✨｜sdxl

[cerulean summit]

sure

[gloomy lark]

https://civitai.com/models/257749/pony-diffusion-v6-xl

they mention it on their page

[hoary saddle]

cascade still my fav

[cunning nest]

sdxl 0.9 is great still

[cosmic quiver]

In my opinion SDXL with LoRAs is still better than SD3 in many aspects

[rustic garnet]

which means it's time for Loras on SD3 😉

[red minnow]

LG7 Concept

[wooden edge]

It's hands down better than SD3 (pardon the pun) I can bang out artistic abstract images all day in SDXL just using prompts and my custom workflow. SD3 just can't produce things like this currently

[vernal cloak]

@verbal coyote Is your AnyLoRa XL model trained with tags? Or is it trained with natural language captions?

[verbal coyote]

both

[glass forge]

not for much longer tho

faces in sd3 are really varied and realistic

sdxl and sd15 look samey

so lol so much for consistency!

which we never had relaly so far to begin with

[glass forge]

Will this channel be archived too like cascade and SD15? 😄 :))))

[glass forge]

holy shit whats going on here, nothing makes sense, is this a michael bay movie? XD

[copper kraken]

it fucking better not

[glass forge]

we are running in all directions sir and shooting and exploding buildings

[solar pier]

/dream,set up a minimalist yet magicaforest

[scenic pier]

Nope, but SD3 channel will be.

[sturdy sinew]

Waw!

[meager canopy]

Stop that!

[tidal patio]

What are ya'll thinking about my realism approach. Think the skin texture should be a little bit less

make sure to watch in full ress

[willow lake]

A Indian student angry on University

[copper kraken]

Here is the image you requested.

Congratulations! As my 1000th prompt, you have won a free upgrade to SD3 for this image.

[past dagger]

Man on moon

[meager canopy]

Did you lose? 😄

[fleet comet]

is anyone here an exepert on installing stable diffusion?

I keepp getting this message

[lilac wren]

you do know that posting a jpeg error is an insult, don't you? There are even countries that sentence people to death for daring to post error logs as images.
(pastebin.com)

[fleet comet]

well i'm sorry but i need help installing stable diffusion and I need to show someone what the problem is

[lilac wren]

https://tinyurl.com/369sm98b

https://tinyurl.com/37b5r7en

[stray warren]

Holmes in search of sharks

[gloomy lark]

I can tell you have the cosxl going. Colors really pop with it.

[stray warren]

No I have not (it's not available on DrawThings), this is Mobius

[gloomy lark]

Wow really ok I guess my phone is just making it look extra special

[stray warren]

It is very vibrant rich colors, contrast... really good stuff

[copper kraken]

i've gotten crazy color with my dora too

really helps when the convolutional layers are trained

[gloomy lark]

Makes me wonder if your stuff is like cosxl where it goes outside the normal bounds of sdxl colors

[copper kraken]

problem with training on base vs cosxl is you have to be really, really thorough to cover all bases if you want that pop

yeah, it def does

it's capable of much much darker images than any finetune i've tried too

the issue is if your training set is limited (as in, 691 images like mine, not millions) and not very broad in scope, you end up picking up the muted sdxl base colors the second you step outside the bounds of your trained captions

is mobius capable of really dark images?

[stray warren]

Yeah, I really want to try your dora. It looks outstanding, very curious how it will do in my DT workflow...

[copper kraken]

https://huggingface.co/ClownsharkBatwing/CSBW_Style/blob/main/csbw_225tkn_r96a48_12epochs_base1-0-fp16_save-fp32_mixed-no_prodigy_decoef08.safetensors that's the best version so far

trained on base

[stray warren]

haven't tried it, I have a darkart mix which can do that

[copper kraken]

good with default stuff but my nodes/workflow really brings out the pop and detail

[stray warren]

thanks, downloading it

[copper kraken]

can do shit like this if you have the nodes too https://github.com/ClownsharkBatwing/RES4LYF (WF embedded)

[stray warren]

I'm not able to run comfyui, don't have a windows pc

[copper kraken]

mac?

[stray warren]

intel mac and ipad pro m2

[copper kraken]

https://github.com/comfyanonymous/ComfyUI You can install ComfyUI in Apple Mac silicon (M1 or M2) with any recent macOS version.

idk if you have silicon

not too familiar with mac stuff tbh

[stray warren]

In don't have apple mac silicom, it's an intel mac

The ipad is silicon

[copper kraken]

ahh gotcha

[stray warren]

That's why I'm using DrawThings which is actually pretty good

[wet nacelle]

@uncut steeple Check PMs please

[sick fulcrum]

Is there an extention like latent coupling for sdxl?

[stray warren]

Hip hip ... for @copper kraken 's lora 😁

dora

[copper kraken]

love it 😄

[gloomy lark]

pixart/aventishorizon - cosxl

[south horizon]

these are cool - did you do more?

[glass forge]

so sdxl still top dog?

[hazy remnant]

What's the best canny controlnet for xl finetunes right now?

[vital ermine]

until a community non merge model and/or sd3 large shows it is better.

[daring marsh]

cat

[solar rune]

for some reason shrek is completely intact

[meager canopy]

4k

[gloomy lark]

jeez... forgot sd3 for now. hunyuan is where it's at.

[stray warren]

Can we run it in DT?

[gloomy lark]

What is DT?

[stray warren]

Oh nm, wrong discord... haha

DrawThings

[gloomy lark]

ah ok

[jade solstice]

is this the distilled version?

[gloomy lark]

I haven't tried that. I assumed that's only for phones and very small devices

Full size is only 3 gigs

[jade solstice]

sorry but where did you see that? the distilled model is 6 gigs in their repo and they say it only reduces inference time, not vram usage

[stone kindle]

What are the best settings to train an SDXL character model with kohya at the moment? All the videos on YouTube are pretty old. For example, how many steps are necessary? What other settings should be made? Thank you ...

[void cape]

any tips on speeding up sdxl generation

im on a 3060m (6gb vram)

[peak dove]

Is RES4LYF a workflow?

[copper kraken]

it's my node pack

[peak dove]

"No matching distribution found for math"

[gloomy lark]

whoa... downloading it now.

that would be huge if it's twice as fast and no performance drop

[copper kraken]

you probably already have math anyway

if not, just pip install it directly

[peak dove]

Do you have a wf png or json?

[copper kraken]

sure, lemme paste one

[peak dove]

Thanks

[copper kraken]

https://huggingface.co/ClownsharkBatwing/CSBW_Style/blob/main/csbw_225tkn_r96a48_12epochs_base1-0-fp16_save-fp32_mixed-no_prodigy_decoef08.safetensors and this is the most recent version of my dora that's used for that image (3 days ago)

np

you prolly have that i think

[gloomy lark]

so the distilled version of hunyuan is like hyper-sdxl or lightning. it lets you run with half the steps, so twice the speed. seems to still look good.

i'm using a refiner...

so the speed is much faster, and the output still looks great.

[jade solstice]

do you see any difference in quality between the two?

[gloomy lark]

distilled / original

although the seeds aren't going to perfectly line up

knowing that it's going through a refiner, i don't think there's an effective difference.

they both come out looking great.

[void cape]

is it possible to use sdxl with a 3060m?

[peak dove]

Too many missing nodes - Manager can find none of the missing nodes. Update.bat was run however no change!

[copper kraken]

those are all my nodes... can you check your console log for errors?

my guess is there is an import missing

[gusty swift]

Create an action-packed scene from the movie "Rampage" featuring Dwayne "The Rock" Johnson. The scene shows The Rock standing heroically in the foreground with a determined expression, holding a high-tech weapon. In the background, a massive, genetically modified ape (George) is rampaging through a city, with buildings collapsing and debris flying. The sky is filled with smoke and explosions, and helicopters are circling overhead. The overall atmosphere is intense and chaotic, capturing the high-stakes action of the movie.

Cinematic, realistic, high-detail, with a focus on dynamic action and dramatic lighting.

[bold mesa]

g

[frosty rapids]

ok but next time do it yourself

[wind fjord]

democracies and other political regimes in the contemporary world

[humble cape]

Juggernaut xl + luma ai = liminal space found footage

[cerulean summit]

cool!

[shy kelp]

Hi🐣

[dense sequoia]

how long does it take to gen one?

[humble cape]

Thanks !

[humble cape]

Depends on the queue

[silent kite]

Best model

[timid garnet]

Old 'something riding something' with SDXL :

[silent kite]

I am just trying to see if I can replicate those more complex prompts from sd3 channel. Never really tried with xl to do stuff like this..

[meager canopy]

3960 x 5080

[timid garnet]

Omost is what you want.

[silent kite]

Seems to work without omost just fine. But I've heard of it.

[timid garnet]

Looks very good. I'm curious about the workflow for this upscale

[meager canopy]

It's from afro_man and should be in the image.

Workflow is on Civit https://civitai.com/user/Afroman4peace

[bitter marsh]

Turn off your room lights and see this image for 10 secs

[copper kraken]

btw give it a shot now... i updated requirements.txt to contain:

matplotlib
pywavelets
numpy```

and nothing else this time

[peak dove]

Manager cannot locate these missing nodes, sad to say

[peak dove]

I have too many conflicted nodes not importing - prolly my bad!

[copper kraken]

i was wondering with the requirements.txt command that doid'nt work for you before

in the RES4LYF folder: when you ran "pip install -r requirements.txt" you got an error

[timid garnet]

Adjusted my upscale to 9.5MP

[peak dove]

No error

[copper kraken]

k cool... do the nodes load now

[grave palm]

wow amazing 🙂 workflow Comfyui?

[timid garnet]

I'm generating from a discord bot using A1111 API

It's basically Hires, then Ultimate SD Upscale with high denoize (0.5+), high padding (256+) for consistency and Soft Inpainting to be seamless.
The one above is not seamless tho, was testing.
The images is always split in 9 tiles for USDU, whatever the ratio.

[grave palm]

thanks for info

[fiery island]

Does anyone here use SD 1.9?

[meager canopy]

There is no SD1.9...is there?

[fiery island]

yes

1.9.0

[timid garnet]

Raw generation / Details++ option

[timid garnet]

Ah, this is the webui, not 'SD'. As SD is the model

[fiery island]

a okay

[timid garnet]

"I believe"

[stray warren]

Just some SDXL cuties for the kids 🙂

[rotund surge]

Amazing, idk how but Luma AI won the race against sora 😁

[timid garnet]

Night sky. (Zoom-in)

[copper kraken]

that's fn awesome

[humble cape]

Yes haha

Now we need a localy and open weight model with the same level

[rotund surge]

True, I hope svd gets updated and could achieve something close (but the requirements would be huge i think)

[limber lynx]

This is from a upcomming SDXL model I am about to release

[timid garnet]

Looks great

[timid garnet]

any suggestion for a retro anime style Checkpoint/Lora for Pony or SDXL??

[limber lynx]

thanks.. this one is from a released one.. (Colossus Project XL 10B) with the workflow I published for it.. I am still working on it though (WF)

[timid garnet]

Hô, very cool !

What denoize are you using while upscaling ?

With a1111 and Soft Inpainting, i can use a higher value (0.48-0.52).
Also always splitting the image in 9 tiles, it helps to keep consistency as well. (between left and right eyes on a portrait, for example).
Before my last webui update, i was also able to run Adetailer in each tile.
Finally, Soft Inpainting puts everything back into the image in a perfect seamless way.

[limber lynx]

The workflow should be inside the image. I use comfy ui

It also works well with Forge though

this one is a raw image from Colossus 10B I made with Forge without upscaling

GN its getting late here

[timid garnet]

Hô wow

*Did you try Omost with your model ? It's pretty amazing

[tropic turret]

I'm keeping things simple, but this is so much fun! 😄

[limber kindle]

SDXL+Deforumation https://www.youtube.com/watch?v=78QZtbduvbM

[placid spade]

Fast food slogan

[livid basin]

Big fan of the CosXL model that Stability AI released. One final model variant for SDXL 😄

Very good bright and dark generation ability

Visibly improved from SDXL original

[tropic turret]

This still makes great images.

[meager canopy]

Using my CosXL merge from https://civitai.com/models/240590/ultimateblendxl

[daring marsh]

绿色眼睛猫咪

[copper kraken]

Here is the image you requested.

[daring marsh]

在哪里创做的，我怎么没看到有创作的地方

[limber lynx]

Because of the current licence of SD3 I made up my mind and do something unreasonable.. As a checkpoint creator I normally holding back some checkpoints. to release later.. this is one of them. Its based on my own research. Today I will release DEMONCORE 4.1 MIDGARDBEAST. Try this with SD3 🙂

[timid garnet]

Megaloceros, nice

I tried one a few dayz ago

[paper parrot]

guys here is a way to add T5 to sdxl by i'm not a programmer so i don't know how could we do it : https://github.com/AIGText/Glyph-ByT5

[glass forge]

so cosxl is good at everything?

those hands are mind blowing

[timid garnet]

@high skiff To not spam SD3 channel
Here is another test with the guy.
Before / After

[limber kindle]

Took 3K film grain against a grey background and re-interpreted/upsampled it with prompts in SDXL. It then renders in 3K straight from SDXL. Those examples are not upscaled. But very hard to control. A 0.01 decimal will change the output.

[meager canopy]

It's a merge with cosxl

[gloomy lark]

Hey. So I learned something about cosxl merging yesterday because I'm a dummy that never did the math before. 🙂

[gloomy lark]

Those input blocks 0-4 should be 0.1, not 0.5 or 1. That means it will still work but give the highest weight to the existing model you're folding into it. Definitely better results.

I mean output blocks.

[meager canopy]

Nice! I merged it a while ago and can't remember what I put in there.

[gloomy lark]

The default workflow has 0.5 and I had done 1 thinking that was max my model, it was the opposite.

Putting 0 is bad as well, so 0.1 is the best of all worlds s

I did it with mobius yesterday and it's massively more prompt following than 0.5 or 1.

[meager canopy]

I didn't want to block out what CosXL is capable of though. It is sometimes better at composition. I'll give it a try later, thanks.

[gloomy lark]

Möbius is probably better at prompt following than most any other model. Cosxl is better than sdxl base, but these other fine tunes are massive better than it. I've now got a workflow that does mobius cosxl and refines with aventishorizon cosxl and the output is great. @meager canopy

[prisma shard]

庆余年网页设计

[meager canopy]

Did you mean input or ouyput, because input looks terrible if I change those.

Output are the only ones defaulting to 0.5, all the rest are 1.0

[gloomy lark]

Yeah output only. The same blocks that default to 0.5 in that workflow. Just set those 0.5s to 0.1

[meager canopy]

These are all with defaults...

Same seed/prompt with 0.1

[meager canopy]

What did you set this too?

[gloomy lark]

I left that alone

Also depends on your input model. Only looking at the hands, your default looks better than the 0.1. But maybe your own model isn't great with that. In side by side with mobius and Adventis, they both turned out better

[meager canopy]

This is Mobius

[meager canopy]

Setting min to 0.001 look better

...or not. I'm not sure if any look better or just different.

[gloomy lark]

Hah that's like 99% of what we mess with. Better or just different hah

[meager canopy]

Big difference between the fp32 and fp16 Mobius too

Those were all fp32

[glass forge]

After the recent developments I foresee a long and prosperous future for SDXL.

[meager canopy]

fp16

[gloomy lark]

So as far as I know, mobius has a fp32, but cosxl only has an fp16, so the resulting merge is fp16. I used the fp32 input. Are you aware of a cosxl fp32?

[meager canopy]

NO, but the output of using the 2 are different.

...after merge

[gloomy lark]

Ah. Ok well I'm glad I used that one then. 🙂

[meager canopy]

Just realised my seed was still on random 😦
Fp16 and 32 are identical in output

[gloomy lark]

Independent of cosxl, there is a difference between those, but the majority of the time it's not something you'd care about.

It's not like a concept is different, just that a pattern on clothing is a tiny bit cleaner.

[meager canopy]

I meant after merge

[meager canopy]

My model after merge

Mobius with same seed/prompt after merge

It's a bit more "dynamic" 😄

[gloomy lark]

Oh, I forgot. I like every other model, mobius hates perturbed. You have to get rid of the nice completely for it not to be messed up.

Unlike

Node

I was getting the same issues until I got rid of it

[meager canopy]

Not using it, that was during the merge.

[gloomy lark]

Can you share the prompt, that merge probably just has lame settings by default.

[meager canopy]

cinematic film still Power Armor, sexy young woman with an angry face, ready to fight, On a rainbow bridge to Asgard, Energy gauntlet, power-enhancing technology, Powered up in brilliantly shining Power Armor, a sexy young woman stands with an intense expression, her fist raised and ready to strike. The armour gleams under unseen magical lighting, hinting at its advanced design. The setting is a majestic fantasy rainbow bridge that stretches across vast distances between two distinct realms

The image above (Mobius) and the one below (UltimateBlend_XL_v2.6) are with the same prompt/seed.

dynamic panorama of A mystical, glowing portal that seems to lead to another dimension, guarded by a group of ethereal, glowing creatures with flowing capes and staffs. by Hayao Miyazaki and Pascal Campion

Negative:

signature, logo, watermark, colourless, featureless

[gloomy lark]

So one with no hands. One with good hands. One with bad hands. So it might just be seed dependent.

[meager canopy]

Yes, I'm not too fussed about that. Mine followed the prompt better for the other examples though.

[gloomy lark]

Trying the other one now

[meager canopy]

I do like the output from Mobius

[meager canopy]

I added the negative I use

[gloomy lark]

That said, some good outputs on that one from hunyuan and pixart as well

[meager canopy]

I'm also using cinematic and enhance in styles. Let me try without...

[gloomy lark]

Do you plan to put the new 0.1 ultimate blend 2.6 up on civitai? I could try using that with my automation for refinement instead of aventis.

[gloomy lark]

Really good

[limber kindle]

Heh 😛

[meager canopy]

I wasn't going to, it'd probably be quicker for you to merge v2.0 yourself than me upload it to Civit 😄

[gloomy lark]

Do you feel that 2.0 with the merge would have the added prompt following you mentioned over mobius?

[meager canopy]

These are originally from that prompt.

[gloomy lark]

Wow, so much flatter colors

[meager canopy]

No, it's already pretty good.

[meager canopy]

That was with a lora and upscaled a lot.

[limber kindle]

The flatter tones look much better imo. Many models do images that look like as if someone just discovered Photoshop filters

[gloomy lark]

Boo this man! 🙂

[limber kindle]

Haha ❤️

[frank wyvern]

its not bot?))

[glass forge]

Ugh, just imagine what SD3 would do with this...

[limber kindle]

Haha!

[meager canopy]

Schizophrenia

[timid garnet]

Reroll of my irridescent squirrel with extra noise

[limber lynx]

I have just uploaded my probably best SDXL model so far.. https://civitai.com/models/155977?modelVersionId=576344

[timid garnet]

Very cool gallery. Neat

[meager canopy]

"Who threw that?!"

[limber kindle]

Can anyone recommend a good outpainter for SDXL? One that doesn't get all morphed and blurry 😛

[paper phoenix]

SDXL is sorely lacking in the in/outpaint department. if its not a gigantic tableau i'd honestly say use something SD 1.5 to get the base down then force and img2img over the top of that to add in details. lotta work but SD1.5 just seems more into in/outpainting than SDXL.

[limber kindle]

Oh hey, you here o// Thanks, will try it out, just to lazy to do stuff by hand. I also have Photoshop, but the results ain't great either.

[paper phoenix]

oh lol. didnt even read your uname.

[paper parrot]

pix art sigma + sdxl refiner. Last one SD3 + sdxl inpainting

[limber lynx]

Awesome work. I love the details

[timid garnet]

ahah thanks, trying to adjust some values

[limber kindle]

Since you're going to train an own lora anyway, go with SDXL, you will just have way more options on how to refine etc.

Well the main advantage that Pony gives you is that it's trained on all kinds of cartoon characters, animes etc. so it recognizes them without using loras. By usng your own lora you kind of overwrite it's advantage. Ther are just way more models out for SDXL to get your desired look.

[grand cargo]

⛏️ ⭐ Amazing but @limber lynx and @freekhitman have done it and we have a new golden pickaxe leader for the SDXL Top 10 Models! Demon Core Midgard Beast (a bit of a mouthful but worth it!) totally smashed on the prompt adherence test). Make sure to follow the recommended settings and you'll be impressed for sure! 🥂 😱 https://docs.google.com/spreadsheets/d/1IYJw4Iv9M_vX507MPbdX4thhVYxOr6-IThbaRjdpVgM/edit?usp=sharing

[shy kelp]

from a blank prompt

[copper kraken]

iterative unsampling

[humble cape]

Jugernaut xl v8 + dream machine

[glass forge]

omg thats wowz

[timid garnet]

Awesome details, which model Is It?

[uncut gull]

Wow these look amazing compared to SD3. It is good to have a good model, whatever else comes. SDXL is good enough.

[limber kindle]

Thank you! realisticStockPhoto_v20.safetensors as a base and realvisxlV40_v40LightningBakedvae.safetensors as a refiner. Lora on 0.2 weight: SDXL_FILM_PHOTOGRAPHY_STYLE_BetaV0.4

Before/after refinement

[glass forge]

Those are some very good models.

[limber kindle]

Yeah for people I usually use stock photo as a base, because it does not screw up faces, then refine it with different models depending on what I want

[serene hull]

generate minimalist logo for a plumbing business named "H2O Plumbing", where a Tap is present and a drop of water is falling from it

[glass forge]

SDXL still top dog.

[timid garnet]

Nice prompt, consistent output.
Zoom-in

[timid garnet]

Dr. Who

General Meow

Sketchy but cool

[wooden field]

/Uncomfortable /Kappa

[glass forge]

So does anyone here remember SDXL launch. I wasn;t around. Was this channel on fire like the SD3 channel is now when XL dropped?

[timid garnet]

I don't think so. People was busy enjoying the model, actually.

One of my very first SDXL pics on Civitai :

[brazen tapir]

When I try controlnet canny SDXL img2img it gives me these weird blotches on the wall. (When I do it for depth, it gives a repetitive fine dotted grainy pattern). What to use for buildings and walls instead?

[glass forge]

the only issue i ever had with controlnet sis they make the image more blurry

[limber kindle]

SDXL+Demofusion 4096x2048 raw render wihiut upscaling. View in full screen

[glass forge]

wonderful but why 2 suns..

[limber kindle]

Haha was just my very first test render. It's crazy slow : D

Will see how good demofusion is

[wary ridge]

well that was a hot mess

[limber kindle]

4096x3072 Freaks out completely lol. Will try an other model

[sturdy dove]

i was told to work in resolutions of 1MP and then upscale thats for sd3 my bad.

[wary ridge]

Resolution too high for my blood pressure

[limber kindle]

lolol

[limber lynx]

Thank you very much for testing it. took me by suprise 🙂

[limber kindle]

4096x3072 raw render

[copper kraken]

#✨｜sdxl message here's a link to the last message on July 25 at 11:59pm Eastern time. Sdxl dropped on the 26th

[limber kindle]

4096x3072

[hushed gyro]

Cat with 4gb vram send help cute bunny in a tea-cup bilateral quad by artist "zentangle",tesselated polygons, afremov hues; by artist "tricolor";

[copper kraken]

Here is the image you requested.

[tawdry current]

pixart sigma because i had my app up while i worked on it(the cat with 4gb ram thing is a user on here, so i omited it)

[tawdry current]

And here's a second one from HunyuanDiT1.1 this time:

[floral summit]

playing with style transfer

[copper kraken]

That's one of the greatest nodes in history

[floral summit]

so fun

[floral summit]

Is there a way to use controlnet with ipAdapter? I'm getting an error with controlnets selective range of applying conditioning

[tawdry current]

i use them together a lot, what is the error?

[floral summit]

hmm thats basically what I was doing, maybe I got some connections mixed up

[tawdry current]

Double check your models. You might have an sd1.5 in your ipa loader

Or in your controlnet loader

Basically, make sure they're all sdxl models being loaded

[floral summit]

seems to be good now that I redid it from the beginning

must have had something in the wrong place

[tawdry current]

Cool beans. On a side note, damn is mistoline+IPA PLUS a strong combo... I didn't expect that troll waifu to be spit out so close to the input one. I had the cnet turned way the hell down too

[floral summit]

im gonna have to try that one out

I really like the dust on the hood in this one

[meager canopy]

SD3? Where did his legs go?! 😄

https://tenor.com/view/dms-slideinto-like-jump-cliff-gif-5848148

Grid-Girl

She's wearing her parachute 😄

Location is good, clothing...not so much

[opal swan]

Lol

[limber kindle]

More SDXL Demofusion tests, watch in full screen/full res

[meager canopy]

What is this?

[limber kindle]

It's this: https://github.com/PRIS-CV/DemoFusion?tab=readme-ov-file It's a highres bucket renderer that runs with SDXL models. The images above have 400 iteration steps each.

[meager canopy]

This is lacking the same detail though

[limber kindle]

It's not upscaling, it renders real details

Very slow though

[meager canopy]

How did you get those so sharp though?

Just seen the embedded flow

[limber kindle]

I run 240-400 steps on each image. The sharpness it does automagically. It's so sharp that it's kind of uncanny super real. If you know what I mean.

It also samples completely differently somehow. Everything looks so clean lol

[limber kindle]

Did you do it with Demofusion?

[meager canopy]

Yes

Could be the model, so I'm using the same Jug as you

[limber kindle]

Hm maybe you didn't install the package right? How did you run it?

[meager canopy]

Installed through manager in ComfyUI

I don't like how it downgraded diffusers 😦

[limber kindle]

Try setting steps to 30 at least

[meager canopy]

It was set to 80

[limber kindle]

😮

I run the cmd line version without gui, maybe it downloaded stuff that comfy doesn't have

[meager canopy]

Ah, perhaps

My GPU is working for this!!

https://tenor.com/view/burning-hot-nvidia-gpu-fire-gif-20813194

[limber kindle]

Yeah it's slow lol. It's just a pity that even though it could, it doesn't add like many objects. All scenes look like super sterile at that res

[meager canopy]

My 4090

It's liquid cooled, and not been that hot for that long before

One image and I think I'm done with this node 😄

[limber kindle]

😛 MY heart tells me to abandon i too, but I just can't resist.

[keen garden]

Any tip for fantasy skin color? If I do get one then the entire scene is that color, clothes, background 😭

[limber kindle]

Heh interesting 😛

[limber kindle]

40 vs. 120 steps, same seed

[meager canopy]

I presume 120 is the overcooked one?

I just tried 20 steps

[limber kindle]

No it's the detailed one. It does 4 phases, so 4x30=120

The low res one is 10 steps per phase

[meager canopy]

Oh! That's not the steps you entered then?

inference_steps

[limber kindle]

No, but I don't know how it handles it in comfy. In standalone it does4 phases per default. Interference steps I usually set to 50-100

[limber kindle]

Oh same for you then

[meager canopy]

The inference_steps is just how many it does each step.

Ah... The following part of your input was truncated because CLIP can only handle sequences up to 77 tokens:

[limber kindle]

It does kind of a weird thing that it decides on what to detail and what not, so steps 50 is kinda required. Here an example where it just detailed the chrome ball

[meager canopy]

40 steps

Just adds too much randomness into the image

[limber kindle]

Yeah some models work better than the others. In realvis I got this:

[meager canopy]

Could be because it's a Hyper model and we're using too many steps...? 🤷🏻‍♂️

[limber kindle]

I have no idea haha. It's just so tempting. I love how it looks, but prompting it is just impossible.

[meager canopy]

10 steps is very bad

[limber kindle]

Yeah. Anything above 40 is fine. 10 doesn't work. https://civitai.com/models/155977?modelVersionId=576344 seems to work bst with it so far

[meager canopy]

Ships in the sky, 50 steps

Sorted!

[meager canopy]

@limber kindle 25 steps

15 goes like this

[limber kindle]

You somehow don't get that clean look I do. It's more of what I'd want

[meager canopy]

Could be this prompt

[limber kindle]

What did you prompt?

[meager canopy]

Two majestic three-decked sailing vessels engaged in an intense maritime conflict, their hulls adorned with intricate carvings and flags fluttering in the wind; the sun casts long shadows across the rippling waters as flames from exploded cannons illuminate the horizon; a heavy mist swirls around the ships, ropes and sails, the ships are silhouetted against an orange and crimson sky as the first stars appear on the horizon

20 steps takes 1 min to run

No control over the noise at all though 😦

[limber kindle]

Hm you get a much grittier look. What model do you use?

[meager canopy]

Boltning Hyper

[limber kindle]

The dragon one is nice!

[meager canopy]

I wonder if AR affects the doubling up and ships in the sky?

You managed to avoid them 😄

Oh, you're not using the same node 😄

I thought it looked a bit low-res...lacking detail

[limber kindle]

Same playground scene, same seed, 70 steps

[meager canopy]

Not just background, the ships are messy too

[limber kindle]

Speaking of HDR. WHat is the best way to save real 10or 12 bit images?

I got 2 HDR screens, monitor and TV

[meager canopy]

Don't compare with an image that I demonstrated as bad! 😄

I'm trying to improve your image, but it's too lacking in detail to start with.

[timid garnet]

The paintings on the wall are okay lol

[meager canopy]

A bit sharper now

[timid garnet]

Extra noise is the secret

To unleash the model

One of you mentionned 400 steps for a pic

How is this amount distributed ?
It's like xx steps per tile on a tiled upscale ?
Or xx steps per pass on the whole pic ?

Wizz sad when good things are standalone and not as an extension/nodes

[meager canopy]

It is a node, Demofusion

You enter steps and it runs 4 iterations over the image.

[timid garnet]

ah, thought it was a standalone gradio gui

[timid garnet]

how much time for a pic ? regarding the card

I'm also running a lot of steps, but result may vary

[meager canopy]

That was 10 steps in the ksampler 🙂

[timid garnet]

Also using Soft Inpainting in a1111, so i don't know how many steps SI is running. But i'm at 220+ without

[timid garnet]

Simple. Lovely !

[limber kindle]

Woooh that'S a difference

[meager canopy]

I have a 4090 and that one takes about 1 minute when I use 20 steps, so 80 steps total. That does need a Hyper model for low steps, otherwise twice as many

[timid garnet]

Hô i see.

It takes me ~850sec per pic with a 3060, kill me

I'm going the slowest possible way on purpose, tho 😂

[limber kindle]

I'm downloading boltning right now, will test with 70 steps for fun

[meager canopy]

You may fry your GPU 😄

[limber kindle]

For some reason my gpu runs at 78%

[meager canopy]

Perhaps because you're using CLI and different drivers?

[timid garnet]

Or cause of the res !?

I think you can reach the maximum usage only with square pics ~1MP

[limber kindle]

Hm will see, maybe I'll install the comfy node too. Currently running standalone, needs 17gb of ram too

[meager canopy]

It can do really large resolutions. The last pic I posted shows it was creating at 3072x2048, but it can do larger.

[timid garnet]

I'm still looking for the best upscale using a1111, limited with 12gb vram
I'm still improving my pics every month

[meager canopy]

You've been posting some great images!

[limber kindle]

70 steps, 2x2k. For some reason it did only 2 phases though

[timid garnet]

With a1111, when using Ultimate SD Upscale, i'm able to run adetailer in each tile, using the API.

I can't anymore, since i updated ...
I can, but i have to choose between ADetailer and Soft Inpainting, can't run both.

I split my pic in 9 tiles, whatever the ratio, and i use a high padding to keep the tiles consistent each others. So this way faces and hands can get 2, 3 or eventually 5 adetailer passes

[limber kindle]

@timid garnet read about this yesterday, but haven't tried yet https://github.com/deroberon/demofusion-comfyui/issues/9#issuecomment-1967333169

Any of you guys tried lumina?

[timid garnet]

Nop

[limber kindle]

Installing demofusion via VoC now, see if it makes a difference.

[timid garnet]

Got this today. I'll just leave this here.

[limber kindle]

ALso @timid garnet will try this today: https://www.reddit.com/r/StableDiffusion/comments/1cbaxsu/introducing_hidiffusion_increase_the_resolution/ looks promising

[timid garnet]

Ah, btw.
My current upscale process is at the limit.
My "only" way to improve the result, using the same process, is to improve the input image. So to improve the base generation and the hires parts.

Do you have a1111 installed ? and are you familiar with the 'noise_multiplier', or the 'extra_noise_multiplier' ?

[limber kindle]

I do, but only used a1111 to run deforum animations, don't really use it outside of that

[timid garnet]

If you can open it, i can show you what i mean.
You'll need to create a decent pic, well prompted, using hires as you usually do, and ping me.
The "how to create details" should be obvious at this moment.

Then, report and improve the logic on any UI

As a Deforum user, you should also understand what's going on with the noise multiplier.

The normal 'default' generation process that we're all using heavily limit the ability for the model to actually draw what it wants to draw.

Comfy and other solutions already use what i'm talking about, to be clear, but it's not on your control

[limber kindle]

I cutrrently have to render stuff for work, will ping you for sure later or DM you. Very interested for sure

[timid garnet]

Let me find some comparisons

[limber kindle]

Yeah would be awesome

[timid garnet]

Found them

These comparison are using hires, then my img2img payload for ultimate SD upscale and cie. Not Raw to upscale

The only difference is a single setting in the Settings tab

Random non realistic pic :

Some prompts are way more impacted than others.

You can feel what part the model is done with, and what part wasn't done yet.

This literally 'unleash' the model.

This setting apply to hires and img2img

Settings/img2img

The reasonable range is 0.075 - 0.1

This is exactly like Deforum, you have the same settings doing the same thing in Deforum

Ah, this is for normal SDXL models, for ~30 steps or so. Need to adjust for lightning or Turbo as well

Here is one without a huge impact :

Particles in the air aren't a bug or noisy artefacts, it's on purpose. The model want to draw them.

[limber kindle]

Very nice! The example with the girl and trees is heavy

[timid garnet]

Yes !

I'm like a 'noise expert' with deforum, i'm playing with SD since like 1.5 year... and i just discovered this a1111 setting last week.

My life is a lie

Also, as mentionned, it's not under your control in comfy or other solution, like Demofusion. It needs to be under your control.

[limber kindle]

I know that feeling. Every time I think I figured stuff out, 10 new rabbit holes appear out of nowhere

[meager canopy]

Are you sure about that? 🤔

[timid garnet]

For my pics, i'm using overweighted prompts, and only hires x1.4 (vram limited).
The pics could be, as i tested, wayyyy better if i coulb hires to 1.6 or 1.7. Wayyyy better

[meager canopy]

Noise injection in Comfy is a thing

[timid garnet]

My current problem is the pre-img2img part, it need a huuge improvment now.

[timid garnet]

this

Underated and not known enough by the pleb.

My other way to improve would be to downscale and redo the img2img process. But would need 25-30 min per pic, this is insane.

I'm using 0.48-0.52 denoize, for hires and img2img. I wanna keep as possible the raw generation unaltered.
So my only way to be seamless is to use Soft Inpainting... adding 50sec per tile. Also blocking the webui, for some shady reason.

If someone have a solution to be seamless, while using extra noise aand high denoize, it would be the best day of my SD life

Raw generation at 1MP / Details++ at 9.5MP

I mean, i need to replace Soft Inpainting, so i can run ADetailer for each tile.

[limber kindle]

https://github.com/megvii-research/HiDiffusion/blob/main/assets/quality_efficiency.jpg?raw=true

[timid garnet]

Huh, stop bullying me 😭 😂

I'm using a Discord bot as UI. This bot is using A1111 API.
I won't rewrite the bot, so i'm stuck with A1111.
To be clear, i love it. But maaaan, why everyone is doing things for comfy 😭

[limber kindle]

😢

[timid garnet]

I can use Comfy in a1111 tho. But it's limited. Never tried

[meager canopy]

Because it's much easier to customise and do what you want.

[timid garnet]

There is an extension for this

[timid garnet]

https://tenor.com/view/no-way-pfff-gaga-lol-gif-18633881

[meager canopy]

@timid garnet Your image upscaled x1.5, a little noise added and sharpened

[timid garnet]

This :
https://github.com/ModelSurge/sd-webui-comfyui
https://github.com/ModelSurge/sd-webui-comfyui/raw/main/resources/front-page.gif

[meager canopy]

That's using your upscaled image as original

[timid garnet]

Damn buddy

And it's only one node. This is so sexy

[meager canopy]

No, that's a whole workflow 😄

[timid garnet]

The better the base image, the better the upscale. But i'm limited. meh

[meager canopy]

Original to upscaled

[timid garnet]

This is really good and consistent

[meager canopy]

Not a huge difference, but it was a good upscale to start with 🙂

[timid garnet]

My last generated pic :

When i see it, i'm like "Meh".
But then i compare :

Still, it's not enough...Wizz frustrated

[meager canopy]

Why do you need more?

Nobody will ever look at it as closely as you, and then it's forgotten about 🤷🏻‍♂️

[timid garnet]

I also tried to run several models in ADetailer, for several classes.
Running after Hires, then in each tile. That was super long, but a cool experiment.

[timid garnet]

True.

But I know

This is actually my third option

Lower the res would improve immediatly the result.

[meager canopy]

I delete all my images after uploading them anywhere.

[timid garnet]

As i'm using discord, all the pics are in my server, and it's easy to search to download the pics from it

I keep them tho.

I'm uploading to Civitai, but the pic is not really the original

[meager canopy]

How's this one?

[timid garnet]

Kind of overcooked Textures are weird, but also awesome.

Neat

[timid garnet]

Also really fast lmao

Compared to me

[timid garnet]

As usual. I'm often "meh-ing", but the comparison is speechless

[limber kindle]

Is that demofusion?

[meager canopy]

No, I can't use demofusion to upscale in Comfy....can I?

[timid garnet]

The best workflow i've (not) seen so far is Zavy's workflow

But he don't wanna send it to meeeeeeeeeee

Riot

https://tenor.com/view/riot-gif-22170733

[limber kindle]

Oh sorry. Standalone has img2img with a scroll comparison , forgot you're using comfy

[timid garnet]

@blazing root

https://tenor.com/view/spongebob-waiting-check-watch-gif-27333543

Good old bad seed

My best friend

[limber kindle]

@meager canopy boltning 4kx4k 80 steps. at many steps it starts dreaming 😛

[meager canopy]

That's its biggest problem...and probably why I won't be using it. 😦

[limber kindle]

Installing https://images-ext-1.discordapp.net/external/zYLnJYwzoZ4-n-6WWNEvRSxoJwOauWb2KGxagVbZoSg/%3Fraw%3Dtrue/https/github.com/megvii-research/HiDiffusion/blob/main/assets/quality_efficiency.jpg?format=webp&width=1762&height=947 right now

[timid garnet]

https://tenor.com/view/noted-note-the-office-gif-6207136

I'm also addicted to semi-random prompting

Trained a GPT2 model for this months ago. Still using it. So small, it's almost instant using the cpu. Text-completion mode, not as chatbot. Lack of variety over time, but still decent i think.

I have a comfy workflow with my prompting group. Need some manual edit to fully work if someone is lazy as i am.
cf. My civitai gallery on my profile.

Found you. Very cool gallery ((:

[uncut gull]

This looks amazing. Might have to learn how to write an a1111 extension.

[timid garnet]

Yes you should.

[floral summit]

miyazaki mustang

[uncut gull]

I notice there are already 4 comfyui node implementations. I'm guessing a1111 integration must be much harder, and I barely know python. 😐

[timid garnet]

ahah jk jk

[limber kindle]

For Hidiffusion I need to downgrade numpy, but get errors. Anyone knows hot to solve it? Attached error log

[timid garnet]

Out of my (small) skillset

https://stackoverflow.com/questions/69458399/numpy-1-21-2-may-not-yet-support-python-3-10

Maybe dumb, but make sense to me

Can find the file here:
https://pypi.org/project/numpy/1.19.5/#files
So
pip install "numpy-1.19.5-cp38-cp38-win_amd64.whl" is want you want ?

[limber kindle]

Hm will reset my voc env and see, but thanks

Had trouble with numpy recently anyway with other tools

Must have messed sdomething up along the way

[timid garnet]

Not a dev lol, i usually shut up when talking about code

[limber kindle]

Jesus voc downloads at 340 kb/s 😢

[timid garnet]

Duuude

.
.

Caption this

[timid garnet]

I have a new workflow

I'll post pics here, and you'll upscale for me.

I can pay with love

[timid garnet]

Interesting fail. (just a swimsuit, the woman is half in the water)
pg13

[dry plinth]

Is there an SDXL model that works really well for rpg battlemaps?

[harsh hinge]

Blue viking tatoo, burning man, futuristic

Blue viking tatoo, burning man, futuristic

[floral summit]

you cant generate images in here lol

beavo and beavs

lol

[gloomy lark]

ella

[limber kindle]

4.8x3.3k SDXL Hyper+Deep Shrink

[meager canopy]

3.5k x 3.5k

[half gyro]

whats the best web ui for SD if I have 6gb vram

currently using foocus and foocus mre

I am pretty new to ai image generation

[floral summit]

which webui wont be your main concern I wouldnt think

but if you are new, automatic1111 is pretty user friendly

[half gyro]

I heard its heavy?

[floral summit]

I know there was a memory leak in it at one point, but the webui shouldnt be using your vram itself

[half gyro]

I see, and it has ample guides on youtube?

[floral summit]

absolutely

[half gyro]

Since I really want to learn

nice nice

a lot of web uis tbh

very confusing which one to start on

[floral summit]

auto1111 will get you on your feet pretty quick. learn how things work, what different settings do, the general workflow, then move on to something more advanced like comfyui. thats what I did

[half gyro]

node based one?

[floral summit]

yeah

[half gyro]

that must be a steep learning curve

[floral summit]

its a slow learning process

with comfy, that is

[half gyro]

whats the real benefit of using node based system

[floral summit]

making your own workflow that runs in one press, rather than having to do it in multiple runs, with each part in a different part of the UI.
also many many custom nodes to do things auto1111 doesnt have.

[half gyro]

I see

so web uis dont have any impact on vram usage? Saw a few videos on youtube claiming that SD next leads to a 60% faster generation compared to a1111

[floral summit]

auto1111 is like a toy car that you can get addons for to change how it looks and rolls.
comfyui in this example would be like a box full of lego car pieces that you can assemble in any way. you can make a limo 6x the length of the toy car if you wanted to.

[floral summit]

running the webui itself wont make an impact on vram, but the settings that affect how the webui handles models can.

[half gyro]

Alright then. Thank you for the guidance. I'll run foocus as well as download a1111 to learn more

about ai generation

[floral summit]

I thought you were concerned about the vram usage of the webui itself eating into you available memory that would be used to hold models and whatnot

[half gyro]

I was just concerned about the image generation time

[floral summit]

but yeah, try em out, follow guides, and change parameters for the hell of it to see how it affects things

[half gyro]

right

[floral summit]

https://www.reddit.com/r/StableDiffusionInfo/comments/126ks7f/limit_vram_usage_at_cost_of_performance/

yeah, I normally just keep this up on my second screen

more style transfer fun

[tawdry current]

Yeah I undervolt my 2080 and lose maybe 5% performance in short benchmarks, but actually gain performance in long throughput because it stays cooler and doesn't have to thermally throttle at all(like if you're generating a bunch of images back to back). Drops my max wattage by ~50w

[wheat gorge]

A girl, walking on the grass, full body, smiling

[floral summit]

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

[tawdry current]

This card is super old, it's a 2080 FE if that gives any clues lol. I've never repasted/repadded it either. It will get to around 70C(the hotspot is probably 80C, but task manager doesn't poll that temp, have to use another app to see it) when doing long generations.

But ironically, the fans never spin up higher than 60 or so percent

[gleaming forum]

Foreground: On the steps next to the circular sports ground of a rural school campus, a young male dressed in a white T-shirt and jeans is walking alongside a young female wearing a white long skirt and a pink top. The two are wandering on the steps, surrounded by many flowers and plants that sway in the wind. Background: A spacious rural school in China around 1997. Styled in Chinese animation. Long-focus lens. Side-tracking shot. ┬─┬ノ( º _ ºノ)

[gleaming forum]

https://tenor.com/tKgaYjwJq16.gif

a cute dog ¯_(ツ)_/¯

[stark maple]

圣旨 龙纹 金色

[storm root]

Maybe it's better to speak english, so more people can answer you 🙂

[stark maple]

@storm rootthanks!

[tawdry current]

i feel you, but i usually just tinker with stuff while i'm busy with other things. oh and here's what my gpu looks like after 20 minutes straight of generating a batch:

so as you can see, the undervolting lets me stay under the thermal limit (dunno if that 88C limit is for the hotspot or the edge though). i think this is a ~250w card btw

not bad. and no, undervolting is different from power limiting. undervolting is hard to tune, but if you do it right, you'll end up being able to hold the same high clocks, for less power. my power limit is set to 100

actually, even higher, 124% lol

see the custom curve for the voltages? it took a while to tune correctly, without causing crashes and stuff under sustained or transient loads

the thin line is the default curve

basically, i can get higher clock rates at lower voltages, which means you draw less power overall. anyways, it's helpful to do shit like this if you're going to be redlining your hardware a lot

i also did the same to my cpu, but just did a flat -0.125. saves a shitload of power under avx loads like rendering (i do a lot of 3d rendering and am in the game industry). and no, i don't OC ever, there's no point. like you said, just buy bigger. not worth the ~5% more performance if it means your shit's going to fail in 1/4 the time

[frosty cedar]

how to set in automatic 1111 to generate one image in several resolutions ? that is, for example, 2 images in resolution 512x512 and 768x768 ?

in X/Y/Z plot there is no option

Does anyone know how it would be done ?

[meager canopy]

You get a different image created with any change of resolution. You'd have to do an upscale of the original image to change the size of it.

[meager canopy]

How were you creating the large image without duplicates? Didn't you say you weren't using upscale?

[meager canopy]

Nicely contained inside the glass 🙂

[gray lagoon]

Automatic1111 is beginner friendly but if you're a bit used to it id recommend SwarmUI as it has a comfy UI backend but also a simple generation interface. So you can usea easy interfaces and if you're up for a challenge can learn comfyUI

I'd post examples of my stuff but i currently got it setup for R34 stuff requests so cant post

[meager canopy]

Fooocus is another option. I don't know how it is for VRAM usage.

[half gyro]

i am currently using foocus

its mostly fine

[gray lagoon]

Did a quick search, focuuus is nice for "i got an idea image but no idea on how to customize the parameters"

Id say for beginners its good

[half gyro]

yeah

its like midjourney

kinda

[gray lagoon]

Automatic1111 is beginner friendly and allows for some customization

SwarmUI is like automatic1111 advanced and optional comfyUI

[half gyro]

I did try swarmUI but I am not trying to learn comfy rn

[gray lagoon]

And comfy ui alone. Lol im not touching it

[gray lagoon]

Oh the comfy is optional, the "generation" tab is basically all you need

Granted its some getting used to

[half gyro]

I see, what does the generation tab have over a1111

[gray lagoon]

Swarm: buildin video, easy resolution, a logical UI, civit ai downloader, active devoloper that constantly fixes or adds stuff based on user input, image history tab that easyly allows you to "reuse parameters" (model, prompt, seed, sampler etc)

Automatic1111: addons, less likely to break but less freedom

I can show you some stuff but im currently omw to a wedding

So if you got time later im happy to hop onto vc and share my screen

Later as in 6 hrs lol

[half gyro]

sure sure that can work

just ping me

have a good time at the wedding

[gray lagoon]

Chilling in a passenger seat

[half gyro]

what about youtube guides

etc

[gray lagoon]

https://www.youtube.com/watch?v=HKX8_F1Er_w has a great guide but the discord is also very active

The disc has more vids but i dont wanna spam

[half gyro]

nice nice, I'll download swarm then and once ur back we can hop on vc

[gray lagoon]

Oh apperantly the dev is now an ex stability ai dev

[meager canopy]

Yes, now part of comfy.org

[gray lagoon]

Gonna watch&read the announcement rn

[gray lagoon]

Do you have random ideas that you generate or are you doing requests for people

I generally just have a few ideas, mess with it for a hour and move on

[meager canopy]

I have a list of old prompts that I revisit to compare with how images used to be 😄

This is the best improvement I've seen yet!

[gray lagoon]

I tend to focus on the "2d" images but yeah sdxl is getting better and better lately

[meager canopy]

It's a new workflow I've been working on.

[frosty cedar]

it's not about that, it's about, for example, in X/Y/Z plot having the ability to generate several resolutions. It is not about scaling. It's about the same thing as with sampler or CFG.

[meager canopy]

I know, but if you change the resolution and keep everything else the same, you do not get the same image but larger.

[limber kindle]

Yo @meager canopy o//

[meager canopy]

Look at the detail in there! 🙂

[limber kindle]

Gave up on Hidiffusion yesterday lol It ate all my vram for some reason even at 2k and was not able to fix it : D

Niice!

[meager canopy]

I deleted it and beuilt my own workflow for these images I've been posting.

[limber kindle]

Yeah was just about to ask, they look good.

[meager canopy]

Thanks 🙂

I tried to break down what Demofusion was doing and recreate it in a workflow.

[limber kindle]

I did some tests too:

[meager canopy]

I saw those

[limber kindle]

So you still using demofusion or replicatin it'S workflow

?

[meager canopy]

I have my own

Gradual denoise with latent upscale in between, also Noise injection in some stages

[limber kindle]

Fancy. Are you using deep shrink?

[meager canopy]

Yes

3088x4000

[limber kindle]

Nice, the focus pops out really well

[meager canopy]

It is really sharp. Even hands and faces come out almost perfect. There's no detailer for either.

[limber kindle]

Yeah it looks $$$ ❤️ gj man

Really want to try hidiffusion but can't get it to run

The standalone kills my gpu and the comfyui node gives an error I wasn't able to fix

Error occurred when executing HiDiffusionSDXL:
'transformer_block' object has no attribute 'use_ada_layer_norm'

YOu know what that is?

[meager canopy]

None

[limber kindle]

😦

[meager canopy]

How did you get these to look like they're inside the glass?

Probably by writing better prompts 🤣

Artistic fail

[meager canopy]

Looks like an artificial horizon

Old father time...

The fight to get a place by the water was getting ridiculous

[shy kelp]

Oh I like the birds. Very nice

[meager canopy]

Could this be a death star prototype?

[stray warren]

From the family album