✨｜sdxl

[livid phoenix]

alright

[deft turret]

Put male or man or boy in your positive prompts. If you still get females, you can put 'female, woman, girl' in negative prompt

[livid phoenix]

Though... How exactly do I edit the prompt so I get something like... Marx from Kirby Super Star?

[deft turret]

Combining things can also be done in other ways like through prompt editing

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing

If you're new to auto1111, I'd suggest you at least skim through this whole page of features and try to experiment with things you like.

[deft turret]

You can write exactly that, and put in things you don't want in negative prompt. But I should tell you that many checkpoints will not even know what characters are you talking about. They will make generalised images. That's why we use Loras and TI for specific characters and styles.

[livid phoenix]

Okay, so how exactly do I make a LoRa and train it to recognize Marx from Kirby?

I downloaded Detail Tweaker LoRA. Where do I put it?

[high skiff]

My GPU is still working, so I am gonna say I think we did in fact fix it

[deft turret]

In you stable-diffusion-webui folder, there's a folder called models, in that there's the lora folder. Put it in there

[livid phoenix]

Alright

As for the Textual Inversion?

[deft turret]

Training a lora won't be practically possible on your PC without GPU. You'll have to use Google Colab for that. You can find videos on youtube for lora training

[deft turret]

In stable diffusion webui folder.. Look for embeddings folder. Put TI im there

[livid phoenix]

Thx

[deft turret]

I'd also suggest to not leave the negative prompt empty. At least put "(worst quality, low quality:1.5)" in your negative prompt. That will make your images wayyy better.

[livid phoenix]

Got it

[deft turret]

Alright. All the best

[peak dove]

Weigh two geau

[peak dove]

Weird Gingerbread Look (like Mulhouse in France) SDXL 0.9

[spring fulcrum]

Can LoRAs, Embeddings, and Models be trained with pre-existing photos or do they need to be done with photos that are being created at the time of training???

[nocturne dove]

You make a dataset of the photos you want to train on and then train a lora or embedding from that. So, yes, you would use pre-existing photos.

[spring fulcrum]

awesome... I was just wondering if me creating like a hundred of several different styles without actively trianing something was wasting my time... This is awesome... Its like I am prepping now.

[nocturne dove]

These videos are older now, but I seem to remember them doing a good job woth explanation: https://github.com/bmaltais/kohya_ss/#tutorials

[spring fulcrum]

I'm waiting until this is all possible in ComfyUI at least for now. I do also have VLAD A1111 and A1111 installed and running though

[nocturne dove]

It's also worth pointing out that the process for sdxl is not really any different than 1.5 except that you would want images at least 1024x1024.

I train LORA using using kohya but then use the loras in a1111 and comfyui. I haven't found a more reliable way than using kohya.

[spring fulcrum]

They are. I have folders for the originals at 10241024 and folders for 40964096 upscaled

I keep my folder structure pretty well organized... every different prompt also gets a new folder for both originals and upscales

Can I train my own models with the upscaled images and have a model that will out of the box use 4096x4096???

[nocturne dove]

It's pretty unlikely that that will work at all since higher res uses more vram and you'll just run out at some point. I used to use 768 for sd1.5 and now use 1024 for sdxl. However that doens't mean that you have to use square 1024x1024 images. The trainer does kind of crop/scale to appropriate sizes for the training.

[spring fulcrum]

well anyways thats a problem for a different day. I'm going to bed. see ya tomorrow

oh and thank you for the info... I've never done any training of anything before so I appreciate the help.

[livid phoenix]

Here's some images I've generated

That last one I'm hoping to make a prompt that has the style implemented properly

[delicate grotto]

what is lycoris?

[delicate grotto]

you're getting lots of bloom that adds the 100% humidity and fog feeling

i think if you remove the bloom, you might perfect this style a bit more

[livid phoenix]

Do I do that in negative prompts or something?

[nocturne dove]

There's like a half dozen different lora versions that work slightly differently and have different names. lycoris is one of them.

[livid phoenix]

How do I remove the bloom and get cleaner lineart and make my prompt look like it's actually from an anime?

[peak dove]

ComfyUI - Saturday's first batch!

[delicate grotto]

interesting ty

[delicate grotto]

i think so, you can also remove the prompt that adds it up alongside it

[livid phoenix]

What words do I use in the negative prompt? Bloom, I got that. Anything else?

And what words do I use that give my prompt a crisp, clean anime look?

[peak dove]

Such Beauty

[nocturne dove]

Anything similar you can think of: bloom, glow, haze, low quality, stuff like that. It is hard to see exactly what the effect of a negative prompt is unless you keep the seed for some "bad" images and experiment with the negative prompt to see what helps the most.

[livid phoenix]

Alright

Keep in mind I'm merely using Stable Diffusion to get some cool references for any projects I'm planning on working on in the future

[nocturne dove]

If you know the name of a particularly bloom-filled anime artist or show or something, you could also try that in the negative prompt.

[livid phoenix]

And I'll definitely train a LoRA bot to recognize certain characters with certain prompts once I get back from vacation

[peak dove]

https://www.reddit.com/r/winxclub/comments/wg8ecc/i_made_bloom_in_anime_style/

[tall lion]

Great rendering

[peak dove]

I need to look into getting a more photographic aspect to my output - qwerty's stuff is majestic!

[tall lion]

@timid sonnet also posted some good renders. I couldn't replicate that workflow thought as there is another model being used after refiner called 'test.ckpt'. No idea what that is.

[peak dove]

ComfyUI SDXL 0.9

[peak dove]

Dope

Elephantarium!

[ionic dragon]

did pseudoterminalx leave the server?

do you know why?

[peak dove]

Boss!

[tall lion]

Excellent

[boreal shell]

Panda with a baseball cap

[peak dove]

Use / dream in Bot Channels Shubhamk

[elfin cobalt]

https://brage.info/GAN/ganbot2/1a64dfe4-49bd-4037-b7c1-aed7e9111312.1.jpg

Here’s some latte art.

You couldn’t mix styles like this with 1.5

[jagged magnet]

Thanks a latte!

[strange mist]

vlad1111 not recognizing the sdxl vae 😭
I put it in models/VAE
refreshed / restarted several times.

[vale eagle]

have a toast?

[west breach]

nice one!

[dense chasm]

holy shit,the sdxl bots renders Elon so real,it just weird there is a portion of certain celebs generating well,the others suck

[delicate grotto]

in the style prompt you can use enhance
and also clean/clean art or anything, neg:shiny/bloom/jpeg artifact etc...
sdxl is quite new so im not familliar with lots of ways to deal with it, 2.1 had a lot of ways that fail in sdxl

rule of thumb is to find the prompt that adds the bloom always, and add description around it without causing it

for example instead of asking for pink hair, you can say
prompt, pink foreground, prompt
neg:pink background

this way you will most likely get the main char in pink stuff, and negative anything that you dont want pink

[thorny frost]

is 1.0 going to be much better than 0.9?

[delicate grotto]

from emad roughly 10% better

that's why they delayed it

[thorny frost]

because they want to make it much better?

[delicate grotto]

not sure if much better, but they did want to finetune it a bit

[thorny frost]

ok. i hope the community fine tune it a lot, because 0.9 results dont seem that much spectacular compared to community fine tuned 1.5 models

[delicate grotto]

it might seem like that at first, but remember that this is the general model... and it competes with the top bangers in 1.5/1.4 and 2.1

and it is not fine tuned to any style yet

in other words, this general model can be trained to be something that surpasses them

[thorny frost]

i know. Appel and oranges.

I dream with having something like midjourney running on my pc

[delicate grotto]

i just hope that training in sdxl is going to be as fast as in 2.1

[delicate grotto]

well, mj is prob a mix of lots of models
it is definitely run on stable diffusion

i noticed lots of hand artifacts that are common in 1.4/5 in it

also, they leave metadata out, so im suspicious of them quite a bit

[thorny frost]

yes i thjink so too. Probably it is some kind of AI improving prompts and selecting SD models.

it must be one of the best guarded secret on the internet.

[delicate grotto]

meh, most likely they just pass it through several different models for a single image

and using the user prompt they selectively choose which to pass it through in the process

anyway, until it leaks we wont know

[clever verge]

@visual glade Sorry to bother you. Tried to sift through your codebase in search for the path to my custom node. Found folder_paths.py but not helping me. Is there a neat way to get the path of a custom_node directory? I want to read a json within my custom node. Maybe the object itself contains it's path?

[west breach]

e.g.

    p = os.path.dirname(os.path.realpath(__file__))
    file_path = os.path.join(p, 'styles.json')
    with open(file_path, 'r') as file:
        data = json.load(file)```

[clever verge]

Sweet, thanks!

[clever verge]

Writing a node that uses the style data we got from SAI they other day, makes it easier to use predefined styles.

[west breach]

check out https://github.com/bash-j/mikey_nodes

[west breach]

nice! looks straightforward. good job 🙂

[clever verge]

Never written python before 😉

[west breach]

The refiner on it's own seems to make decent images at 512x

Or should I say it composes well at 512x and then doubling to 1024 with a hires fix ends up with a pretty decent image

[sharp robin]

Wonder how we can exploit that

[sharp robin]

But also makes no sense why they trained sdxl base on 1024 and refiner on 512

Wait… is this why latent upscale delivers poor results?

[west breach]

well that image was upscaled to 1024, if you try do a 1024 image from scratch you get a mangled mess

[sharp robin]

Exactly!

So inherently they trained refiner on 512 whole sdxl on 1024.

Seems like pretty big oversight

[west breach]

so maybe try to use the ultimate upscaler at 512?

[sharp robin]

And reason for problems

[sharp robin]

Yeah

[west breach]

hmm.. but 512 the output is not great

[sharp robin]

Hmm. I wondered if they realized this.

[ionic dragon]

is it 0.9?

or 1.5?

[west breach]

refiner 0.9 with no base beforehand

[ionic dragon]

sdxl only supports few res properly, just make sure to use them, and why are not using the base?
refiner is not good at offsetting the noise

[west breach]

I'm just experimenting

[ionic dragon]

oh ok

@west breach

[west breach]

Thanks. I know how to use SDXL properly. I'm just playing around with things, I noticed if you drop the resolution for the refiner to 512, it can compose images properly like a 1.5 model, although the overall look is not great

[sharp robin]

I don’t think u understood. Testing refiner alone yields good images, but they are trained at 512 so there’s a little bit of an issue there.

[indigo carbon]

yo i think my SDXL is fucked, ALL my images are tillable. help?

[west breach]

did you tick tiling checkbox under the sampling method?

[indigo carbon]

nope

I turned it on only once and now it stays in that state, even after a restart

[west breach]

sorry, I'm not sure 😦

[indigo carbon]

bro

it's not fully tiled, but wtf is going on

[west breach]

did you try a different checkpoint?

[indigo carbon]

what other SDXL checkpoint is there????? did 1.0 release yet?

[west breach]

I mean a 1.5

[indigo carbon]

oh, that problem doesn't occur on 1.5

[clever verge]

@west breach NODE_DISPLAY_NAME_MAPPINGS does not seem to be doing what I except it to. Any idea?`NODE_CLASS_MAPPINGS = {
"SDXLPromptStyler": SDXLPromptStyler,
}

NODE_DISPLAY_NAME_MAPPINGS = {
"SDXLPromptStyler": "SDXL Prompt Styler",
}`

[west breach]

I just use the NODE_CLASS_MAPPINGS

[clever verge]

Trying to mimic build in nodes but fails.

I'll try that.

Well, then the node name changes and also title. Weird.

[indigo carbon]

got it =]

[glad fulcrum]

Whats the cause of granular gens? Its like textures made of dots

Is it happening to you in some prompts?

[west breach]

Do you have an example?

[glad fulcrum]

Yes hold on

[west breach]

it should, we just need new controlnet models

[azure oxide]

Curious, what was the fix? Or did it just randomly fix itself lol

[glad fulcrum]

you see the dots patterns

its not every image like this. depends on the prompt

[ionic dragon]

SAI said they have them ready, they are gonna release them with 1.0

[west breach]

Try a lower CFG value at 7 or less, set the scheduler to simple

[glad fulcrum]

ok

[indigo carbon]

yeah, I found 6.5 to work best with SDXL, well, at least on A1111

[glad fulcrum]

Oh I had it on 10 for some reason

[west breach]

A separate CFG of 3.5 can work well for the refiner

[indigo carbon]

idk about the refiner. it's not implemented in A1111 yet

[glad fulcrum]

Im using confyUI

Refiner should work on a fork of auto1111

[ionic dragon]

straight out from Emad himself

[glad fulcrum]

You need to change to sdxl branch tag 1.5

[indigo carbon]

I'm on dev branch, it doesn't

[glad fulcrum]

Tag 1.5?

[indigo carbon]

I am, no option or dropdown menu for the refiner

[glad fulcrum]

Img2img

[indigo carbon]

but heck, I'm making the best AI images I generated rn lol, so I ain't complaining

[glad fulcrum]

You load the refiner there

The workflow is not as clean but it should work

[indigo carbon]

I'm not up for clicking 4 buttons after every gen I do. also it's not the same.

[glad fulcrum]

Nah

Just make a bunch of base

Then refine the one you like

Maybe that doesnt work for you

[indigo carbon]

IDK man, Idk if this even needs the refiner

[glad fulcrum]

Some artstyles dont need it for surr

Some get the image even worse

But for some others it does a reslly good job

Like photorealistic

[indigo carbon]

Idk, using the refiner on IMG2IMG isn't like using it on comfyui, and I don't like comfyui

I guess I'll be patient until A1111 implements it without the need for img2img

[glad fulcrum]

👍

[glad fulcrum]

What sampler are you using?

[indigo carbon]

I find DPM++ 2M Karras to work extremely good for SDXL

[west breach]

dpmpp_2m_ancestral for base and dpmpp_2m for refiner

[glad fulcrum]

Thanks

[indigo carbon]

yeah, the DPM family work surprisingly well with SDXL

[fresh path]

the head is so overexposed

[indigo carbon]

wdym

[delicate grotto]

gotta say, the refiner does its job well

but not always

[delicate grotto]

i can see you there

[west breach]

depending on the image you might need to tweak the amount of work the refiner does

[delicate grotto]

yep, im mumbling atm between 0.5 to 0.25 in 20 steps

[clever verge]

@west breach Found the fault, didn't import everything in __init__.py

[west breach]

oh yeah, that caught me off guard too

[delicate grotto]

yeah seems like 0.5 X 20 steps X unipc X kerras works wonders

so basically the stock sdxl is great for anime and art, and the refiner to make everything look more realistic

[indigo carbon]

is it just me, or is the contrast too high here?

also, why is this

[clever verge]

Probably time to push it to github. 😄

[strange mist]

anyone running sdxl with vlad1111? how did you get sdxl vae to work? 🙂
I put it in models/VAE, but it isn't being detected for some reason.
I've redl'd the vae, restarted, reloaded etc. nothing seems to work.

[indigo carbon]

I use it on A1111 and it works on the dev branch

[strange mist]

dev or release candidate? but A1111 doesn't support the refiner, I think?

[indigo carbon]

yeah, no refiner yet, but damn, it sure does work lol

[delicate grotto]

high CFG

[indigo carbon]

I used 6.5, what do you recommend?

[delicate grotto]

how much steps?

[indigo carbon]

60

[delicate grotto]

for 6.5 ~20 is good, 7 is good at ~25, 8 between 20 and 30

damn 60 is a lot, good for details though

[indigo carbon]

HEY

4.5 worked damn

this is with 6.5

[delicate grotto]

lower cfg gives the model more "freedom" but it starts to slack in details at a certain area

[indigo carbon]

and this 4.5

that's a massive difference

[indigo carbon]

idk man, the detailing got insanely better with 4.5

[delicate grotto]

try it with some other stuff
its interesting, it is as if sdxl reacts completely differently to CFG than the rest of the models

this is what you'd usually excpect with 3.5 CFG

an amobea like feeling

this can usually be fixed only in img2img or higher cfg passthrough

[west breach]

try doing the first few steps with a low CFG like 5, then pass to a 2nd base sampler with like a 9 CFG

[delicate grotto]

mmm didn't try 9 cfg yet, i did see it completely burned at 16

[west breach]

it doesn't burn if you do a few steps with a low CFG first

[delicate grotto]

ill try a few combos with the refiner, 2.5->9

[paper phoenix]

I mean you can use high CFG and not get burn if your prompt is basically a novella, but for short prompts a low cfg generally works best.

[delicate grotto]

ok, seems like the refiner has a certain bias

[paper phoenix]

low gravity mode.

[delicate grotto]

inflation is hitting this nation hard

[indigo carbon]

dude, SDXL is damn close to nirvana with a low CFG

[delicate grotto]

try doing the crysis character

i think that what the refiner does is add stuff that look more cool and more expressive

[visual glade]

look at how to get the path to the current file with python

[west breach]

8 steps

[delicate grotto]

oh damn, 8 steps?

[dry valley]

where is SDXL 1.0??

[west breach]

yep 8 steps

[cold mica]

Check out our weekly event #1087493421209485393

[quartz sequoia]

is there a node that would allow me to save images in their own folder? and the images generated names are based on the type of model or lora i switched to?

just for extra organization

[visual glade]

the current one can do it if you do: output_folder/prefix

[upbeat summit]

@comfy I tried this (I know it's not an official documentation) but it didn't work. was it ever implemented? https://blenderneko.github.io/ComfyUI-docs/Interface/SaveFileFormatting/

[visual glade]

yes that docs should work

[trim orbit]

are there any codeformer nodes available for comfyui? roop uses it on automatic1111 afterwards to fix the low resolution transplant that roop does. i can't seem to find any way to achieve this same post restore fix on comfyui.

not the most important thing. i should just be training faces with loras

[upbeat summit]

hmm okay. since the search and replace feature would be super powerful. the filenames always just gave out %variable% instead of populating it

[trim orbit]

beauty

[visual glade]

make sure you do: %node_name.widget_name%

[upbeat summit]

I'll try that -thank you!

[visual glade]

I think there's a codeformers node on civitai somewhere

[peak dove]

can I trace back my exact prompt and settings in ComfyUI as I can in A1111 (Image Browser)?

[visual glade]

yeah drag the image on the UI

[peak dove]

Wow! Does SDXL do "proper" text already?

[quartz sequoia]

not sure how i can do that, save image only lets me change filename_prefix to a node or a widget

[indigo carbon]

around ~60% of the time, yeah

[peak dove]

So drag each image and it'll reconstitue itself on the UI? Good to hear

[indigo carbon]

it follows text with all caps very easily

[sly kelp]

wasn't sdxl supposed to release a few days ago?

is it out yet

[indigo carbon]

0,9 yes

[visual glade]

click on the "ComfyUI" written in the widget and it will let you change it

[quartz sequoia]

its still baking

[indigo carbon]

IDK about 1.0 tho

[jade hill]

the real comfy ? the dev ? 😮

just wanted to drop a big thanks then, you revolutionized my way to use SD
Sytan told me so, but really happy they took you in !

[quartz sequoia]

you dont mean this right?.... cuz it only lets me change the value

[trim orbit]

i got two big node packs off civit that said they had face restoration stuff. was pack and impact pack. they're chalked full of nodes but their face restores don't seem to be applicable. they're set up for upscaling and i don't think they use codeformers. I can't search civit outside of popular packs though. The search functions succcck

https://civitai.com/?query=comfyui nodes&view=feed

how am supposed to find nodes on civit?

[quartz sequoia]

like this?

[visual glade]

https://github.com/Haidra-Org/hordelib/tree/main/hordelib/nodes/facerestore

[trim orbit]

there's a reason i come to an active community to ask.

[jade hill]

so many new nodes, nice

[trim orbit]

nice. maybe that might work.

i'll have to figure out how to install though. should just have to copy the facerestore folder into custom_nodes? or is it more involved. that repo looks more involved

[visual glade]

yeah copying that facerestore folder should work

[trim orbit]

https://civitai.com/?query=comfyui&view=feed even this shows zero results. civit just sucks for discovery. we really need a new central host. these guys have capitalized on getting in early and are phoning it in

[visual glade]

yes

[peak dove]

I am using text to speech and making a file (available soon) of many Prompts I have in a Gen AI Magazine ...

[autumn forum]

Shows a bunch of results for me lol

[quartz sequoia]

intresting

[trim orbit]

they're pandering to the anime / nsfw audience mostly and nothing useful seems to get popular there because they've designed it, intentionally or otherwise, so waifus are maximum signal boosted rather than anything quality. i mean shit. the search function doesn't even do it's most basic job.

[west breach]

Clear your filters

[trim orbit]

i'll clear cookies. filters are clear

shit i'll just make a new account. i criticize civit for waifu shit all the time and my account there is my discord account. i wouldn't be surprised if public opinion got my account shadowed

logged out and tons show up. gfigure

[peak dove]

I'm torn between ComfyUI's simplicity and capacity to work straight out of the box; and the easier extensions available at A1111 - especially Latent Mirroring and Image Browser - but both are superlative software packages!

[clever verge]

Is it possible to use date-functions to create subfolders too?

[trim orbit]

sorry for ranting about civit's lack of nodes when they're clearly all there. feel a little gaslit by civit admins or whatever bug causes the search to die when i sign in

[indigo carbon]

almost

[visual glade]

yes

[quartz sequoia]

well that actually worked

[delicate grotto]

its interesting that you left the last step 6/7 with noise and sent it to the refiner with additional 21 steps

this is how the refiner supposed to work, but im unable to send a 128x128 latent into the k-sampler with different resolution

any idea how it is supposed to look in comfy?

[elfin cobalt]

Comfy would have a VAE Decoder step between the refiner and image.

[eternal fog]

The way this is presented is a bit strange, because the base can't produce a 128x128 latent without it looking like complete shit.

And the Refiner doesn't make images bigger

[trim orbit]

depends on the layout. many of them just feed the latent from the base to the refiner. see sytan's

[eternal fog]

Is that latent resolution different to how we create an empty latent in comfyui

Because if you put in a 128x128 latent, it would just produce a mess

[elfin cobalt]

Right. I did say between the refiner and the image. 😉

[elfin cobalt]

The numbers you give to the 'empty latent' node are divided by 8.

[trim orbit]

mb

[eternal fog]

Oh ok, so that is doing exactly what the main workflow is doing

[delicate grotto]

yep, seems the same to me
it looks weird

wait, divided by 8? mmm so that's it

[indigo carbon]

hey, isn't SDXL1.0 supposed to release tomorrow? I don't know why I think that, but I remember something like that.

[tender timber]

I can't remember where I saw it (it was on this Discord though), but is there a node/workflow that can customize the file name output in ComfyUI?

[eternal fog]

no

[paper phoenix]

said a week so tue. i wouldnt hold my breath for dates.

[eternal fog]

They've not said but there is an SDXL 1.0 event on the 26th

So probably then

[autumn forum]

Probably Wednesday the earliest

[indigo carbon]

wait, so what's happening tomorrow? I remember Emad or someone saying something is happening on the 23rd

[eternal fog]

The Save Image node has a box to change the output filename

[tender timber]

Can it use variables like time and the first few words of the prompt?

[eternal fog]

Maybe you are thinking about Emad tweeting 3 on the 20th? https://twitter.com/EMostaque/status/1671121009817034759?s=20

Emad (@EMostaque)

3

Likes

182

Twitter · Jun 20, 2023, 11:41 AM

[mystic cosmos]

Tomorrow Sunday happens

[fresh path]

It'll be like a week or so so nearly there and all the variants are being ready for the big launch where we will have a nice chat as a community.

dates moved around a lot

[indigo carbon]

so, next week is promised?

[eternal fog]

Not sure about that. If it doesn't you could probably modify the node to do it

[fresh path]

"week or so"

[indigo carbon]

or just another way of them saying it's delayed again

[eternal fog]

Nothing has been promised

[fresh path]

is either a week up until infinity

[tender timber]

OK thanks, I was hoping to get the equivalent of the AUTO1111 naming ability, like below

"samples_filename_pattern": "[datetime<%Y-%m-%d %H-%M-%S>] [prompt_spaces] ([seed])",

[eternal fog]

Doesn't look like it does that at the moment but you can edit the SaveImage class in the nodes.py file.

It's line 1119

[indigo carbon]

honestly, I'm more exited for the A1111 support for the refiner than for SDXL1.0, I think 0.9 is already better than expectations

[visual glade]

that's up to the devs of that ui to implement it

[indigo carbon]

I doubt they won't, A1111 will be forced to do so, most of this community uses his WebUI

[eternal fog]

I imagine its a pain in the arse to do it because of how hacky the A1111 WebUI is

[visual glade]

I don't think the devs of a1111 get payed for their work so

don't blame them if they don't have it implemented

[eternal fog]

It sounds like SAI have been giving them a hand though

From what's been said in here by staff

[wicked frigate]

i've been offering help wherever i can but for the most part auto's just goin at it himself

[indigo carbon]

they did confirm they will implement the refiner in an efficient way, but the question is when

[wicked frigate]

we've gotten some fixes merged into SGM to make auto impl easier

[tender timber]

Thanks, I found the repository, it was https://github.com/bash-j/mikey_nodes and the python file has code showing the addition of date/time to the filename. I'm about as good with Python as I am at levitating, so I'll see how little damage I can do 🙂

[peak dove]

😄

[indigo carbon]

also, will using the refiner when it's implemented cause performance issues?

[visual glade]

if it's properly implemented not really

[indigo carbon]

or will it just unload the base then load the refiner

[peak dove]

I am using Base in A1111 - are u telling me that Refiner does not yet function inside A1111?

[indigo carbon]

nope

I know, right?

[visual glade]

you can try it right now in comfyui, the performance is pretty much what you are going to get if it gets implemented properly

[peak dove]

You mean the excellent results I am getting out of Vlad A1111 SDXL are Base and Base Only?!

[visual glade]

vlad has the refiner properly implemented because he's using diffusers

[eternal fog]

There's a little delay in passing off from the base to the refiner, but I think that's more because of VRAM

[noble shoal]

You could change Line 1152 of nodes.py to: file = f"{datetime.now():%Y%m%d_%H%M%S}{filename}{counter:05}_.png" also add "from datetime import datetime" to the top of the file.

[indigo carbon]

idk about vlad, i'm talking about A1111

[tender timber]

Thanks, would an update from ComfyUi override that though?

[noble shoal]

Yes

[eternal fog]

Base only is fine for most things, the only place I've seen the refiner provide a decent improvement has been on eyes and fingers

[peak dove]

It keep crashing because 100% VRAM usage (8Gb) - but even using -medvram or --lowvram it still gives memory warnings

[visual glade]

you can copy the save image node to your own custom node and do the changes there

then it won't get overwritten by an update

[tender timber]

💩

[tender timber]

OK, I'll have a go once I'm brave enough 🙂

[eternal fog]

Personally I think having the date on the filename is a little pointless, when you have the file creation time anyway

The prompts I get, but if you use long prompts you can't fit it all on

[clever verge]

@west breach Pretty happy with it now. Thanks for the help. https://github.com/twri/sdxl_prompt_styler

SDXL responds well to the templates

But I guess that's expected. 😄

[spring fulcrum]

hee's sleeping

[patent badger]

tried adding it but getting no option to connect other nodes to it

[clever verge]

right-click on the node and convert text_positive and negative to input

fiddling with a few custom styles too, might come later

[peak dove]

Just d/loaded - will have fun using this!

[shy kelp]

Cheers!

[clever verge]

Here's GTA-style:

pretty insane what that templating can do. 😄

[high skiff]

I do. He left for several reasons, however I don't think I want to stir the pot by mentioning any of them, and just know that he left the server because of negative interactions

[shy kelp]

prompting with 'Tilt-Shift Photography' is awesome 🙂

I just prompted: "Mediterranean city port with a flying spaceship attacking. it is firing a laser beam to the city causing many large explosions, Tilt-Shift Photography"

[compact obsidian]

Is the thought right now that SDXL might be just one model? or will it more likely be the base and refiner for 1.0

[azure oxide]

Is pseudo the one who constantly leaves the discord and reenters under a new name or was that another regular that I'm thinking of

[ionic dragon]

i asked him on the other server

but he was a valuable asset for this community

[clever verge]

Pushed a few new styles to the repo.

[urban fjord]

There is a new refiner in the bot now so I guess they're going to keep having base + refiner.

[indigo carbon]

Also, @wicked frigate , does Automatic1111 have access to SDXL1.0 so his UI can support it right after it's released?

[visual glade]

it's the same architecture as 0.9

[indigo carbon]

Ah, I see. I thought it's slightly different?

I don't remember who said that

[wicked frigate]

We had experiments with bigger changes that would need code changes but seems the big changey ones won't be 1.0

(a shame, I like some of the bigger changes, but perhaps a future model)

[dusk mica]

i thought , the refiner was baked into the file or smth

[visual glade]

yeah the big changes that break everything are going to be a future release

[dusk mica]

any hint what the big changes could be?

[visual glade]

it depends what works

[autumn forum]

Well I’d say it’s for the best to not make super big code changes for 1.0 yet sense people will get upset lmao. 1.0 will most likely be the new standard. At least I hope it is.

[visual glade]

1.0 is the same model arch, the only thing that will be different from 0.9 is the recommended workflow

[dusk mica]

i did found a diff when using SDXL and a 1.5 model (e.g ghostMix), i think SDXL makes the propt much varied, when i type " a decorated hallway" SDXL tends to generate more varied versions of a hallway, where as ghostMix has like 1 hallway theme/color scheme, is this a training thing or just 1.5 ?

[autumn forum]

Which I’m guessing will be discussed in the super stage?

[visual glade]

I'm just saying that because it behaves a bit differently vs 0.9

you'll see when you get your hands on it

[urban fjord]

McMonkey: Oh, what kind of big changes are you experimenting with?

[errant yarrow]

a weightlifting panda with a serious era doing a lot of effort, muscular and heavily tattooed through all of his fur with tattoos of Vietnamese warrior symbols leaving traces of their designs through his fur, in a dark and ultra modern weight room with a red background and light effects, lifting forcefully, as high as possible, above his head a weight bar with large cast iron discs, showing the effort in the gesture and squeezing hard the helm, ultra realistic, fully respecting physical laws for more realism

[autumn forum]

us the bots, this is for sdxl chat not generating images

[urban fjord]

Is that 0.9, 1.0 or something else?

[visual glade]

open the image in comfyui and you'll see

[errant yarrow]

#bot-x

[visual glade]

it's SDXL + a few tricks + WD beta3 illusion

[urban fjord]

What is the broken FennecEarSizeMultiplier node that looks like a LoRA?

[visual glade]

it's something I'm experimenting with, it's named that way because that's one of the effects it has

[clever verge]

Damn, SDXL realy responds well to style prompts.

[errant yarrow]

Mediterranean city port with a flying spaceship attacking. it is firing a laser beam to the city causing many large explosions, Tilt-Shift Photograph

"Mediterranean city port with a flying spaceship attacking. it is firing a laser beam to the city causing many large explosions, Tilt-Shift Photography"

[visual glade]

go to the bot channels

[errant yarrow]

Hello, where can i find sdxl 1 ?

[clever verge]

Didn't have tilt-shift style but could fix that. 😄

[dusk mica]

holy

that looks so good

[clever verge]

@errant yarrow Here's tilt-shift photo. Is that what you had in mind?

[errant yarrow]

It' better

[clever verge]

Yeah, I feel the same way. And this is one single try.

[urban fjord]

Tried updating ComfyUI and LoRA's are still quite slow sadly.

[clever verge]

I commited the tilt-shift and a few new styles. If anyone else want to fiddle with it it's here: https://github.com/twri/sdxl_prompt_styler/

stability ai's prompts have prefix sai- and my custom ones are gen-

Thinking about giving the possibility to block either negative or positive prompt templating, could increase flexibility.

[urban fjord]

Nvm, without noticing it I see that I compared LoRA and Base with different steps.

[wicked frigate]

stuff~

:3

[ashen coral]

guys, how do i use sdxl lol

[wicked frigate]

(you'd have to beg that info out of Dustin or Joe, not my work and I don't want to steal other people's announcements from em)

[wicked frigate]

A: download ComfyUI, or auto webui's dev branch, download the model, put it in, use it.
or B: pop over to the bot channels #1100170312106127410 and do dream commands there
or C: clipdrop/dreamstudio

[ashen coral]

thank you very much ❤️

[azure oxide]

Just take a screenshot of their discord messages and spread that around without telling them about it, its as easy as that

[clever verge]

Styles are grunge, dystopian, cubist, constructivist, pixel art and film noir

[clever verge]

and a simple prompt "beautiful woman, age 35"

[gusty saddle]

when is sdxl released?

[autumn forum]

july 26th, or so. maybe. . you can use the bots right now tho!

[livid cradle]

im already getting insane results with sdxl 0.9 cant wait to see how 1.0 changes this.

will the 1.0 be available over the API customers?

[ionic dragon]

is this channel dying bcuz of psuedo left?

its getting spammed with images

[wicked frigate]

wat

this channel has had 500 messages in it today alone

[ionic dragon]

500 is too low

on 18th, there were too many

[wicked frigate]

yeah the day that was the initial target for release that we missed and delayed a bit longer, where there were tons of messages from people talking about the release target, the leak of the delay, etc.

[ionic dragon]

I sent around 102 alone

[autumn forum]

there are very specific times this channel gets REALLY active lmao

[ionic dragon]

before that too, we had too many msg's too

today its too idle

[autumn forum]

just imagine whats its gunna be like wednesday haha. this channel is gunna blow tf up

[ionic dragon]

when?
on wednesday?

[autumn forum]

theres an event. https://discord.com/events/1002292111942635562/1130919343941758978

[ionic dragon]

Emad's gonna join?

[autumn forum]

no idea.

[ionic dragon]

damn, i wouldn't be available, I would be watching oppenheimer 🤦‍♂️

[formal quarry]

@livid cradle may I ask you how you achieved such results on the beverage commercial photos?

[livid cradle]

im building my own magic prompt to give me this results.

yet, due to a bug in SDXL masking, it keeps fucking up my originial image... can't make this consistent and keep the brand intact... anyone can help with that?

using this (SDXL generated can) and try to position it in various scenes. yet sometimes i get lucky, sometimes i just get freaking frustrated.. im working on the prompt only to generate the backdrop and palcement...

Wold be nice if i can feed my prompt a variable with the image so it only builds around it and take that as the main subject of the canvas.

but as you can see in the last iamge, it clearly generates another can behind my original init_image... waiting on someone from SD team to address this masking bug in the image to image api

[stray depot]

I'd love a can of 'coocunt' where can I find one?

[livid cradle]

everyone loves coconut, that's why my marketing photos will be so appealing 😛

[stray depot]

No I don't want coconut, I want coocunt as advertised.

[clever verge]

Maybe not as good as above

[livid cradle]

see... that's waht i keep saying... it modifyes the text on the init_image which is fully masked... can't understand how this is not a big problem and immediately addressed by SD developers?

[stray depot]

They need you on their team.

[stray depot]

Strange, thought. But probably 1.0 it is addressed?

[livid cradle]

if i would of known how to fix i would of passed them the data... but unfrotunatelly i don't know where the issue is, i just reported my findings.

now i started a cocunt trend on this room...

[stray depot]

haha, it's a great word.

[livid cradle]

anyone from this room in good relation with any developer from here? at least they can take a look at the masking problem... 😦 im so frustrated that this keep happeneing.

[stray depot]

But you're using 0.9 right?

[livid cradle]

yes

you can't get same level of results with 0.8 i've tried...

[stray depot]

It's a bit like using the BETA, it ain't gonna be quite right. I find little oddities with 0.9 which I am confident won't exist in 1.0

The basic results are 10x better than anything I would get from 1.5 base. So it is encouraging for an unfinished product.

[livid cradle]

yeah i know it's beta. but hopefully this masking issue is going to be address with the release of 1.0 (i don't think it will, i reported it a few days ago only) so hopefully some big hearted developer will look into it and try to fix it. 😄

[clever verge]

This is better, had to change seed.

[stray depot]

Is that a can of cu...oh dear.

[jade hill]

👀

cuud of course

[stray depot]

Gothcha, cuud. Sounds the same, taste the same?

[jade hill]

like coconut 😉

[stray depot]

Truly delightful.

[jade hill]

always loving a lick

[stray depot]

As long as they don't describe this beverage as dry, then I am in.

[jade hill]

And as long as we keep on running around the terms and not triggering autobot, I'll see myself out xD

(really good wordplay mate though)

[stray depot]

You get good at when dealing with finicky prompts.

[clever verge]

English is not my native language so I must miss out on the fun. 😄

[stray depot]

Oh what is your native language?

[clever verge]

Swedish

[wicked frigate]

idk exactly what you're currently doing or the API stuff, but, from the perspective of just, how to do the thing you're doing in SD (in general, not even really SDXL specific) what you can do in this case is txt2img a plain background within any can mentioned in the prompt, then just apply the can over top, and used masked img2img to redo the background with the txt2img as a low % init, so it's less likely to get confused by the can

[stray depot]

And a seemingly sensible answer has emerged.

[clever verge]

Here's a new flavor, got a bit sharper too.

[livid cradle]

I was thinking about this workaround as well, it's a good idea, but its going to require multiple api calls , longer wait time, and more expendive on me... so i need this to work as initially designed.

[formal quarry]

dude i would pay you 20$ for just this

[stray depot]

The designs are lovely. Definitely something you would expect to see on some Mediterranean billboard

[clever verge]

Given how fast I got here I'm amazed.

[formal quarry]

no joke I do a lot of ai for over two years but those image creations get me every time. this is truly art

[stray depot]

I'm a photographer (for fun) and these days I think I might as well sell my camera. AI is it.

[clever verge]

My workflow was to give GPT4 some context about what a good prompt is and then suggest to have commercial themed prompt styles. Feed that back into json and use it with my prompt-node along with my custom prompt.

[livid cradle]

im actually building a product placement, product photoshoot ai app , would anyone be interested in early access to such thing? 😄

im actually starting to think i should build this in public and ask users feedback as we advance the development 😄

[clever verge]

Yeah, looks nice. Bit sharper than my style but im certain it could be improved upon.

[stray depot]

Yes.

I worked with a company that sells high end spirits and we did a lot of lifestyle photography for those products. Seems to me there could be a market for such an app.

Had a dedicated studio for it.

[formal quarry]

@livid cradle what platform is this? never seen this ui. is this a homebrew :)

[livid cradle]

yes, its not just a SD webui like all the others, it's going to be a platform dedicated for ecom owners, influencers, marketing experts, product photographers and small business owners specialized in product photography. it's work in progress.. hoping i will have the MVP by the mid August.

[clever verge]

Can in a can, pretty meta.

Nice colors though. I would buy it if it was on sale.

[stray depot]

A can of Colonial? Me too.

Such a cool design tho.

[formal quarry]

thanks and best of luck with it

[stray depot]

I never considered this for product photography, and by extension social media, but it works really f'in well

[formal quarry]

sorry for the stupid question but is there a special hf space or notebook you are using for this. i'm often too lazy to get my own setup running haha

[clever verge]

Most of us are running it locally on our computers

[spring fulcrum]

What prompts do you use for those bottles?

[stray depot]

A bottle of Dalmore whisky, sitting on a wooden table in a scottish pub, with a lit fireplace in the background.

SDXL responds quite well to 'conversational' prompts I find

[spring fulcrum]

That is awesome

and yes it responds very well to it. I really like how it has been improved so much

[livid cradle]

ive coded my own prompt generator that is focused on generating the best prompts for photography scenes and shots. called it Magic Prompt 😛

[spring fulcrum]

I'm excited to see 1.0

[stray depot]

Yeah, I hated so much that I have to specify not wonky eyes. Tf is that.

So this is a massive leap.

[spring fulcrum]

Is your Magic Prompt Public?

[livid cradle]

so inside the app you just have to say your product type a few words like "a can of coconut flavored energy drink" and based on this prompt it will enhance it to this:

Mountainous Set-up: for a can of coconut flavored energy drink: Create a scene with a granite rock in the midst of a clear mountain stream against a majestic Alps mountain range under a clear blue sky. The scene is illuminated by the vibrant, high-altitude daylight, characterized by dominant greens, whites and blues, with a focus on creating a refreshing sensation. Incorporate surrounding alpine flowers and ferns to enhance the product's natural sensation. The scene composition should be a low-angle shot, capturing the product emerging from the water against the mountainous background with a low-angle to emphasize the product and the towering mountains, creating a refreshing, invigorating, and exhilarating atmosphere. The overall style of the scene should be realistic, vibrant, and breathtaking.

[stray depot]

Solid prompt. Let's all use it for free.

[livid cradle]

not yet it's not i will give early acces to interested people soon.

:))

[spring fulcrum]

are you going to have it hosted or is it going to be open source and on github

[livid cradle]

feel free to use it.. i generate unique prompts each time in various styles, scenes, backgrounds and a whole other more categories not shared here 😛

[stray depot]

I'm gonna put that prompt in, see what comes.

[spring fulcrum]

I would be happy to test it and provide feedback

[livid cradle]

paid and hosted applciation.

[clever verge]

Good prompt.

[stray depot]

Same issue with some what messed up text, but yeah, workable.

[clever verge]

Text is not what SD is perfect at, but I guess nothing that photoshop won't solve.

[livid cradle]

some early results.. using same image as base image.

im not interested in the text on the can... it's best with no branding so no copywrite issue on my promotional page. anyone will use their own products to enhance.

this is just a sample i use to test.

but the results are promising... various styles and scenarios same product.

i just wished someone from the SD team would say something about this image to image masking bug i reported. (meaning they are working on a fix for 1.0) 🙂

[stray depot]

Have you guys tried doing prompts with a 0 CFG?

It's terrifying.

[clever verge]

Yikes

[spring fulcrum]

It might not make it into a fix for 1.0 It seems like its too close to final release for any major bug fixes... It could become a hotfix after release but it takes a lot to fix big bugs like that I would imagine.

I used to work for a company that made proprietary software and fixes used to take months after I would submit them... It would first come through a customer. I would submit a jira bug ticket.. then tier 2 would pass it to dev. dev then goes to QA... then back to me for final testing... then it would be staged for implementation in the next release.

[patent badger]

anyone knows what are the cfg and steps which are used for the bot / clipdrop?

[spring fulcrum]

Not sure what they have set on the backend... but theres a good chance they are either set low to reduce server load and offer better speeds, or there is some sort of "rule of thumb" guide they use for optimal results based on whichever sampler they use for the image.

[clever verge]

I wonder if one can make a good photo for fashion magazine cover

[spring fulcrum]

probably.... its all about the right prompt

there are some folks in here that are geniuses with prompting. I have seen so many great images in here lately.

[somber hill]

How To Use SDXL in Automatic1111 Web UI - SD Web UI vs ComfyUI - Easy Local Install Tutorial / Guide https://youtu.be/eY_v5IR4dUQ

sdxl working great

but no refiner suppor yet

[patent badger]

I mean doing it only with SD / SDXL - probably not a high chance, but if your'e mixing in some photoshop skills and something similar - for sure

[spring fulcrum]

that too

[patent badger]

wanted to say I found out about your channel like a few days ago and I'm more than amazed and grateful. is there any way we can donate to show some appreciation? (I don't have too much but if there is something like this already lmk😂 )

[spring fulcrum]

Photoshop is great for finishing up final works

[patent badger]

especially with the new AI tool in photoshop

got a chance to use it and while it isn't as powerful as SD / SDXL - it is so much more useful for specific tasks

[clever verge]

I'm on it. Still a simple prompt, so If you want another style probably easy to adapt.

[eternal fog]

For the bot it's deliberately variable so they can gauge ratings.

[clever verge]

Shoulder pads are the latest fashion this year I've heard. So much so that Kanye wears them.

[livid basin]

You can do unconditional samples from SDXL by using an empty prompt

Gives you interesting random images from the training set

[clever verge]

It can do some pretty insane stuff that's for sure.

[livid basin]

Is that from an empty prompt?

[clever verge]

No, custom style and a simple prompt.

[livid basin]

Ah I ran like 30,000 images from different seeds of an empty prompt

[clever verge]

This was the first image I got.

[high skiff]

My uncle got me hooked on RVC, and I've been using it all day lol

I'm taking a day off from SDXL research, and then I will be back at it again tomorrow with some new LoRA tests

[autumn forum]

lol like the audio thing?

[high skiff]

@paper phoenix I'm not sure if you saw, but I fixed my 3090

Yeah, the real-time voice changer, it's incredible

It works so damn good, and it sounds so real

[autumn forum]

its pretty good. ive been messing with it the one that changes vocals for songs. it does work very well

[high skiff]

Yeah, and on my 3090 it runs fairly fast too

1/3rd second buffer

[spring fulcrum]

Where can you get the changer?

[shy kelp]

Did the model on the bot change recently? I got some amazing pics a few days ago but now with the same prompt they are all really, really bad

not sharp, over contrasty, not even following the prompt fully

[eternal fog]

They are constantly changing things, so they can figure out what's the best when people vote on them.

So different model candidates. different steps and CFG as well

[shy kelp]

I see, is there any way to know when/what changes or do they keep that quiet?

[eternal fog]

Keep it quiet on purpose so it doesn't introduce bias in voting

Once 1.0 is released I imagine it will be more consistent

[urban fjord]

Hopefully, I felt the images got worse after 0.9 in the bot.

[shy kelp]

hopefully, because this change is like night and day. It went from 'actually good' to 'my first prompt'

[eternal fog]

Could just be a change in CFG to be honest

[shy kelp]

you wouldnt even guess these were the same prompt

[eternal fog]

You can get massive swings from just steps in CFG

[autumn forum]

watch 1.0 suck and everyon stick with 0.9

[shy kelp]

Hey I already moved my main lora model to 0.9 so xD

[shy kelp]

seeing this 'cooked image' look worries me about the models they got for sure

[eternal fog]

It's fine, they'll be using "bad" settings on purpose

[high skiff]

What changed your?

*what changer?

[eternal fog]

I think he's asking for the realtime voice changer

I know what it is but I don't remember the link

[autumn forum]

yeah they dinfintely using bad cfg's and sample steps on purpose

[spring fulcrum]

that

[eternal fog]

I've got the links for RVC and the other one

[high skiff]

Oh, it's literally just called real-time voice changer, it's an AI that you can run locally on your hardware, and there's a huge community that supports AI voice LoRA's for it

[eternal fog]

It's quite scary how good they are.

[high skiff]

Yeah it really is lol

Especially for being able to run practically real time

[autumn forum]

https://www.youtube.com/watch?v=pHhjg2JwdPI

[eternal fog]

I think it's this one - https://github.com/w-okada/voice-changer

[shy kelp]

Makes it kinda hard to vote lol. "Is this garbage better than also garbage?" I dont like either what do they want from me lol

[eternal fog]

I couldn't get the realtime one working well, but RVC was good and then the other one based on Tortiose for normal TTS was decent too

[autumn forum]

you choose the best of the two. lol

[clever verge]

@high skiff What fixed your 3090?

[eternal fog]

If they are both shit just don't vote

[spring fulcrum]

ya what fixed it?

[high skiff]

The current idea is that there was something wrong with the mounting pressure on the back plate

The seller and I went through Helen back doing everything, so I decided to consider the card completely dead and tear it apart and rebuild it, we tested various different things with thermals and none of them yielded any better results, but then I completely disassembled it a second time and put it back together, and then the issue stopped

[shy kelp]

thats usually the route I go down

[high skiff]

The backplate does use very short thread spring screws, so it's possible that through shipping they got jostled out and the card wasn't making proper contact, or the card was making two hard contact on one side which could have been causing issues

Whatever it is, GPU's been solid for 15 hours now, and I've been hammering the hell out of it

[high skiff]

I have talked to Joe penna directly, and he said if both results are bad, do not vote

He said voting for bad images just hurts the algorithm, so be selective with the votes you give, and make sure they are for results that you are happy with

[clever verge]

Glad it fixed it, a weird one for sure.

[high skiff]

That was a huge concern that I had as well, so I brought it to his attention and he informed me that we are not supposed to vote for all images, only the good ones

[high skiff]

Undoubtedly, however the GP was running a rock solid, and it's performance is incredible

[spring fulcrum]

What channel do we go to to vote on images?

[high skiff]

Before, generating two images on it back to back would instantly blue screen it, now I've generated over 400 images, trained a LoRA, been messing with real-time AI voice changers for several hours, ran nearly a full hour of stress test and heat test benchmarks, and all sorts of stuff

[high skiff]

Any of the bot channels, or showdown

[spring fulcrum]

ok

will do

[shy kelp]

good to know, looks like I will be not doing anything for a day until they tweak the settings. Its just generating cooked image after cooked image with the same pose over and over again.

[high skiff]

Yeah, it very much seems like SDXL has 5 days of being bad, 1 day of being good

[eternal fog]

I do wonder how people vote on it though. Are they actually voting on if it was close to the prompt, or if it just looks good.

[high skiff]

It's very weird, their results are very inconsistent and you think that they would at least have an idea of what they are doing at this point, and cut out all of the things that are obviously bad

[shy kelp]

Considering it once gave my feral lion tits
I think we know the answer

[high skiff]

I talked to Joe, and he said the most consistent voting alignment is people legitimately just clicking random buttons

He said something about the bot being able to throw out votes when they're cast too fast and succession

[eternal fog]

But some people will just vote on a nicer looking picture, even if it's not exactly wht they asked for.

[autumn forum]

im guessing they vote on what looks good. bad for prompt adherence but good for model output

[high skiff]

He said a large group of users form these side biases, to where they will vote almost undividedly for either A or b all the time

Almost like they think that one of them is one setting and the other is the other, and they form a bias to one of the sides

[eternal fog]

I know some people have been voting for bad ones on purpose for some reason

[spring fulcrum]

AI h8ters

lol

auto ban

[eternal fog]

I think it was someone that had just decided that SAI were doing it "wrong"

So was being an ass

[lusty raptor]

i get self-conscious about that when i'm voting in my channel, and it just happens that a series of A or B is better

[shy kelp]

Maybe my luck is just shit today lol. I did the 'resize' and it did a tall one, so I did it again and it did the same size, and then the same size again, and then finally a different size. Then I did the restyle and it did a sticker, then I hit it again and it did another sticker...

[high skiff]

Side note, Kandinsky 2.2 should be pretty concerning for SAI lol

It's not as good as SDXL, but it's a pretty big leap over 2.1

And it can generate some pretty damn good looking images still

[shy kelp]

The average SD user is not smart and doesnt look more than 5 min ahead in time, I dont think they need to worry lol

[clever verge]

Yes, seen that model too.

[high skiff]

2.2 is a huge leap

[shy kelp]

if all the people going 'no tits = not using sdxl >:(' doesnt say as such lol

[urban fjord]

I've had instances where I've voted one side a lot more than another but that's because the images were better.

[high skiff]

A cinematic photograph of a white tiger standing in the middle of a forest looking through the leaves of a shrub with dappled lighting from above

Kandinsky 2.2 vs SDXL

Obviously SDXL is doing better, but Kandinsky is improving far more rapidly

[shy kelp]

I remember when Kandinsky 2.1 came out and I was all for it

but no one wanted to impliment or use it

why?
Because russia made it

[indigo carbon]

not even close. plus, kandinsky is tuned, SDXL isn't yet

[spring fulcrum]

what prompt did you do for those?

[high skiff]

While I do agree that SDXL looks way better than Kandinsky, my statement was that Kandinsky is catching up really fast

Kandinsky has made a hell of a lot more progress to the quality of their images in the last month than SDXL has

[indigo carbon]

do "realistic, Ice dragon, desolate, intricately detailed, artistic lightning, particles, beautiful, amazing, highly detailed, digital art, sharp focus, trending on art station,"

[indigo carbon]

yeah, SAI is now in a slumber lol

[high skiff]

Kandinsky vs SDXL for:

A portrait photograph of a beautiful woman with long and curly blonde hair and red lips smiling in a white gown in a chapel with gorgeous and colorful stained glass out of focus in the background

I think STXL looks more polished, but I think condensky has an overall better vibe to it personally

*SDXL

Granted, the resolution difference is huge

SDXL has significantly more detail

Also, I accidentally generated them at different aspect ratios

Let me try SDXL over again

[shy kelp]

your average SD user will not move to anything new unless it can do boobs

[high skiff]

God damn some of the settings on SDXL look terrible today

[shy kelp]

even if the model would be better in the long run

[high skiff]

WTF is this bro

[high skiff]

This one looks better, but still looks pretty off

[shy kelp]

yeah I stopped genning because its literally not following my prompt any better than 1.5 now

[high skiff]

I'm curious, would you mind sharing what words you use for that sort of teal / turquoise leaf color?

[spring fulcrum]

What's off about that one... It looks pretty good to me

[shy kelp]

"oh you want it in the style of X and Y artists, both people I did recently? here is a photograph style image"

[high skiff]

It does look good, it's just not as good as I would be hoping for this late in the SDXL process

It honestly looks like something that I could still pretty easily generate using 0.9 on my own computer

Although I will say there are sometimes when the 1.0 candidates really do just hit it out of the park

Likely just not the best settings for that image

[clever verge]

This is my prompt: positive_prompt: cinematic photo a white tiger standing in the middle of a forest looking through the leaves of a shrub with dappled lighting from above . 35mm photograph, film, bokeh, professional, 4k, highly detailed negative_prompt: drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly, headdress

[high skiff]

Interesting, I wonder if it's the term cinematic

[spring fulcrum]

What was your prompt for that one?

[clever verge]

Dunno why I still have headdress in the negative lol

[shy kelp]

just do a style, it just removes any negatives from your prompt for some reason lol

[high skiff]

A portrait photograph of a beautiful woman with long and curly blonde hair and red lips smiling in a white gown in a chapel with gorgeous and colorful stained glass out of focus in the background

[shy kelp]

dont need negatives when it wont let you do t hem for some reason

[upbeat summit]

greetings

[high skiff]

Howdy

[clever verge]

A gothic variant

[spring fulcrum]

Ice Dragon

[clever verge]

I'll try ice dragons tomorrow. They look great!

[high skiff]

Oh wow lol

The image I was criticizing just landed me my third pantheon win lol

[spring fulcrum]

!!Warning!! This is what happens if you put a duck in the microwave

[indigo carbon]

also, I made a prompt template that works surprisingly well on SDXL, well, at least on the tuned version I made of it

[eternal fog]

I'm not sure what this is lol

[livid basin]

SDXL 1.0 finally coming out next week

Exciting to see what the community eventually votes on

[spring fulcrum]

after 30 images rendered of ice dragons these are my favorites

[eternal fog]

Dragons you say

[spring fulcrum]

I really like that electric looking glow

[eternal fog]

Dragon + Movie heavily biases it towards old east asian movies

[eternal fog]

This one is a bit freaky

[upbeat summit]

but the style could be promising!

[eternal fog]

Yeah going to try it in a sec with something else

Didn't keep it for a different animal

[upbeat summit]

nice

[eternal fog]

Hmm, using "Anime Movie" seems to go very Ghibli

I guess I can see why though

It seems to have decided this is how you sit in a chair

[spring fulcrum]

Getting better. Slightly changed the prompt

[eternal fog]

omg, what has it done to this character lmao

[spring fulcrum]

lmao

[eternal fog]

I think all her insides have been pushed up

[sudden cliff]

There's almost no hype, almost no discussion, but foremost you have to prove you can at least do something better. Realistically you have to do everything better, most of the time. People dismiss LLMs that aren't the 'overall' best even if they're best at one thing or another.

[high skiff]

ohhhhh

my god my heart just sank

came back to my PC OFF

Its a new windows install, the screen timed out while I was getting food lmfao

[flat ibex]

Anyone know why in the world SDXL can't generate a racoon? Like, it doesn't upset me but it's weird that it'll do literally anything else but that 😂😂

[wicked frigate]

works fine when i try it in local - quite probably there's an overeager word filter accidentally blocking you there, as the name of that animal is also a racial slur in some obscure contexts

[upbeat summit]

it's probably not the style you're looking for but I got this one 😉

[flat ibex]

That's fair but it was strange cause that last one was literally, "prompt: racoon style: photograph aspect: 1:1" and it made a picture of a camera. Just super strange. It being a racial slur is a good point, could be a weird word combo. Was just super strange.

[flat ibex]

I. Love. This. 😆😍

[sharp robin]

bots?

[flat ibex]

It was Bot 8.

[sharp robin]

it's the nsfw filter, that is too hard core, has word "coon" in raccoon, so it blocks it from generating raccoons, try the local version

or try something like trash panda

hopefully it understand u mean raccoon

[flat ibex]

Ahhhh that makes sense.

[flat ibex]

Trash panda might get some good results.

[spring fulcrum]

Question:

How can I train something that will help SDXL generate images at higher resolutions out of the box... Is it possible to train a model using upscaled images at 4096x4096 so that the model I use will generate images at 4096x4096 instead of 1024x1024... Is this something that is possible to do or is there a reason that it is normally 1024x1024.... I guess my real question is how does stable diffusion evolve the resolution of generated images??? If my method is possible is there a way I can help you guys get it there over time? I have an RTX 4090 and a fairly beefy PC setup.

[autumn forum]

generating at 4096 is gunna be insanely hard on the gpu and theres no reason when you can just generate at 1024 then upscale or using something like ultimate sd upsaler or hiresfix to get 4096.

[spring fulcrum]

ya but... it is progressing over time... sd1.5 was at 512x512 and so on... My question is how can I help with that progression.. They seem to be developing a way for this to evolve over time. I just want to see if there is a way for me to help them get there.

They have been doing it within graphics card abilities

[autumn forum]

for some comparison, 512x512 is 262,144 pixels, 1024x1024 is 1,048,576 pixels, and 4096x4096 is 16,777,216. thats a huge jump.. i understand a 4090 can probably do it. but im just saying at this point. it is a heck of a lot faster to generate 1024x1024 then upscale and get same quality.
if you want a good 4096 model youd have to train one yourself and you wouldnt be able to do that on a 4090 at all at that scale of resolution.

youd have to code something completely new if you wanted an efficient new 4096 model. or just wait until we have the tech and graphics to handle such massive sizes

[flat ibex]

@sharp robin trash panda worked 😂

[sharp robin]

upscale?

[sharp robin]

almost always theres a way

[spring fulcrum]

I do use upscalers

[autumn forum]

huh. i was saying 1024x1024 image then upscale to 4096 is faster than generating a 4096 image itself

[sharp robin]

oh thought i was replying to Scorp

[ember hare]

when is sdxl 1.0 open source release coming out?

[upbeat summit]

that's super cute 🙂

[flat ibex]

Lil guy made it to the showdown 😆

[nimble heart]

does comfyui have prompt editing like [woman:goblin:0.5]? Readme doesn't mention it

[visual glade]

no but you can get very similar results with either conditioning combine or concat

[nimble heart]

hoo boy that's a lot of node

[spring fulcrum]

Some of my recent projects favorites

Where can I put some of these where they can be voted on?

[timid sonnet]

I've only been using SDXL on Comfy, if I use A1111 do I use Hires fix as well?

And is the refiner for img2img after txt2img is done?

[upbeat summit]

yeah a1111 styled prompt editing would be wonderful to have. In SD2 I mixed a lot of TI embeds and other tokens together to create interesting things.

really powerful stuff

[upbeat summit]

there's this but I haven't tried it yet: https://github.com/taabata/Comfy_Syrian_Falcon_Nodes

[nimble heart]

so I found that the less human more "other species" the better it works. Dragonoids/dragonborn are easy, government lizard people are doable, "woman with green skin" is very hard. I kinda managed to make it slightly more doable with a super cursed dual-prompt node layout that I 100% wouldn't recommend to anyone, and the color still bled into her hair so that's fun.

maybe if there's some fantasy races with a really strong weight that's basically just "human but purple" it'd work a little better. Like a Twi'lek but with hair.

[eternal fog]

Is there any documentation on what these do?

I kinda get what combine would do, but not sure about how concat would work.

[upbeat summit]

interesting - thanks for sharing your research!

but at least you can do hybrids now

[nimble heart]

it's like the balance meter in a tony hawk game where the left is feral monster and the right is normal human and it's constantly trying to throw you to one or the other

[visual glade]

concat is like BREAK in a1111, combine is like AND

[eternal fog]

ok, never used BREAK in a1111

[upbeat summit]

yeah I noticed this as well. when I did a lot of prompt editing with SD2 I always tried to find this sweet spot and then try to push it in the direction of the mix or style I wanted to explore. but it's not working the same for me right now.

right now my no. 1 issue with my prompts are the hands . I'm stuck with this since 2 days lol

[nimble heart]

I just gave up and try to compose the shots in such a way that the hands aren't visible

without controlnet you're basically fighting rngesus

with cheats

[upbeat summit]

yeah - I tried so many things now. In some prompts it works alright, but I got so many images that I find interesting but the hands...

[nimble heart]

I think it's a fundamental problem with diffusion. It interprets fingers more like scales or hair and not like a constrained structure.

[upbeat summit]

this is one of the better ones (1/50?) of the prompt I'm currently working on but it's still problematic.

[nimble heart]

Also it doesn't understand kinematics, so it may internally consider things like a closed fist and an open hand two completely different structures.

[timid sonnet]

No one wants to answer my questions?

[timid sonnet]

?

[nimble heart]

I just get one where the arms look decent and run the hands on a fast inpaint for like 50 batches

[nimble heart]

Only if you want a hi res fix

[true wing]

I think the problem is that hands exist in this wierd space between fluidity and rigidity, as well as having many small parts which don't take up significant space so the amount of pixels it has to figure it out isn't that high.

[timid sonnet]

I tried using hires fix on A1111 and it took a long time to generate

Do I use the refiner in Img2img only

[upbeat summit]

Sorry, I haven't used SDXL with A1111 yet

[trim orbit]

hands have millions of possible poses and configurations. very complex little sausages with very indiscernable patterns

[true wing]

Exactly. If they were more fluid 'flaws' wouldn't be so noticeable, if they were more rigid they'd be more predictable. But they're neither

[nimble heart]

faces while having many expressions only have one layout so they're not bad. The pattern is so simple you can see faces on everyday objects or the moon :^)

[autumn forum]

i hate how the refiner trys to make a face out of everything. wish it didnt. lol. cuz it works well, but damnit it trys to hard to maake faces that are not meant to be faces

[trim orbit]

https://tenor.com/view/im-not-sure-what-to-do-with-my-hands-ricky-bobby-will-ferrell-talladega-nights-confused-gif-21688469

[upbeat summit]

yeah, that's definitely part of it. so I guess no spell casting for now till we have ControlNet or some fine-tuning that helps. I tried all the tokens and prompt techniques I can think of right now

[nimble heart]

but hands have many different layouts, so I'm guessing it treats them more like a "pattern fill" such as fur or scales and just kinda adds fingers until the gap is filled

[true wing]

Honestly it won't surprise me if at some point we end up with a refiner that identifies problematic areas like hands and then passes them off to a specialized model that only does hands or whatever the problem is

[nimble heart]

think they mentioned investigating a "many experts" model like that

[upbeat summit]

the latent space is trolling again

[eternal fog]

Hmm, this keeps randomly happening when using the latent composite

[true wing]

Yeah there's been some speculation that's how GPT4 works, no confirmation yet but it makes sense

[trim orbit]

giving you the middle finger

[upbeat summit]

yeah. I guess image segmentation is the way to go anyway. background, foreground, single objects and subjects

[true wing]

Coincidentally the 'many experts' model is similar to how the brain works with individual regions assigned to types of tasks. They're not completely separate but they are specialized

[upbeat summit]

no joke. generated a couple of minutes ago

[trim orbit]

that aint bad

[true wing]

Agreed. I also think image segmentation will likely play a role in the next round of training. Instead of just training the image dissect it into it's parts so the ai can 'understand' what each component is and how they fit together

[trim orbit]

one of those so close you can taste it kinda results

[upbeat summit]

yeah there are many pieces to do that already available. it just needs to be in a nice package now 🙂

[nimble heart]

sdxl that's not what a fucking woman looks like

[true wing]

be honest, did you type horny woman into your prompt?

[upbeat summit]

okay after 300+ test images to fix hands - time to move on to a different project and return some other time 😄

[trim orbit]

sounds similar to the crop conditioning they did in sdxl

[trim orbit]

with big pumpkins

[nimble heart]

"woman with red skin"

oh I just dragged a workflow into comfy and firefox quit existing that's nice

[eternal fog]

When I used Latent Composite it turns everything into flowers for some reason lmao

[upbeat summit]

no prompt or when it doesn't understand it does mostly nature images for me

[true wing]

I'm not completely familiar with that but I think that's more about breaking the image into chunks where as I'm talking more about defining whats actually in an image for the AI

like Meta's new segmentation model would let you pull an object out of an image entirely

[eternal fog]

Yeah I don't know if I'm doing something wrong, or it's SDXL, but Latent Composite just goes to shit

[trim orbit]

yeah. there are a few conditioning layers. its about attention conditioning and i think what you're proposing could tie in with the same kind of work they're doing with the crop conditioning attention goals.

the sdxl report is a pretty neat read

i know what segmenting is already. Its one of the coolest arenas of computer vision imo

segment anything is huge and the disruption waves it has caused haven't made landfall yet

[upbeat summit]

in some 16:9 sizes that are outside of 1024x1024 like 1600x904 or 1820x1024, I get a couple of pixels with a white gradient either on the bottom or right of the image - in almost all generations

[true wing]

Definitely. Even the basic application of identifying objects and/or parts of objects are huge. Then you add the potential to increase the amount of data AI can pull from already existing images because it now has context and I think it's almost certainly going to lead to big gains

[trim orbit]

with textual inversions, you can already kind of do that. it takes the alpha mask as an input to put attention only on the subject left unmasked

[timid sonnet]

AI is going to end up out of our comprehension, our brains have not evolved enough to be able to even grasp some of the ideas or understandings that AI will eventually present

[trim orbit]

by that point we'll have become augmented intelligence and the ai machinations will be our own

[true wing]

Honestly I haven't looked for an actual example of how accurate textual inversions are at it but I'd be shocked if it's anywhere near Meta's model

[trim orbit]

oh TI embeddings have nothing to do with segment anything. it's for training embeddings for stable diffusion i mean

[timid sonnet]

We've only recently discovered how the human body continues to go through puberty and a girl can begin menstration (her period) even with zero brain activity

[trim orbit]

just an example of how a mask focuses training already

it's not that far out of an idea

[true wing]

Interestingly this is what Isaac Asimov speculated the end game of robotics to be, a state where robots and humans were indistinguishable. A lot of people take the wrong message from his works because they had to have conflict and inevitably included 'hostile' AI/robots, although really the point was more that it was the people that were using them wrong.

[timid sonnet]

If that's the case then technically our mind is essentially a separate entity from our body so why wouldn't we be able to extract the human consciousness into something else?

[trim orbit]

yeah i'm a big fan of asimov and the other greats of scifi past

[trim orbit]

if it exists and happens in the universe, it's not magical and it can be engineered

[nimble heart]

well it's still finnicky but I think those new nodes are on the right track. Color doesn't bleed into the background and it's so green it looks photoshopped

[upbeat summit]

You know how this is going to go - Matrix, Skynet - you don't give the AI rights and it will rebel

[autumn forum]

i think a problem for our future could be that most of the world is ran by older people that regect new concepts of technology therefore they will percieve it as hostile even though it is not

[timid sonnet]

@upbeat summit 1824 x 1024

[nimble heart]

combine node seems to alternate between conditioner A and B for each sample

[trim orbit]

suppose we gonna give toaster rights now too then! rarwararawrar

[nimble heart]

you say that like your toaster doesn't already have rights you slave-owning colonizer

[true wing]

Eh those types of stories always attribute a level of emotion/malice to the AI. Like there's no reason skynet would care about if it exists or humans wipe it out. The much scarier part of AI, at least to me, is that it actually doesn't care at all.

[timid sonnet]

Any form of life that is essential to our ecosystems or any form of life that can display higher level intelligence should be given rights

[trim orbit]

my toaster loves it's life in my home! i treat it right!

[nimble heart]

much better

now you're ready for Roko's Basilisk