✨｜sdxl

[urban fjord]

Blip2 captions are great, just look here:

[thin nova]

i'm also training blip2 with another lora. will see if that helps

[inner ruin]

instructblip is the shit

based on vicuna

worth checking out

[urban fjord]

I originally hand captioned that dataset, I just got suggested that I tried clip-interrogator on them to see what it gave me

The results I got from just training on a white image, nice brightening of the scene.

[delicate grotto]

you can train in comfy?

[trim orbit]

no. you can infer in comfy

[delicate grotto]

ah

[halcyon tusk]

Just a friendly reminder, we need everyone to go and vote on the bot channels, please select what image you like best: "A" or "B", this is for helping improve and finetune what will be the gold ⭐ SDXL 1.0 model.

[thin nova]

sometimes it's really hard to pick which is better

[high skiff]

wow lmfao

[autumn forum]

Do we get a prize for the person that has the most votes? Edit: actually scratch that, if we did get a prize people would just vote with out actually looking at the images

[high skiff]

oh, Diodotos is one of the people in our research server

[kindred plinth]

absolute gigachad

[thin nova]

do you accept lurkers?

[high skiff]

its not my server to link to, tho I wouldn't think so as we are testing things that likely lead to fails to try and find good results, and our whole idea is to not let that unfished info out to the public

[hasty axle]

anyone get this error on hugging face?

{"error":"module 'diffusers' has no attribute 'StableDiffusionXLPipeline'"}%

[eternal fog]

grrrr, it's so annoying. with the upscaling at the moment it seems like it's a choice of upscale enough and get good eyes, but wash the detail out of everything else. Or keep the detail, but have bad eyes.

[eternal fog]

That looks like you don't have the newest version of diffusers that supports the XL Pipeline

[hasty axle]

I'm just using the hosted inference api here https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9?text=A+summer+day+in+new+york

is this outdated @eternal fog ?

[eternal fog]

I think so as the how to section says "from diffusers import DiffusionPipeline

pipeline = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-0.9")"

[halcyon tusk]

exactly! You win an image of a complete A100 card for this realization 😜

[eternal fog]

I've not forgotten about you. I'm trying to fix a few issues in the workflow.

[ionic dragon]

How can we get into the research team?

[ionic dragon]

Oh cool, that's better

[high skiff]

Not too sure, we were all just kinda together already

[eternal fog]

The img2img pass is really annoying, if you go too far it doesn't fix the face and hands enough. But if you don't go far enough they it fucks the background.

and its between 2 steps where this change happens

[shy kelp]

What about LLAVA or minigpt4 for captioning?

[sudden cliff]

IMO Llava hallucinates too much

even at lowest temps

KOSMOS-2 is out

For something like image training captions KOSMOS-2 would probably be pretty great

[ionic dragon]

Minigpt4 is verbose

[tribal lantern]

I see the good old learning by repeating the failure of others, something you guys critized stability about not explaining why they made the choices they made, now you use the same fluff speak to do the same, god I hate the immature attitude the people involved in these AI fields have about open source, it's gtab and take for them, so glad that the (non ai) software fields I'm more active in are so much more open to sharing

[high skiff]

The information is shared with people who know how to use it. With all due respect to the community, stable diffusion users are not the most educated on the tools they use

The goal is a robust and effective tool/workflow for the masses, not the people that have the ability to make their own things as they please

And besides, I also do research documentation for things I don't properly understand for people to mess with, so please don't come at me/us like we are doing it for some form of walled garden or something. We owe nothing to this comminity, and dedication to make something functional for all is harder to do than just throwing out poorly documented and untested findings, if I do say so myself

[azure oxide]

this is a lot of text to say ur mad that u cant join someone's personal discord lmfao

[tribal lantern]

Congrats, you made sure they'll never be educated this wat, it's really a dumb argument, but as everyone says, you do you bro 🤡

[azure oxide]

its going to be fine

[lament rune]

how's the new card @high skiff

[high skiff]

also, as an addition, I have nearly 3000 words of documentation for my workflow specifically on how AND why it works

So if people wish to read it and understand exactly why these choices were made, they can feel perfectly fine and supported doing so

[high skiff]

dead

not completely, but effectively

[tribal lantern]

not mad, just baffled by that general attitude, it's totally contrary to what i'm used to in opensource development

[lament rune]

after all that!?

sorry to hear that

hope you get a refund

[high skiff]

yes, let me just put out information I do not understand with improper documentation and then expect people who know less than I do to fix it for me, rather than running around with my poorly made/optimized product and spreading misinformation on behalf of my name. How baffling an idea

[shy kelp]

it was nice not having a condescending self centered person around for 24h

[high skiff]

24hr? I thought you were only muted for 1

[shy kelp]

ive never been muted

[high skiff]

its not like dead dead, but its looking like its headed that way

I have a few more extreme measures to test first, but my hopes are not high

[lament rune]

Chaz the raver denies allegations of improper conduct online

[high skiff]

lmfao

A response from one of the members in the server about letting randos in lol

[stray mantle]

ComfyUI SDXL 0.9 personnal workflow. Can't wait for 1.0, damn ! 😄

[warm hazel]

1.0 has been the first model in almost a year that I've been eagerly awaiting getting my hands on.

Can't believe it's almost been a year since 1.4

[trail bay]

yeah it has been forever since 1.5 release. especially in AI time lol

[stray mantle]

This was among the 1st images I generated in Midjourney beta a year ago almost day for day...

[eternal fog]

This is what I have so far, with a custom node I've modified for aspect ratios.

The 2nd pass after an upscale isn't needed for all images. But it will fix faces if needed, with the downside that it tends to smooth out other features. I don't believe that's something that can be fixed though and is intrinsic to the model. It effects certain images more than others.

[shy kelp]

wild to see how far we come

[high skiff]

@azure oxide do you mind if I DM you?

[high skiff]

That is so insane, I remember using image generators before they were even diffusion based, back when Wombo dream was still running its clip master or whatever

*masher

[sudden cliff]

@visual glade did you know there's a major determinism difference between images generated under --gpu-only and normal vram?

I was worried it was a code regression but it's gpu-only or not

Reproduced back and forth, bigger than gpu non-deterministic sampler differences

[stray mantle]

Yeah I probably have some VQGAN+CLIP images from 2 years ago somewhere lol, back in the days it was close to dreams and imagination was doing the rest 😄

[high skiff]

Basically lol

Although I will say, it was impressive how much it could capture the essence of certain things, while being extremely abstract

[visual glade]

which samplers are you using?

[azure oxide]

nah go for it

[visual glade]

how big of a difference is it?

[high skiff]

I need to add you as a friend, that's why I wanted to ask

[sudden cliff]

dpmpp_sde_gpu, which varies from inference to inference but not as dramatically

will give an image compare, one moment

[visual glade]

oh on that one it's normal

[high skiff]

When trying to send a friend request, it says that you're not actively accepting friend requests from other people, so you have to send me one lol

[sudden cliff]

but specifically gpu-only has a different determinism for that one?

[high skiff]

Trying to DM you has more steps than my entire workflow lol

[azure oxide]

oh strange

[stray mantle]

french revolution on VQGAN+CLIP around june 2021 I think, generated on a Google Colab in several minutes lol, 400x296

[high skiff]

I think it's a privacy thing

[sudden cliff]

same workflow / seeds etc. Only varies slightly between inferences. gpu-only ALWAYS gives a nose, normal VRAM always gives the lack of a nose, for example

different arm placements

[high skiff]

That is extremely strange, that's a very big difference

[sudden cliff]

you get tiny differences normally but the gpu-only flag makes it very different, euler_a also changes between gpu-only and not

but slightly

[visual glade]

yeah that's why I split them into _gpu and regular

[sudden cliff]

is euler_a also non-deterministic?

[visual glade]

I don't think I made that one deterministic yet

[sudden cliff]

OK I thought it was because of the per inference difference, not the wider difference there

[stray mantle]

and this is now, a couple of minutes on local machine, ComfyUI + SDXL 0.9

[azure oxide]

sent

[high skiff]

thanks!

[high skiff]

what GPU if I might ask?

[hearty ginkgo]

Ain't just you sytan

Idfk why I'm using windows insider build tbh 😭😭😭😭

[high skiff]

g-g.... GREEN SCREEN!?

[eternal fog]

Yeah insider builds do that

So people can tell at a glance

[high skiff]

oh wow, I never knew that haha

@hearty ginkgoalso, same phone moment

[hearty ginkgo]

Lol I have to switch it to normal or sum I had insider build because it was a work around since I had a i7 7700k and it wasn't compatible but I upgraded to the i9 13900k so idk

[high skiff]

I also just came from a 7700k-

are you

ME?

[desert copper]

About time everyone gave Windows the boot and switched over to Linux hides

[high skiff]

I can understand the sentiment, but I am just not there yet

[hearty ginkgo]

lol it was soooooo bad djskanfjhnadsfjhnbdf the 4 cores killed me

[eternal fog]

Would if I could play games on it properly

[high skiff]

yessssss

I just switched to a 12600k, even that is a monumental improvement

[hearty ginkgo]

i would use linux but this is my gaming pc

[arctic bloom]

what stops you to play games properly? 😄

[hearty ginkgo]

lol i had a 1080ti and a i7 7700k switched to 3070 last year and just now i got the i9 13900k

[eternal fog]

Anti-Cheat, lack of proper support for Force Feedback Wheels

[high skiff]

bro lmao

I had a 7700k and a 1080, then went to a 3060ti, then this PC lmao

[desert copper]

If you must game, dual boot is never a bad idea, though I hear Steam/proton is getting pretty good these days.

[hearty ginkgo]

prob gonna get the 5090 or 6090 and still have my i9 13900k lol

[arctic bloom]

online gaming?

[high skiff]

we are just alternate universe versions of each other lol

[eternal fog]

Yes

[sudden cliff]

Side note- it's wild how much minute boilerplate prompt stuff that we've gotten used to affects SDXL outputs compared to SD... A few term swaps you think are benign can change an image from CG to hyperrealism

[hearty ginkgo]

lol have i have no choice to do it in parts lol dropping 1k on a gpu and then 1k on a cpu at the same exact time is hard but 1k on a gpu then 1 year later on a new cpu is easier

[high skiff]

for sure, 100%

[desert copper]

I hear you can drop $100 dollars on an apu and get similar results to a $1500 gpu.

[high skiff]

yeah, GPU's are improving much faster than CPU's at the moment in terms of capability improvements

[high skiff]

maybe $1500 GPU from 12 years ago haha

you could pull an apple say say your iGPU is as fast as a 4090, but really mean just in tools that can't use the 4090's compute lmao

[desert copper]

@high skiff you may laugh, but we're talking about an APU, not a GPU 🙂

[hearty ginkgo]

i have this motherboard https://www.amazon.com/ASUS-Gaming-Intel®-Motherboard-Thunderbolt/dp/B0BG6KQPWD the i9 13900k all the same lol and i got 2x32gb ram stick ddr5 at 6400maz

[desert copper]

https://level1techs.com/video/amds-hidden-100-stable-diffusion-beast

[high skiff]

second it can leverage the 4090, the M2 GPU looks like a graphic calculator lol

[high skiff]

I know what an APU is lol

tho that is very cool to see

[desert copper]

Good. So when I said a $100 APU, I didn't mean a $1500 dollar GPU from 12 years ago. Meant what I said 🙂

[high skiff]

I was assuming you meant APU as its more common consumer term definition, my bad

[hearty ginkgo]

Finnaly got koyka or what ever it'd called to work lol one Greenscreen and reinstalling python is all it took

[high skiff]

lucky ass

[hearty ginkgo]

idfk no more

[fair crow]

Anyone here ever tried running sd (training) on a tpu?

[delicate grotto]

ah yeah, part of the reason that i never update nowadays

[desert copper]

@hearty ginkgo This line: Tried to allocate 20.00 MiB (GPU 0; 8.00 GiB total capacity; 7.16 GiB already
allocated; 0 bytes free; 7.30 GiB reserved in total by PyTorch)

[hearty ginkgo]

i aint got 20 mib lol

[desert copper]

Your VRAM is exhausted, in other words.

[hearty ginkgo]

like always

[high skiff]

@desert copperI just watched the video you sent, and thats really cool, tho nowhere near a $1500 GPU

3it/s at 768x768 is far from the near 10 I get on my 3090 when its working 😅

[hearty ginkgo]

Can I not trian with 8gb of veam?

Vram*

[high skiff]

glad to see I am not the only one that edits messages multiple times and still has typos lol

[high skiff]

for SDXL, I don't think so

[hearty ginkgo]

Oh rip

[high skiff]

I know its been more optimized since I last tested it, but it was using 17GB VRAM for me at BS1 before

tho I have a friend that i think said its working on a 3080 now

[desert copper]

It's near enough when you don't want/have $1500 to drop on a GPU, but sure, it's AMD, and we're still not there with AMD drivers quite yet from what I understand. But price/performance ratio is about as good as it gets.

And it's also, what, 5 years old? So it does pretty well, considering.

[high skiff]

oh for sure, I totally agree

[hearty ginkgo]

Rip

I'm over it

[high skiff]

its a dope option for those with not much money, but a lot of skills

@desert copperoh my god the background of their website is satidfying

[trim orbit]

i figured out linux and if i can do it, hope for anyone

[desert copper]

Protip: don't be afraid of the command line on Linux. 🙂

[deep torrent]

I am getting this weird points/pixels in all of my images while using sdxl 0.9 with automatic1111. does anyone know why this happens?

[sudden cliff]

If you're into AI and still afraid of the command line it's going to be rough for you

[high skiff]

Thats the SDXL 0.9 watermark, its in all images (unless you use comfy UI, cause Comfy learned how to turn it off)

[trail bay]

can consider renting gpu or just using cloud -- especially if you only plan on training a little bit. or trying it out. cards will keep getting better. price-- hard to say since there is massive demand for the AI-capable hardware

[high skiff]

I saw a site with outrageously good prices

[hearty ginkgo]

how do i enable low vram for kohya

[high skiff]

it was like $0.40 an hour for a 4090 IIRC

[indigo carbon]

alright, i just tested out sdxl0.9, it's not as good as I thought it would be

[high skiff]

with like 12 cores and 128GB RAM

[sudden cliff]

I didn't know that one either, that's a really rough watermark

[high skiff]

you have to know what you are doing

[high skiff]

yeah, it is

[indigo carbon]

oh, i know, trust me

[high skiff]

it can show up as slight pixel discolorations, or in diffusers it shows up as a bunch of red pixels

[deep torrent]

I wasted so much time trying to fix it 😄

[desert copper]

speaking of watermarks, I hope we've got rid of all of those images with watermarks from the model.

Probably best to avoid anything with a watermark, not least those stock image libraries that are now suing Stability 🙂

Let them keep their walled garden. We don't need anything from it 🙂

[sudden cliff]

do you know which commit he did it in? Just curious what that looked like

[indigo carbon]

SDXL

1.5 model

[high skiff]

no clue, from what I know, it was stripped out from the start

bro, the new 4090 version of my 3090 is HIDEOUS LMFAO

my nice clean and sleek 3090

[indigo carbon]

I don't think I'm doing this correctly, it doesn't add up that a 1.5 model outdone SDXL0.9

[high skiff]

and the new 4090 version of it lmao

[high skiff]

let me have a peep

[eternal fog]

AssertionError: network for Text Encoder cannot be trained with caching Text Encoder outputs / Text

[sudden cliff]

tbh it's kind of hard to top something like that, but...

[indigo carbon]

hmm...

[high skiff]

oh, you need to use a special command

[hearty ginkgo]

ok thanks

[high skiff]

just a sec

you can't train text encoders yet for LoRA's

(we are working on that in the research server)

at least not well

[hearty ginkgo]

rip i turn that off and it jumps up to 20mib usage again ejhfbsdjhbfsdhjbsd

[indigo carbon]

Idk man, I was almost confident that SDXL could be better than 1.5 finetuned models in this kind of thing =\

[high skiff]

@hearty ginkgo--network_train_unet_only

[hearty ginkgo]

were do i use that?

[high skiff]

you need to add that to the optimization args in kohya

[hearty ginkgo]

like that

[high skiff]

it should decrease VRAM a bit as well, and make training timmes way faster

let me double check, just a sec

[hearty ginkgo]

k

[high skiff]

yes, there

for me, it was about 2x faster with that (as it didn't waste time on the TE)

[visual glade]

SDXL is great if you use it correctly

[high skiff]

for sure

[high skiff]

oh man, the stylization on that is fire

[sudden cliff]

here's something a little fancier

[visual glade]

it has to be used as part of a pipeline with the refiner

[high skiff]

this is one of the results from one of my LoRA training tests for SDXL

[hearty ginkgo]

egh you got any other recomandations to try to lower vram?

[high skiff]

I haven't done anything with that at the moment

you can try cache latents, cache latents to disk, and enable gradient checkpointing

[upbeat summit]

and na'vi do not look stretched in that format 😄

[high skiff]

yeah haha, they are very lanky as is lol

so this LoRA quality is with no text encoder, just pure unet

[indigo carbon]

SDXL

1.5 model

[high skiff]

it looks even better with my refiner I made which is the same LoRA trained in 1.5 as a very small pass

[visual glade]

are you using the refiner?

[indigo carbon]

yeah

[high skiff]

SDXL vs 1.5 fix pass

[upbeat summit]

1.5 looks great! the styles are not 100% comparable. I guess the overall style is interpreted differently

[agile quarry]

what kind of lora is best for SDXL training? just normal LoRa or ...

[high skiff]

I have only tried LoRA

but some of my colleagues have suggested LoHa's

[visual glade]

#✨｜sdxl message
Try that workflow or this one:
https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[high skiff]

omg that meeee

[indigo carbon]

yeah, this is kinda sad, because SDXL has about x5 more parameters and it didn't beat a 1.5 model for the most part =\

[sudden cliff]

for those that want to take the beach with them

[upbeat summit]

well is it base sd 1.5? because a fine tuned model would definitely react differently

[indigo carbon]

nope, it's a 1.5 finetuned model

but still

[upbeat summit]

let's see what a even more fine-tuned SDXL model can do 😉 I mean for a base model it is really great

[indigo carbon]

SDXL

1.5 model

[desert copper]

prompt: multiple mason jars on ornate table containing the cosmos, (containing dioramas of wonders of the world), ultrarealistic, ultradetailed https://cdn.discordapp.com/attachments/1101178530865352815/1131358264270000258/multiple_mason_jars_on_ornate_table_containing_the_cosmos_containing_dioramas_of_wonders_of_the_world_ultrarealistic_ultradetailed_steps-51_style-Photographic_seed-0ts-1689807218_idx-0.png

[urban fjord]

If you're testing things you know work well on 1.5 and then redo it on SDXL then of course 1.5 is going to look good.

[visual glade]

I think there's something wrong with what you are using because it's not supposed to look like that

[indigo carbon]

I used the workflow @high skiff made

[visual glade]

your images have a1111 metadata though?

[indigo carbon]

wha- I'm currently on ComfyUI, I think there is a save format or smth

[eternal fog]

Probably using that horrible extension that puts ComfyUI into Auto1111

[high skiff]

the one he linked IS mine

https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[indigo carbon]

yeah, that's the one I'm using rn

[high skiff]

it has yet to be updated, as I am fighting with my GPU

[high skiff]

can I see how you are prompting it in that case?

[indigo carbon]

yeah, hold up

[amber fulcrum]

SDXL - considering that it is 0.9 not-finetuned and we still wrapping our head around about prompting - I am happy with this model

[hearty ginkgo]

i give up sytan lol

[sudden cliff]

I think you can only go so far with 'simple concept, photographic'. Probably using an established realism upscaler would benefit more..

[urban fjord]

If you want to compare them ask for actual scenes and not simple concepts.

[shy kelp]

30 year old Julie with blonde hair and roots showing is emotional and breaking up with her raver boyfriend 31 year old chaz with short brown hair at the Leeds City Centre Rave Memorial in a scene from the BBC drama A Raver's Last Chance, Chaz pov looking at Julie, hella emotional, dramatic scene

[trim orbit]

optimus break dancing

[shy kelp]

do those with the bot?

[trim orbit]

in automatic1111

[shy kelp]

noice

[agile quarry]

my colab crashes allot using SDXL with comfys colab is there any startup argument that might help? i can do a highram runtime but it doesnt seem to fix anything on its own

[indigo carbon]

Idk man, this whole ComfyUI shtick isn't working for me very well. I think I'll wait for the full release and A1111 support

I don't like how it wants 2 prompts

[hearty ginkgo]

You don't need 2 prompts hyperthetically

[trail bay]

how is auto1111's support coming along (For SDXL)?

[jolly creek]

Using it right now

[shy kelp]

31 year old raver with short brown hair Chaz finds himself in a neo-sci-fi-futuristic neon Leeds in the year 2099 in a scene from the BBC drama movie Cyberpunk'd

[trail bay]

yeah seems very active

[indigo carbon]

extension?

[jolly creek]

Last two days was using extension, yes (running localy)

But from today tried out A1111 1.5

https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/11862

If you are interested

Working well

[spring fulcrum]

Is that using the 1.5 SD Model or is Automatic1111 at v1.5 itself?

[jolly creek]

Automatic1111 at v 1.5

using sdxl 0.9

[spring fulcrum]

nice... what improvements have you seen. did they make any changes to the UI?

[jolly creek]

For now UI is the same. Pretty much only improvement that I have payed attetion to is that I can load and use sdxl model

Some samplers are not supported when using sdxl

[indigo carbon]

With refiner?

[jolly creek]

Thing about refiner is that I don't see that beeing used in comfortable way yet. Only way could be using it throught img2img right now

But base model itself giving good results

[indigo carbon]

Oh, then I'll wait some more time. The base on it's own isn't as good as some 1.5 finetuned models and it's way slower =\

I guess after the full release this wouldn't be the case

I hope, at least

[jolly creek]

I believe so

[indigo carbon]

Also, when is the full release?

[jolly creek]

Maybe together with sdxl 1.0 😄 your guess is as good as mine

[visual glade]

SDXL should be almost as fast when sampling 1024x1024 as SD1.x

[indigo carbon]

It's about 30%, even on ComfyUI.

I usually get about ~16-20it/s. On SDXL with all optimizations I get about 5it/s

Idk, I think it would run better after A1111 works with it in the full release. For now it's all too experimental to get results as good as 1.5 models stretched to it's limits.

It does follow prompts way better than 1.5, but the detail isn't quite there yet

And it's slower. So I think I'll stick with 1.5 until the full release

[upbeat summit]

those are terrific!

[azure oxide]

is this at the same resolution for both?

[visual glade]

comfyui is pretty close to the peak performance you can get with a pure pytorch implementation, the way to go faster is stuff like AIT or Tensorrt but they are not that easy to implement in a way that is transparent to users

[eternal fog]

Looks like some weird advert for strawberry shampoo or something.

[trim orbit]

the refiner doesn't work as designed in automatic. the way sdxl was designed is you do enough steps in the first model to get an image built, then pass the latents to the refiner model to finish the denoising. you can use it other ways, but that's how it was designed.

Automatic's pipeline is limited to one model one prompt, so you're forced to do things the other way.

do the whole image from the base, let the vae cook it, pass it to img2img, load the refiner, set the denois to 0.4 or 0.5 and run it again. higher tends to cook too much. it's really sloppy this way. it should be done with latent space

[shy kelp]

yeah refiner for img2img gives you very diminished results compared to leftover noise in latent space

inb4 dumbfucks nagging at me once again for saying this

[indigo carbon]

will A1111 not make it like, load both the base and refiner? that's not how it's supposed to be

I think there should be another dropdown next to the checkpoint dropdown, but for the refiner

[shy kelp]

I wonder will it be possible to run this on auto next week by just doing a normal git pull?

[visual glade]

yeah the a1111 codebase might have difficulty handling multi model pipelines

[indigo carbon]

people will be so mad if they will have to click 3 buttons for using the refiner

no way A1111 will look at the refiner and just go ''nah, fuck that''

[shy kelp]

i just meant in general

[sweet bane]

im ready to answer any questions

[trim orbit]

git switch release_candidate
git pull

maybe update venv

[visual glade]

there's at least 2 UIs that can handle the full pipeline and there's probably going to be more next week

if your favorite ui doesn't handle it you can just switch

[trim orbit]

comfy i know. What's the other?

[indigo carbon]

Idk, he has no reason to. some 1.5 finetuned models are somewhat better than SDXL0.9, I already tested that before

[visual glade]

the vlad fork I think has it properly implemented or if it's not it's going to be because he's using diffusers

[trim orbit]

right ok

i don't think his ui loads both the base and refiner for a prompt still

[urban fjord]

Not sure exactly what you're referring to but please stop making bad GUI's a badge of honor.

There is no shame in wanting a good user experience.

[indigo carbon]

Idk about Vlad's UI. I tested it in comparison to A1111 with all the same parameters, optimizations, sampler, etc.. the image I made with Vlad's UI was not as good as A1111

[trim orbit]

i dont think that was his what he was implying. think it was more "auto better fix it" instead of welcoming the bad ui

[trim orbit]

a1111's seems to prompt to the openclip layer and vlad's seems to prompt towards the clip layer. that is my assumption from testing

good ui's on comfy prompt to both

[shy kelp]

cheers

i'll ask again when 1.0 is out lol

just playing with the bot until then

[urban fjord]

But sounds like it is still too early to switch to Auto1111 then if they still got issues.

[trim orbit]

that branch will be live by 1.0 release. wont need to switch

[shy kelp]

seems hella powerful to me, i mean so did the 0.9 but all the channels are solid imo

[indigo carbon]

what other UI is there that does that is there except Vlad's?

[trim orbit]

i should say good layouts for comfyui

[shy kelp]

did emad say theyre releasing their own ui? maybe not there were so many things

[frozen inlet]

are dynamic prompts and generate forever available for comfy ui?

[urban fjord]

They did open source the thing DreamStudio is using.

[visual glade]

generate forever is the unlabeled checkbox in the extra options

[urban fjord]

But currently SAI has no good UI of their own.

[trim orbit]

i think so . i'm not a comfy guru . just started

[trim orbit]

already is a stable ui released https://github.com/Stability-AI/StableStudio

[nimble heart]

If you have the tools inspect their "repasting" work. Also clean it with 99% isopropyl if you have it and let it dry.

[shy kelp]

is it new? or does it suck lol. never heard anyone talk about it

[trim orbit]

it's the FOSS version of dreamstudio. i think it's good. no updates since release though

[urban fjord]

It has no dedicated backend as far as I'm aware other than SAI's API

[indigo carbon]

it can't run anything. it's just a frontend. but a good frontend though

[high skiff]

I have been peenged

[high skiff]

I've actually taken a look at it, it's fantastic

[mild garnet]

When I try to train a model I get this error: src/tcmalloc.cc:283] Attempt to free invalid pointer 0x594800d99740. Anyone knows why ?

[high skiff]

I inspected all the thermal pads and everything

They make excellent contact with the active back plate, great surface area, great finish

[nimble heart]

also isopropl the pci fins and even the slot on the mobo if its used

[high skiff]

The GPU core and the VRAM temps are perfectly acceptable

[high skiff]

I did that first thing

[nimble heart]

I've fixed smoker & dusty cards like that

hm

[high skiff]

It's likely everything that you're going to list I've already done, but I'm 100% open to suggestions

[nimble heart]

is the PC still running fine when the screen blacks out or does music and stuff stop playing

[visual glade]

if it was an old gpu you didn't care about I would suggest the oven trick as a last resort

[nimble heart]

Wtf is the oven trick

[trim orbit]

did you pray to the elder gods?

[agile quarry]

i Can't use sytan's workflow with out it crasing it tried with --use-split-cross-attention and i got two images instead of one before the crash. any advice?

[visual glade]

oh you are on payed colab, you should use: --highvram

[agile quarry]

ok will try ty

[visual glade]

the oven trick is removing the heatsink and anything that could melt from the GPU and putting it in the oven at 380F for 8-10 minutes

[lusty raptor]

i thought that was thoroughly debunked?

[indigo carbon]

this isn't working. I just finished setting this up

[visual glade]

I have fixed a total of 3 GPUs this way

[trim orbit]

used to do that for poorly manufactured xbox 360's

it only fixes hardware failures where solder isn't seated right

[lusty raptor]

huh, nice. i finished two off. I know it was always controversial, but i still have a distant memory of "here's why it doesn't work blahblah"

[visual glade]

it's something you should only try if the GPU is broken and not under warranty

because at that point it doesn't matter anyways if it doesn't work

[mild garnet]

When I try to train a model I get this error: src/tcmalloc.cc:283] Attempt to free invalid pointer 0x594800d99740. Anyone knows why ?

[trim orbit]

and if it works you feel like a mad scientist

[hearty ginkgo]

Tey chatgpt

Try*

[mild garnet]

good idea

[shy kelp]

31 year old short brown hair raver Chaz sitting in the cold morning light in his Leeds apartment after the Leeds City Centre Rave Disaster, dramatic scene, hella emotional Chaz, powerful moment, from the BBC drama A Raver's Last Chance 1989

31 year old short brown hair raver Chaz from Leeds UK is guest starring on the new Game of Thrones series, dramatic moment

pretty accurate

anyone know how to use negatives on the bot

[urban fjord]

I'm testing out some LoRA training and I'm happy so far. (Top is with base only, bottom is with LoRA)
With the goal of being able to prompt X as Y a lot more consistently. So here Seth Rogan as Pikachu.
Neither in the dataset I used for training.

[shy kelp]

though i never know what to put for negatives lol

[mild garnet]

I tried using chat gpt, but it doesn't give me any specific answer and since I don't know anything about programation I can't fix my issue

[urban fjord]

You don't really need negatives other than if you want to remove something from your image.

[high skiff]

Thats the only thing my documentation for SDXL doesn't include, because I self admit not really knowing what I am doing with the negatives lol

SDXL and negatives are weird

[trail bay]

lol that error probably needs lots more information before being able to figure it out. like all the previous errors you have had up to this point that you might have ignored, the system, how you are training, what libraries you are using, etc.

[mild garnet]

okay I see

[shy kelp]

31 year old short brown hair raver Chaz from Leeds UK is the newest Avenger in a scene from the Marvel movie End Game Part 2: The Return

[fresh path]

I put my motherboard in the oven regularly because my room was so damp idk what was wrong with it but putting it in the oven or heating it up with a hair dryer worked

[trail bay]

based off the error it is from code executing in tcmalloc.cc line 283, tcmalloc is Google's custom implementation of the (C) malloc

did you do any manual compilation?

are you using up to date libraries during the compilation or runtime? sometimes bugs are fixed down the road

knowing the full stack you are using would help, but not sure

[mild garnet]

I followed a YT video and I used DreamBooth

And TheLastBen

From what I understand

I'm new to this thing so I don't understand everything really well yet

[trail bay]

okay looks like fast-stable-diffusion does have a dependency on libtcmalloc

[hearty ginkgo]

Can u ss the error u get

[trail bay]

are you using the most up to date fast-stable-diffusion?

can you update libtcmalloc? not sure if it that will help or hurt things

not sure if it really that library vs some code base that is just fucking up lol.

[mild garnet]

src/tcmalloc.cc:283] Attempt to free invalid pointer 0x594800d99740

[trail bay]

is this you: https://github.com/TheLastBen/fast-stable-diffusion/issues/2347

[mild garnet]

Nope

[trail bay]

are you using Collab?

[mild garnet]

But I get the same error

[mild garnet]

Yeah

[trail bay]

you might want to comment there and see if there is a resolution

[urban fjord]

While you're waiting you might just want to try another training program to see if that works for you. Given how long since update I don't think it would support SDXL anyway.

[mild garnet]

Since I'm new, I don't know how to properly train a model or a LoRA. That's why I've been following yt tutorial. So right now I don't have other training programm

[urban fjord]

There should be plenty of guides to using Kohya-ss gui and it's not too tricky to setup.

[mild garnet]

If it is of any help, this is the video I’ve been following: https://youtu.be/c6r25rT8DV0

[trail bay]

what model are you fine-tuning? is it even SDXL?

[mild garnet]

I don’t know XD

I don’t really know what I’m doing

[agile quarry]

i couldnt get the kohya gui to work in colab i use https://github.com/Linaqruf/kohya-trainer he has new xl notebooks if you look in the files

[simple thistle]

ehy guys, what i need to try sdxl in a 4g vram?

till now i did miracles with the gc

i run stable diffusion with this set COMMANDLINE_ARGS= --always-batch-cond-uncond --opt-split-attention --xformers --medvram

[shy kelp]

31 year old short brown hair raver Chaz sitting at the pub having a pint reflecting on the Leeds City Centre Rave Disaster, hella emotional scene, dramatic, focused on Chaz, from the BBC drama A Raver's Last Chance

pretty f'n solid for just a bot with a dummy prompting imho. hands not perfect but close enough lol

[green python]

(not cherrypicked 1st gen) combining sdxl and illuminati diffusion to achieve more detailed results

[simple thistle]

don't make jokes about his deformed hand, he is pretty sensible on the argument

[trail bay]

he has an amputated finger but that would be from a farming accident

[green python]

discord compresses the image a lot

[urban fjord]

I don't think Discord is doing any compression here.

[green python]

idk why but the image went from 768x1152 to 436x655

[urban fjord]

You need to click on "open in browser" to get full version.

[green python]

it's the same for me

wait

srry, yeah thx

[trail bay]

open in browser is the original

[mild garnet]

Thanks for trying to help me

[trail bay]

it might be just worth trying another tutorial/other software

[mild garnet]

Yeah I’ll try that tomorrow

Today was the first time I tried to train a model

[trail bay]

kool

[mild garnet]

I’ll do more research later

[green python]

MOOOOOOOREEEEE

[high skiff]

I have a question for the community

in regards to how you would all prompt a specific thing

sun rays going through foliage on a subject, I can't think of a proper way to tag that

[hearty ginkgo]

Linguistic: An enchanting photograph capturing sun rays filtering through lush foliage onto a subject.

Supporting: nature, ethereal, magical, backlighting, dappled light, atmospheric, dreamy, natural beauty

Negative: Blurry details, distracting artifacts, overexposed highlights, lack of clarity, lack of focus on subject, uninteresting composition, excessive noise, flat colors

[azure oxide]

sun rays going through foliage

[high skiff]

dappled lighting!

thank you!

I knew it had a term, i just forgot it

[fresh path]

god rays

[high skiff]

dappled lighting was exactly what I was looking for

[urban fjord]

cinematic shot of sunrays hitting Yann Lecun through foliage, god-rays

[high skiff]

I am actually using it for 1.5, so I am trying to find terms that work well there

dappled light is the actual correct term for it, though 1.5 does not seem to know what it is

[urban fjord]

You should use clip-interrogator if you haven't already.

[hearty ginkgo]

try
Speckled light
Streaming light
Radiant beams
Golden glow
Sunbeams
Illuminated foliage
Mottled light
Light shafts
Sun-drenched ambiance
Luminous patterns

[high skiff]

trying various things, none of them seem to be affecting it much

hmmm

[high skiff]

alright, I think this specific model just does not understand this as a concept

oh well

I am sure I could train a LoRA on it at some point if need be

[hearty ginkgo]

I have no idea why I got a painting vibe with that lol

thats much better

[urban fjord]

Training a LoRA is often the fastest way.

[spring fulcrum]

I honestly have no Idea how to train LoRAs or embeddings. I always seem to find a way to mess it up.

[agile quarry]

crepuscular rays

[placid coral]

Wait sdxl is a censored model???

Is it true?

[trail bay]

yes the training data is filtered. you'll have to wait for tits and ass until the fine tunes.

the vanilla model might be able to do some stuff

[urban fjord]

No, SDXL is not censored.

Popachu. (Base only as running the refiner is too heavy with LoRA)

[agile quarry]

i think the bot rooms here gave me nudity (not intentionally)

[trail bay]

but even with filtered training data it should be able to get a basic sense of the anatomy

[placid coral]

Great so I’ll have to wait until two weeks until sdxl is fully unleashed

[trail bay]

actually let me try local NSFW prompts. see what it knows

[placid coral]

Because it’s currently on a leash

[trail bay]

2 weeks? I wouldn't assume anything

[placid coral]

The internet is fast

[trail bay]

based off of what happened, there is no release date. just hopes the 'event' is next week

[placid coral]

1 week I give ppl

[placid coral]

Oh so 1 week delay is not official?

[trail bay]

event is something like 'super stage' . I have no idea what that even means.

just hints wink winks of the release

just don't get your hopes high. assume nothing

but likely it'll be within the next few weeks (the release)

[placid coral]

Well whatever, I think in 1 week after official release that’s when sdxl will be unleashed

[trail bay]

but you are correct, just wait, the model is great, it'll be fine-tunable, and you'll get what you want

[placid coral]

Because it’s currently the on a leash

[urban fjord]

It is not on a leash, apart from licencing of course, but nothing prevents you from finetuning it for yourself already.

[placid coral]

It’s gonna be SDXL unleashed after ppl take off its leash

[trail bay]

the person is talking about where 1.0 candidate inferencing is available.

and if those places are censoring. no big deal. we are on the same page

[placid coral]

Yup, basically this.

[trail bay]

the training data does have filtering too but Emad has hinted at 'no worries' by saying stuff like 'it learns fast'. it should understand anatomy well enough. I agree, let the community go wild soon enough

[urban fjord]

Yes, the bot channels have rules, but the models are not censored.

The training data has some quality control and should really have been filtered more to avoid too many duplicates and watermarks. But that's not censorship.

[trail bay]

are you suggesting the training data in no way has filtered out NSFW? lol

[agile quarry]

i just made boobs 😊

[placid coral]

Huge if true

Maybe SDXL is not on a leash after all

[trail bay]

it has. just not enough to where the model gets stunted into infantilism forever. It is a strong model.

[urban fjord]

Sure it has filtered some stuff, but that's quality control. You do not want the things it has filtered away in the dataset.

[trail bay]

@placid coral easy answer is, if NSFW cannot be done [even with fine-tuning], it is dead on arrival. and I feel like it won't be dead on arrival. rejoice

[placid coral]

Okay so it still has a filtering leash on… seems like we need a fat cat to help unleash the beast, so to speak

[trail bay]

I mean to include fine-tuning

base model might not be able to do stuff, but it should be very strong foundationally

to where you can make it do sooo much, more than it already can do, which is already very strong

[placid coral]

Yup, what this guy said.

[trail bay]

where it can do styles well, just via prompting

just patience. the tooling is being worked on simultaneously

[urban fjord]

Training tools are already ready so go ahead and train it if you want, just remember to follow the license.

[trail bay]

though it would be training 0.9

[urban fjord]

Just note that you will need to retrain on 1.0 when that comes out if you want to release stuff, but you can test the waters now.

[trail bay]

which I Think is fine honestly, especially if the wait is forever, but if it is just a week, then waiting for 1.0 is good

if I had more time off I'd def practice training on 0.9, a lora since pple have been doing that. Also curious about dreambooth, but not sure what the requirements would be for that

[urban fjord]

I just want people to test it for themselves so you can stop spreading the idea that SDXL is censored.

Please no more meaningless Dreambooth models that should be a LoRA model instead

[trail bay]

arguments revolving on semantics only go so far on the internet anyways, especially in chatrooms

[upper tangle]

genitals are not in the training data by the looks
I've been running it locally since release
tits and ass it'll give you all day, but genitals are replaced by blank skin or clothing you didn't ask for
I've never done a lora before so it was probably bad, but I trained a lora on 10 images of 1.5 outputs and while it artifacted and sucked (my fault probably) it DID add back genitals properly

so there shouldn't be any concerns

once people who know more than me start working on it, it'll give you whatever you want

though you might have to do a refiner lora too, since the refiner 're-censored' everything

[trail bay]

yeah in that case, the filtering is defended by 'safety'. like I Said, semantics. What matters is what matters -- practically speaking. The model will do what people want it to re NSFW [with fine tuning]. People shouldn't worry about that IMO.

[trail bay]

ohhhhh re the Refiner. very interesting

[upper tangle]

I didn't make a refiner lora (idk if you can? I assume so?) but I assume that would fix it

[trim orbit]

refiner doesn't change a pic much if you set it up right

[trail bay]

kool

[upper tangle]

refiner is very good at adding small details
but when some of those small details had to be lora'd in, it'll remove them too

[trim orbit]

all datasets are culled. even what runway ml created. omg censorship. its so cherrypicked. the self righteousness is .. just.. ugh. we know why you want the boobs. it's not any glorified freedom ideal

just lay off on the censorship thoughtspeak. it's dum

[upper tangle]

?
we aren't complaining
we are just talking about whether it's hard censored like 2.1 or just not in the training data

[trim orbit]

i'm over it

[trail bay]

2.1 was 'hard censored'? what does that mean

[upper tangle]

at no point have we complained that it's censored

[upper tangle]

loras/finetunes etc couldn't get it to do nsfw properly

[trail bay]

the training data was so filtered that the model couldn't be fine-tuned enough?

wow

[urban fjord]

2.0 had a mistake in the data filtering that was reversed in 2.1, it was never intentional.

[trail bay]

no wonder why 1.5 has all the attention

understood

[upper tangle]

which is the reason it's such a big topic around sdxl

[trim orbit]

big oopsidoodle there

it was filtering out boats

the embeddings i tested in comfy worked

in the refiner too

last i tried to load them in auto they failed though

[shy kelp]

31 year old short brown hair raver Chaz from Leeds UK is going mental rocking out on the stage with a glowing electric guitar at Glastonbury 2008, absolutely epic solo, mental, good vibes

[upbeat summit]

here we go again

[trail bay]

hey how often is the huggingface diffusers library used anymore?

[trim orbit]

https://tenor.com/view/roller-coaster-staring-funny-gif-22587939

[hearty ginkgo]

does anyone know why i get the grid lines with the ultimate upscaler

[upbeat summit]

35mm color photograph of 81 year old short grey hair male furious raver Chaz from Leeds UK going mental in the supermarket trying to buy bagels, mental, good vibes, still from 1972 british comedy film

[hearty ginkgo]

also if you zoom in on the black you can see another lion face showing

[urban fjord]

Yes, the ultimate upscaler splits your image up in several chunks and upscales them seperately and you see the lines between the chunks.

[hearty ginkgo]

how do i not split it into chunks?

[upbeat summit]

lion invisibility cloak

[azure oxide]

u dont use the ultimate upscaler

thats sorta the point of it lol

[visual glade]

you have to use mega ultra ultimate giga upscaler

[hearty ginkgo]

oh ok

[runic hatch]

hey, i am a little out if the loop and tried searching the discord and reddit already.
What is the best way to run sdxl when you need to use a cloud service? is it supported by google collab?

[shy kelp]

😂

[upbeat summit]

35mm color photograph 60 year old raver Chaz from Leeds UK going mental in the streets because he forgot to buy milk, mental, good vibes, still from 1972 british comedy film

[shy kelp]

I can feel that one coming

[agile quarry]

i use https://github.com/comfyanonymous/ComfyUI#colab-notebook

[visual glade]

yeah with that you should be able to run the base SDXL model on free colab but the refiner might be a bit too much

[hearty ginkgo]

are you messing with me?

[urban fjord]

I guess you run out of normal RAM on Colab?

[urban fjord]

But yeah ComfyUI works well in the cloud with the right hardware.

[visual glade]

yes but something with that name might exist soon enough

[upbeat summit]

@shy kelp I can't
35mm color photograph 40 year old raver Chaz from Leeds UK going mental because he stole pizza from the ninja turtles, mental, good vibes, still from 1972 british comedy film

[ionic dragon]

Thank you Arron

[urban fjord]

3 seconds per image in the cloud is really nice.

[upbeat summit]

seeing those images appear in front of me - priceless

[upbeat summit]

that's an amazing image!

[west breach]

finding some interesting images from the overnight run

[ionic dragon]

Where's pseudo I can't find him

I had to ask him something

[upbeat summit]

damn - amazing stuff

[west breach]

there's plenty of broken images in the batch too 😄

[upbeat summit]

bottom row center

[upbeat summit]

of course - that's part of the process 🙂

[west breach]

😄

[upbeat summit]

that portal wasn't good for the dino

great colors 😄

[west breach]

soldier in a tropical storm 😄

[upbeat summit]

now we are going terry gilliam style 😄

[west breach]

the glam style is inspired by kitsch quirky movies from the 70s 80s

I've found using RealESRGAN_x2 to upscale can cause similar colours to blur together into one blob

[upbeat summit]

who is @shy kelp the raver from Leeds UK? I tapped into his latent space life and he's so consistent lol (probably the british comedy token)

[peak dove]

SDXL 0.9 VladAUTO1111

[upbeat summit]

very pretty

[peak dove]

So, SDXL Base is txt2img; then the image produced here is fed to refiner - so the 2nd stage of SDXL is img2img

[peak dove]

Thank you 🙂

[agile quarry]

the refiner still takes prompt input. it can be distinct from your main prompt too. also it can be setup so that you dont really get a image in between models the refiner can take over and run the last steps before sending it to the decoder and outputting an image.

[tame osprey]

Any tips on trying to get it to edit NSFW content? I uplosded an image and it was immediately flagged for being 'inappropriate'. Probably because it was haha.

[dense chasm]

anyone knows how to adjust the denoise value between 0 to 1.0

[spring fulcrum]

I know this is an SD server but did anyone notice that ChatGPT using GPT 4 has upped the limit from 25 messages every 3 hours to a limit of 200 messages every 3 hours?..... That is awesome

[west breach]

I thought it was 50

[dense chasm]

code interpreter is awesome

[spring fulcrum]

mine says 200

[hearty ginkgo]

WTF HOW

BRO THIS 25 cap is such bs

[spring fulcrum]

log out and log back in see if it changes

maybe even clear your browser cache

they are also well known for rolling things out in small batches at a time

[hearty ginkgo]

i cleared site data and all

did you apply for a higher cap?

because i legit have the gpt 4 api but i cant get a higher cap its bs

[spring fulcrum]

I would say give it a week max and you will have it... Nope... I had no Idea they were even going to do that... I just happend to notice when trying to get chat gpt to help me with some error code for this LLM I'm working with

[hearty ginkgo]

bs

[spring fulcrum]

If you are directly using the GPT 4 API the only limit is your wallet... as long as you have a large enough amount set and deep pockets to go with it they let you use it as much as you want.

[hearty ginkgo]

oh ik that part

i got aproved for 120$ a month limit but the api playground and chat gpt itself is much different resualts

[spring fulcrum]

Right now I'm trying to setup the Oobabooga UI with the following picture and it doesn't like it. It keeps saying I'm out of RAM not VRAM so I'm doing a fresh install:

[hearty ginkgo]

what gpu you have?

[spring fulcrum]

It's metas new llama 2 70B model

I have an RTX 4090 and for ram I have 64GB DDR5

[hearty ginkgo]

if you want to try it you can always make a nat.dev account

[spring fulcrum]

Whats that?

[hearty ginkgo]

all the chat api's into a open ai playground and chatgpt chat box

[spring fulcrum]

That looks interesting... Does it cost anything?

[hearty ginkgo]

5 dollar min

[spring fulcrum]

Cause I just put the last 15$ I had in the gas tank today

well its all good... I'll just keep trying to wrangle this beast into working for now... all it costs me is time and electricity

[hearty ginkgo]

i would just let you log into my account and finish the credits i have left off but i connected it through gmail rip

whats crazy is the claude ai is 100k token limit

[spring fulcrum]

Its all good I wouldn't do that to ya....

[hearty ginkgo]

lama is 4096 token limit

[spring fulcrum]

Ya claude looks awesome

Still nothing produces code like ChatGPT on GPT 4 with Code Interpreter

[hearty ginkgo]

give me something to ask chatgpt and the new lama i can do a comparision response

[spring fulcrum]

hmm

[spring fulcrum]

ask it the woodchuck rhyme and see if either of them answer correctly

[hearty ginkgo]

hp;y shit this lama ai is slow it did 19 chars a seccond gpt 4 did 69

[spring fulcrum]

for math ask them to solve 2x+6=12

They should get x = 3

[hearty ginkgo]

lama si stupid

[spring fulcrum]

lol

[hearty ginkgo]

but claude get it right i think

Michael scores a 95, 87, 85, 93, and a 94 on his first 5 math tests. If he wants a 90 average, what must he score on the final math test?

Possible Answers:
86

88

96

84

90

ill ask it that

claude got it

[upbeat summit]

later all!

[spring fulcrum]

So I was able to get the 13B paramerter model to load with no problem. I tried the 70B parameter model. I think its just too big for my little 4090. I'll have to wait for the 34B Parameter Model to drop in a few weeks. They said they needed to work on that one for "safety"

[dense chasm]

chatgpt supports different languages best so far,LLaMA is far way behind

[spring fulcrum]

they just dropped LLaMA 2 today... maybe its better

I barely speak english well enough so other languages are a mute point for me

[hearty ginkgo]

This Greenscreen bs is killing me

I can't wait till I get my gen5 m.2 drive and can just do a clean boot 😅😂😅😂🤣

[hearty ginkgo]

Based of that you will never be able to run got 4

It's just so big

[spring fulcrum]

Well if it was open sourced and someone created a 4-bit quantized version you might be able to run it on a single H100 with 188GB VRAM or maybe even 48GB A6000 who knows

I need some Ideas for Pictures to generate.... Any suggestions are welcome

[hearty ginkgo]

A lion

[spring fulcrum]

This would be one hell of a hitchhiker

[hearty ginkgo]

Bro he'll nah

Try sum like that

[hearty ginkgo]

Is this in sdxl?

[spring fulcrum]

yep

[hearty ginkgo]

What workflow are you using 😭😭😭😭

[spring fulcrum]

Just try to steal from this candy store

[hearty ginkgo]

Bro your shit is so much better then mine

[hearty ginkgo]

What's your prompts?

[spring fulcrum]

Install this... https://github.com/ssitu/ComfyUI_UltimateSDUpscale

Then press the clear button....

Then drag the picture of the lion into your comfyui after closing it out and reloading once you install the thing I mentioned.

When you drag the picture in it will give you exactly what I used

[hearty ginkgo]

mine is the same as yours i think

[spring fulcrum]

that looks really good to me it might just be arranged differently than mine

[hearty ginkgo]

this is the best true black i got today

[spring fulcrum]

That looks really good

[hearty ginkgo]

by anychance do you know why i get those black grid lines?

in the black area?

aslo what is your ultimate sd upscaler settings just ss them to me if you can please?

[high skiff]

Did somebody ping me?

I saw a ping in here and I can't find it

[hearty ginkgo]

Somebody did but then removed it they said you should bring your pc to best but or sum

[high skiff]

been super busy all day, haven't even been able to do much for the GPU

[hearty ginkgo]

Dam

[spring fulcrum]

I'm sooooooo bored tonight

and I can't sleep

I'm not even tired and its 2:30AM

[hearty ginkgo]

Same lol

I have work at 9 am and I never sleep so I usually end up going to bed at around 4 am

[spring fulcrum]

same

ADHD FTW

[autumn forum]

Lmao same

[hearty ginkgo]

Lol

Scorp I got a challenge for you

[spring fulcrum]

Whatcha got

[hearty ginkgo]

Generate a car carrier it's legit impossible

[spring fulcrum]

ok out of 16 images these 2 were the closest to some sort of messed up reality:

[hearty ginkgo]

Lol told you it's impossible

Never got anything close on stable diffusion but an hour ago after like hours and days of trying I got one semi good on midjournery

[spring fulcrum]

if only ComfyUI had some sort of image to image then it would probably work

[hearty ginkgo]

those are the 2 best i ever got but it took prob 2 months of trying over and over on midjournery

either way its fuck

aint no way i got a watermark using midjoruerny

[hearty ginkgo]

try doing truck carreir with cars in the back

[hearty ginkgo]

Wtf did you use a prompt for that lol

[spring fulcrum]

A photo of a demon a dark demonic scary landscape

[hearty ginkgo]

Question what's the difference from those 2 different batch counts

[spring fulcrum]

the one on the left will generate multiple images at once and put them in a grid. this takes longer and seems to use more RAM... the one on the right will generate images one at a time but will give you the amount of images you put in there. I always use the one on the right to generate them one at a time

I think its faster and never crashes my system

[hearty ginkgo]

If I edit the batch count that's right under quee prompt with it still show them all in the save image node?

[spring fulcrum]

it still saves all of them to whatever folder you had it setup to save in but it won't show you them all at the same time

try it and you'll see.

[hearty ginkgo]

I'm in my bed lol got tired at sitting at my desk xd I have no glass on my pc case so that section of my room is hot asf but my bed is nice and cold yk

[spring fulcrum]

I'm pretty lucky... It's nice and cool here tonight.

Last week sucked

[high skiff]

just a heads up, if you have the PC resources to do it, doing multiple images in the same generation does infact save time

[hearty ginkgo]

Well I live in Florida and I'm convinced fpl is turning my ac off and on

[high skiff]

but if you do not, then turning up right side batch size

[spring fulcrum]

I should give it another go. I upgraded my setup since the last time I used batch size instead of batch count

[spring fulcrum]

I feel like hes about to say "Beuller", "Beuller" lol

[sharp robin]

@high skiff whar happened to pseudo? And did ur 3090 get fixed?

[high skiff]

Pseudo left the server, and I have not at this moment

[sharp robin]

[lusty raptor]

i'm getting Forbidden Zone vibes out of this

[hearty ginkgo]

@spring fulcrum try to recreat this

[west breach]

yeah, nice

[high skiff]

just casually genning 32x 600x904 images at the same time lol

comfy UI is so efficient man

[spring fulcrum]

ComfyUI Is soo good I can finally show pictures of my mountain girlfriends lol

[trim orbit]

going to refine this terran style lora more with 768 resolution overnight. just as an experiment . see y'all on the other side. goodnight

[spring fulcrum]

@high skiff Ya doing the batch size is better.... However when I tried to do 12 batch size My system went OOM when they hit the upscaler

[high skiff]

oh yeah, upscaling at high BS is intense

[spring fulcrum]

Is that a RAM issue or do you think thats VRAM or maybe page file? I have 64GB DDR5 and an RTX 4090 with an i9 not that that really comes into play but its a pretty decent setup and it still goes OOM

[high skiff]

likely VRAM

its very intense

what res were you going to?

[spring fulcrum]

4096*4096

with BS of 12 images

[hearty ginkgo]

scorp i got a photo idea for you

[spring fulcrum]

whatcha got

[hearty ginkgo]

a trex dinorsour eating a human

:)

as realistic as possible

[west breach]

even a single 2k image using up all my 24gb of vram when decoding

[hearty ginkgo]

idk why you got that issues im 4x with my ultimate upscaler and i can do batch of 12 on a 3070

[spring fulcrum]

are you upscaling to 4096*4096 too?

[hearty ginkgo]

i think so

i think the max i tried with the upscaler was 10 let me try 12 right now i think i did 16 before though so im not sure im do a 12 real quick

[spring fulcrum]

Your setup is on the left, mine is on the right... Maybe that makes a big difference in how it uses the resources... I haven't adjusted mine. I got my setup from someone else and just left it as is.

[hearty ginkgo]

yes 12 works for me

the blur isnt the difference but use linear for mode type based of @high skiff said to me if its none it dosnt do any upscalling

[spring fulcrum]

your denoise, mode type, and mask blur are different than mine

[hearty ginkgo]

the blur was because i got some weird grid lines when doing all black images

[high skiff]

denoise 0.5 is way too night

I try to never go over .1

[hearty ginkgo]

denoise sytan said should be 0.075 and linear for mode type

[high skiff]

thats my findings at least

still not ideal

[west breach]

it's super sensitive, even really low numbers I was seeing eyeballs show up in dark areas

[hearty ginkgo]

Is that because of denoise?

[west breach]

yes

[hearty ginkgo]

So I should just make it 0 technically

[west breach]

then it's not doing anything

[spring fulcrum]

let me give it a try with a high batch count now... Ill set it to 24 just to see if it breaks

[hearty ginkgo]

It's just Denoise though

[high skiff]

that makes it do nothing

[hearty ginkgo]

But is denoise nneeded?

[high skiff]

yes

denoise is how much noise is added in order to refine new detail

0 means 0 noise, and no changes

[hearty ginkgo]

Oh

Isn't this noise and denoise removes it

[spring fulcrum]

so far so good with a BS of 24

[hearty ginkgo]

Are you trying the trex eating a human?

[spring fulcrum]

trying but I think its no dice on the eating a human part

[hearty ginkgo]

You have the 4090 you can prob do a 49 batch max without it crashing tbh

[hearty ginkgo]

Rip lol

[spring fulcrum]

lol

[hearty ginkgo]

Those are the best I got last night in midjournery in stable diffusion I got pretty much just drawing of a trex lol

[spring fulcrum]

I mean this dude looks like he's about to be eaten

[peak dove]

Is this "the greatest of all time" or is it a GOAT?

[hearty ginkgo]

Lol

[spring fulcrum]

totally both. Those were my GOAT GOATs

@high skiff I'm not sure if it hung up here or if it just takes a long time to see the progress bar move

[high skiff]

honestly, you are playing with fire, so just be careful lol

[spring fulcrum]

with the 24 image BS? my cpu and gpu temps are like 45 degrees Celcius

how long does it take yours to process that many?

Well anyways folks I'm going to let this batch sit overnight and bake.... I'll let you guys know if I fried my GPU in the morning.... well... later this morning anyways.

[uncut steeple]

Gave it a quick try with 0.9, not perfect

[magic frost]

1girl,hat,longhair,bag,walking,solo,blackhair,shoes,simplebackground,orangebackground,(cat:1.4),backpack,bluefootwear,profile,socks,shorts,baseballcap,longsleeves,blush,whiteheadwear,fromside,leash,hood,fullbody,bluesocks,lora:tuyafengge_20230707170048:0.9,

1girl,hat,longhair,bag,walking,solo,blackhair,shoes,simplebackground,orangebackground,(cat:1.4),backpack,bluefootwear,profile,socks,shorts,baseballcap,longsleeves,blush,whiteheadwear,fromside,leash,hood,fullbody,bluesocks,

[uncut steeple]

Hey there, you can generate images in #1100170312106127410 for example

[uncut steeple]

Another quick and dirty try

[mortal fossil]

hi anyone knows how to do img2img using sdxl? It seems can only refine a picture, but I want to change the whole style of original one

[nimble heart]

More denoise

[peak dove]

Is your GPU working now?

[high skiff]

Nope

Although I haven't really been able to work on it at all today, I've been very busy with other things

[livid cradle]

has anyone tried placing a real product and generate the surrounding scene around it with SDXL masking REST API function??

[high skiff]

I already submitted a goods and services claim with PayPal, they haven't covered

*they have me covered

[livid cradle]

i get quite amazing results, but sometimes the masking not doing the trick, and it generates another bottle behind my original product. can anyone suggest some tricks to make the original product blend with the generated environment?

i know it's working to geenrate okish results with SD 1.5 and 2.2 but im interested in this new engine to do it as the results are better.

im building a product photography application, do you guys find it interesting ?

[ionic dragon]

the text is ai generated?

[livid cradle]

no, that bottle is the original bottle of bvlgari man in black eau de parfum

i got some generations with text in it, not the best, so i must have the original product integrated in the scene.

this one is fully generated by sdxl. text included.

yet that is not the orignal bottle of the brand..,

[uncut steeple]

Whats your setup for generating with masks? Im intruiged

[livid cradle]

what do you mean the setup?

im sending the init_image, the mask source and the mask_image the mask source im using black and white

[peak dove]

So you were sold a lemon! Sorry to hear!

[vale eagle]

https://civitai.com/models/111435/latentbyratio-comfyui-jnode-sdxl-sd15-sd2x It contains a img2img and upscale workflow and two costom nodes. For the load upscale model, you could replace with original one.

[mortal fossil]

Thanks man, I will try it!

[uncut steeple]

I mean can you share your comfyUI setup for example?
I never used masks so I kinda wanna try it

[livid cradle]

i don't have this made in comfyUI, don't know how to do it there... i coded everything on a backend.

if anyone could share a workflow with sdxl masking in comfyui id be glad to give it a try there maybe i can spot some parameters to help me fix this masking issue.

I already reported this to SD team that a bug is there with the masking in sdxl.

they noted, and hopefully we can have a fix for it

another thing that's interesting, even if my mask is fully black on the subject, and i only want SD|XL to modify my white area sometimes i get strange modifications on the product subject, with a mask applied fully black, no spots left.. it just don't care about it.

it does not fully change my original image portion, but it tends to do freaky stuff on the text.

im testing with Alpha_channel mode see if that improves anything.

if any smart people here want to join this conversation and give me some good ideas to achieve the perfect background scene change without adding extra artefacts to the orgiginal subject, id highly appreciate it.

[stray mantle]

anyone knows if in ComfyUI there's a custom node existing that acts like a branch killswitch ? Like I have multiple branches on the output of a node, for example different upscale models that are followed with different flows and I want to test a single one, I'd like to mute the other ones. I'm trying to create it myself, I have a node with a boolean and upscale model inputs, and upscale model output. If bool=1 it continues, but I can't stop the branch if I put it at zero as it seems it absolutely needs to continue.

[soft zealot]

yup can Mute Nodes with CTRL+M but I find that deosnt stand out.

My method is to use what I call "jumper" (cos they remind me of old school dial up modems)

Ie Rerout nodes that can be used to break the flow

and with the IMG2IMG branch "jumpered" on

[royal fern]

you sometimes get errors with jumpers... but I tend to do the same

[stray mantle]

Ok so you just cut the link between 2 nodes ? There's not really need for a special node for that I guess ? If you click on the disk in the middle of a link between 2 nodes you can delete it

(it's probably cleaner visually using your method though)

[soft zealot]

yup , just delete the link if I dont want that branch to run

[vale eagle]

usually Ctrl m is enough

[stray mantle]

Yes I didn't know this shortcut, time for my crap dev talents to stop 😅

[soft zealot]

I just dont find that stands out as clearly personally but yes it does work

Another example of where I use it is to switch between "Standard" and "Enhanced" Prompting

[peak dove]

I was getting so many Xformers errors with Vlad AUTO1111 - I now have Cumfy - it is so easy to setup 🙂

Where to put sdxl Diffusers in CumfyUI?

[elfin cobalt]

Can you share the configuration of your 'prompt style' selectors? I'm adding a function like that to my bot, and it'd speed things up a lot.

[peak dove]

My first CumfyUI Output (not yet sdxl)

[uncut steeple]

But this is the SDXL Channel

[peak dove]

Yes I agree - where do I place the SDXL Diffusers inside CumfyUI?

[magic crescent]

also completely random question but how did you get your nodes to connect with straight lines? xD

[wicked frigate]

use the regular SDXL model, not the Diffusers version

[peak dove]

Safetensors? OK

[uncut steeple]

Its in the settings

Link Render Mode

[dusk mica]

whats the best comfy ui config rn for sdxl?

[uncut steeple]

Im using this with Searge's SDXL nodes

[dusk mica]

thx

[uncut steeple]

Not saying its the best, but I had the most success with it

[dense chasm]

stability updates their API website with sandbox

but still with the xl beta model

[peak dove]

I got SDXL in CumfyUI up and running in 10 minutes 🙂

[edgy pollen]

What video card is recommended for SDXL? I want to generate 1920x1080 images with it. I can cope with slow generation speed but I think that my current card RTX 2080 8GB may throw CUDA OOM errors

[elfin cobalt]

More memory is better. A 3060 12GB would be the cheap option.

Just make sure you don't confuse it for the 3060, which is a completely different card.

3090/4090 are ideal, if you can afford it. More for future-proofing, but if nothing else, having the memory to run SDXL and a web browser is nice.

[edgy pollen]

3090 is too expensive for me. Thanks for recommendation.

[soft zealot]

Budget Option 1080ti , has 11Gb VRAM and just works

[peak dove]

I have an RTX2070 8GB VRAM - generates good quality SDXL, and then I upscale using Topaz GigaPixel - much cheaper and quicker than a 1920x1080 native off-the-board generation

[edgy pollen]

Up scaling is nice but original generation can generate more details

[peak dove]

Agreed - but how much detail is enough?

[soft zealot]

theres no benefit in generating much past the trained sizes (512 for SD1.5, 768 fo SD2.1, 1024 for SDXL) as it starts generating mutations & errors

it may work but..........

[peak dove]

GigaPixel does a fine job - fit that against the expensive hardware to get 1920x1080?!?!?

[soft zealot]

Bog standard 1024x1024 in SDXL on my 1080ti takes around 60-70 seconds, on a 4080 would be 5-10 seconds probably

[peak dove]

But I guess the way things are going, it'll be all native 8k video soon ... ! 🙂

My RTX 2070 8Gb VRAM is getting me one 1024x1024 every 12 seconds

[uncut steeple]

4090 with 30 steps dmp sde takes about 7-8s

[dusk mica]

4090 is also lightyears ahead of a 2070

[dusk mica]

nice

what workflow?

[peak dove]

With the batch processor on Vlad A1111 - I can set up 100 or 200 different prompts, leave it grind away overnight, and by breakfast - hey presto!!!!

[peak dove]

You know some peace!

[stray depot]

Takes 21 seconds on my 2080ti. So that's good to know.

[floral island]

mild variation of my prompt from yesterday, just trying out artist names 🙂

[halcyon tusk]

nice work! @peak dove

[peak dove]

Thanks, just putting CumfyUI thru its paces

[halcyon tusk]

cumfyUI is the nsfw version? 😂

[peak dove]

Tee hee

[halcyon tusk]

is it all comfy?

I like the color pallets and the african motifs before

[ionic dragon]

something that @visual glade should address

[ionic dragon]

the one I posted?

[peak dove]

ComfyUI is better than I thought it would be - and so easy to setup when compared with Vlad AUTO 1111 SDXL

[ionic dragon]

its way faster

[peak dove]

Its way faster to setup for sure

[ionic dragon]

we can only unlock SDXL to its full potential using comfy

[uncut steeple]

It also launches faster, im impressed every time

[ionic dragon]

exactly

A1111 takes around 2mins

[peak dove]

You want the prompt?

[ionic dragon]

comfy take 3secs

[halcyon tusk]

if you want to share the setup with the community great! I was just liking the work, it's looking nice

[peak dove]

a vivid watercolor depiction of diverse Rococopunk Afrofuturist women with beautiful and bold head wraps walking toward a huge moon in the background, holding hands, walking away from the camera, collaborating in a creative and productive environment, women empowerment poster, photography, inspired by the styles of victo ngai and vladimir kush

[uncut steeple]

Its art, I swear

[halcyon tusk]

now I want to know how those whisky glasses are prompted!

hahaha

[uncut steeple]

should have comfy meta data in it

[halcyon tusk]

oh really? i thought discord erased all that

[uncut steeple]

They reverted that iirc

but here is the prompt anyways : (Photomontage:1.2) A jarring image where elements of various photographs collide and coexist. A cat with the wings of a dove leaps from a toast, above a sea teeming with swimming umbrellas. It's a realm where the laws of nature are suspended and the impossible becomes real

[indigo carbon]

hey, I used the A1111 branch that supports SDXL0.9 base. I tried figuring out how to use it myself, but it gives me this error, help?

[peak dove]

Whenever you get overwhelming errors like this just delete the VENV folder and run webui again

A1111 uses diffusers - u must set the diffusers checkbox in settings

[eternal fog]

Discord doesn't. But there is a bug in chromium browser's were it changes the encoding of it if you copy paste from the browser window and it stops it working.

[peak dove]

Add this argument at the start of webui.bat file - webui.bat --upgrade

[peak dove]

Chrome has its own proprietary "brand" of jpeg called .webp

[peak dove]

A lot of places (like Discord, Photoshop) hate .webp!!!!

[uncut steeple]

I do too

[peak dove]

Kewl

[sonic furnace]

Doesnt discord use webp for all of their guild icons etc?

[floral island]

good old hokusai 😄

[timid sonnet]

Im not sure whaat hokusai is 😛 but this is a tattoo I have, I was looking for new ideasa

That's an old pic though, half of it is shaded in

[floral island]

hokusai tidal wave, very famous painting

[timid sonnet]

Ohh! nice

that pic is dope

[floral island]

that's the original

[timid sonnet]

oh shit AI definitely used that as the basis for the image it generated me

the top wave is like identical