✨｜sdxl

[molten gull]

one more try with the tobasicpipe removed completely

ok, trying that

still not working

[meager canopy]

No idea, works fine for me. Start with the example workflow (in the Aura repo), but if that doesn't work it's going to be a local issue.

[meager canopy]

That's good, but there's something screwing it up in your workflow. Now build up on that basic workflow, until something breaks...or it works 😄

[molten gull]

i m confused 🙂

it maaaay be a thing that it doesnt want 1532x1024

yeah, i think it hates not squares

[meager canopy]

That's too large, but it can do other ARs

[tawdry current]

id assume its like most 1024 models where they dont like being pushed too far beyond 1 megapixel. if you want a 2:3 portrait ratio like that, try something like 832x1216

[meager canopy]

I'd recommend this node, the sizes should work

[tawdry current]

comfyroll also has a pretty good one like that as well

[meager canopy]

The "magic" of SDXL 😄

[limber kindle]

Pixart Sigma base + SDXL refiner and bucket upscaler

[peak dove]

Definitely overrated - looks like from all the promo pictures "that it's good at text!!!" It Ain't! Join the FAL Discord if need be!

[meager canopy]

It's an early beta, but my very first image from it was this #✨｜sdxl message

[peak dove]

I mean it's good - is it exceptional at all?

[meager canopy]

Is any early beta release?

[peak dove]

Good pictures - still not the visual acuity of SD3!

[meager canopy]

It has potential...but it's probably too latge for most

[peak dove]

I've been using it online ... its OK, I don't think remarkable ... but that's just me 🥳

https://fal.ai/models/fal-ai/aura-sr

[limber kindle]

I played around with SD3 a lot and tbh just don't like the look most of the time. It has this Crysis screenshot look to it oftentimes

Plus It's only good at landscapes and close up portraits. Going outside those boundries it's not as good as guided SDXL imo

[peak dove]

Someone said that "SD3 The ReMix" is in the making?!

((((SD4))))

I'm into ipiv's AnimateDiff Master of Motion lately ... it's fun!

[limber kindle]

Master of Motion? Can you link?

[peak dove]

https://civitai.com/models/372584/ipivs-morph-img2vid-animatediff-lcm-hyper-sd

https://www.youtube.com/watch?v=AUKQfJZX-78

[limber kindle]

Ah I know that, played around with it when it came out

Stuff lik that : D

[peak dove]

Cool

[copper kraken]

makes sense, yeah i just see that a lot in images around here

it's distinctive and def worse off for it

[tawdry current]

Oh he changed it "Note

PAG may produce striped "noise", setting sigma_end to 0.7 or higher may reduce striped patterns."

Used to specify sde/a samplers, guess it's all of them

I'll have to keep that in mind, but I don't really use it anymore. I use an attention preset in automaticcfg that also does pag

[copper kraken]

gotcha

i think the issue mostly manifests when ppl have it cranked naively

it's usually accompanied by a vague blown out look to some parts of the face

[fathom surge]

anyone know why comfyui keeps giving me mlp generations when using pony v6 xl? (in a1111 im getting humans with same prompt)

hmm i think i figured it out, comfyui uses -2 clip skip

but it didnt change anything 😐

[gloomy lark]

wow really neat

[limber kindle]

Thanks \o/ Should actually revisit that workflow again

[gloomy lark]

awesome stuff

i should really try getting that going with ella

[peak dove]

Is Ella a cut above? Is it like Ollama at all?

[gloomy lark]

It gives sd1.5 t5 llm level prompt understanding instead of just clip. I don't know how well it'll work with animatediff but it's worth a shot.

[gloomy lark]

This is ella refined with kolors. Even single subject images benefit because you can use longer llm (gpt4 or Claude) enhanced prompts.

[vast galleon]

if only he didnt have that dumb sword going through his hand ...

[vapid flint]

Really? I didn't see anything about this on the page for it

Also, using comfyui, is it possible to set the denoise for different parts of the image to different values? I'm trying to do img2img but I don't want some parts of the image changed as much as others

[vapid flint]

Also pony does not know what a poncho is

I'm going nuts because pony keeps giving my character glasses even though that's not in the prompt, what the hell?

[vast galleon]

Yes, you would need to set a mask for each area then chain the nodes so that it batches the masks together and generates for each mask at the different denoise ratios and then combine the images back on the original.

[wheat rose]

all you have to do is look at the pony prompts

[vast galleon]

something went wrong lol

[vast galleon]

these are from 2300 loras. Thinking I might release this one.

[vast galleon]

Somehow got a house-train lol

[vast galleon]

Almost got a pose right. Close ...

... nope lol

[gray dock]

guys is there such thing as the ReActor but for short form videos?

[vast galleon]

totally didn't prompt for that. lol for some reason it likes to double up on dragons

[vital ermine]

Vegetarian Dragon, or vegan?

[vast galleon]

It's a trap. He's hunting vegans.

[vital ermine]

Lol, hahahaha

[vast galleon]

Portlandic futurism 😂

The cosplay in Portland is something else.

[vital ermine]

Well, Portland has always been a bit weird

[vast galleon]

this model is refusing to do anime and it's a little disturbing.

[vast galleon]

that's an interesting effect. the details are nice too

[vital ermine]

yeah, I was thinking the same

[glass forge]

Taste my Gom Jabbar.

[vital ermine]

I like this effect

[frigid willow]

A close-up portrait of a puffin with its beak open, holding several fish in its mouth, set against the dark background of ocean depths. The scene is captured from behind and illuminated in the style of studio lighting, highlighting intricate details on both the bird's face and tropical fish scales

[full bluff]

Draw an American man and control a puppet dog with strings

A spiritual creature

A puppy

[stray warren]

🐉

[shy kelp]

I started using sharksampler 🙂

also I use this as my latent for everything now

do you know if there is a better one?

I want a latent that makes small details

velvet noise is so spiky

that it kinda works

and I only denoise at 90%

[copper kraken]

not sure, there's a few different things out there

that one is pretty good

i always denoise to the finish line myself but mileage may vary

[shy kelp]

I found that it depends on the noise

this is my current one

maybe for this I should go all the way

but some of the other noises for special effect

they lose the special effect if you denoise too much

a certain way of doing pink noise, denoised to 90%, basically adds colourful clouds and smoke and fog

its cool

noise more like this

and it gives you special effects like this

[copper kraken]

awesome

noise is a ton of fun

[shy kelp]

I am pretty sure we still have the sampling 100% wrong

spend the last week reading some books about solving ODEs

I am finding that DPM++ 2M at 150 steps

[shy kelp]

is easily losing to even just Bosh3

and DPM++ 2M is normally one of the best

SD3 seems different

[copper kraken]

yeah

ODEs kinda level off in gains pretty early

those high step counts are where SDEs shine

[shy kelp]

I'm worried that SD3 is stiffer than we thought

and so we should all actually be on implicit methods

there is an old method where
each step it switches between explicit linear multistep and backward differentiation formula
depending on the local stiffness

I think we need that

[copper kraken]

interesting, what's it called

[shy kelp]

LSODA

the explicit linear multistep in this case is Adams

there is a fourth order implicit Adams in the ODE sampler node
I think this is probably the best we currently have that I can find

but it will be slow because it can't switch

[copper kraken]

gotcha

if you find pseudocode for that lemme know, the lsoda

[shy kelp]

yeah it might have been brought somewhere

ok so I just ran RK4 to the same tolerance as Bosh3 and it performed much worse
this is very bad news because
that means stiffness

[tawdry current]

No, it's the best trade off between time and quality. That's why it's one of the most commonly used

For actual higher quality, SDE samplers tend to be superior most of the time, but the trade off is that they are slower

[shy kelp]

oh I agree
my preference for SDXL is to split the sigmas and at least the first bit is done with SDE
then ODE to finish
sadly no SDE on SD3 though

[tawdry current]

As long as they use the same noise schedule, it's usually fine to do that. Just watch out for samplers that do different things on the first and last step

Some samplers do an euler last step iirc

But yeah sd3 is very different

So you can't use all the same samplers with it

[shy kelp]

if I used less steps then splitting the sigmas would be really bad
because the sampler has to start from scratch
but since I like to generate at like 300 steps or more
the switching cost is relatively small

TBH I have to slowly go through to code of everything in my workflow
cos I don't know exactly what each part is doing

like yeah doing euler or heun on the last 1-2 steps

I've seen that

its hard to know exactly what's going on

[tawdry current]

You can also do stuff like having an overlap between samplers. Like let's say you're doing 100 steps. Sampler A might go to 50/100, but sampler B can always start at 40 and go the rest of the way to 100. Those ten steps of overlap can sometimes smooth things out if you're shifting sampling methods

[copper kraken]

you can just change the noise scaling/eta in most cases as you go

that's what i did with the "clownsampler" node and a few others... you can schedule eta to be an arbitrary value at each step

[vast galleon]

I released V4 of Mangled Merge XL if anyone would like to mess around with it. 🙂

https://civitai.com/models/447902/mangled-merge-xl?modelVersionId=645241

[vast galleon]

Still can't seem to get a horsefly on the back of a horse...

Makes for some interesting horses though

[vast galleon]

Juuuust over the line with the 2D/Realism bias.

[thick obsidian]

#1237459938901491852

[copper kraken]

blood, ketchup, or paint?

[shy kelp]

fruit syrup

[shy kelp]

thanks this is a great tip
gonna try this

I have to learn how momentum sampling works
because momentum sampling lets the second sampler carry on from the previous one's work
it might not work on SD3 though (might be random elements)

[meager canopy]

I bet those headless dinosaurs can't wait until evolution kicks in for them

[peak scarab]

Please I need some help with an art editing

Who can help out

[tawdry current]

Yeah it can help a lot in some scenarios because the second sampler doesn't have to have the same seed even and the overlap helps blend any jarring composition changes if the sampler switch happens at a low percentage of the way through diffusion. Usually most primary forms are all blocked out by ~30% of the way through though. The momentumized samplers are pretty good as well like RES, but yeah, sd3 doesn't really like anything that does any sort of ancestral/SDE sampling

[shy kelp]

yeah this sounds good

I don't wanna spend much more time optimising sampling

because what I have found is

at longer steps the difference between methods sort of goes away

since I use high step count I don't have to optimise sigmas as much either

I will try Clownshark's Clownsampler as well because that lets you dynamically switch from SDE solver to ODE solver LOL

I already use his Sharksampler 😄

Sharksampler is coloured noise injection

although I tend to do it manually because I split the sigmas so much, I figured I may as well inject some coloured noise before next sigma starts

[lusty temple]

white cat

[limber kindle]

Dat water $$$

[copper kraken]

that's only gonna matter with anything that has each step depend on the last one somehow (ancestral etc)

overlapped steps can absolutely smooth things out - i've got another little hack that can do that in another way, the "eulers_mom" option which basically uses the euler method to project removing a bit more noise at each step

if you want more detail, skipping a step somewhere can be pretty interesting

[glass forge]

terrifying

[vast galleon]

I know, right?! 🤣

[hoary saddle]

anyone have a good solid IP adapter + controlnet workflow that yields good results for combining 2 images to make something new while maintaining some elements from both original images?

[tribal lantern]

Tried in kolors

[vast galleon]

Nice! Haven't played much with the new models but the fact that it can do that has me interested. Sadly I'm only running a 3060 with 8g vram so i'm limited. Can it run on 8g?

[tribal lantern]

No idea, i use a cloudgpu, it has a nasty llm that's huge https://github.com/kijai/ComfyUI-KwaiKolorsWrapper https://github.com/MinusZoneAI/ComfyUI-Kolors-MZ

[vast galleon]

Damn, 16g. Of course it can make a horsefly on a horse. LOL... but it didn't get the macro photograph or size factor 😆 . I'm sure with the right prompt it would be possible though.

I kinda like the ability to make really weird looking creatures though.

[tribal lantern]

Didn't promot for that 😉 But i love weird mixes too, sdxl is best for that, mashing the embedding weirdly together is lost in these newer models :/

[tawdry current]

Yeah easily with comfyui. I run it just fine on a 2080 with 8gh vram

[vast galleon]

I might look into it. I wonder if anyone is training on it. At this point it seems like there are so many options

[tawdry current]

Just make sure to use the quantized 8bit text encoder

The comfyui wrapper addon has instructions on how to use the Q8 text encoder

[vast galleon]

how does it compare to pixart? I know a lot of people are training on that architecture now.

[tawdry current]

Visually? Far better than pixart sigma. Prompt adherence? About the same, but kolors does seem to have some concept bleeding if you make things too complex. Usually with colors, like you name a bunch of things in a scene and describe more than a few colors in it

At least that's what I've noticed

But all models will break like that if you go too crazy on specific details

[fleet zinc]

o_o

[frail pawn]

Swamp denizens:

[tawdry current]

looks like comfyanon added in a node for the union model to pick which mode you want. hell yeah

before, people could get results a lot of the time, but it wasn't perfect since the cnet was missing the vital extra information tensor responsible for picking the mode. probably was just defaulting to auto internally, but that might bug out in different circumstances

[tawdry current]

seems to work alright, enjoy the android fish scientist waifu 🤣

[hoary saddle]

send workflow fren!

[tawdry current]

@hoary saddle oh it's just a standard workflow. i just took a couple of images i made in the past and ran them through depthanythingv2 and anyline, to demo using two different union cnet modes at the same time. the only thing you need to is to put the setunioncontrolnettype nodes between and pick which mode you want it in

[hoary saddle]

managed to somewhat duplicate it 🙂 tx

[wet vine]

1

[hallow ore]

confrontation between Patrick Bateman and Anton Chigurh, manga style, scribble art, black and white, serious, showdown, japanese text in the middle saying the final showdown, angry, screaming, furious, final boss fight, samurai in the background, dead bodies in the background, in the style of Yoji Shinkawa

[meager canopy]

Not sure we want "confrontation" 🤔

[high spear]

controlnet union (auto)

[flint scarab]

Native Americans seeing Columbus's 3 ships arriving for the first time

[high sand]

draw: al aqsa mosuqe

[sturdy willow]

hey

[half bear]

Future Photo of John Paul Jones Memorial, with a new bronze donald trump statue addition on top.

[crimson stirrup]

((best image quality、8K、masterpiece:1.3))、1 girl、smile、Full body Esbian、slim face、cute woman、(dark brown hair)、super detailed face、fine eyes、double eyelid、blurred background、slim face、in the hustle and bustle of the city＿Super long brown hair、big breasts、full body shot、((sitting on a chair, night view terrace background:1.3)),Hair with loose waves on the inside、Wearing a satin blouse, ultra-thin fabric、smooth、See-through fabric:1.3、gold necklace＿Larger earrings、((shy smile:1.3)),full body shot、Standing gracefully,

[smoky shard]

That's a bizarre way to type commas, my foreign friend

[shy kelp]

wow this is incredible

I love sci fi cities

[meager canopy]

Thanks 🙂

[shy kelp]

how did you get so much detail

[meager canopy]

Not much space left to fly in here!

[meager canopy]

I have a couple of detailer/upscaler workflows

One for t2i and one for i2i

These are t2i

They're on my Patreon

[shy kelp]

ok thanks

I will join and check it out some time
I haven't gone through the different patreons yet that post workflows

[meager canopy]

That'd be great! It's only in the "Creators Lounge" tier, so choose the right one 🙂

Wondering how 2000 piece jigsaws may look 😄

[shy kelp]

LOL I spend the last week making star wars images but these are so much better

the best I could do so far was this

I forgot how I made it, maybe kolors

I haven't learnt yet how to do massive detail

[meager canopy]

It was with kolors, I remember you posting it...possibly in another Discord 🙂

[shy kelp]

ah thanks

I'm sure kolors just stole midjourney images etc

but they give that style very well

maybe even a bit too much

also he's meant to be firing a laser instead of looking at a brass pipe

[meager canopy]

A bit less "crazy"

[shy kelp]

yeah a bit less is good

[lone ocean]

hjtgj

[limber kindle]

Anyone played with DMD2 sdxl? So far it's insane oO

[meager canopy]

What is it?

[limber kindle]

https://github.com/tianweiy/DMD2

It runs at around 1fps at 1024x1024

And quality is completely on par with other sdxl models

[shy kelp]

ah I don't rly like distilled models

its cool though

[limber kindle]

Tbh one upscales raw sdxl outputs anyway, and so far I don't see a difference in quality

It's mad with auto queue, it's basically realtime

[shy kelp]

8.35 FID though

[limber kindle]

Sorry, but what does that mean? 😛

[shy kelp]

ah its a measure of image score don't worry

this does look cool I will try it

I am mostly the opposite kind of guy though I aim for 30-60 minute generation time instead of 1 second

the other end of the quality-speed trade off

[limber kindle]

I know what you mean, I come from VFX, sometimes one frame can take a day depending on what one does. I don't mind long gen times as long as it's worth it

But since I upsacle and post pro normal sdxl images anyway, I don't really care about being pixel perfect on the initial image

When I can just iterate in realtime, it'S kinda cool

[shy kelp]

yeah its not actually worth it yet
my workflow needs more developing

[limber kindle]

Installed it on my laptop with a 3080Ti, still crazy fast oO

[shy kelp]

wow that looks nice

[lusty wolf]

Are you talking about DmD2?

[limber kindle]

Yep

[limber kindle]

I mean it works with all the loras and normal stuff, but iterating prompts in realtime while still having a very good base is kinda cool

[lusty wolf]

Just stared playing with it.

[high spear]

For a first pass, it works remarkably well, and the times are record-breaking

[limber kindle]

Absolutely! My way of thought, I run all standard gens through upscalers and buckets anyway, so having this is very nice

[lusty wolf]

These look very good. I suppose that is not a basic workflow?

[high spear]

It's a basic workflow, now I am applying upscale with SDXL Lightning, but those images are a batch of 6, with the basic flow

[lusty wolf]

Cool. I will figure it out... This is a SDXL mix still with 4 steps

[high spear]

• sdxl upscale (x1.5)

[limber kindle]

😛

[lusty wolf]

Sorry, rookie question. Do you feed the latent into a second ksampler, and that is connected to an SDXl model. How much denoise?

[high spear]

Set the steps to 4; I set it to 8 if I don't want to upscale

[lusty wolf]

@high spear Thankyou so much. I'm trying on my own... But failing that's why I ask. See what is happening. Like its jiggeling the image...

[tawdry current]

Reminds me of the random architecture in dreams and your brain will be like "this is completely normal"

[limber kindle]

Haha true : D

[tawdry current]

Still really cool though, I like that pic

[limber kindle]

Sometimes can't resist to break stuff, just love abstract high frequency stuff for a change to "normal" images

[tawdry current]

Yeah I feel you there, but I'm a big fan of surrealism and absurdism. Normal stuff is too boring

[hoary saddle]

what are you doing at work today.......

must be bored hah

curse you Comfy and you're fun tool

[high spear]

DMD2 - 2688x1536 ~ fast

[rotund surge]

What model is this? so many objects and particles! epic! 👏

[rotund surge]

Which one of this did you download bro?

[meager canopy]

This is much more about the workflow than the model.

[clever cave]

its nothing new, but still good. lightning merges are better imo and are similar speeds

[tawdry current]

finally managed to get inpainting to work with the new union promax controlnet. i guess it just wants the image to actually have black where it needs to be inpainted. also, it doesn't work with gradients, so no differential diffusion, but from my testing, it's right on par with it

but for real, it's pretty f'in rock solid in my limited testing so far

[shy kelp]

nice

I haven't tried it yet

[tawdry current]

well follow what i did here or manually paint black in mspaint and import that

[shy kelp]

so far I do all my inpainting with SEGS detailer

which has the downsides of YOLO's limitations

it is too censored to know stormtrooper or gun/blaster/laser

[tawdry current]

the one drawback im seeing though for union promax vs differential diffusion(all i use, deleted all my inpaint models because they aren't needed with diff-diff) is that it's image only really and doesn't seem to operate with regular latent masking. so it will be great to taking already made images and inpainting, but if you're trying to do inpaints in latent space between ksamplers, it's useless, unless you want a bunch of vae artifacts from constantly encoding and decoding

[shy kelp]

didn't know people inpainted latents

[tawdry current]

thats the only real way to do it on complex workflows. encode->decode is a lossy process

[shy kelp]

oh I didn't think of that
I encode and decode and back loads

I should stop doing that

[tawdry current]

if it's a presaved image, you only want to vae encode once. if you're generating the image from scratch and plan on inpainting, keep it in latent space and only decode at the very end

[shy kelp]

hmm yeah thanks

I am definitely losing too much quality

I also kinda learnt today that clownshark-style noise injections lower your quality a fair bit

for photorealistic

[timid garnet]

👋 .

[tawdry current]

Oh and another thing about what I said: you can still pull off vae decoded previews at various inpainting steps to see how it's going, but when going from ksampler to ksampler, you want to stay in latent space

[shy kelp]

ah yeah I like that method

branching off to make VAE previews

[tawdry current]

There's a latent bridge node that works really well for doing previews and masking between stages, it saves you a lot of the hassle

Forget which common pack it's in

Probably impact or inspire or comfyroll, can't remember but I'll check

[shy kelp]

its ok i have seen that one before

[vast galleon]

I think the CFG might be too high. LOL

[timid garnet]

👋 .

[spiral lintel]

Is SDXL strictly for generating at a minimum of 1024, or can it do it lower resolutions just as effectively?

[stone kindle]

I use everytime 768/768 .If I like a picture, I increase the resolution

[vast galleon]

I'm not sure if it depends on the model but I've generated some decent stuff at 64x64 testing for pixel art.

[limber kindle]

I just used this https://gist.github.com/comfyanonymous/fcce4ced378f74f4c46026b134faf27a

[limber kindle]

Really nice! Would probably be $$$ with something like https://civitai.com/models/229213/extremely-detailed-no-trigger-slidersntcaixyz

[keen mason]

Bro steam gift 50$ - steamcommunity.com/gift-card/pay/50

[lusty wolf]

UltraPixel

[gloomy lark]

ella/kolors

[rotund surge]

Thanks bro, it worked!

[vast galleon]

Set/Get and Bypass nodes make workflows so much more organized. Dream come true for OCD.

[vital ermine]

yes, as I hate noodles with a passion.

they drive me insane and makes it so hard to track anything down

Though I use everythinganywhere as a lot of the time I could not get those to see each other.

[eager prairie]

What can I be doing differently?

[tawdry current]

Replacing the woman with Willem Dafoe

[eager prairie]

No I mean generation wise. This prompt is random from some prompt generator, and I know its long, but why does it just make gibberish in the form of a photo

[hoary saddle]

nightmare for those that work solely in the API 🙂 way too many lines of python

but it does look much nicer

[tawdry current]

Yeah, plus it causes massive headaches for tracking shit down when you have complex workflows. It's good for beginners and simple workflows though

[eager prairie]

I had chatgpt shorten the prompt significantly and I think the results are better

That was after 6 failed attempts where the woman grew triple the amount of arms

[tawdry current]

Well that's the basic workflow that can get images to completion and get you 99% of the way to what you want. You can inject controlnets between the prompts and ksampler if you want more control over composition. Or ipadapter.

Other than that, Google and YouTube will be your best friends for learning

[eager prairie]

But I want it to be fast. I have an rtx 3060 so I thought everything would've been super quick anyways non-turbo wise, but it wasn't so I've been using this and if I give more than 6 words, it just freaks out

Ok apparently if you remove illustration it makes them naked

I don't like that I get horrible anatomy but then make a typo "hat in cat" and get this masterpiece

LOOK AT THIS

[tawdry current]

look, i don't want to be that asshat, but there are hundreds of people that have already spent absorbent amounts of time creating tutorials for beginners. this isn't something that can easily be explained in a few replies. i highly recommend you do some basic googling and youtubing.

[eager prairie]

I have, I've watched so many tutorials and use the exact same settings and get these weird photos generated

[tawdry current]

well in those videos, they'd also likely suggest using different models other than base models

[eager prairie]

I tried before and got really weird things with their recommended settings

But can you explain why there's such a huge quality gap between those two images I just sent? Even at the same length of prompt and formatting, they are wildly different

[tawdry current]

if you want speed, look into lightning or hyper models. they are 1024x1024 based models and will run just fine with 8gb of vram. the 8 step ones are the best tradeoff between speed and quality

[vast galleon]

I haven't tried it in the api. Time to test it out and see what you mean.

[eager prairie]

I'll look into it I guess

[tawdry current]

https://huggingface.co/ByteDance/Hyper-SD is a good start

you can use the lora with other sdxl models

[eager prairie]

Can I use it just as a checkpoint?

[tawdry current]

no, it's a lora that augments a model. https://civitai.com/models/139562/realvisxl-v40?modelVersionId=361593 is a really good model as well. and i'd avoid 1 step loras and models, they are almost always garbage

[eager prairie]

I have that installed but I don't like how slow iti s

[shy kelp]

are you really using the settings from that screenshot?

[eager prairie]

Me?

[tawdry current]

well learn delayed gratification then or get on meds for ADHD

[shy kelp]

it says CFG 1 and steps 1

and its the wrong resolution for SDXL too

[eager prairie]

But how come it's so slow on a 3060? I saw someone say either it was a similar card but with like 6gb of vram or it was like a 20 something and they did it so fast

With like 30+ steps too

[tawdry current]

oh yeah forgot they had a lightning version of it

[tawdry current]

"Use Lightning models with DPM++ SDE Karras / DPM++ SDE sampler, 4-6 steps and CFG Scale 1-2"

[eager prairie]

These are the recommended lightning settings. dpmpp_sde karras with 1-2.5 cfg and then 4-10 steps

Oh that's turbo

[tawdry current]

read what i just said

[eager prairie]

No I read that but I wrote the turbo one to you by accident

But I'm still within both

[tawdry current]

and since it's a lightning based model, change your resolution to 1024x1024

[eager prairie]

Why it so bad lol

[tawdry current]

youre using 512x512

[eager prairie]

HOLY

Wait do I make seed fixed and then just add something new to the prompt and it stays consistent?

[shy kelp]

fixed seed yes

[eager prairie]

I dont think it really did but still impressive

[tawdry current]

every seed will be different

[shy kelp]

would recommend getting off lightning models
unless it really has to be fast

[vast galleon]

it is pretty ugly in api format. Good thing I set my api up *before * I did all that. 🙂

[eager prairie]

It does trust

I am a very impatient person when it comes to 1AM ai image generation

[hoary saddle]

ultrapixel is scary to install, damn it's huge with lots of dependencies
wasn't sure comfy was gonna restart or my gfx card drivers still work

[tawdry current]

one of the top rated models is a lightning model https://imgsys.org/
it's second only to realvisxl 4.0

[eager prairie]

Ok I changed it back to blue sweater and it looks like the first girl again

[hoary saddle]

holy crap tho, ultrapixel does make nice results on the control net

[eager prairie]

Yo whats the prompt to the 2nd image smuz

I wanna see how this model makes that

Wish I could reverse google search this to find my future gf 😭

[tawdry current]

well youre not going to find a future partner sitting around making virtual waifus all day...

[eager prairie]

No I also make dogs and cats

Anyway

Why this?

Little random error

I made seed +1 and it works, does the seed like run out of uses?

[tawdry current]

same seed, it doesn't create it again because it's already created

[eager prairie]

Ok another question get ready

[tawdry current]

set the seed to random

[eager prairie]

What nodes do I add to upscale my images?

[tawdry current]

google.com

[tawdry current]

https://letmegooglethat.com/?q=comfyui+upscale+tutorial

[eager prairie]

Thank you good sir

I was listening to some weird ai youtuber yap

[shy kelp]

yes but this is testing just a prompt on its own
for a big comfy workflow lightning/hyper/turbo don't really work that well because they have that really limited useable CFG range
the issue is there is no room for IP adapters or control nets to work

[eager prairie]

THATS SO COOL

[tawdry current]

true, i don't like working with models that have strict 1-2cfg ranges. also, yeah, cnets/ipa become extremely difficult to use.

[shy kelp]

I don't know the science of this but they also break noise injection effects

[tawdry current]

well a lot of things are based on step counts

[shy kelp]

the SD3 2B also doesn't like noise injection
but I hope the final SD3 8B will

[tawdry current]

when you only have like 4 steps, that means each step is walking through MASSIVE amounts of the noise schedule

[shy kelp]

oh yeah that's probably why

I didn't think of that

[tawdry current]

but speed models have their time and place. they are fun for prototyping and then resampling with a regular model using much high step counts

[shy kelp]

yeah some people just have the opposite goal to me

and want a decent image in a few second

I am looking more for a workflow that makes the image over the course like like an hour

[tawdry current]

waste of power

[shy kelp]

it is yeah

[tawdry current]

there are really strong diminishing returns on things because the actual model precision is only goes so precise

and each pixel still only ends up with 256 different potential values per RGB channel

[shy kelp]

for a single pass yeah

but within that hour

you can fit like 3-4 additional passes with refiners

and then 2-3 upscalers

[tawdry current]

i mean whatever floats your boat and all, but still, it's a waste of power and time. that same hour could be spent creating dozens of other potential cool things

[eager prairie]

@tawdry current I followed one of the tutorials, took a little long to figure out how to get the manager thing, but I did and now I got good upscaling images now.

[shy kelp]

that's so much better than your earlier images

massive improvement in quality

[eager prairie]

Huge improvement between these two as well

Thats screenshotted when I zoomed in, not full image btw

Besides the # of toes and fingers, this looks awesome

Yo Im actually addicted to this

[eager prairie]

Yo last question

If I was to do this hypothetically in comfyui or another local web app for sd, could I just install everything, then turn off internet and run it to save battery in like a car setting?

[shy kelp]

you can turn off external internet but you need internal network

ComfyUI is a web UI

so it uses a local host address

if you use something like Diffusers then you can avoid that

[eager prairie]

Can you dumb that down

[shy kelp]

I don't know how to
maybe someone else can explain

I'm not good at teaching

[eager prairie]

I tried asking chatgpt but it started contradicting itself and then said something about a tomato I don't know. I think that it needs a connection to just run the stuff locally as like a website kinda ui, but it doesn't actually need internet

No I got no idea

This is scary cause of how real it is

[tawdry current]

Bruh... Wi-Fi usage on a laptop would be like <1% of the power being used to generate the images. I don't think they are even a single watt and it wouldn't surprise me if they were in the 100 milliwatt range. Basically, it would be the least of your concerns for saving battery

[eager prairie]

Well the question was more like aimed towards could I use this in a car

With a bad connection or no connection at all

[tawdry current]

The answer is yes

[eager prairie]

So then how does the whole 127.0.0.1 thing work without internet?

[tawdry current]

You can test it right now by putting your laptop into airplane mode

That IP is your localhost

Aka your PCs own virtual self address to itself

[limber kindle]

\o/

[raw wolf]

Create a vibrant and detailed festival poster with the following elements:

1. Background- Dominant color: Deep red with a subtle pattern resembling woven straw or traditional textile. Large bold text at the top: "FESTIVAL" in dark red with a slight shadow effect for depth.
   2.Main Character- Central figure: A smiling farmer wearing a traditional Vietnamese conical hat (nón lá) and an orange-brown traditional outfit.

   • Position: Slightly tilted forward with arms crossed, holding a bundle of golden rice stalks. 3. Supporting Elements:- Surrounding the main character: Abundant golden rice stalks with detailed grains and leaves.
   • Green swallows with dynamic wings, positioned to the left and right of the character.- Geometric shapes: A bar chart with upward trending green bars and a cogwheel, both integrated into the design around the character’s feet. 4. Text and Labels-
     5Details and Style:
   • Art style: Vibrant and semi-realistic with a touch of traditional Vietnamese art.
   • Colors: Rich, warm tones with bright highlights to emphasize the festive and agricultural theme.
   • Lighting: Soft, natural lighting to enhance the realism of the illustration.
   • Texture: Slight grain to simulate a printed poster.

2. Additional Elements:

   • Behind the character: A stylized mountain range in green and yellow hues, blending into the background.
   • Foreground elements: Parts of a gear and a foot wearing a traditional Vietnamese slipper, partially covered by rice stalks.

[bright ice]

My first SDXL lora

[graceful egret]

Anyone have any recommendations for some models and/or workflows for outpainting with SDXL?

[hushed forge]

Looks wonderful any plans to release it?

[hoary saddle]

ultrapixel

it's so slow to generate on a 4090 24 tho lol

[shy kelp]

ultrapixel is a cool project

[sharp vortex]

Hey hey people!

I wanted to generate some landscape images with SDXL, and so I found a nice picture on civitai and tried to replicate it as a base to go from, so I loaded the metadata from it into SD WebUI, downloaded all the models and LORAs used, but still mine looks horrible in comparison. I didn't expect 1:1 reproducibility, but it doesn't even look like the same style at all. Any idea what is wrong?

Theirs:

https://civitai.com/images/11496743

Mine:

https://i.imgur.com/Tw6d7kz.png

[dusky jay]

/HelloWorld

[bright ice]

Maybe.

[bright ice]

It was supposed to be pencil artwork. Not eating pencil...

[glass forge]

so is ultrapixel worth it

the 10 gigs of space and vram

[lusty wolf]

If you have at least a 12GB GPU, yes.

[glass forge]

welp it freezes for me on stage C

i guess back to Kolors

too bad coz it doesnt max out my vram

it just craps up

[stray warren]

Hand and nails!

[glass forge]

That is indeed an excellent hand.

[stray warren]

It's a Kolors hand

[glass forge]

Kolors is very good AND FAST. It's the SD3 we never got but deserved.

[stray warren]

Yeah, looks like it but there's still hope

[glass forge]

Real open source to the rescue!

[shy kelp]

Kolors is so distinctive

I had a feeling it was Kolors as soon as I saw it

its kinda cool to have a model that goes in a different direction

[bright ice]

Forgive me...

[hoary saddle]

kolors

[bright ice]

Say hi to Reimi.

[copper kraken]

old sd15 generation on the left. result of being passed thorugh my 16 channel unet on the right

runtime about 3-4 sec

[rose canopy]

You gave them souls.

[empty oyster]

Hey, not sure where to ask. What would be the best approach to try to apply a "pixel art" effect to a real image? I have a portrait picture of myself and I want to make it "pixel art".
I am trying with Pixel Art XL Lora (https://civitai.com/models/120096/pixel-art-xl?modelVersionId=135931) but I am getting very poor results. I am doing img2img with 0.2/0.3 denoising strength, prompt including only the lora activator (<lora:pixel-art-xl-v1.1:1>). I have tested multiple combinations of sampling methods, sampling steps and also cfg scale. But I get either completely different faces or just the same picture with some "pixels" (like a super-high quality pixel art).

Any tips/suggestions on how to get this? Ideally I am looking for something like this: https://i.kickstarter.com/assets/044/139/979/c10a1e2f4769063ae87b79167d31ed4b_original.png?fit=scale-down&origin=ugc&width=680&sig=NLWqti%2Fxyd3wfDxLvB7NrKCqOElhzz2xwilywW9wKH0%3D

[gloomy lark]

https://wgmimedia.com/midjourney-pixel-art-prompts/

[gloomy lark]

I haven't tried to do pixel art myself, but you may need better promoting in addition to what you have to invoke the full pixel art style of the base model as well as the Lora. Stuff like 8-bit etc

Most of these models have pixel art training even without a Lora

[empty oyster]

well, the thing is I want to create a pixel art from an existing image. I know how I can create a pixel art from a prompt. Just not sure if (using SD) I can "apply" an style like pixel art to an existing image (using img2img).

[gloomy lark]

Img2img needs just as much prompt coercion as an image from scratch.

Just make sure to only include adjectives about the style and not about the subject

Kolors does a good job with pixel art with 8-bit and 'pixel art' as well.

[rustic garnet]

if you train the model beforehand on your face then it should be simple

[empty oyster]

Unfortunately I don't want it only for myself. It's more like a feature I want to provide to my users. So they can "pixelize" their profile pictures.

[hoary saddle]

made a gallery that pulls from the output folder

[rustic garnet]

you can try IpAdapter but for real person portraits I never found it accurate enough

[unreal agate]

https://tenor.com/fKwAeF62Cs.gif

A peaceful forest morning with sunlight filtering through the trees, illuminating a young girl. Her hair shines brightly under the gentle morning light, creating a sense of tranquility and beauty in the scene. --version 6 --quality 1 --chaos 0 --stylize 100 --aspect 3:4

A peaceful forest morning with sunlight filtering through the trees, illuminating a young girl. Her hair shines brightly under the gentle morning light, creating a sense of tranquility and beauty in the scene. --version 6 --quality 1 --chaos 0 --stylize 100 --aspect 3:4

[limber kindle]

Nice, what model?

[meager canopy]

No idea! 🤣
I've been changing model with almost every image, sorry.

[limber kindle]

Haha all good : D

[meager canopy]

Mainly been trying different Pony variants and my own models.

[limber kindle]

Tbh I rarely play with new models. The last one I tried is https://civitai.com/models/432244/forrealxl

Which is very decent, but have to give it more time

[meager canopy]

Some of the guests were surprised at what he could do with a glass of water.

[reef notch]

vevy god

fluentlyXL_finalHyper8S super

[reef notch]

hi res

[cobalt pewter]

a cat

[final estuary]

/help

https://tenor.com/mwsCT1Udtrk.gif

[vital ermine]

https://civitai.com/models/597313?modelVersionId=667361

[kindred shoal]

does anyone know why this would be happening? the colors are way too neon...

[tawdry current]

That's usually a cfg related issue, but there are a ton of other factors that can cause that.

[young prawn]

Looks a little deep fried, can be cause your steps or cfg numbers are wack.

I’m a little more familiar with 1.5, but you generally want your steps between 15 and 60 with your cfg between 5-12 although you generally want to be somewhere in the middle of those numbers. Too low or too higher number can cause a deep fried output.

[void cape]

hi would it be possible to run sdxl on an rtx 3060m? (6gb)

[hushed forge]

Yes should be possible might be a tad slow. I was using a 2060 mobile for a long while

[covert willow]

Sdxl seems to put out the same kinda image every single time is there a way to have it change it up with out having to change prompt every single time? im using styles but still same thing same angle same expression with slight diffrences. help plz.

[flat acorn]

who likes 1600x1600 native SDXL gens?

90 steps, but I mean

[crimson dragon]

Hi. I'm doing some tests involving image to image workflows in ComfyUI where I want to generate urban architectural images and I keep having this issue where I get really splotchy or "dirty" looking textures, especially on wall surfaces that are supposed to be reasonably clean. Does anyone of you have some insight as to why this happens that you can share? Or can someone point me to some tutorial/resources where I can investigate how to improve my setups in this regard? Help is appreciated! 🙂

[flat acorn]

https://civitai.com/models/347097?modelVersionId=670028

Part one, there's a positive

Truthfully I just need to merge like 4 of these

[valid niche]

[lethal stone]

Big Cat eating burger

[rancid shuttle]

https://tenor.com/bCgxW.gif

[meager canopy]

https://tenor.com/view/nhech-big-lips-eyebrow-raise-stare-filter-gif-16626258

[hallow furnace]

有谁能告诉我,contrelnet 现在进化到哪个版本？如何和sdxl 一起使用？

[tawdry current]

アーマードーコア

[edgy tree]

Thanks！

[toxic wasp]

Reminds me of King Tut and Cat Woman from the 1966 Batman show, except Tut is wearing an Egyptian batman costume.

[timid garnet]

👋 .

[lilac wasp]

hello

how to use?

[austere elbow]

#1237459938901491852 there is a waterfall in the middle of a Blue green to pink gradient landscape, volumetric lighting,beautiful fantasy cave scene,fantasy matte painting,cute,beautiful fractal ice background,cascading iridescent waterfalls,pillars of ice background,floating waterfalls,dream scenery art,ice sculpture,cotton candy trees,pink waterfalls,mystical fantasy landscape,neon light and fantasy,fantasy landscape,3 d virtual landscape painting,frozen waterfall,photoshop water art,fantasy beautiful,crystal forest,lost in a dreamy fairy landscape,colorful otherworldly trees,beautiful fantasy painting,fantasy realm,matte fantasy painting,aqua volumetric lights,epic dreamlike fantasy landscape,volumetric lighting. fantasy,detailed matte fantasy painting,fantasy landscape painting,3d landscape

[limber kindle]

Custom LUTs + 12 exposure step tonemapping

[shy kelp]

the contrast is really high, looks good

[shy kelp]

it would need tiled upscale to fix the cars

I blurred them instead

[lilac wren]

new toy

[sharp scarab]

New audio-reactive geometry system!

[Architecture - Archaic / Ancient - AI]

https://www.youtube.com/watch?v=IOD-eTIm9g0

[rancid shuttle]

#1237459938901491852 taiga

[raw prairie]

1 year of sdxl

[lilac wren]

Guys, what is the best optimizer with LoRA creation?

AdamW8bit, Adafactor, Dadaptation?

[copper kraken]

there is no true best

[shy kelp]

I recently read some papers on this
it doesn't seem to matter too much

[copper kraken]

it's important but you won't find one that's just superior

[lilac wren]

man, It's a nightmare, I've tried so many times to get a good result...
Can I share with you my dataset and profile.json (Kohya_ss)?
If you could please take a look and tell me if I've done something wrong, I'm trying my best to follow good advice on the net, but nobody agrees with anybody.
I try to train a pony model, but it screws up all the time, the results are not clean whereas it was trained in 1024²...

Look :

[boreal bough]

is there a lora or model for ultrawide?

those look really damn nice!

made a fancy lora for DND Elementals :3

[lilac wren]

you trained it yourself?

[boreal bough]

yeah

https://civitai.com/models/603042?modelVersionId=674080

[boreal bough]

Thanks to DoRA I can do a bit more complex things without the need for 2k datasets

[lilac wren]

Do you train with Kohya ss locally?

@boreal bough
Until a few hours ago, I was using a train profile I'd made from data I'd found here and there, but the results were... Questionable.
I've just taken this profile from Civit.ai, it looks much better, what do you think of these settings? I'd like to have the opinion of someone who makes good LoRA.

[boreal bough]

yeah. until OneTrainer finishes the DoRA implementation I'm currently stuck on kohya

[boreal bough]

are you on super low vram?

cause your setup looks like it

you're making a lot of tradeoffs to save vram

[boreal bough]

bucket_reso_steps = 64
cache_latents = true
cache_latents_to_disk = true
caption_dropout_rate = 0.3
caption_extension = ".txt"
clip_skip = 2
dynamo_backend = "no"
enable_bucket = true
epoch = 200
full_bf16 = true
gradient_accumulation_steps = 1
gradient_checkpointing = true
huber_c = 0.1
huber_schedule = "snr"
keep_tokens = 1
learning_rate = 0.0008
loss_type = "l2"
lr_scheduler = "constant"
lr_scheduler_args = []
lr_scheduler_num_cycles = 1
lr_scheduler_power = 1
max_bucket_reso = 2048
max_data_loader_n_workers = 0
max_grad_norm = 1
max_timestep = 900
max_token_length = 225
max_train_epochs = 500
max_train_steps = 6425
mem_eff_attn = true
metadata_author = "Caith"
metadata_description = "DoRA Trained on 257 images. "
metadata_title = "Fairy (for Autism Pony)"
min_bucket_reso = 512
min_snr_gamma = 5
min_timestep = 100
mixed_precision = "bf16"
network_alpha = 1
network_args = []
network_dim = 32
network_module = "networks.lora_fa"
no_half_vae = true
noise_offset = 0.0375
noise_offset_type = "Original"
optimizer_args = []
optimizer_type = "AdamW"
output_dir = "X:/AI/MODELS/lora/pony/character/dnd/fairy"
output_name = "fairy_v1"
pretrained_model_name_or_path =
"X:/AI/MODELS/checkpoints/pony/autismmixSDXL_autismmixPony.safetensors"
prior_loss_weight = 1
resolution = "1024,1024"
sample_prompts = "X:/AI/MODELS/lora/pony/character/dnd/fairy\\prompt.txt"
sample_sampler = "euler_a"
save_every_n_epochs = 10
save_model_as = "safetensors"
save_precision = "float"
seed = 12345
shuffle_caption = true
text_encoder_lr = 0.0008
train_batch_size = 8
train_data_dir = "X:/DATASET/fairy"
unet_lr = 0.0008
xformers = true

is my current setup

This is ideal for a 3090/4090 on a windows machine
(clip_skip = 2 is for pony models only - so clip_skip = 1 is correct if you do sdxl)

can be used with a minimum of around 16gb vram
where you'd have to adjust batch size, and learning rate only.
(learning rate / batch size = 0.0001)

[lilac wren]

No, 12GB, i have a 3080Ti

[boreal bough]

oh. yeah. that does count as super low in terms of training x_x

[lilac wren]

:\

[lilac wren]

Meaning? If you can explain a little I'm lost

[boreal bough]

• training unet only
• using adafactor
• low batch size

all of those impact your final quality, but save a lot of vram

[lilac wren]

adafactor is not better than adamW8bit? And i don't know why, with this settings, i can't enable adamW8bit, cause i have a bug with LR warmup, i can't set on 0, and i have an error when i try to learn, that says, LR warmun have to set on 0, and its greyed

[boreal bough]

adamW would be ideal. adamW8bit makes a lot of tradeoffs already.
adafactor is fine as well - just more things that can go wrong, since its super easy to overfit on it. Resulting in loras where the output gets eerily close to the training images. For example faces always pointing in the same direction, or skin color of faces not matching the rest of the body

emphasis on can go wrong
it can be solved with the right settings, and fixing issues in the dataset. also min snr 5 helps A LOT for this

so top improvements for you would be:
• use min snr 5

your network weight of 32 is perfect for this type of lora.

[lilac wren]

I've just realized that the WD14 caption does a much better job too. How do you formulate your prompts for images? Is it like this? :
sebl4rdpony_v1, solo, looking at viewer, shirt, black hair, 1boy, white shirt, male focus, parted lips, food, blurry, black eyes, blurry background, facial hair, portrait, beard, realistic

[boreal bough]

I have a really really aggressive blacklist, which removes like 95% of all booru tags, and only leaves those that exist in natural language. (and I use: SmilingWolf/wd-convnext-tagger-v3)

https://github.com/jhc13/taggui

this tool is perfect for tagging stuff quickly, with various taggers

[lilac wren]

Hmmm k

[boreal bough]

trigger word, <tags>

or if I'm fancy about it:
trigger word, <caption>, <tags>, <background>

[lilac wren]

and what if, i have sebl4rdpony_v1 and sebl4rdpony_v2.
If i call sebl4rdpony_v2, it will be a risk that call sebl4rdpony_v1 too?

[boreal bough]

in A1111? no

I do that myself xD so I'd know

[lilac wren]

🤣

[lilac wren]

it's always a worry when you've done dozens, and 99% are failures.

[boreal bough]

it gets better with time. I'm currently at a roughly 80% success rate

[lilac wren]

me 80% of failures

[boreal bough]

tbf, adafactor is harder to get right. especially when you cant train the clip models. so thats not all on you.

if you ever get a better gpu, you'll be real damn good at it, since you actually understand how to the train the unet properly 🤣

[lilac wren]

Sometimes you don't know why either, like my head (my avatar), 10 crappy photos, with people around sometimes, blur, bad lighting, and it comes out better than other setup much better prepared Oo

[boreal bough]

most finetunes on civit are just glorified clip loras, that also happened to train unet a bit

[lilac wren]

I bought a 3080ti yesterday, because my poor 2070S (without BF16) was clearly not made for it x)

[boreal bough]

AHHH. right you have a 3080

you can enable full bf16 training

that saves A LOT of vram

maybe enough for you to do clip training as well. you should def try it

[lilac wren]

yeah, in same conditions, but in 768² instead of 1024², 10-15 more hours on train, imagine...

[boreal bough]

if you also train clip, it cuts down training time by about 4 times

so an 8hour lora, trains in 2 instead

[lilac wren]

my 2070S was around 20-30s /it, horrible

3080ti with full bf16 : 1.60~2s/it (1024)

I don't regret my purchase, and I couldn't afford a 3090 anyway. not to mention the hellish power consumption, and the roar of the Concorde aircraft.

[boreal bough]

yeah. I bought two desktop fans, and frankensteined them on top of my 3090, in my training rig 🤣

that fan sound was hell before that

[lilac wren]

Dude, look my 2070S xD :

2x chromax black, from Noctua

before that :

https://tenor.com/view/plane-yay-gif-24049337

[boreal bough]

OH MY GOD'

NO U DIDNT

[lilac wren]

Of course, I imagine you've also been through the process of making a fan curve on Afterburner that will save your ears the most.

[lilac wren]

hahaha

[boreal bough]

my 3090

[lilac wren]

I imagine it must change everything too. People don't realize the crappy fans they put on our graphics cards.

[boreal bough]

2x beQuiet xD

[boreal bough]

0 audio. it runs in sub 30db, meaning you wont hear it unless there's absolute quiet.
also temps in the 50 degrees, instead of 88

yeah. standard fans that built in cost around 1€ in manufacturing x_x

I hate how exactly thats where they save money

[lilac wren]

So very much, and they make us pay the price for this crap

[copper kraken]

It looks like you fried the text encoder

[boreal bough]

"unet only" :' )
he fried the unet

[lilac wren]

??

This one i suppose? Its not my profile, and i don't know what it change

[boreal bough]

thats fine. what your image showed is typical issue if you use a too high learning rate on the clip models. you dont have that issue though

[lilac wren]

fuck this pony in my trigger words...

[copper kraken]

no coffee yet 😄

[lilac wren]

what do you suggest guys, if i want train a pony model, ponyRealism_v21MainVAE is good for this?

i'm trying with adamW this time

[copper kraken]

these are ultrapixel with cascade. just don't use the safetensors they provide, they're fake safetensors and contain potentially malicious pickles

i converted them to real safetensors and forked the repo with the necessary changes to ensure it's safe

[boreal bough]

that explains it 🥲

I remember you talking about your improvements to cascade. was just hoping you found such a good workflow for sdxl, cause of the sdxl chat xD

[copper kraken]

well, there's a sdxl checkpoint node floating around in that WF somewhere lol

not being used though

however, i kinda wonder if the concepts behind this model could be applied with sdxl/kolors

it can even do 1440x5120 no prob

time to churn out an infinite supply of wallpapers

[shy kelp]

the only issue I have with the compressed latent methods (including cascade, ultrapixel, koyha deep shrink and hi-diffusion)
is they reach a much higher resolution but at that new resolution their detail level is lower than what is typical for that resolution

[copper kraken]

depends what you mean by detail tbh

the complexity will be lower, but the resolution of the detail can be comparable with a finetuned version of stage B

in many cases that's for the best imo... myself included, we all have a tendency to spit out images that are too busy sometimes cuz those intricate details/complexity is addictive lol

[lilac wren]

looks better 🙂

[lilac wren]

working very well with me. With the worst dataset xD

[boreal bough]

making a cover image for a lora is weird type of fun 🤣

building the ultimate collage of generations