec i'm a 24 yo adult cx (i didnt read the other half of msg send context of other models as long as legal cx (rhetorical but yee 🥂💕))

https://tenor.com/view/rikka-takanashi-rikka-takanashi-math-numericals-gif-22074535

i'mma h^ntai model to h^ntai model

cx

[ionic gulch]

@visual glade @indigo carbon i can confirm that the latest version is using vram too heavily on my 24GB RTX4090. the windows os gets very slow and comfyui doesn't respond very well (heavy delays) while generating images and vram is used completely (23-24GB).

after generation vram usage goes back to ~18gb and everything is responding fine again. i like that you want to keep more stuff in memory, but it seams you have to be a little bit more conservative and keep more buffer to avoid negative side effects.

at least "--disable-smart-memory" works for me and solves the problem for now.

[visual glade]

are you using any custom nodes in your workflow?

[ionic gulch]

@visual glade yes, but no ait - mostly very common stuff and a few of my own nodes, but those are more for input/menu purposes and very light weight.

[hardy cipher]

hmm, I wonder if comfy could tell me how to keep the widget slider on on odd numbers

[native knot]

🤣

Got frustrated trying to improve the "OTHERWISE workflow...so went with something much simpler.

[soft zealot]

so the first image (if these post in the right order is using your OG prompt in copaxTimeless same as you with the same seed and no syles or other enhancements applied.

Second image is my revised prompt, agan with no styles or enhancements applied

3rd image is my prompt with a single Global Glamour Style applied

4th Image is my prompt with ProtoVision XL as the Base Model and DynaVisionXL as the Refiner model , no styles etc applied

5th image is your Prompt using ProtoVision & DynaVision as in #4

Draw your own conclusions ;o)

[idle zenith]

can somebody tell me why those white specs appeared in the photo? and why the faces are similar even if two different girls prompt used

[native knot]

The things in the red circles look like some kind of jewelry trying to be rendered in. Try adjusting the negative prompt to say earrings.

[idle zenith]

or maybe i should raise sampling steps?

[native knot]

Since there are so many things that change the results, it's worth trying. Ultimately, you could get the results you want by any number of combination of changes.

[idle zenith]

ok and what about the face similarities anyways? how do i get rid off similar or twin faces ?

[native knot]

Same changes could impact that. But personally, I'd play around with the prompt more by using more words that distinguish that there are two different people.

[idle zenith]

can you give me any example prompts? i really can't figure it out

any example photo link with multiple people with different faces

[native knot]

No offense, but I'm in the middle of doing some other things, but I would just encourage you to try stuff.

[idle zenith]

no problem mate

[hardy cipher]

if you're using a similar prompt along with similar cfg, similar model, etc, it's just going to happen that you'll get similar looking people. that portion of latent space is being activated

[steady grove]

those look like over training artifacts, like purple splotches, but after they've been refined through a second denoising pass. that's what my eye says anyways. i can only speculate since i dont' know how you made that image

[clever verge]

I have reported this a week or so ago. There has been some improvements since then but sometimes I OOM like you do. Mostly when doing larger batch sizes or during upscaling. And yeah, disabling smart memory fixes the problem with a small performance hit.

I'm on a too new NVIDIA driver (536.67). Could be worth checking if running 531.79 would fix it.

[ionic gulch]

no oom for me - just the whole os is getting super unresponsive. doesn't even show virtual vram usage - just usage of the full 24gb which results in the unresponsive os. (i'm using an older nvida driver - the newest ones would shift things to vram even earlier to avoid oom errors)

[clever verge]

Yeah, that's OOM of VRAM.

I bet it starts to use your RAM too when that happens.

[ionic gulch]

531.79 gives you real oom errors if you use too much ram, but doesn't tend to use shared vram. newer drivers shift things to shared vram if you use 80-90% of your physical vram - that prevents oom errors, but reduces performance in ai applications a lot - this is a feature meant to help games and doesn't work very well with ai and some other applications like video editing. nvidia knows about the problem and want to fix it in a future update.

[native knot]

System RAM no problem when you're running 64GB+ 😉

[ionic gulch]

i have 96gb ram - the problem is the vram.

[native knot]

I'm just being fecetius.

[steady grove]

96 is a weird configuration though. mismatched pairs or 4x24? 8x12? hows it work?

[clever verge]

Makes sense, we are probably seeing the same issue but with different drivers.

2x32, 2x16

[ionic gulch]

2x16 + 2x32 - as memory is dual channel that doesn't cause much of a problem.

[steady grove]

i dont know about that. but alright

in my experience, that causes a TON of problems in this virtual memory management realm. timings being mismatched are not great for high performance

[ionic gulch]

@steady grove only some server hardware/configurations "bundle" more than two ram modules. so using more than 2 ram modules or mixing 2x2 different modules can give you a little bit of performance loss (espeically if they have different timings), but nothing major as long as you use the right memory slots.

[steady grove]

yeah that's different from mismatching pairs

i bet it goes away if you drop the 16s

[clever verge]

Servers have different ranks too, many more slots, more than one CPU, it gets complicated fast. 😄

[native knot]

It's not that weird on the newest motherboards. In fact, there are now kits with 96GB. The reason is that it's more about matching up the amount of RAM to cores. Since there's a bunch of processors now with the # of cores divisible by 3, they're doing these kits now.

[ionic gulch]

i don't have performance problems with the ram. i have problems with latest comfy that uses 24gb of vram instead of unloading unused models, etc. i have no shared memory usage - just straight 24gb vram usage - so ram isn't involved.

[steady grove]

96gb ddr5 kits i think are actual 24gb dimms though.

[native knot]

Yes, I believe that's correct.

[steady grove]

mismatched dimms will cause the cpu to compromise and find a compatibility mode across all of them

[eternal fog]

Have you updated your Nvidia drivers, I've just updated my drivers and the memory usage has gone to shit.

[steady grove]

maybe it won't have performance problems. i bet its related though. this is the realm of swapping and low level memory management afterall

[ionic gulch]

@eternal fog i'm using 513.79, because nvida drivers after that shift things to vram if you hit 80-90% of vram (explained in more detail above). so for ai and video editing all newer drivers cause huge performance drops if you use >80-90% of your vram.

[eternal fog]

Yeah I know that, but I wasnt having issues on the previous ones, now I am. So was wondering if you'd recently updated.

It might be comfy as well though, it seems like it's not unloading things from memory after gens and then doubling up.

I can do 1 gen fine and then next one it's going into shared memory and being extremely slow.

It should not be doing this for a single 1024x1024 render

[ionic gulch]

@eternal fog if you have problems now, this is probably caused by the comfy changes. use --disable-smart-memory when starting comfy and the problems will be gone.

[eternal fog]

I thought it was that change that made it better, because after he added that it seemed to work a lot better, did he change it again?

[ionic gulch]

my problems started with the latest comfy updates. and using that option solved it (by not using the new memory management)

[steady grove]

the new drivers more intelligently use the shared gpu memory bank. the system memory. It doesn't automatically start using them at high vram useage. i know this becaues i use all 16gb of my 4080 very often without it edging into shared memory. When it goes over it slows down to my mainbus speed, since memory swapping is going on.

i bet when you hit those edge issues, problems occur because you've got mismatched memory banks running at lower timing's than they're meant to. I also think the newer versions of comfy might be tuned more for the newest drivers. You might be using the old drivers in an effort to disguise the problems of your mismatched dimms better

i've built systems for decades now. i've never seen mismatched ddr working at peak performance

[clever verge]

I have the same issues as JPS but with matched memory. I doubt memory is the issue here.

[eternal fog]

The old drivers don't use shared memory at all they just OOM if you go over, the new ones do instead of going OOM, but it slows down a lot.

[ionic gulch]

i'm using the old drivers to avoid a problem confirmed by nvidia, not to cover the ram issues you think i have. there is no shared vram usage according to task/performance manager when the problem occurs. and the "more intelligent" usage of newer drivers is fine for games, to hide the fact that many nvidia cards have not enough vram for latest games, but is poison for ai and some other tasks.

and yes, my system behaves even worse than some newer high end systems if shared vram is used, but not mainly because of my ram combination, but because i'm using a 5 year old mainboard and ddr4.

but the real problem is that the latest comfyui update tries to use every little bit of vram and fills up the full 24gb - which is enough to make other vram depending apps like the webbrowser slow down heavily. newer nvidia drivers would make it even worse and push 1-2gb to shared vram.

[clever verge]

If it starts using RAM it will be bad.

[clever verge]

I just reverted to 531.79. Saw that NVIDIA released 537.13 but not tested.

On a normal SDXL base + refiner 1024x1024 smart mem is 6s vs 10s with smart mem disabled. But with a 4x upscaler -> downscale 50% and a base pass it takes 27s on both.

[steady grove]

the driver issues nvidia "confirmed" are overblown and often misrepresented. you stating a margin of 80-90% for example. everyday i'm pushing my 4080 to the edge of it's 16gb and it doesn't touch the shared memory. Some problems do occur when shared memory is dipped into, which i mostly by going back to python 3.10.9 from 3.10.11, but at the 80-90 margin ? i dont know about that. i've never seen shared memory open up at those thresholds

I'm on the newest drivers with my system configuration and i don't see these problems in comfy. running the default out of the box windows portable build

well with updates

[eternal fog]

The main issue I have with the Nvidia drivers is in Training.

It doesn't just tell you that you can't handle it and OOM. It tries to do it with RAM and takes a million years

[steady grove]

heh yeah i've hit that too. before it would just crash and i'd adjust

now it just keeps going. darnit

[ionic gulch]

that shouldn't happen with the old drivers. old drivers should give you a real oom error and allow you to use more of your vram, as it doesn't start early to push things to shared vram.

[clever verge]

This could be specific to 4090.

[indigo carbon]

It's not, it's specific to the entire 4000 series

@ionic gulch the fix was to merge the latest commit with a commit from a week ago

[clever verge]

Talking about the smart memory in Comfy

[indigo carbon]

Yeah, I know

[steady grove]

smart memory isn't causing me issues and i've a 4080

gigabyte

[ionic gulch]

maybe your models / workflow doesn't hit the limit - if there is nothing to fill up the full vram, even the new comfyui update can't cause problems.

[steady grove]

pushing the limit is something i revel in doing. often theres a soundtrack in my head when i'm doing it

[indigo carbon]

It's not the limit. Because Nvidia being drunk when they made the 4000 series, when VRAM is over 95% full, speed divides itself by over 10

[ionic gulch]

@indigo carbon are you talking about the problem with latest drivers or in general? never had problems since i've gone back to the old drivers. with the newer drivers i had that behavior.

[indigo carbon]

No, the latest drivers actually made it a little better. It's an issue with ComfyUI

[uncut gull]

Running SDXL and LLama2 13b at the same time on my PC makes both of them slow to a crawl. 😕

Switching between the two isn't too bad. I have Llama2 churn away inventing new prompts.

[indigo carbon]

The old commits never fills up VRAM over 90%, which seems to be the threshold on 4000 series

[native knot]

There we go...no other text.

[uncut gull]

The hands don't look too great though.

[clever verge]

What version are you running now?

[native knot]

The sign tells you they're bad hands. 🤣

[indigo carbon]

Merged the latest commit with a commit from a week ago fixed it

[clever verge]

Driver version that is. 😄

[ionic gulch]

that's the problem i think. the new version really tries to use 100% vram, which is problematic with old and new drivers. there has to be some buffer left. the idea to keep more in vram is great, but not filling it up to 100% as there are other vram depending applications you will affect by doing that. one of them is the webbrowser you use to control comfy.

[indigo carbon]

No, it's not the drivers, I tested it with 10 different versions of drivers, that's an issue with ComfyUI

[wet nacelle]

Me and my

[steady grove]

heh

[eternal fog]

Yeah that's what I'm saying, old drivers OOM and new ones try use shared RAM and it takes ages

[indigo carbon]

That's a good thing on the 3000 series because when NVIDIA made them they didn't use memory busses from over a decade ago. I think there should be a mechanism that uses the old memory management on the entire 4000 series.

[clever verge]

I'm going to try 537.13 now.

[indigo carbon]

The current fix to use the old memory management is to merge a commit from a week ago with the latest one

[ionic gulch]

shared vram will always be a problem even with much faster bus speed. we need more "real"/phyiscal vram and/or avoid filling it up to 100%.

[strong field]

80gb consumer cards please

[indigo carbon]

Yeah, the threshold seems to be 90%. The old memory management seems to know that.. when I use the new memory management I get 30s/it, when I use the old one I get 10.5it/s

[pure crystal]

gpu with an m.2 slot would be nice

[ionic gulch]

for me the switch @visual glade mentioned fixed the problem. not the perfect solution, as the new "smart" memory management would have some performance advantage, but i don't want to merge things manually with every future update.

[idle zenith]

i figured out.. the issue came with the model itself so it recommended using xl vae
after using vae i got clean faces

[strong field]

Fire and not the good kind lol

[pure crystal]

haha

[uncut gull]

Llama2 coming up with image ideas and writing prompts.

I feel like it's not tapping into its inner artist yet though... Hmm.

[strong field]

@soft zealot would like to have words with you

[steady grove]

the pcie4 complete spec, something b series amd boards don't support, have a dedicated pcie lane from the ssd to the gpu. without the cpu needing to pass data. Directstorage is supposed to support it but i've seen nothing coming for that yet

[soft zealot]

eh

[pure crystal]

ahh I see, not my magical thinking version then (haha)

[steady grove]

once the api's are out for gpus talking to ssd's directly, direct storage and whatever other variants there'll be, we'll likely see gpu boards with m2s directly on them

[strong field]

I meant you wanted to show off your baby. I know I would 😛

[uncut gull]

Have you tried encouraging Llama2 to pursue an art career?

[pure crystal]

this tech will be developed and brought to market fast imo. It's not hard to do, it just didnt have a killer app

[soft zealot]

theres a scroll bar and a search function :p

[uncut gull]

Search functions are awesome. I've been developing fuzzy search algorithms since high school. Not sure what we're talking about though tbh.

[soft zealot]

my chatgpt assisted workflow using Omars Quality of Life schizzle

[strong field]

#✨｜sdxl message

[strong field]

That link was for you I fudged up

[ionic gulch]

ip adapter has problems if your source image aspect ratio doesn't match the target ar - seems it tries to crop it, but handles the resolution in a wrong way - guess i have to prepare the source images myself in some way:

[eternal fog]

Can't wait for my more RAM to get here lol

[clever verge]

What was your prompt to llama? Try to give it a few examples of a prompt you like.

[strong field]

Yup many issues if your AR doesn’t match, I don’t like IPA chaining for this reason

[uncut gull]

Prompt setup I was working on yesterday.

[strong field]

I think you can get good results if you fiddle with blending and strength/weight but why mess with all that when you can stack revision on top lol

[soft zealot]

@uncut gull what @strong field was pointing at was the way I've integrated Omars QOL ChatGPT node into my workflow using the standard Prompt boxes & logixc switches.

thsi is the broken out view of the integration

[uncut gull]

Sweet Christmas.

[clever verge]

Looks like a good start but would be interesting to see what happens if you add a few examples.

[strong field]

Mostly cause I still need to do this in my workflow

[uncut gull]

It had examples, but only a couple. And my examples were pretty lacking to be honest, to which it responded with similarly lacking prompts. I'm trying to think through it again now.

[soft zealot]

thats just the daily driver view

This is the helicopter overview with straight spaghetti on lol

[clever verge]

What llama2 model are you using? Running locally?

[uncut gull]

Local.

[clever verge]

Feels like there are more llama models than there are ants on earth.

[uncut gull]

I love programming and coding, even assembly language. I've been trying to learn nodes programming ever since it came out in Blender, but I just keep putting it off. I have to get over my dislike of 2d navigating... 😐

[strong field]

I’m here for it. Except for that kill switch

[soft zealot]

oooh maybe we need 3D Node Building, that would be cool 🙂

[uncut gull]

I guarantee it will come out as soon as I learn 2d nodes.

[indigo carbon]

you can also do 4 by the way

[uncut gull]

All Llama2 images. It's being very diverse, but not creative, you know? What am I missing in my prompt? I can't just tell it "be creative".

[indigo carbon]

can LLaMa2 see images? or that's something only GPT4 can do

[clever verge]

I'd say the older driver 531.79 is better than the most recent 537.13 with smart memory. Both have issues.

[indigo carbon]

fix was to merge an older commit with latest; but I'm sure Comfy will get on it

[clever verge]

Yeah, I guess I'll wait for that.

[shy kelp]

One message removed from a suspended account.

[uncut gull]

It's writing prompts for me. It can't see. I just want it to get more "creative" with its subjects, but I don't mean more "diverse". Even focusing more would be fine. Just more creative somehow.

[strong field]

Have you done any fine tuning, training, or embedding?

[uncut gull]

I tried for 8 hours yesterday to get Lora training working in Oobabooga. Errors all the way through, and I'm all burned out on that. I'm just prompting today.

Trying this now. 🙂

[strong field]

Ah yeah…. I hear that

[soft zealot]

@visual glade just a thanks from those of us at the "budget" end of all this (1080ti 11Gb VRAM driver 531.68 do I dare upgrade I wonder??, 64Gb System RAM), recent updates haven't degraded performance at all and indeed looks like COld Start perfromance has improved

[shy kelp]

One message removed from a suspended account.

[vital wraith]

Anyone else use the SDXL ultimate workflow on civiai? Anyone have insight to why the imgtoimg includes a downscaler?

[upbeat orbit]

Anime-finetuned XL is indeed better than 1.5 with details, and with understanding of different artstyles and cultures.

[uncut gull]

All Llama2 prompts. 🙂

[strong field]

You could try to install superbooga extension to ingest your own documentation with prompt examples. Otherwise Next steps would be api > lang chain > database / embedding

[upbeat orbit]

Are there any users of mst.ai? Their discord is closed, so I would want to know, if there is any progress towards XL support.

[uncut gull]

The roadblock here is my understanding of prompting and creativity. What, exactly, do I mean by wanting it to be more "creative"? And how do I put that into words?

[keen nimbus]

4090 pumped those out

[shy kelp]

One message removed from a suspended account.

[strong field]

That’s why using a data set of “creative” prompts would be a good solution. Go collect a set of examples you like and call creative and then tell the llm to use that dataset as a knowledge base when prompting.

[indigo carbon]

you can try increasing or decreasing temperature, that tends to mess with the creativity

[strong field]

Lol try this. Wonder if something this simple would work

[uncut gull]

I forgot about this. I played around with it back when I was doing creative writing. It definitely makes a difference.

[hardy cipher]

it'll pretty much break it going to either extreme as far as temperature is concerned

but then you can couple it with other parameters to go all sorts of directions

[shy kelp]

One message removed from a suspended account.

[uncut gull]

I tried just doubling the temperature from 0.7 to 1.4. Its prompts look more diverse style-wise, but still pretty on-theme for the imaginary artist I gave it. Llama2 has clearly never read Reincarnated as a Slime though. Generating images now.

[round walrus]

Anyone knows when will be released 1.6?

[indigo carbon]

"firefighter" took too literal..

[uncut gull]

These look much more creative! Well done, Llama2. I think it might pair better with a different SDXL model though.

[hardy cipher]

1.4 is high, but I guess still coherent depending on the other settings

[vital ermine]

That is precisely why I did it, lol.

[keen nimbus]

can some of you use my custom prompt

and see what it spits out for you?

Stranger things.
Walking outside in a nostalgic town from 1984. Looking through the front door of a building, on the other side of the door is another town, but that town has rainy mist ambience, the sky is grey and blue from the misty rain. probably a cool afternoon in september around 7pm. you notice the prototypical red glowing light ambience that is synonymous with stranger things. the town looks very familiar, like you have been there before

thats the entire prompt

@upbeat summit see above

[uncut gull]

Since its goal is to mash together subject and style words while ignoring grammar, I think it's pretty robust against errors even at high temperatures (for this particular task).

[keen nimbus]

did the 1.2 channel get pulled

or 1.4 or whatever it was, or 2_1

for xl?

i guess people are using comfyUI now as the standard, no more "oogabooga"

[uncut gull]

Most people are still using A1111. And oobabooga is for text generation, not images.
(I'm using A1111 and oobabooga at the moment.)

[keen nimbus]

i used that gradio ui before for image generation

but this comfyUI is just as effective

[uncut gull]

Doesn't work on tablets / phones.

[indigo carbon]

yeah, if anyone hates nodes, they can just use SSW

[keen nimbus]

i use comfyUI on 4090 rtx, and 7950x

win11

my gen speed is slow though

on 4090

[hardy cipher]

have it make up hypothetical words

[keen nimbus]

C:\Users\xt3\Desktop\ComfyUI_windows_portable>.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
Total VRAM 24564 MB, total RAM 65457 MB
xformers version: 0.0.21.dev581
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 4090 : cudaMallocAsync
Using xformers cross attention
Starting server

To see the GUI go to: http://127.0.0.1:8188

[strong field]

So are you saying I have a problem because I run a local host just so I can access via tablet?

[keen nimbus]

is there someting to use to speed it up?

[uncut gull]

Slightly better. :-). Llama2 pretending to be an AI prompter paired with CounterfeitXL.

[keen nimbus]

is Using xformers cross attention not ideal for 4090 w/ comfyui?

[uncut gull]

The ComfyUI interface glitches every time the keyboard opens and closes on my phone and tablet. (Android.)

[keen nimbus]

different params to use to speed it up?

[hardy cipher]

something to consider, stable diffusion doesn't really know if a word is made up right?

so wouldn't it just try to connect the dots if it get's fed a new word?

[strong field]

Interesting, I had to zoom out a bunch on my browser but seems to work ok

[uncut gull]

Also, it understands languages other than English, and the sense is not the same. "Forest" is not the same as "bosque", even though they're both forests.

[soft zealot]

as I understand it it will atrempt to break the word up into words it can recognise and then tokenise

[indigo carbon]

If you want speed; you can use AIT. it increases speed over 200% on 4090

[keen nimbus]

show the flags

[indigo carbon]

but slightly complex to set up atm

[keen nimbus]

url?

tutorial

[eternal fog]

Where is this model, I've seen people mention it, but I can't find a model with that name.

[indigo carbon]

https://github.com/FizzleDorf/AIT

[hardy cipher]

yeah. doesn't really look at things like us. so if you think about it, you could put different tokens together that don't really have real world counterparts

[uncut gull]

Yep. There's a gem called "carbuncle". I tried to get SD1.5 to generate an image of it, and I got pokemon. 😕

[upbeat summit]

you can use hgsksdfzg as a token or just leave it empty and get nice nature photography 😉

[strong field]

Split text > chunk > tokenize

[upbeat orbit]

I am using CounterfeitXL v1.0 along with some other finetuned checkpoints.

[hardy cipher]

if you type frantic weird random letters it will probably not make beauty

[eternal fog]

Ah ok, I could never get CounterfeitXL to look as good as 1.5

[hardy cipher]

or just put in something like "destitution" and have it loop over itself

[upbeat summit]

but it does

[keen nimbus]

@indigo carbon so basically , with how fast changes occur in this space, is that url some new optimization technique that vastly speeds up IT/S generation speeds

on 4090 for image gens?

[strong field]

Eventually it will tell you to stop acting crazy, but until then it will just tokenize with closest weighting

[soft zealot]

nope , all good post driver upgrade to 537.13 here at the cheap end @visual glade (well mostly, cold stat performance has slowed sligtly but that could be margin of error, same for warm starts)

[indigo carbon]

for 3000 series and 4000 series. however; you can already get a somewhat nice boost already by using CUDA11.8 and the latest cuDNN software stack

[soft zealot]

Supercalifragilisticexpialidocious

[hardy cipher]

but like dest-itution. two tokens. you can swap for other things.

[strong field]

Returns picture of umbrella

[hardy cipher]

just interesting

[indigo carbon]

@keen nimbus the workflow linked in the repo already has an explanation for how to infer in it

[uncut gull]

In CounterfeitXL, "forest" vs. "bosque". (Bosque is Spanish for forest. SDXL understands most languages I've tried, but the words have different feels to them.)

[shy kelp]

One message removed from a suspended account.

[soft zealot]

I (really did lol) got this first time around (no enhancements enabled)

[indigo carbon]

if your not looking for an anime checkpoint, there is a checkpoint called copax_colorfulXLV2 and it's insanely good

[eternal fog]

I've found Nightvision better than Copax personally

[indigo carbon]

copax colorful is way better than normal

[uncut gull]

I have CopaxTimeLessXL that I still haven't tested. I've been using CrystalClear 99% of the time, and CounterfeitXL for anime stuff. Thanks for the reminder. I will at least test CopaxTimeLess now. (As soon as SDXL gets done with its long slow churn through a measly 10 images...)

[soft zealot]

and when I turned on the ChatGPT assistant it spat this out from the same seed
lol

A picture of a whimsical landscape with vibrant colors, showcasing a sprawling meadow full of wildflowers and a sky filled with cotton candy-like clouds. The image exudes a sense of enchantment and joy, with every detail capturing the essence of the word "Supercalifragilisticexpialidocious."

[indigo carbon]

copax timeless is not copax_colorful; I also used to like CCXL best: but copax_colorfulV2 has a league of its own

[strong field]

Much better than I could of done

Scary smart

[uncut gull]

Okay. Downloading now. I will totally delete it for HD space if I don't like it though. 😛
You mean Copax Vivid, right?

[eternal fog]

This should be Civitai's catch phrase or something

[indigo carbon]

no, copax colorful

[shy kelp]

One message removed from a suspended account.

[indigo carbon]

hold up, I'll get the link

[uncut gull]

Civitai link?

[indigo carbon]

https://civitai.com/models/118111?modelVersionId=137533

[uncut gull]

That is literally Copax TimeLessXL. That's the one I downloaded yesterday and was about to test.

[eternal fog]

It's not

Look at the top

There are different versions

[indigo carbon]

it's the colorful version

[uncut gull]

No, you're right. I see it now. Okay, restarting download.

[soft zealot]

I just tried "llun o goedwig" which is the Welsh for "picture of a forest"

Nope lol

[eternal fog]

Not even the Welsh know Welsh so I doubt the model will

[uncut gull]

"forest" vs. "bosque" in SDXL base model, not fine tuned.

[soft zealot]

I understand spoken Welsh better than I speak it 🙂

[uncut gull]

Yeah, never tried Welsh. Also Latin doesn't work. (Not many latin-labeled images in the dataset. 🙂 )

[eternal fog]

But are you Welsh

[soft zealot]

Yes, Welsh was compulsory in school until the age of about 15

Mind you that was nearly 45 years ago lol

[eternal fog]

lol, I lived in Cardiff for Uni for 3 years. I found that people from South Wales used much less Welsh than those in the North. And the North had their own little group where they would speak Welsh regardless of who was around.

[soft bone]

i think it only knows those words thanks to them being in the captions of the original data. images of a forest labeled "bosque" are likely going to be from a certain area or certain forest (spanish speaking people), giving it a unique look of the forest since its a specific one.
It's not like there was any extra effort made to make the model bilingual (afaik), it just knows the word because spanish people caption their forest pics with it, so yeah it wouldn't look the same as "forest".

[idle zenith]

if i copy data from a generated image from someone and use exact same setting with exact seed.. why do i get different result?
like in original the face is front facing but in mine it always comes as side facing or back facing

[soft zealot]

I#m a Cardiff boy and yes it is true

As I student I presume at some point you visted TWAT cafe ?

[eternal fog]

I do not remember, it was a while ago lol

[soft bone]

But it would be really cool to use multiple sets of captions in different languages for every image. idk if the model can learn like that

[soft zealot]

I have a new use for the chatGPT assitant

On the fly translation

[uncut gull]

Yep. That's my theory too, anyway. Don't see what else it could be.

[soft zealot]

The Warm As Toast on Salisbury Rd

[eternal fog]

ChatGPTs ability to translate is... interesting lol

[idle zenith]

not same in all similar settings?

[soft zealot]

its not just translated it literally, its done what I asked it to an provided an embellishment

but its impressive I put 3 Welsh words in and it cme up with something sensible

[uncut gull]

"Fire" vs. "fuego" (Spanish) vs. "feuer" (German) in SDXL base model, not fine tuned.

SDXL, why are you burning animals???

[soft zealot]

If I thorw "fuego" into my assitant it responds with
A picture of vibrant, fiery flames dancing within a fireplace, illuminating the room in a warm and inviting glow. The high quality image captures the intricate patterns and intense colors of the flames, making it feel alive and dynamic.

[eternal fog]

This isn't how you ride a motorbike

[uncut gull]

Is this what it would feed SDXL? I gave mine example non-grammar prompts to fix its tendency to use full sentences.

[soft zealot]

but remeber the "rules" I've given it are:

Act as a creative problem solver that answer in prompts.
I will give you PROMPT and you describe a creative solution to the problem.
Use terse concise terms to describe the answer.
User descriptive details, answer in one or two sentences in one response and keep to under 40 terms starting with either the phrase " a picture of" or "a photo of".
you can use commas between terms.

Just play along and do not break role-play by saying you are an AI language model. Just guess at the answer.

[soft zealot]

yes based on the rules I provided it

[uncut gull]

Mine: AI image generators are prompted by describing the subject as a list of attributes and style factors. Rather than exact and fluent language, the AI does best with short concept fragments in a list. Style factors include key words associated with images of that style. For one style example, artistic illustrations are associated with oil paints, watercolors, colored pencil, acrylics, etc. And for another style example, photographs are associated with key words such as depth of field, bokeh, lens, focus, clear, etc. The AI image generator may accidentally ignore parts of the prompt, especially if part of the prompt describes something uncommon or out of the ordinary. To mitigate this, features that should definitely be present must be repeated with different phrasings. For example, a "suit" commonly has a black jacket, a white shirt, and a red tie. What if you wanted to prompt for a completely black suit? That is an uncommon and unusual thing, so the prompt "all-black suit" would most likely be ignored in favor of the more common visual representation of a suit. Instead, the prompt would be extended with rephrased repetitions: "a black suit, a suit with black shirt and black tie, completely black suit, all black clothes, black fabric". The prompt has a maximum length of 75 words, so "fluff" words such as conjunctions, prepositions, and pronouns must be avoided at all costs.

Examples it gave me for the reincarnated as a slime gothic artist persona:

2. medieval manuscript illuminated initial P depicting knight holding sword before castle ruins moonlight faint distant church bells sound foggy grass ground. gold leaf silver paint illumination calligraphy golden ratio
3. futuristic cyberpunk 3d animation of humanoid slime monster with laser beam eyes moving through smoky alley full techno equipment stacks neon signs buzzing sound effects motion blurs distortion effects
4. abstract pixel art style of swirling multicolored fluorescent gas planet in endless void space, cosmic particles emanating energy beams, intergalactic travel feeling of loneliness cold mystery
5. classic impressionist oil painting capturing carefree moments young boys climbing cherry tree blossom branches in park. light spring sunshine fleeting moment joy happiness children laughing playfulness innocence```
(Edited, I pasted the wrong thing there.)

[eternal fog]

What an earth is wrong with CounterfietXL, I just downloaded it and all its renders are like this

[native knot]

How many steps and what sampler?

[eternal fog]

Wait, fuck sake. I must have dragged accidentally. It's changed my steps settings and fucked it all up.

It had moved my steps to 257, but my end at step was on 28

It must have dragged over when it was lagging

[soft zealot]

@uncut gull

It made a good stab at

Gwlad, Gwlad, pleidiol wyf i'm gwlad
Tra môr yn fur i'r bur hoff bau
O bydded i'r heniaith barhau

The literal translation of which is

Country, Country, I am faithful to my Country.
While the sea [is] a wall to the pure, most loved land,
O may the old language endure

whilst feeding the following in as my positive prompt/

A picture of a peaceful Welsh countryside, with rolling green hills, a serene coastline, and clear blue skies. The high-quality image showcases the beauty of the landscape, capturing the essence of Wales' proud heritage and the everlasting love for its language. Tags: Wales, countryside, greenery, coastline, skies, heritage.

and then spat out what looks like part of the Welsh coastline

[uncut gull]

but that stuff like "capturing the essence" and "proud heritage" is hard to get rid of. It was challenging to get the model to just put visual-related words together without using grammar.

[soft zealot]

I wonder what its fgoing to make of the openingverse of La Marseillaise ;o)

anyways thats for another time

[uncut gull]

I'm just waiting on my PC running tests on different Copax models. Can't even prompt right now anyway.

[indigo carbon]

did you get the one I linked? it's really incredible

it even does text like base does; which most XL finetunes lost the ability to make text

[uncut gull]

Yep. I'm comparing it to two other Copax models, and to CrystalClear, Juggernaut, and Counterfeit on the same prompts list. (The last prompt list I got from Llama2).
It'll be a while before all the tests are done rendering though.

[indigo carbon]

if you would compare text in images, that model would easily be miles better than the others

I get full sentaces about 70-80% of the time like this

[uncut gull]

I could try that later, but I've never had a need for text. Not the main thing I'm looking for, definitely.

[uncut gull]

That's crazy. I'll at least give it a shot. (Still have a long wait though...)

[indigo carbon]

yeah, it's almost as good as DF IF with text. while being incredibly clear and detailed

[uncut gull]

sign saying: "WELCOME TO NOWHERE" on it in the middle of the forest, desolate, intricately detailed, artistic lighting, particles, beautiful, amazing, highly detailed, digital art, sharp focus, trending on art station
Do you always use sign saying: "..."?

[zinc cargo]

how to use more than one CN with SDXL?

[indigo carbon]

if I want a sign saying something yes.

but always with :, "", and CAPS

[zinc cargo]

where does these stacks go?

[lilac raven]

Is someone else having a problem with a1111, where it's possible to run SDXL once and any subsequent run or parameter change will cause out of memory errors, requiring to restart a1111? Interestingly, the behavior is different between a 3060 and a 1650: the 3060 can generate many images, as long no parameter changes from the initial run. The 1650 can generate one image and that's it.

[thorny frost]

no. I just used a1111 xl for the knight

[glad grove]

better ?

[upbeat summit]

row #1 model: base sdxl (1-3)
row #2 model: nightvision XL (4-6)

[lilac raven]

Are you using Linux?

[thorny frost]

nop, win11

[lilac raven]

Hmm... this may be a Linux issue
Maybe https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/12478

[wind storm]

Made this using Deforum using SDXL if anyone is curious what that might look like.

https://youtu.be/gcxPRuvicPY

[soft bone]

futuristic terracotta army

[glad grove]

midas entire family (he touched them)

[soft bone]

terracotta from krypton

[autumn forum]

the stone men

[uncut gull]

Model comparisons: CopaxColorful, CopaxVivid, CopaxTimeless, CrystalClear, Juggernaut, SDXL Base

[indigo carbon]

I kinda like copax_colorful best, but that's just my opinion.

[uncut gull]

I'm testing their text-writing abilities now, but for all 6 of those, I found images I preferred from each. Even SDXL Base had the best image on one of the prompts.

[indigo carbon]

the writing abilities present in my images are also caused due to my optimized AIT workflow using complex stuff, it might not be as good as the quality I get to

[uncut gull]

So far I'm pretty happy with all of them, actually. 🙂 But I've only glanced at results. I'll count typos and get %s when the tests finish running.

[late hill]

I have a question about the Clip Vision models in SDXL if anyone knows. Is there much of a difference between using the 9Gb open_clip_pytorch over the 3Gb clip_vision_g.safetensors?

[autumn forum]

tried to outpaint... but i failed. dont know how to outpaint in comfy lol

[terse pewter]

Hello! I am not able to generate this image, I have been trying for two days now. Will be trying on comfy UI tomorrow.
Can someone help me with this scene?

"view from outside of a big garage with snowmobiles, gas station in front of the garage, a man sleeping on the ground near the gas station, outdoors, nightsky, north pole"

It always misses out on the dead person on the ground, I have tried on all the ai image generators

[steady grove]

they let you adjust with mouse wheeling too

[vital ermine]

@upbeat summit more reason he is so meme worthy.

[uncut gull]

Writing comparisons CopaxColorful, CopaxTimeless, SDXL Base, CrystalClear (best).
Prompt sign that says: "WELCOME TO SMALLVILLE"
Plus a bonus, 1 image where CrystalClear spelled the entire prompt 100% perfectly! (Although still with some unwanted text underneath, so not usable.)

[glad grove]

ngl id love to try the taste of berfatinn

[native knot]

The one sign tells you what to tune-in to in order to get the most from your visit... "JAM 79.9"

[glad grove]

now imagine vid2vid + seagal lora + the matrix

[strong field]

It I can get a matrix remake where every character is a seagal interpretation but true to the movie I will tolerate every seagal meme until that time and it will have been worth it

[native knot]

I'm sure if you stay alive long enough, that'll be possible.

[upbeat summit]

close-up shot of (Agent Smith:1.25) played by Steven Segal wearing a grey suit with sunglasses walking in downtown Manhattan, looking at viewer, detailed suit, detailed outfit, detailed hair, detailed sunglasses, (from The Matrix 1999,:1.3) heavy rain, soft light, IMAX, 4k remaster, insanely detailed, rain everywhere, green tinted color grading

incredible likeness made possible with https://civitai.com/models/135544/segalsdxlv2 by @vital ermine

[vital ermine]

I can only imagine The Matrix with him as NEO, lol.

[heady vale]

Seagal is the ONE

[vital ermine]

Hehehehe

Well, he does make a good Jesus as I found out.

I am very pleased with how it turned out based on the really shitty , low quality, small pics I could find. FULL of noise and jpeg blocking nonsense.

took me longer cleaning them up than the training or even collecting them (which was no small task as the dude doesn't have all that many pics of him on the net only the same ones).

[heady vale]

Trinity has been replaced

[strong field]

Holding hands like good buddies

[vital wraith]

Is it possible to alter a Openpose image in ComfyUI? OpenPoseEditor only seems good for creating entirely new poses.

[eternal fog]

lmao I mixed your cube with a person

[indigo carbon]

Huh, clip vision can fuse images together?

Or did you combine tokens

[eternal fog]

Put your cube into an IPAdapter, then a woman into Clip Vision

[indigo carbon]

IPAdapter? Can you clarify, didn't hear about that

[heady vale]

revision can do backgrounds in another picture pretty well

[eternal fog]

https://github.com/laksjdjf/IPAdapter-ComfyUI
https://github.com/tencent-ailab/IP-Adapter

[indigo carbon]

Thanks. I wanted to make an image fusion version of my workflow, this might prove useful.

Does it need a prompt though?

[eternal fog]

There's lots of weights and settings you can move around to get different effects

[upbeat summit]

@ionic gulch has already build a workflow that contains clipvision and IPAdapter and many more features https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[eternal fog]

If you want to just get variations of the input image then no prompt needed.

If you want other things as well as the image then it needs a prompt, or a clip_g image

[strong field]

A true believer. Amen.

[indigo carbon]

No AIT though, I might make a version of my workflow that uses that principle

[south frigate]

I just gave up with this ComfyUI, the only advantage of Comfy for me is only reneder time - classically in A1111 everything looks better and easier. I hope they will introduce AIT quickly to A1111. Thanks for those comparison photos, I spent some time on it to try with this Comfy but no, it doesn't make sense because the A1111 contains everything I wants without the annoying look and unnecessary fiddling with it.

[indigo carbon]

What images did you fuse here?

[eternal fog]

That was the single image with no prompt, but at a lower weight

[indigo carbon]

Peculiar.

[eternal fog]

And this is when you go too far

[strong field]

Stableswarm uses comfy as a backend so best of both world with a gui you are familiar with

[strong field]

The possibilities are endless

[indigo carbon]

can you share the gen params used here? that would be very useful.

[south frigate]

I tryed. Still its not the same. Still problematic is that in order to add something to the program I have to deal with "noodles" to make the various plug-ins work. This is not for me.

[eternal fog]

My workflow is a complete mess, But I can send you it.

The weights make huge differences so you can just play about with it.

[strong field]

Godspeed my guy

[eternal fog]

That one should have the workflow in

[pure crystal]

red house looks like the town from The Goonies (astoria, oregon)

[indigo carbon]

thanks alot, I'll credit you if I would implement that into the AIT workflow.

[pure crystal]

lower denoise

[south frigate]

Thanks. My hope is in A1111. I hope that they will not abandon it in favor of the fact that Comfy is becoming popular.

[eternal fog]

I think you might have to compile it for this, as the model needs to go through the IPAdapter, and I doubt it can understand the AIT Model.

[upbeat summit]

totally. but stranger things is basically goonies 😉

[indigo carbon]

I can compile everything. the compile script I made to provide the precompiled ones for batch sizes are universal.

[vital wraith]

Does ComfyUI have an OpenPose editor like Auto1111 where I can adjust it overlaid on the orginal image for accuracy? Really the one thing I've ran into where I can't get the same functionality out of Comfy.

[ionic gulch]

Testing AIT and if it works well an AIT version of my workflow is already on my todo list.

[indigo carbon]

do you want to work together on this then? I'm planning to make a version of the AIT workflow that you just imput 2 images and it fuses them.

[eternal fog]

It's not a direct fuse btw, it can require messing around for certain images, there's loads of weights you need to tweak and sometimes it just doens't do what you'd expect.

[upbeat summit]

@indigo carbon Have you heard anything about the LoRA + AIT issue?

If not I guess I will file a bug report on the AIT repo. That LoRA delay with AIT is really frustrating. Every image takes 5-8 seconds longer and if you use more samplers you can double the time for every sampler again. This keeps me from really using LoRAs at all.

[indigo carbon]

for me LoRAs work fine, likely an issue on your end

[strong field]

Side note, is there an easy way to turn off any base + refiner workflows?

[upbeat summit]

I talked to other 4000 series users. they have the same problem

[ionic gulch]

this would already possible with my workflow (it has up to 2x revision + 3x ip adapter as input) and should behave like that if you keep the txt prompt empty. unless you want a less bloated, smaller workflow. but during the week i don't have that much time, so it may take until the next weekend before i can check AIT.

[upbeat summit]

you have no delay?

[indigo carbon]

might be a driver issue then

[indigo carbon]

nope, same gen time

can you try LoRAs with the batch size modules

[upbeat summit]

the gen time is the same. it doesn't start generating. the delay is before the gen starts when you have a LoRA in the pipeline

[heady vale]

AIT + Lora for me I have roughly a 4 second delay before each sampler starts

[upbeat summit]

yeah I got 5 - 8s

[indigo carbon]

@upbeat summit can you try this? I suspect it could be because the old modules weren't compiled with the compile script I made.

[ionic gulch]

you mean the 4 versions (70-100%)? not at the moment. i would like to make them optional / switchable, but haven't found a good solution. would prefer a real switch instead of just cutting and connecting wires.

[upbeat summit]

sure. where to get those?

[indigo carbon]

just use BS4/2 with the newest version of the node, it will get them automatically for you.

[strong field]

Yeah, no rush I can disable manually

[strong field]

Will it replace bs1 or do you have to manually remove bs1

[indigo carbon]

yes, it's within replace the BS4 modules with the BS1, this is due to the new modules being flexible.

[strong field]

Sweet

[upbeat summit]

good idea. batch size 2, new modules were downloaded, but the delay sadly remains

[strong field]

I’ll grab ‘‘em when I get back

And tests some Lora’s @upbeat summit

[indigo carbon]

=\

[upbeat summit]

it happens with all LoRAs

if I bypass the Load Lora node it goes back to instant

[indigo carbon]

how are you implementing lora @upbeat summit ?

[upbeat summit]

Checkpoint Loader -> Load LoRA -> AIT -> sampler

Checkpoint Loader -> AIT -> Load LoRA -> sampler
also works but same effect

[indigo carbon]

that's how you should do it.. does the delay occur without AIT?

[upbeat summit]

nope

only when AIT is in the mix

[indigo carbon]

can you try checkpoint->AIT->LoRA->AIT->sampler?

[upbeat summit]

first you win 100% speed than you get 50% taken away because of the delay, that is what is so frustrating 😄

[upbeat summit]

I think I tried it before, but let me try again

[upbeat summit]

no change. delay remains. I think it has something to do with sampler nodes. because when it hangs it highlights the sampler node with the green active border

[heady vale]

same delay here

[indigo carbon]

I don't get it, it shouldn't effect speed

[heady vale]

gen speed is the same but start to finish makes the time slower

[indigo carbon]

LoRAs have same architecture as the model they were built for, it should do it

[upbeat summit]

AIT + LoRA loaded
AIT wihtout LoRA

AIT + LoRA takes 10 seconds longer because of the delay at the start

[ionic gulch]

new version 7.0 of my workflow keeps the aspect ratio of the source image in IMG2IMG and adjusts it to one of the recommended SDXL resolutions. also ip adapter source images get resized to reduce resource usage for very large input images.
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[upbeat summit]

So I'm 100% speculating here, but maybe it has something to do with memory management? because the behavior goes away for a couple of minutes when launching ComfyUI with --gpu-only.

Than after a couple of minutes I get really bad performance issues like 1 step every 20 seconds with the GPU running at 100%. I have to restart ComfyUI to fix it. So I can't run --gpu-only because it's not stable for me.

[upbeat summit]

@heady vale could you try starting ComfyUI with gpu-only and see if the delay is gone for you? just be beware it could get unstable after a couple of minutes

[nimble heart]

what ait nodes is everyone using?

[upbeat summit]

https://github.com/FizzleDorf/AIT

[nimble heart]

damn

[nimble heart]

that one just mistakenly thinks my GPU is sm80 architecture then dies

[upbeat summit]

it should work for 3000 and 4000 series cards

[nimble heart]

ya I was hoping it'd hot-compile for whatever arch

instead of just dl-ing

I'm using radeon's AIT fork specifically for my gpu

[indigo carbon]

if its an issue with memory management it's a problem with ComfyUI, not the node.

[nimble heart]

I mean the node could have a bunch of clones

[indigo carbon]

sm80 is the only one supported

[upbeat summit]

yeah I don't know how to determine what it is. I just know that --gpu-only solves the problem for a couple of minutes, but it's get unstable with or without AIT. also gpu-only increases my performance even more with AIT

[nimble heart]

was comfy still working on official AIT support or should I begin working on a redneck patchjob for gfx1100 arch

[glad grove]

do the redneck thing

[steel mist]

I'm getting a missing node warning for RecommendedResCalc -- I don't see it in the manager?

[ionic gulch]

https://i.imgur.com/Vz8nFW3.png

[steel mist]

ohhh duh, missed it in the manager because it's got spaces in there

[indigo carbon]

@ionic gulch is your workflow capable of fusing images without text input?

[steel mist]

Not sure why it wasn't showing up in missing nodes

[ionic gulch]

yes, in various ways: img2img + revision, img2img + ip adapter, 2x revision, 2x ip adapter
for example 2x ip adapter:

[strong field]

Do you keep model loaded on AIT? Curious if this is in the mix

[indigo carbon]

can we talk in DM? I have some questions.

[ionic gulch]

sure. but i don't have much time left today. already 3:40 am here and i have to work tomorrow. :)

[upbeat summit]

tried enable and disable on the "Load AITemplate node" - no change

[strong field]

Ok

You have 4000 gpu right? I have 3000

[upbeat summit]

yeah

[shy kelp]

is ait support on colab?

[upbeat summit]

do you have the delay issue when a LoRA is in the pipeline?

[vital ermine]

I still can't get two loras, or more, to even work

if I want that I have to use auto

[nimble heart]

nah. not re-downloading all the models from HF as the dorf docs seem to imply is necessary
https://github.com/FizzleDorf/AIT/blob/main/docs/compile.md#unetcontrol-unet

[strong field]

Not at desk just yet

[nimble heart]

if the navi 3 branch gets merged into the upstream facebook repo then maybe

[strong field]

That’s strictly for compiling

[nimble heart]

yes

which I have to cause I'm not on sm80

[strong field]

Oh duh nvm. Thought you meant to use it

[nimble heart]

I was gonna try this fork specifically made for my gpu arch
https://github.com/ROCmSoftwarePlatform/AITemplate/tree/navi3_rel_ver_1.0

but too much effort

[strong field]

You rocm?

[terse pewter]

what is lora?

[nimble heart]

7900 xtx here

[strong field]

Yeah unfortunately everything harder with ROCM

[nimble heart]

AutoGPT-Q works fine

diffusion works fine minus xformers

[strong field]

Gave up my 6950 XT for a 3090 because PyTorch no likey

[nimble heart]

my text generation performance was comparable to Sytan's 3090 using autogptq

SD hurts though still using sub quad attention

games completely rip. bought Shadow of War for $8 and it runs 4k 150fps maxxed

on linux

Navi 3 cards are getting "official" rocm support for use in pytorch and all that in the fall.

[lilac raven]

So I discovered it's possible to use more than one LoRA in one generation. Does the order on which the LoRAs are loaded affect the result on ComfyUI?

[nimble heart]

navi 2 and earlier can just get fucked I guess

[upbeat summit]

A LoRA is a fine-tuning for SDXL. Think of it like a model add-on or extension that can introduce new and enhance existing concepts, styles, objects, subjects and more.

[nimble heart]

Benchmarks using pytorch nightly if you were curios. Not sure how they stack against 3090
#✨｜sdxl message

[strong field]

I couldn’t wait. Sad panda

[heady vale]

That pushes past my 10gb vram, so no benefit with that flag

[strong field]

Just reviewed, of the top of my head my 3090 is about 28it @512
5.4it @1024

[upbeat summit]

I see - but is the lora delay gone?

[pure crystal]

muppet spirit quest

[heady vale]

will test it

[nimble heart]

yea seems about right. Think mine's closer to the 3080 scores I've seen when they're using xformers.

supposedly navi 3 is getting SDP (xformers) suppport eventually™️

[strong field]

Damn sorry to hear that about your card. Can’t wait till you’re up again. But enjoy that mom time!

[nimble heart]

so see if that helps

[nimble heart]

pretty sure he got a different 3090

he got an evga FTW earlier this month

and got refunded for the zotac

[vital ermine]

my comfyui manager just disappeared and when I try to git pull

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

• branch HEAD -> FETCH_HEAD
  fatal: refusing to merge unrelated histories

[nimble heart]

checkout master

[vital ermine]

F:\ComfyUI_windows_portable_nightly_pytorch\ComfyUI\custom_nodes>git checkout master
Already on 'master'
D extra_model_paths.yaml.example
D models/checkpoints/put_checkpoints_here
D models/controlnet/put_controlnets_and_t2i_here
D models/loras/put_loras_here
D output/_output_images_will_be_put_here
Your branch is up to date with 'origin/master'.

I deleted it and git clone

[nimble heart]

do you have a detached head?

[vital ermine]

last I looked, no

[nimble heart]

that's the main reason pulls fail

if you haven't modded shit you could just git reset --hard

[vital ermine]

No mods from me

[heady vale]

Its difficult to say, because even with 1 step for each sampler its still very very slow to start due to vram

[high skiff]

I-

That was like a month ago

[native knot]

Finally getting a chance to mess with the new 2nd image capability with IPAdapter and I gotta say that it's insane how well it does pulling the concepts from the pictures. No text applied on this:

[high skiff]

Actually, that was more than a month ago lmao

How did you even get all the way back there?

[nimble heart]

he's just really really sorry to hear it

[upbeat summit]

yeah - I understand. thank you for trying it out though!

[high skiff]

Lmao, I guess so haha

[nimble heart]

got two whole months worth of sorry backed up

[high skiff]

You know I'm still dealing with PayPal to get my $600 back?

[nimble heart]

tf

thought that was settled

[high skiff]

I'm getting ready to follow up with them again tomorrow

Oh no, not even close

[nimble heart]

bruv

[vital ermine]

git clone got me back

[nimble heart]

didn't the actual seller site side with you?

case resolved gg

give muns

in fact make it $800 for the inconvenience

[high skiff]

They canceled the appeal five separate times, illegitimately

Tried to state that I was outside of their return window, of which I proved them wrong

I got moved up to an account supervisor who is extremely rude with me, to which I have a recorded phone call of her trying to basically talk me out of going through with the refund and also strong arming me

I proved that somebody else in her department told me that I had 30 days rather than only the eight that they had given me. She tried to call my bluff to which she listened to our recorded phone call and heard that person tell me that, so she agreed to give me the refund

PayPal closed the case so she can't do the refund through that, so she's had to push it up to a supervisor payout, of which she's been postponing day after day for the last 10 days

So I'm going to politely call her up tomorrow and tell her that if she doesn't get me my money, I'm taking her to court

[strong field]

lol my bad, i read it tonight so....

[high skiff]

I have a huge grounds for a lawsuit right now, so if they really want to keep testing me, I'll go for it

[strong field]

tought it was related to this fiasco still

[nimble heart]

hell yea hit her with the fucking OBJECTION

[native knot]

Nuts

[high skiff]

I can get them on withholding a funds, emotional distress, and damages

All because in the end of the phone call she apologized for all the shit that they put me through for a month and a half, and stated that it was completely unnecessary

Of which means that all of these problems that I've had, including flare-ups with my autoimmune disease, of which I'm still dealing with currently, are all grounds for emotional distress and damages

[strong field]

glad to see you still using that monstrosity

[native knot]

Hah...what, that crazy steampunk hoodie? 🤣

That thing's hilarious.

[nimble heart]

ask for 8M settle for 400k buy Ford GT

[high skiff]

Bro, I strong armed the fuck out of her. I had her on the call for an hour and 45 minutes of her trying to weasel her way out of the refund, to which I had receipts, recorded calls, and their own emails that they sent me as evidence

She tried several times to tell me that there was some reason that I was not going to be awarded the refund, to which I proved it wrong every single time

I even made her sit on the phone for 25 minutes until I got her confirmation email saying that she was a supervisor, what her name was, when we talked, the conclusion that she reached, and that she was going to be awarding me $600 because of all of this.

[nimble heart]

"awarding" you mean returning what's yours

[high skiff]

Yeah, miss speak on my end

I'm going to get into contact with them tomorrow, to which if they do not settle, I'm sicking my credit union on them

[native knot]

From:

[nimble heart]

brb tacos

[high skiff]

Now, I've been told by several people who have dealt with these problems before with PayPal, that if they don't side with you and then you pull the funds away from them, they consider it a debt, and they'll send collections after you.

To which I say, try me, cuz I'll escalate it to small claims courts to where they will have to pay for my legal fees, and I'll bury them in evidence

[glad grove]

they did the same to me,i sold something to a guy,2 months later the scammer goes to the bank and does a chargeback,paypal says they know hes a scammer but i still gotta pay the whole ammount because he did a bank reversal

[vital ermine]

Yes, main reason I don't sell shit these days as people are being scammed like that left and right.

[nimble heart]

did you paypal the money straight to the guy or did it go through the site? If it went through the site I'm confused why paypal isn't just billing them

[vital ermine]

Reason I would never sell my vid card unless something drastically changes. They get the card (item) and their money which a lot of sellers complain about happening far too often to them.

[strong field]

@upbeat summit Just updated, AIT and Comfy, here's my runs without AIT, then with Lora. I am finding the same behavior as you with AIT+Lora and only after updating both Comfy and AIT

[upbeat summit]

Thanks for doing the tests! Very interesting. On my setup I have the delay issue since the beginning - so for a couple of weeks when AIT was first released.

[vital ermine]

@native knotYou have shit in there I have never seen before

IPA? Only IPA I know about is Inidan Pale Ale.

[native knot]

Bro... IP Adapter.

[vital ermine]

oh

so you aren't local then?

[native knot]

Local?

[vital ermine]

no other need for ip stuff

unless you aren't gen locally

[native knot]

I am generating these locally.

[nimble heart]

I think it's a different ip

not internet protocol

[vital ermine]

beats me, as I said, I have no idea about that ip thing

[strong field]

@upbeat summit this is so dumb, i can gen BS4 faster with Lora, still slower tha without tho.

[vital ermine]

whale fuck is the best kind of fuck. :/

[nimble heart]

soooo

you're a furry?

[native knot]

https://github.com/laksjdjf/IPAdapter-ComfyUI

[vital ermine]

Damn, Flava Flav V2

[vital ermine]

No, and if a furry whale pops by anytime lately I am running.

[glad grove]

tbh a furry whale is less scary than a whale furry

[vital ermine]

Ahhh shit. click link. Hit wall of Asian text. Yep, whale visited me so time to grab a smoke and get the translator

[nimble heart]

even a scaled whale would still make you a furry

furry degeneracy is like a tree where the main clade "furry" has all the subspecies like "scalie" and similar.

but they're all still furry

[vital ermine]

Not my thing tbh. Just never went that way with no want to do so.

after looking I will skip ipa

[glad grove]

maybe someone should make like a furry genealogical tree to see all of its branches and species and types of stuff etc

[strong field]

@upbeat summit I got lots of digging to do, but it seems this issue stems from these commits, separation of controlnet and state, i mean lora

[upbeat summit]

hmm... yeah that is strange

[strong field]

i can confirm before the updates this was not occuring

[high skiff]

because they keep saying my tracking info is illigitimate, even tho he has had the GPU for over 2 weeks now, and I got an email from them on the day arrived saying they SAW it arrived, and they are processing my refund

[pure crystal]

amelia bedelia

[upbeat summit]

do you have 3000 or 4000 series card?

[strong field]

3000

[upbeat summit]

that is very interesting to know. I have a 4000 series card and have that delay since AIT was first released. so it started for you on a 3000 series card with the latest updates

[nimble heart]

have AI write a script explaining the entire furry fandom phylogeny in the style of a Clint's Reptiles video

[strong field]

the timing is suspect, not sure why 4000 series wouldnt have worked prior

[upbeat summit]

they didn't. the delay was always there

[strong field]

ive had loras in my workflow since AIT released

[nimble heart]

email gpu dude's email to paypal and say "he has it go after him"

[strong field]

I have to check what Fizzledorf changed, (forced change due to the Comfy change in lora.py) Ill let you know if i find anything

[upbeat summit]

yeah maybe that will help to identify the problem. still not sure if I report it to the AIT repo or ComfyUI repo - maybe both need an update to fix this

[high skiff]

They know he has it lmao

[upbeat summit]

cool! sounds good

[strong field]

My path to resolution starts with identifying Comfy change (commit came first) and then review AIT

[high skiff]

she legit told me they knew he had it, but I sent it back past their return date, which is different from what I was told, which is who now she is "honoring" it

[upbeat summit]

yeah good thinking. if it happens on 3000 series cards now that may help find a fix for 4000 series cards too

[high skiff]

she basically told me "You decided to send it back without pay, so he gets the money AND the card"

[vital ermine]

was this your old 3060?

[nimble heart]

what tf site was that again?

[high skiff]

paypal

[nimble heart]

I mean the card

[high skiff]

hardware swap

[nimble heart]

most 2nd hand sites have buyer protection terms for that

[high skiff]

they banned him and barred him from the other sites

[high skiff]

Their buyers protection IS PAYPAL

[vital ermine]

lot of good that does you

[high skiff]

180 days

I opened a claim on day 3

[nimble heart]

ahhhhhhhh

great

[vital ermine]

yeah, PayPal is shit and began to become shit after the forced break up ebay and paypal. Just getting worse by the day.

[nimble heart]

so wait are they just giving you $600 or are they gonna dock him the $600 he owes you?

sounds almost like they're just going to credit you $600 and he still gets his $600 too

[high skiff]

no idea

I think what you said

[nimble heart]

I said two things which one

[high skiff]

we both get $600

[nimble heart]

Ah

well i guess it still works for you

[vital ermine]

Sounds about right

[high skiff]

even tho they ruled in my favor cause he scammed me

and he also agreed to a refund if I sent the card back

[vital ermine]

you are lucky as a lot of the times they now will tell you too bad, sucks to be you.

Anything else we can help you with?

[nimble heart]

my justice boner is extremely upset at this however

[high skiff]

I am not lucky, I have been dealing with this for 2 months now, my stress is high, and my medical issues are the worst they have been in years

[vital ermine]

2023, there is no real justice