#✨｜sdxl
pure crystal
half ivy
i am here
high skiff
well, let me see if I can share it actually
pure crystal
my uncle works for nintendo, they're working on ultra nintendo switch
high skiff
ok, yes, I can
lusty moss
my uncle works for nintendo, they're working on ultra nintendo switch
Nintendo 128 I presume..
bright valley
My dad is mr nintendo
half ivy
bro's, im looking for a collection of Comfee templates to play with, where might one find such a directory o0
bright valley
high skiff
@bright valleythere you go, linked what I was talking about. He just dropped it apparently lol
bright valley
this is all of SDXL retrained for some reason?
high skiff
Hes the leader of our research group, and is also responsible for most of the progress on my LoRA
high skiff
this is all of SDXL retrained for some reason?
it added in Zero-Terminal SNR, a paper that SAI themselves said they couldn't train into it, and he figured it out
with the hel of hugging face
bright valley
dude if y'all got a team of people to work on shit at least make a checkpoint or somethin damn 😂
pure crystal
high skiff
he writes all his own training tools, all of his own bucketing code, all of his own everything. I don't even know how to run it, and its for testing for hugging face ATM
bright valley
I mean, I figured out how to train something into SD that they couldn't too 
or migjourney
mid
high skiff
an entire research paper vs a single concept is a big difference 😅
I don't even know how long the whole thing took him
several months
bright valley
eh, I think text is a pretty big concept
it was like the holy grail of generative AI
besides hands
high skiff
I feel like you are playing up what you did a lot... lol
A concept the AI could already intermittently do that you refined
Vs a whole entire way of handling early step diffusion and noise schedueling across an entire model not made for it
bright valley
I'm just being honest about it
high skiff
like the text thing is dope, something I don't know how to do, thats for sure, but SDXL could already do text
just like how SDXL can do realism already. You and I training LoRA's are nothing compared to what he did here lol
glad grove
thats cool and all
bright valley
if it could, I, nor anybody I know, knew about it
lusty moss
Sounds like a good talk for the general chat.
glad grove
Sounds like a good talk for the general chat.
i thought i was in reddit tbh
crisp owl
I'm just sitting here trying to make a giants glove fallen in the forest
bright valley
i made a forest giant yesterday
lusty moss
Trying to get rid of the double letters unfortunately.
noble shoal
Good morning. Unfortunately I haven't worked on it lately due to a shift of interest into other styles. A big problem is, that SD doesn't seem to understand what part of the prompt is actually the text. One letter words often need a repeat in the prompt to get them showing up. So, yeah, currently the Status is "on hold"
crisp owl
lusty moss
yeah thats a knoen issue
All good, just messing with it.
bright valley
Good morning. Unfortunately I haven't worked on it lately due to a shift of inte...
you may think my lora is interesting https://civitai.com/models/176555/harrlogos-finally-custom-text-generation-in-sd
bright valley
All good, just messing with it.
sry didnt finish that thought lol. I'm working on it, I think it may be a training issue, but I'm definitely trying to fix it
there may even be some prompting technique that can help
people keep finding those every day
noble shoal
you may think my lora is interesting https://civitai.com/models/176555/harrlogos...
I saw it and it looks promising. But how many words can you chain together before it falls apart?
noble shoal
Ok, I was at 5-6.
bright valley
but I felt like I had enough to correct with just single words first before I try and increase it
Well, you said yours was busted yeah?
noble shoal
There was a high chance of a Missing letter or a double, but at least it tried to write it 😅
bright valley
oops i didnt mean to send the top left
ignore that
but two words is pretty achievable
there we go
lusty moss
So close!
bright valley
Trying to get rid of the double letters unfortunately.
Moonster 😂
Mononster
I actually love that O too!
bright valley
actually how it does that to the bottom of a few letters
thats a neat effect
my buddy ral was doing some crystal shit with it
lusty moss
(PCMonster:1.2) (Text Logo:1.2), (Green:0.9), ([Glass | Crystal]:1.1), (Caustics:1.1), (Refraction:1.05)
bright valley
lusty moss
Trying different things.
bright valley
make sure you at least try the prompting structure on the model page
it is built to operate that way at least
here's that sarah one i just made
lusty moss
Just added it as some notes.
pure crystal
Santa vs Aliens
lusty moss
Santa vs Aliens
If it is the one from Futurama we will be alright lol.
bright valley
This year, Santa's bursting from chests to deliver the gifts
pure crystal
bright valley
lusty moss
Followed the prompt instructions this time so see what we will get, also trying my new custom nodes.
bright valley
oh no sry you just need the colors
not any of the rest
so its basically Text, colors, style, extra shit
lusty moss
Will try it that way to.
bright valley
it's trained to pick up the first color as the color of the text for example
pure crystal
"home for the holidays"
lusty moss
"home for the holidays"
noble shoal
so its basically Text, colors, style, extra shit
That must be hell to explain to civitai users.
lusty moss
With the prompt i shared.
pure crystal
do tiddies
bright valley
I've done many dumb words, but tiddies is actually not one of them, I'll get right on that
noble shoal
They actually get it pretty well, surprisingly
You have always the "that's not how it is supposed to look like" image submissions. I can clearly see that with my MS Paint LoRA. Some of the Images are like masterpieces, but it's supposed to look like shit. 5/5
lusty moss
Tried any cutoff with it to dial in colors etc?
bright valley
Tbh I've been blown away by some of the shit people have been making with it
bright valley
Tried any cutoff with it to dial in colors etc?
What do you mean exactly
lusty moss
bright valley
Tbh I've been blown away by some of the shit people have been making with it
^ and this is a weird thing for me, as it's trained on my own artwork, so now some people are better at my style than I am 
lusty moss
Kind of a workflow, can use it tailor specific aspects of images.
bright valley
Oh interesting
Haven't tried it yet
I've been meaning to learn more of the comfy stuff, like conditioning and all that
lusty moss
Node sanity lol
bright valley
I have one like that too haha
Problem is I don't know what half the shit does
It's what I made this with tho
lusty moss
Can see the steps basically here
bright valley
I want to build a workflow that makes a bunch of base renders and shows me previews of all of them, and I can select which to pass on to the rest of the workflow
I think that would be the best way to use the lora
lusty moss
Can send you this, would have to download my CSV Loaders custom node though.
bright valley
Is that what it does?
lusty moss
Drop down options for prompts.
bright valley
Oh yeah I wouldn't need all that
Just 1 prompt, but like a batch of base gems, then pausing
And you can select which of the images to send to the rest for upscaling and all that
lusty moss
You use the reroute nodes as "switches"
bright valley
As you've seen in its current state, it can sometimes take a bunch of attempts to get the spelling right
I think there's actually node(s) that provide that type of functionality
I've seen somebody with it, I just forgot what they're called lol
lusty moss
Can use boolean nodes, not sure after that.
bright valley
It's sucks because we were in voice chat on the civit server when they showed me, sharing their screen, so I can't go back and find an image or something
lusty moss
Probably like this, not to sure though. I use the re-routes.
noble shoal
I just use Custom nodes that pass the input. So, latent in -> latent out. They can be muted fine.
Is it already possible to mute reroutes?
lusty moss
You mean like right click bypass?
noble shoal
You mean like right click bypass?
For reroute you mean?
I mean Ctrl+m
To break the chain ⛓️
lusty moss
Yea same thing as bypass i belive
noble shoal
I am unsure, but I think that doesn't stop the flow
bright valley
pretty sure mute and bypass are different
one stops the workflow the other doesnt maybe?
not sure
noble shoal
And reroutes could not be muted. At least it had no effect
lusty moss
I do it more like this, if I want to use Image to Image on the next step I connect i2i to ON, if I want to use Ultimate SD Upscale i connect USDU to the ON. More of a selection.
bright valley
thats pretty neato
lusty moss
If i want to skip that section entirely I switch it to OFF
So my flow goes txt2img, img2img, ultimate sd upscale, after detailer, inpaint. Can switch on and off and interchange models/loras along the flow
Can pass things along as needed and route them through the re routes,
bright valley
yeah that gets wayyyyyyyyyyy to unruly for me
i just have workflows of specific tasks
lusty moss
Have to admit its a lot, can download the base off CIVITAI i think its called Ultimate ComfyUI workflow or something. I have adapted it for my need and updated some of the nodes.
noble shoal
Can pass things along as needed and route them through the re routes,
Nice subway map you have there 
bright valley
yeah i mean if I have a different use case, id much rather click Load than a million nodes and buttons and logic gates and whatnot
lusty moss
I come from Blender 3D where nodes are life lol.
But they have much better nodes for node management. ComfyUI is strong but needs some QOL updates.
bright valley
man i left programming to do art full time
i thought i was done with nodes
these worklows look like goddang db schemas
lusty moss
If anyone wants to poke around with it.
noble shoal
But they have much better nodes for node management. ComfyUI is strong but needs...
You can check out what's possible with the nodes backend on the original repository. Pretty neat. https://github.com/jagenjo/litegraph.js
There's also a demo
lusty moss
You can check out what's possible with the nodes backend on the original reposit...
Will give it a look. Thanks
Check this out for a cleaner ComfyUI front end.
https://github.com/Stability-AI/StableSwarmUI
bright valley
looks kinda similar
crisp owl
lusty moss
I don't mind my noodles though, what you gain on the front end you loose power on the backend sometimes.
bright valley
sekiro vibes
lusty moss
Loving these!
heady vale
bright valley
lusty moss
Going to try yours the the Cutoff to see if I can dial in the colors etc.. last one before bed though
lusty moss
It did mostly, just playing around.
high skiff
@bright valleyso I uhhh, messed up big time with those images I was generating for comps earlier lmao
I had my LoRA ontop of realistic vision, not base lmao
that explains why it was better at some things than I remembered, and worse at others
this is what it looks like PROPERLY compared
it suddenly got better at eyes, and a lot worse at foliage, and I was so confused lmao
bright valley
Looks like you've got a ways to go to be the best realism model if you couldn't tell the difference 😂
high skiff
no I COULD tell the difference, which is why I was confused
I thought the foliage looked a lot worse, but the eyes looked better (which made no sense cause my model didn't finish training)
my LoRA didn't finish training which is why it has splotchy small details
altho I will say, it looks like it plays well with real vision which is nice
lusty moss
And now I have a new logo lol.
high skiff
you can see the difference here
I was trying to figure out why my LoRA was darker than normal
bright valley
Now you're getting it
high skiff
base, real vis, my lora on base, my lora on real vis
bright valley
You wanna see where that style came from
lusty moss
The text style?
bright valley
high skiff
my loRA does seem to fix the issues with real vis a lot tho
I could demo it on base and ontop of realism finetunes, might work even better on realism finetunes
lusty moss
Pretty awesome!
bright valley
Thanks! Apparently it works great as training data too after seeing your gen
The dark purple and neon green, ya can't beat it
Altho- don't sleep on gold
That's the sleeper hit of the bunch
lusty moss
PCMonster text logo, black, green, purple, slime, skeltons
prompt for it.
bright valley
I told you to try the prompt structure 😂
It's like midjourney, it takes the simple prompt and just goes to the moon
lusty moss
Can say it worked lol, mostly lol. Was trying to get black text but close enough.
bright valley
Black is a tough one. If you do black and white it works, but I guess I haven't made many logos in those styles with black text
Here's one I made with gold
lusty moss
Good enough for me.
bright valley
Lol dude I love the pfp! That looks super dope
If ya have the chance I'd love it if you uploaded your favorite(s) to the model page
I'd totally want some of that on display
lusty moss
Will get a few up there in the morning, gave it heart so will be able to find it
bright valley
Awesome I appreciate it
I guess civit is making a training video for the site and their social media's for it, and in the next couple of days I'll be hosting a contest through their bounty system
20,000 buzz on the line 
bright valley
I mean, it's not hard to do it
I can't say I know all that much about getting it right
This was my first model
I think I honestly just got REALLY lucky with all the parameters I guessed on
Yeah man the gold is littttyyyy
It knows dragon and wolf dragon
lusty moss
My name is harder I think because it keeps wanting to put monsters in the image. And can't really put it in the negative prompt lol.
bright valley
Yep I've noticed that
Things that are actual things or objects or whatever can confuse the text encoder
crisp owl
What if you put "monster" in the negative?
lusty moss
Removes the word from the logo.
crisp owl
bummer
bright valley
Monsters not an activation word so you're at least good there 😂
bright valley
PCMonster text logo, purple, yellow, cyan, neon, lights
Dude I'm telling ya, I'm so happy I didn't listen to anything people told me about captioning. It responds so well to the activation words and order
It's actually built to know single or plural too
Like Skeleton or Skeletons
Heart or hearts
lusty moss
Keeps brining up this little dude lol.
high skiff
wow, Real vis and base react very different to my Realism LoRA lol
bright valley
somebody made me do an on/off comparison thing today because they thought it couldn't be my LoRA doing it
high skiff
I tend to like both loRA outputs for different things
bright valley
this was the picture
and when I took the LoRA off, same everything else (Seed, etc), I got this
high skiff
Mine listens more to what I ask for (makes sense, its the model it was trained for), but thew realistic vision one gives interesting results for sure
lusty moss
What Sampler are you using for yours?
bright valley
honestly I feel like Euler is the easiest
high skiff
comfy made Euler and DDIM identical now, so I guess I also use Euler haha
bright valley
for me it gives the best results most often, but if you look on the gallery for it on civit, people get insane results from all KINDA samplers
high skiff
yeah, skill matters a lot more TBH
bright valley
this is what i mean man, people better than me at this shit
high skiff
I just choose DDIM cause its the fastest
bright valley
and at my own style 😢
lusty moss
bright valley
noble shoal
and at my own style 😢
I wonder how the other living artists feel 👀
bright valley
No man like I totally get what theyre talking about now
in a visceral sense
but im doing it to myself on purpose tho
noble shoal
but im doing it to myself on purpose tho
I try to avoid training on art from living artists for that reason.
bright valley
there's a guy zodiac, mod on the civit server, he makes tons of cool pictures with it without any text in them lol
bright valley
I try to avoid training on art from living artists for that reason.
I only use images I made myself so I get it
high skiff
I use any images that are free to use, and nothing copyrighted, so I also get it
bright valley
Logos were what I got commissioned for more that anything else I'm pretty sure
so I've made a ton of damn logos over the past couple years
high skiff
and my own stuff as well, but I have a pretty limited dataset cause of where I live lol
bright valley
So that part, all the work was already done, when I decided to try I didnt have to get or touch a single image
high skiff
I get part of the Anti-Ai arguments and stuff, but most of it is just immense fear mongering and a select few artists feeling threatened when they aren't actually being targeted. I say that personally as an artist and Ai researcher. I have also changed the consensus around AI with all of my creator friends after showing them how it actually works
The amount of false info in the internet made just to scare people about AI is insane
bright valley
yeah i mean im a full time artist now 
noble shoal
I get part of the Anti-Ai arguments and stuff, but most of it is just immense fe...
True that.
bright valley
So I guess technically "artists" includes me
high skiff
Some of my artist friends used to give me major shit for using AI generated content, back when they didn't understand how this stuff really works. And now they all hit me up constantly asking for generated references, imposes, styles, and environments that you just cannot possibly find online
crisp owl
It didn't help that nearly every "prompt" guide that got popular with 1.5 was based on including an artists name (looking at you frank frazatta)
crisp owl
lol
bright valley
lol are you being for real
crisp owl
so much fluff
bright valley
watermark, signature, all that in the negative
you know what you're doing at that point
lusty moss
high skiff
It does depend on what models you're using
bright valley
dude these are wild
high skiff
I make almost no money off of anything related to AI, and absolutely no money off of anything related to the intellectual property of individual creators
Most of my stuff is for personal use, or references for creative individuals in my life
I get really pissed at the share amount of individuals who know nothing about AI acting like they know everything about it online, utilizing their large follower counts to fear monger how the technology actually works
*sheer
Several months back there was a video by a pretty major creator on YouTube, one that I can't remember at the moment, and also one that I don't want to give any attention to, where they effectively said that once an AI sees anyone image you've created, you can basically be replaced instantaneously
This was back well before SDXL was even in beta
high skiff
That looks pretty sick
high skiff
I think it would look cooler if the pentagram was upside down personally,
high skiff
Very splatty
high skiff
The problem with pentagrams in SDXL is that they associate occult pagan pentagrams and satanic pentagrams under the same term. And there are a lot more occult pentagrams out there than there are satanic pentagrams
And as somebody who tries to generate a lot of demonic / satanic imagery, I have almost never gotten a properly oriented pentagram in any of my generations ._.
high skiff
Interesting, I didn't realize you trained your LoRA for tagging instead of captions
I have alternated between both for my trainings, but I found that even if I tag my training images, SDXL seems to respond better if you caption your prompts. At least for what I'm training
glad grove
It didn't help that nearly every "prompt" guide that got popular with 1.5 was ba...
theres also loras for almost every japanese artist out there (for 1.5)
high skiff
Yeah, and a bunch of douches in the community decided to go after specific and iconic artists to try and attack them. Like everybody who decided that it would be a great idea to attack and go after Sam does art
glad grove
sorry to say this once they posted their works online thats when it became public property
high skiff
Like sure, I don't like the guy at all, and I think he was very rude in the way that he handled the whole AI stuff, but I don't think that going out of their way to try and attack his livelihood was the way to do it
It sure as hell didn't give a good light to the community
bright valley
yeah it works in a specific order, which equates to front-to-back if you think about them as layers
every term is in front of the one behind it
high skiff
The most impressive model I've ever used to date is still a 1.5 model, absolutely blowing SDXL out of the water in terms of how well it listens to what you're prompting. And I just recently found that it listens the most to what's in the middle of the prompt
As weird as that sounds
I did a whole bunch of tests and found that it responds differently depending on what concept you place at what part of the model
I mean the prompt
The whole model is just the most insanely saturated and capable model I've ever seen, and it's trained on over 77,000 unique tokens just for its own content, but it's really crazy the way that you can prompt it
bright valley
Like PC Monster was doing what would be quality prompting regularly at first
high skiff
I don't see any SDXL models coming close to even a fraction of its capability, diversity, and consistency anytime soon.
I just seriously hope that the people who trained that model get their nose into SDXL soon, because oh my God, that will be one of the greatest things generative AI has ever seen
bright valley
got stuff like this
first time (that I saw) he used the prompt structure from the model page,
high skiff
One of the most interesting things that I have found about my LoRA is specifically that the original versions were trained to follow a very trigger tag Have you prompting style, and the yielded pretty bad results.
But when I ended up dropping the trigger tagging in favor of simplistic linguistic prompts, the whole thing performed monumentally better
So then since that moment, I've simplified my tags/captions, and found that it's performing even better, although I do have a little less granular control over it. Which leads me to believe that if I mix the simplistic tagging with a ton of images that still cover a broad range of tags, I will allow it to have the highest quality generations, while still having clear control over certain aspects
Generally speaking, when it comes to realism prompting in anything related to SDXL, I 10 out of 10 times would recommend using a linguistic prompt over a tagged prompt. I have yet to see a single model/LoRA That reacts better to tagging than captioning when it comes to photographic images
bright valley
yep that's what everybody told me to do
lusty moss
Going off the rails again lol.
bright valley
Nah man, this stuff is awesome
love how that black paint looks on the wall
hmph, looks like I still need to work on neon a bit more
the neon around the heart tho, shyyiiittt 
peak dove
Harrlogos LoRA "will do great things!!!" 😄
bright valley
Ayyyyyy!
peak dove
... it's getting there!!!
bright valley
That's cool, what is that like children's book illustration style?
bright valley
... it's getting there!!!
Yep we're only on v1 so hopefully I have nothing but room to improve
peak dove
The prompt in SDXL ComfyUI was Neon text "DAFFODILS" fibonacci psytrance
bright valley
Lol! Wow would've never guessed that
peak dove
Same, but "HIBISCUS"
bright valley
I hate to ask you but, did you try structuring prompts the way it says on the model page?
This was the difference for PC
peak dove
OK, will go back to the Model Page ...
heady vale
bright valley
Its kinda hit or miss with anything that's not listed, most things work, some not really
peak dove
Just for fun!!! Oink!!!
lusty moss
Trying to do some inpainting but not much luck
bright valley
Dude I swear every night somebody tells me they're going to bed, or 1 more picture
And then they just spam for another hour straight
late marsh
Pretties with one button prompt
lusty moss
Lol i got 2 min my router auto resets at 4am for updates lol
bright valley
Aww close!
bright valley
Pretties with one button prompt
Damn love this look
peak dove
Gotcha!
peak dove
bright valley
They have 2 tails but we can deal with that
upbeat summit
nice colors
bright valley
Also that kinda stringy style, it's very cool
peak dove
Zentangle is that style
crisp owl
Before
peak dove
Pigs Will Fly!!!
crisp owl
after a bit of photoshop touchup
bright valley
Ayyyyyyyyyyyy!!!!!!!
With the panoramic joint
Damn you understood the assignment 😂
peak dove
Other interesting styles - fibonacci, cubism, vexel, prismatic, pentimento, sfumato, sgraffito, alebrijes, greebles ...
bright valley
Also wtf man that photoshop job is so good
You have them back connected to eachother
peak dove
after a bit of photoshop touchup
Way to go on using PS - Generative Fill?
FYI PS is now available in a WebUI
Some interesting garbled text ...
crisp owl
Way to go on using PS - Generative Fill?
Mostly brushes, clone stamp, eraser I believe. I kinda just go with it. Been using photoshop for not even sure how many years now lol
bright valley
Really a great edit
I have to check them against eachother to find out what changed
crisp owl
I connected it to my webui at one point, but removed it as it was when it was first released and didn't work too well
The U was quickly done, most noticeable to see it wasn't super nice. But it's late, and i"m tired and getting ready for bed
crisp owl
symmetrical at least lol
bright valley
Could've had it through the O though
crisp owl
queue 100
nimble heart
well the rocm guys just disabled the multi query attention unit tests for Flash Attention so i feel like it's gonna be a while longer before it's running smoothly lol
bright valley
Nooooo!!!!! Take anything, but not my multi query attention units tests for flash attention
peak dove
I connected it to my webui at one point, but removed it as it was when it was fi...
When I use 3 inputs on IPAdapter and have Desktop PS open, PS robs too much memory, so IPA x 3 is very slow. The Photoshop WebUI is a godsend as it takes much less memory!
heady vale
upbeat summit
using @bright valley Harrlogos
masslevel text logo, cinematic scifi film still of 3d motion graphics letters made out of metal. fire particles. blue light streaks. floating in space, dark energy, in the background stars and galaxies
nimble heart
Nooooo!!!!! Take anything, but not my multi query attention units tests for flas...
if you dont be quiet ill take something else
interesting idea with the roses
directly rendering 960x1440
heady vale
if you want a softer pink like I have, try Sakura rose
noble shoal
Amazing how SD can't handle rain in front of things. Might be worth a LoRA. But it's like the "behind window/ behind Glass" problem.
nimble heart
I have "blurry" in my negative which might dissuade foreground raindrops
fierce hollow
dalle does rain really well 🥲
noble shoal
I have "blurry" in my negative which might dissuade foreground raindrops
Nah, it just can't handle that.
So, an impossible prompt would be: a crowbar in the rain.
Good luck 🤞
fierce hollow
it's not terrible surprisingly, other than the cane? pickaxe? shaped crowbar
a kitten in the rain shows how it's putting the rain everywhere but in foreground better I guess
high skiff
bro, why does Kohya's Bucket res work so ass
I am gonna have to hand crop and downscale all of these images since kohya's built in image bucketing is so fucked
these buckets are trash, WTF Kohya
1024x640? what the fuck
yes, let me just train at almost half my target resolution
also, might I add, thats below my minimum bucket res???
like how does it get that fucked up
the only thing I can understand is if its some form of bugged cached latents from a test with bad bucketing, but it seems to be inconsistent
shy kelp
Euler not Euler a?
high skiff
I found the issue with the bucketing, after WEEKS of issues
apparently, it only follows your settings if you do not enable "Do not upscale bucket resolution"
So I have to have bucket upscaling on my 4k-8k+ image dataset in order for it to listen to my bucket settings
noble shoal
apparently, it only follows your settings if you do not enable "Do not upscale b...
Wait a sec. Is "upscale" a mistranslation and actually needs to be "extend"?
high skiff
I am not sure
all I know is that after looking through all of the startup statements, I found one that said "Bucket resolution sizes ignored due to "do not upscale bucket resolution" being enabled, please disable to enable custom bucket resolutions"
I seriously think that the fucked buckets have been why my LoRA has had such trash fine details, which makes sense, cause it was training on images as low as 600k pixels, instead of the 1,040k goal
noble shoal
Hmmmm, I might look into OneTrainer today. See what that has to offer.
high skiff
these new buckets are all 1,030k-1,110k now, which should be much better
unfortunately, the increase to proper res also means that BS8 no longer fits in my GPU's VRAM
so BS6 it is in guess
man, this should hopefully increase image quality monumentally
along with the lower LR and the better captions AND training fully, man, this should be a huge leap in all regards
indigo carbon
Are you training with DeepSpeed? that'll probably help you if you don't
high skiff
no I don't use deep speed
indigo carbon
no I don't use deep speed
you should; it's a bit funky to set up, but it's basically AIT but for training
high skiff
maybe some other time then
high skiff
@indigo carbonhow would I even go about setting it up? Is it something that can run in Kohya?
indigo carbon
<@418848098316451863>how would I even go about setting it up? Is it something th...
I'd assume so, it's just a training optimization
indigo carbon
what do you us it im?
I did use Kohya a while ago, it did support DeepSpeed if I remember correctly
it would just need you to either install the pip package or build from source
indigo carbon
but you've never used it yourself in Kohya?
I used it in Kohya_SS a while ago, it should still compatible
high skiff
That doesn't sound super confident 😅
I'll have to ask you a little more about it some other time if you don't mind
indigo carbon
<@418848098316451863>how would I even go about setting it up? Is it something th...
anyways; it should be as easy as pip install deepspeed if you're on Linux, if on windows- it's not as easy to do
high skiff
I am on windows, of course lol
indigo carbon
I am on windows, of course lol
yeah, you'd need to build from source
indigo carbon
it's so dumb that microsoft's OS isn't compatible with their own optimizations
fierce hollow
apparently, it only follows your settings if you do not enable "Do not upscale b...
it's looking at min_bucket_reso, resolution, max_bucket_reso, bucket_reso_steps and bucket_no_upscale, no upscale means it will only downscale images if they are smaller than resolution, resolution is the mean pixel count, max is max width/height, same for min
if your image is 4096x4096 but your resolution is 1024x1024 and max_res is 1344x1344 the output will be 1024x1024
high skiff
it's looking at `min_bucket_reso`, `resolution`, `max_bucket_reso`, `bucket_reso...
that doesn't make sense, as its still downscaling the whole image. It should really be named something different, cause there is no upscaling involved at all
fierce hollow
if you have bucket_no_upscale set to True, images will only be downscaled if needed, if you have it set to False they will be either downscaled or upscaled to match bucket reso
I think it makes sense
fierce hollow
If your image is 768x768, bucket_no_upscale set to False, it will be upscaled to 1024x1024 (with res being 1024)
high skiff
ok, then why would that affect images that are way higher res than the target? they are still downscaling, not upscaling
that would make sense if that decided whether or not to upscale images that have a dimension below your max size, but mine do not
fierce hollow
bucket reso steps being 64 also matters, it's probably looking for the nearest matching w // 64 * 64, and h // 64 * 64
high skiff
I am using 128, but my point is, if my images are several times higher resolution than my bucket, then why would "upscaling" be a part of it at all? that setting should either be renamed, or changed to only affect upscaling input images lower than your target res
I personally don't see how downsampling a 60MPX image to 1MPX has anything to do with upscaling lol
fierce hollow
uhh I don't think we're on the same page, nevermind
high skiff
I am not sure what there is to be confused about, but alright
I personally think that getting a 4000x6000 image to 900x1200 res doesn't involve upscaling, and thus, should not be held back by turning upscaling off
now getting a 450x600 image to 900x1200 does involve upscaling, which I could understand in that case
vital ermine
fierce hollow
in any case just remove --bucket_no_upscale or disable the corresponding gui setting and it should be fine
high skiff
as stupid and irrational as it is, that did in fact fix my issue.
Don't understand why you can set a max bucket res if a single setting makes the max bucket res default to the same max you pre-specified. Might as well just get rid of the button, and set the damn thing to your max square res lol
Works the exact same
oh well, its only just a few months worth of trouble shooting and training down the drain cause of a poorly worded and redundant setting :/
fierce hollow
In Kohya's defense it's not enabled by default
high skiff
if anything, it lead to me getting great results with fucked up training, so I can only hope that the new training runs improve massively now that they aren't being cucked
vital ermine
Been down similar rabbit holes with Kohya myself. We so much need new tools but they are beyond me to make. 1.5 the porn, anime, NAI people have said they went back to over XL. I really wish they didn't have 1.5 to go back to and we would probably have some stellar tools by now.
fierce hollow
1.5 is still objectively better for a lot of use cases so not happening, blame XL if anything
peak dove
I am just blown away by 3 picture IPAdapter!!! Just wait until I branch out to 4!!! Mebbe get another 64Gb DDR5 ... ?
vital ermine
1.5 is still objectively better for a lot of use cases so not happening, blame X...
Plenty of blame to be tossed around but it is all directed at SAI tbh. They have yet to not just pump and dump and with XL it is far too much for the same tools that people used to train with actually work. Amazed it works at all.
NAI has a lot of blame too, per the porn creators. They have flat out said that if it had not been for that leak none of this would be where it is now.
fierce hollow
Yeah, pretty much, was just trying not to be too mean
vital ermine
I am mean when facts are obvious.
Dumping and moving on is just wow to me. Where are the tools at? We will figure it out? No, we will not those who can just go back to 1.5
I saw this in 2.1 and it came back in spades with XL. Same old stone tablet and chisel being used for 2.1 was rough, but doable. In XL with 2 te's and lord knows what else happening the trainers know we need two sets of controls for each TE then go from there.
OT is working on that for a few months but nothing close to releases from what I gather.
vital ermine
high skiff
@bright valley I Just got done talking with a representative from one of the companies I reached out to for research funding. They ended up being so interested that I have an appointment scheduled on November 3rd. I'm really hoping that this goes through.
Proper funding/support is exactly what I need to bring this to the mainstream, because I'm confident that I can create something truly incredible with the proper resources
lusty moss
@bright valley
Dropped you a review and post for the text lora.
vital ermine
delicate kelp
I have something new coming today, new model, home-grown, home trained. CineVisionXL - Keep an eye out for it, will be dropping on Civit this afternoon/early evening 🙂 🙂
crisp owl
I have something new coming today, new model, home-grown, home trained. CineVisi...
pics look good, what's the intended usecase and/or training method for it?
delicate kelp
Cinematic output. Like my NightVision, but for when you want that movie scene feeling
rather than photographic which NV does
crisp owl
Nice, I may use that heavily for a project I'm working on. Been using NV a lot for it, but maybe this will work even better
delicate kelp
as for training, it saw lots and lots and lots of cinematic themed synthetic output. put several hundred hours total into this one, lots of creating movie scene style output in MJ or in NV, working on it in LR to get the feel I want, then into the training pile.
crisp owl
Nice, looking forward to giving it a go 🔥
delicate kelp
I'm also working on an animated model in the pipeline (I've been calling it AnimeVision, but that's not quite right, it's different than anime) that will be another 100% homegrown SCG model. been having fun with that one, it's getting close to a public release as well
clever verge
@high skiff If you have issues with bucketing in kohya, just take care of it yourself. way better control over the dataset. I can probably share a few python scripts that does it for you if you wish.
vital ermine
<@296720689468145664> If you have issues with bucketing in kohya, just take care...
Yes, as here is one for you.
Ran into this a few weeks back but I asked for 1536x640. Standard XL rez and it bucketed my 1024x1024 as 960x960. I had to manually go in and change everything.
don't upscale was unchecked
heavy zinc
Oh okay, could you please tell me how to properly upscale images as you mentioned using several platforms?
vital ermine
indigo carbon
Dumping and moving on is just wow to me. Where are the tools at? We will figur...
I'm not sure if SAI are even moving at the moment, they haven't announced anything about diffusion since the original SDXL release
vital ermine
I'm not sure if SAI are even moving at the moment, they haven't announced anythi...
All I heard was they had started on 3.0 but that was 3 weeks ago. I was amazed since I think it is far too soon to release another flop. Of course emad is on twitter talking about it so they are moving.
vale eagle
I don't trust emad any more
indigo carbon
All I heard was they had started on 3.0 but that was 3 weeks ago. I was amazed ...
oh really? I gotta read all that, link?
vital ermine
I don't keep links. Just find him on X
vale eagle
Conclusion of that tweet: a lot of things are coming
vital ermine
LOL, yep
indigo carbon
could they finally be moving on to a new text encoder?
or just a mini version of SDXL
vital ermine
Sounds like the former
vale eagle
Many new things come after SDXL released. AI fields change too fast.
vital ermine
Honestly, does it really matter in the end if all they do is dump what they have and move on without the tools for what they dumped? Honestly, it doesn't.
indigo carbon
I don't keep links. Just find him on X
last post from Emad is from January 21
vital ermine
last post from Emad is from January 21
Nah, maybe he has another account but it was mentioned in here. I don't keep him in my subs he isn't worthy of that as he just spouts nonsense when he tweeted anyway. I am sure someone will get ya what ya need just realize they are not idle.
vale eagle
vital ermine
There he is
indigo carbon
and no technical details?
vital ermine
and no technical details?
From emad? LOL LOL LOL godd luck
indigo carbon
damn, here I was ready to go and read more papers
vital ermine
ex hedge fund manager is never ex.
indigo carbon
anyways, I am pretty sure whatever SD3 is it will probably either have image conditioning/different text encoder. I'll keep track of this
vale eagle
It is so limited with current architecture. It works but just could do it better.
steady grove
why do so many people hate emad so much here?
whenever i see him brought up it's just a slew of toxicity
i wouldn't be surprised if internal policy was "don't engage with the toxic community" and that's why it's so silent
indigo carbon
i wouldn't be surprised if internal policy was "don't engage with the toxic comm...
nah, I just don't really know. my only complaint is the lack of papers/technical details when announcing about new models
steady grove
as soon as theres SD3 news, the misinformation toxic army will start talking about how it's beyond censorship and they're infringing on human rights
its better that they don't say anything for as long as they can
toxic community members have managed to weaponize most press releases against them . its ridiculous to watch occur
indigo carbon
I mean, it would be nice to know what architecture/changes they've made on the new models but idk.
steady grove
yellow journalism youtube would take any announcement and make a 15min clickbait video about how its the death of free AI if they even mentioned architecture changes
indigo carbon
yellow journalism youtube would take any announcement and make a 15min clickbait...
who cares? I'd just read the original paper
steady grove
you saw how much youtube hate about how censored SDXL was going to be before it released? we all remember that don't we?
the lead up to 0.9 was nothing but toxic hype cycles
stability cares. they've gone silent since
indigo carbon
whatever, all I'm saying it will be easier to make all the optimizations compatible with it before they even release it. AIT took a month to adapt to SDXL because they released it out of basically nowhere
vital ermine
whatever, all I'm saying it will be easier to make all the optimizations compati...
Not their style, as much as we would love for it to be.
steady grove
"sdxl released out of no where" is the kind of toxic hyperbole. Case in point. 0.9 came out and allowed people to test and build before the 1.0 release. but, that doesn't count because of reasons.
they're better off with no communications since everything gets twisted to toxic proportions anyways
honestly expected better from a developer. toxicity i can expect from the flavaflav hype men. but for a developer to claim they had no leeway? the toxic strands weave and twist
i remember when the sdxl 0.9 report got published and people were tearing it to shreds and syaing it wasn't good enough. big members of this community
indigo carbon
well, if there would be an SD2.9 and it'll have the same architecture as SD3 I probably could have modules/compilation scripts ready
steady grove
there was an sdxl 0.9 and you've denied it ever happened. why should they bother to help you at this point?
i'm sure through academic circles, machine learning students and researchers have private access to the research in stability labs. Wider public releases haven't been received well, historically
to me it makes sense that they're a lot more quiet these days.
indigo carbon
there was an sdxl 0.9 and you've denied it ever happened. why should they bothe...
AITemplate wasn't a flexible optimization then, I never denied SDXL0.9's existance
steady grove
whatever, all I'm saying it will be easier to make all the optimizations compati...
i misunderstand what " AIT took a month to adapt to SDXL because they released it out of basically nowhere" because you just contracdicted it there
if sdxl released out of no where, that means you think there was no beta release
what could that statement be if not a denial of a soft release?
communication is a two way street and if anything said is twisted and lied about, people are going to leave that street
vale eagle
One thing that might be announced, stability text to video. They are testing it on bot channel. Just wait for their announcement.
steady grove
they trickled a little bit about that stuff out and did you see what happened?
100% expect them to remain tight lip and potentially not even release that model, after the initial reception
bright valley
using <@145754764414550017> Harrlogos
`masslevel text logo, cinematic scifi film...
Awesome stuff dude! Before I even saw your prompt I thought it looked straight out of a movie trailer or something. Really cool!
bright valley
<@145754764414550017> I Just got done talking with a representative from one of ...
Yeah man if you can find a company to pay you money to train your own LoRA, without expecting anything in return, power to you! I know I'd consider you a master of business negotiation.
bright valley
<@145754764414550017>
Dropped you a review and post for the text lora.
Thanks a ton homie! I had a great time seeing all of your creations last night, you got the hang of it real quick
indigo carbon
i misunderstand what " AIT took a month to adapt to SDXL because they released i...
well, due to using components from FX2AIT, it won't be too hard to adapt it to anything at this point. I'd say it won't take too long to have modules ready after they release it anyways
steady grove
so, tooling improves regardless if stability does a soft release or not? Seems like it only benefits them to stay quiet. All pros and no cons
vale eagle
Don't expect anything until it released. It good for sai and the community.
steady grove
we've seen how the past year has unfolded. i fully expect them to be liberal with information from now on. even emad's last post, there's all sorts of conspiracies forming aorund it all ready. reddit calls it a deleted post but its still there on his x profile
deft coral
how are people even traing SDXL models
steady grove
very obvioulsy it was only deleted to change which gif he used, but accurate informaiton isn't the name of the game. conspiracy theorists gotta make up stuff
deft coral
i can't even get one to train
steady grove
how are people even traing SDXL models
using nvidia gear HEYOHHHHHHHH hehe. but yeah, i don't bother training models because i don't have 24gb. i focus on loras instead
steady grove
you're using youtube guides aren't you?
steady grove
https://rentry.co/59xed3 read this whole thing and go through it over again
vale eagle
It could be trained with 24GB with reasonable amount of time for lora
steady grove
its for anime but it works
indigo carbon
so, tooling improves regardless if stability does a soft release or not? Seem...
the only thing that would need some heavy changes is if they are using something like T5, but that's as simple as changing which pipelines the compilation script imports
deft coral
i wounder how someone managed to get 2 million steps on a single card
steady grove
the only thing that would need some heavy changes is if they are using something...
sd3 i don't think will use t5 since that'll require 24gb gpus
deft coral
steady grove
i wounder how someone managed to get 2 million steps on a single card
don't worry about htat. get it working first. those people got a lot of time and experience on their side
indigo carbon
sd3 i don't think will use t5 since that'll require 24gb gpus
perhaps a quantized version of T5? I experimented the other day on DFIF, quantizing the T5 text encoder barely did anything besides lowering VRAM usage. I was able to go down to 6bit with almost no difference
steady grove
i don't know how that would translate to actual models though. often is said you want higher precision for training. Pixart-alpha seems like the state of the art, they use t5, and claim it needs 23gb. I think it needs more than 23 which is why it's not released yet.
indigo carbon
i don't know how that would translate to actual models though. often is said yo...
I tested that on DeepFloyd IF. barely effected anything besides memory usage
steady grove
using 70gb of vram is why.
indigo carbon
hold up, I'll send the results I got there
vale eagle
This is fine tuning with 2400 images with A100.
vale eagle
You could get 4 batch size using pageAdamW8bit
and even more to reduce the resolution.
deft coral
steady grove
since its batch size 1
batch size of 1 shouldn't be shooting up to 70gb of vram usage
deft coral
im using the wrong optimizer then
indigo carbon
i don't know how that would translate to actual models though. often is said yo...
FP16 T5
6bit
indigo carbon
nice results. what's the vram use like?
of just the encoder?
steady grove
<@575744962310701069>
theres your problem. you're using the dreambooth extension which is the most unoptimized training software
nobody uses that. does the author even maintain it?
steady grove
doubt. it's using 70gb for batch size of 1
deft coral
he updates the software but lazy for optimizing
steady grove
i think i've told you you need to use kohya , like weeks ago.
You're kind of standing in your own way
steady grove
he updates the software but lazy for optimizing
i honestly think he's just grifting his patreon donators
indigo carbon
nice results. what's the vram use like?
the text encoder on its own used 8gb on FP16, 3gb on 6bit
steady grove
https://github.com/d8ahazard/sd_dreambooth_extension the github doesn't even exist
deft coral
wtf
steady grove
the text encoder on its own used 8gb on FP16, 3gb on 6bit
not to bad at all
steady grove
he actually delete it kek
i'm not surprised
it's all just been a donation scheme. everyone saying it works i bet were just bots trying to astroturf hype
indigo carbon
not to bad at all
right? I feel like the degradation isn't even noticeable with T5. so maybe SD3 uses a 4bit/6bit version of T5?
steady grove
right? I feel like the degradation isn't even noticeable with T5. so maybe SD3 u...
maybe that kind of research is why pixart-alpha isn't out yet too. they're working on 4bit versions
steady grove
its stll here
cached copy
indigo carbon
maybe that kind of research is why pixart-alpha isn't out yet too. they're work...
it was pretty easy and didn't need me to do any adaptations to DeepFloyd for it to use a 6bit version of T5, I doubt that's their reasoning
deft coral
hoyah better be good for SDXL
indigo carbon
anyways, 6bit T5 easily fits as a text encoder on normal cards.. maybe that's what SD3 uses
steady grove
theres another problem. you use steamunlocked. means your system is rife with malware and that will screw with your gpu drivers
before you don't listen to me about piracy, remember that you didn't listen to me about the dreambooth webui extension too
steady grove
But pixart-alpha released their captions tool
oh? like, using t5 to caption?
steady grove
anyways, 6bit T5 easily fits as a text encoder on normal cards.. maybe that's wh...
yeah i think sd3's aiming to run on more gpus sdxl does. like it may be another 512 model with bucketing up to 768
steady grove
i probably won't. you'll have another problem because of being root kitted and a virtual machine is using your gpu for mining or something.
steady grove
even for koyah
your browser bookmarks show steamunlocked which tells me you've got horrible security policy and have multiple rootkits running
indigo carbon
yeah i think sd3's aiming to run on more gpus sdxl does. like it may be another...
it also certainly either has image conditioning or a different text encoder, maybe not even T5- idk
we'll see eventually
steady grove
pirated pc games are the #1 vector used to get botnets with powerful gpus. black hats love releasing scene cracks
if you pirate pc games, you're owned. bottom line
deft coral
like if you never pirated before
steady grove
grew up in the sneaker net days
know all the words to dont copy that floppy. used to know the dance
steady grove
youll run into more problems, even with a properly configured kohya. you'll see
deft coral
ya we'll see
steady grove
its because of the piracy
stone fossil
Paradox SD XL 1.0: https://civitai.com/models/188001?modelVersionId=211105
Introducing Paradox, original by Design.
Paradox creates unique content while offering you more control over your creations by generating exactly what you request.
The model doesn't rely on merging; it generates purely original content and is finely tuned to follow your instructions.
It draws influence from over five thousand original images and trained on 65.000 steps, resulting in a meticulous approach that leads to a creative model.
We invite you to load the model onto your favorite stable diffusion application, give it a spin, and share your thoughts with us.
- Special thanks the people who help me and who I care a lot about:
@markorez
@upbeat summit
@Kamikaze(Elon Musk)
@osiworx
@mix
@polar gale
crisp owl
ah hah, that's where sushi kitty came from
native knot
Paradox SD XL 1.0: https://civitai.com/models/188001?modelVersionId=211105
Int...
Downloading post haste, my dude! Can't wait to try it.
indigo carbon
Paradox SD XL 1.0: https://civitai.com/models/188001?modelVersionId=211105
Int...
Great job! I might be using it as part of general model block equation in the future if you won't mind
stone fossil
🙂 ❤️
stone fossil
Great job! I might be using it as part of general model block equation in the fu...
Yes sure the model is allowed for merge and anything.
cyan crown
heady vale
cyan crown
sturdy abyss
very nice tune
stone fossil
Thanks mix love u man. 🙂
high skiff
<@296720689468145664> If you have issues with bucketing in kohya, just take care...
Generally speaking I would love the functionality of that, however I know basically nothing about implementing my own code and programs, and the last thing I want to do is jeopardize the training of my future models because of some faulty code I potentially write
heady vale
indigo carbon
cyan crown
indigo carbon
steady grove
totall same face syndrome
stone fossil
glass forge
So is SDXL worht it? Ill just stick to 1.5 for now.... I;m fairly new to all this, I just upgraded to be able to run locally...
abstract depot
So is SDXL worht it? Ill just stick to 1.5 for now.... I;m fairly new to all thi...
SDXL is worth it
crisp owl
I haven't touched 1.5 except for like 3 images since SDXL came out
lusty moss
So is SDXL worht it? Ill just stick to 1.5 for now.... I;m fairly new to all thi...
Just switch to SDXL everything is moving over an I'd say decent Loras now.
glass forge
ok thanks guys
cyan crown
So is SDXL worht it? Ill just stick to 1.5 for now.... I;m fairly new to all thi...
Yes. Far away from 1.5 and also better than MJ
steady grove
sdxl has 4x the base resolution so its prettynice
cyan crown
sdxl has 4x the base resolution so its prettynice
not only. SD 1.5 base model was trained from ~90 million images. SDXL base model was trained from 6.6 billion images
it has everything inside
glass forge
i trained one lora for 1,5 haha took 20 minutes or so with like 25 images
not sure what awaits me if i want to make loras for xl
worthy lotus
Whats the difference in base_1.0_0.9vae? Should I stick with the default 1.0.safetensors? Noob here haha
worthy lotus
👌 thanks for the heads up
steady grove
i thought they rereleased 1.0 and thats all old news
cyan crown
Lora for XL takes way longer
steady grove
1.0 thats available now has the 0.9 vae right?
cyan crown
I do on runpod because with 4070 it takes too time
crisp owl
Well for the base model on the website. Standalone vae is rereleased. Not sure if they rereleased the base+vae. I just stick with manual vae anyways
steady grove
yeah i use ollins half precision one
cyan crown
is it possbile to train SDXL models
yes
deft coral
but what settings tho
steady grove
cyan crown
I train Lora not model so can't help
deft coral
grate
cyan crown
but SDXL has really everything inside
cyan crown
oh wow
deft coral
when i tried SDXL earlier i blue screened my gpu
cyan crown
wich gpu?
deft coral
RTX 4090
cyan crown
it's good also for training
deft coral
i can only find lora guides online not model ones
cyan crown
start training a Lora
deft coral
i can get one trained in 4 hours
cyan crown
how may images for 4 hours?
deft coral
700
cyan crown
good result
with mine settings on 4090 (runpod but it's the same) 2hrs with 50 images
deft coral
whats runpod
vital ermine
What you get after eating Taco Bell
cyan crown
since I've 4070 that is good for SDXL but not for training sdxl, when i need a Lora I use Runpod 0,74$ per hour
vital ermine
16GB is really the min and 40GB is comfort
cyan crown
Just do Loras
vital ermine
you could do it in 14GB but not all that easily. Lots of quality things you have to ditch. For instance on a LoRA (forget about a dreambooth) you would need 16 dim cause more the mem requirements shoot up
cyan crown
vital ermine
Anyone trying to train XL in 8GB is a masochist.
12GB is doable but no less. have to use all sorts of tricks
cyan crown
12Gb is a pain fopr training sdxl, but good for using SDXL
vital ermine
remember XL is 4 times the size, datawise, than 1.5 and 3 times over 2.x
one trick people do is train the lora XL at 768
can work for styles
Hell, for DB I'd kill for 32gb, or 40GB for xl. 24 I barely eek by if I attempt prodigy.
worthy lotus
XL should only be used with 1024x1024?
vital ermine
XL should only be used with 1024x1024?
Or one of its predefines for best results, yeah
lusty moss
vital ermine
and get an RTX A6000
More ram slightly slower lower wattage.
My 4090 stays with me until it, or I, die.
deft coral
renting another card will get expensive in the long run
vital ermine
look, something is wrong with you, or the card if you have a 4090 and can't train. just being 100 as I have a 4090 and train DB as well as lora and locon. I struggled with the shit but if you want a lora just use the predefineds in Kohya_ss gui as they work you just need to tweak to fit your dataset.
lusty moss
About how much ram do you need to train?
vital ermine
23.2-23.6 but I use BS10-14
much less for lora
if I have like 6 images I can BS16
I mainly DB then extract
lusty moss
Ahh have a 3060 12GB so out of the question for me.
lusty moss
Can you run 2 cards at once? Like seperate PCIE lanes but still train?
lusty moss
Would it really have to be sli could you half batch it?
steady grove
i dont think you can train ai with sli
nvlink you can share memory between both, but sli you can only use one memory bank
sturdy abyss
just started testing LoRA training SSD-1B using a modded kohya
@heavy gulch thanks for adding the category, and @drifting jackal for the quick response as well!! Expect a new version soon, but uploaded the current epoch.
vital ermine
lusty moss
sturdy abyss
also very exciting is @stone fossil 's Paradox model tonight:
https://civitai.com/models/188001/paradox-sd-xl-10
Have fun! Make art.
high skiff
newest training test for my LoRA looks more promising. Hope its enough to impress my potential sponsors
I have until Monday to come up with some new tests as well, which could be very useful. I have an idea for a new test as well
vital ermine
torn river
Wtf do I have a wish-version of SDXL or why my outputs are like this regardless of the level of detail in my prompts (and negative prompts)?
vital ermine
heady vale
Wtf do I have a wish-version of SDXL or why my outputs are like this regardless ...
wrong image res. use these only
#✨｜sdxl message
vital ermine
torn river
wrong image res. use these only
https://discord.com/channels/1002292111942635562...
Huh, interesting.Thank you! I picked one, but still kinda bad. What else am I missing in my settings here?
vital ermine
heady vale
Huh, interesting.Thank you! I picked one, but still kinda bad. What else am I mi...
best results are 1024 x 1024. try that
torn river
best results are 1024 x 1024. try that
Got slightly better but still kinda rough
solar herald
Anyone have any kind of image to share generated with the base SDXL 1.0 (31e35c80fc). I am testing this SDXL but I can't find any image online to replicate to test that I get everything working. I am looking to do the asuka test of sorts for SDXL.
I am getting weird results and I just need someone to show me a proper render
heady vale
Got slightly better but still kinda rough
It will always come down to how good your prompts are in the end. if you want a photo, start the prompt with photograph, dont add hyperrealism, or realism.
Just have "photgraph of sand dunes" or similar
solar herald
All the images I find have the parameters stripped. I legit can't find a single SDXL image online with parameters so I can verify my settings
Very frustrating
heady vale
All the images I find have the parameters stripped. I legit can't find a single ...
you can find prompt examples on https://civitai.com/
torn river
It will always come down to how good your prompts are in the end. if you want a ...
I tried some example prompts I found with example images included and they came out nothing close 😄
heady vale
in order to have the exact image you need the same resolution as well, which isnt provided on civitai. sometimes the device or software you are using will be a different result
vital ermine
torn river
in order to have the exact image you need the same resolution as well, which isn...
I see. Well I keep on exploring. Thanks for your tips though!
delicate kelp
Hi everybody, happy Friday! I'm happy to announce a release of a brand new trained SDXL model, CineVisionXL, a new trained model I have built with several of the awesome folks from my Discord channel SCG-Playground. This model has been trained on widescreen (16:9, 21:10) formatted images and can produce stunning "straight out of a movie" scenes that are coherent and beautiful.
Read the full patch notes and what went into it, along with the dozens of samples for you guys to play with in the gallery!
bright valley
Hi everybody, happy Friday! I'm happy to announce a release of a brand new train...
Nice, been looking forward to trying this out, thanks for your awesome work as always
Also Dynavision is my absolute favorite checkpoint to use with my LoRA, it's like a match made in heaven.
solar herald
you can find prompt examples on https://civitai.com/
They have bunch of nodes and stuff. I just want a extremely simple image rendered on the automatic1111 to test my installation.
I don't have the comfyui or ability to do that node stuff.
delicate kelp
Also Dynavision is my absolute favorite checkpoint to use with my LoRA, it's lik...
Love the look!
vital ermine
heady vale
They have bunch of nodes and stuff. I just want a extremely simple image rendere...
many people using SDXL are using comfyui, but if you dig through a bit more you should be able to find some A1111 gens with metadata
delicate kelp
I've got another model in the works too, started as an anime focus, but it's kinda turning into it's own thing. it needs a lot more work, tho if you keep an eye on my discord, you'll see it a lot.
bright valley
Love the look!
Hey I can only take half the credit 😉 guess we make a great team
solar herald
many people using SDXL are using comfyui, but if you dig through a bit more you ...
Is there any way someone has any render. I've been literally searching for 1 and a half hours for an reference image
bright valley
What do you need?
solar herald
Getting quite frustrated that there is not a single "try this image to verify that you have installed correctly" type thing
bright valley
On what? Usually there's default values for almost everything
solar herald
I need extremely simple image generated on base SDXL 1.0, can be anything, but I need to see the parameters, seed, etc. I then try to render the same image and verify that I have installed correctly
delicate kelp
Few more. I still need to give it some more training, but I like where its going
solar herald
I have stuff behaving very weirdly and I get bad results
bright valley
What UI
solar herald
Automatic
bright valley
Okay, and you have an XL checkpoint loaded, and try a text to image gen?
solar herald
Yes, but I am not quite sure if there is even a problem. That is why I would need a reference. Known good result
To compare
bright valley
Uhh.. well does it look good or not
solar herald
Ahh, forget it
glad grove
we cant know if theres a problem if u dont post the images
lusty moss
Ahh, forget it
I can make you one real quick I have Auto1111 and ComfyUI
bright valley
That was a genuine question, I'm only trying to help lol
lusty moss
Give me your prompts (pos/neg) and settings and we can compare
heady vale
Yes, but I am not quite sure if there is even a problem. That is why I would nee...
your gens dont need to be exact. just enter all the same details into A1111 and you should get a similar result
bright valley
If there's something "wrong" with your installation/setup, you'd be able to tell pretty easily I imagine
lusty moss
@bright valley Few more i turned out today for my son and girl friends online handles.
bright valley
So if you create a decent looking image, you're good, just need to tweak parameters
bright valley
<@145754764414550017> Few more i turned out today for my son and girl friends on...
Ayyyyyy!!! Love it man! Cat is a very cool activation word
lusty moss
And some for a friends band, more did it out of practice but offered them to him to use.
bright valley
Made these for a friend last night as well
lusty moss
bright valley
Dduuuuudddeeee
Super dope!
Man I'm such a sucker for that neon green and purple
lusty moss
Man I'm such a sucker for that neon green and purple
Same and the font reminds me of the pop punk bands lol.
bright valley
I used to play in metal bands for a few years
So my style is heavily influenced from those days
lusty moss
Makes sense. Do kind of wish there was a way to pick the font a little better though.
bright valley
It works on more of a style basis. I tried to get as many styles as I could in v1 (pixel art, graffiti, tattoo, anime, etc)
There will be more to choose from in future versions
worthy lotus
How fast is everyone's s/it? I'm just trying to find out if mine is slow or something
glad grove
depends on gpu,whats yours
worthy lotus
Just my old a2000. Usually between 1-3s/it which is pretty acceptable to me but sometimes it'll have a process that's doing like 6s/it and I'm trying to figure out why it's doing that
Might be throttling or something
glad grove
with that 12gb card you should get like 4it per second
worthy lotus
O wow ok I'll have to look at the installation again
bright valley
Depends on the render too
glad grove
yea 4it/s at 1024x1024
bright valley
Some operations I get 5it/s, some are 10s/it
glad grove
if u are using auto1111 just switch to comfyui or foocus,its faster for sdxl
worthy lotus
Yeah I'm on a1111, I'll give the other ones a spin
How do I know what renderer it's using tho? Noob me I just pick model and everything is pretty default
bright valley
I just mean it depends on what's involved with the render you're doing
worthy lotus
Ah gotcha
bright valley
For example if I'm generating animation frames with Deforum, it's also doing optical flow cadence, and using a depth model, etc
So those will be more resource intensive operations
Also if you're using control nets, those would effect it
Stuff like that
lusty moss
Post apocalyptical Sweeney Todd, time for your cut.
vital ermine
LOL
lusty moss
Pretty cool though
vital ermine
bright valley
What is it/s? Iterations when it comes to samples?
They're the same, it/s is iterations per second, but if iterations are taking longer than a second, it'll switch over to s/it
vital ermine
lusty moss
After your cut lol
lusty moss
vital ermine
lusty moss
vital ermine
lusty moss
How does she breathe lol
vital ermine
Just think when she gets to 70
glad grove
she doesnt breathe shes made of plastic
lusty moss
Thinnest nostrils oxygen breather I've seen lol. Must be a robot.
vital ermine
vital ermine
https://tenor.com/view/hypnotoad-futurama-toad-gif-12116346
Yeah, girl out in the middle of a field of shrooms with a nice plate of them. I dunno man.
lusty moss
Actually really like this one.
kindred fable
I noticed the price of SDXL 1.0 on https://platform.stability.ai/pricing has reduced to 1/10th ($0.002). There is no news about the price drop though.
bright valley
Yeah, for their API, and likely their sandbox too
lusty moss
Ooof.. local or bust for me.
bright valley
Same.
lusty moss
Homie got bills to pay.
bright valley
Like I love to animate, and for a great while Warpfusion was the cleanest and smoothest render you could get it AI animation
And I wanted to use it so bad, I didn't care about the monthly sub cost or anything either. It's just it works through a colab notebook.
bright valley
I see that XL is $0.20 now that's wild
$10 = 5000 SDXL images
How do they even turn a profit from that 😂
lusty moss
How do they even turn a profit from that 😂
bright valley
Either I'm a dumb idiot or their margins have to be infinitesimal on XL
lusty moss
Or the audience is LARGE. A lot of people cant afford GPUs i guess.
vital ermine
lusty moss
vital ermine
Actually really like this one.
Thanks
lusty moss
Messing with Chibi
lusty moss
vital ermine
crisp owl
weeeeeeeeeeeeeeeeb
lusty moss
The WEEBiest
vital ermine
lusty moss
It's official, need a weeb model lol
lusty moss
do [32-bit pixel art | voxel art] in the same prompt. Might add more 3D pop. Interested to see the results.
bright valley
once I get a full training job that doesn't fail, 3D is going to be an activation word
lusty moss
If you don't mind lol
vital ermine
lusty moss
same seed?