✨｜sdxl

[vast galleon]

i had to try it 🙂

[smoky shard]

Creature

[brave inlet]

looks cute and scary at the same time 😭

[eager prairie]

I just found this model called SSD-1B, it's amazing. Super fast, I loaded it for the first time so already some time spent there, and then even at 24 steps it did the whole image (combined with loading model) in 30 seconds

[vital ermine]

lol

[vital ermine]

https://civitai.com/models/637987?modelVersionId=713411

[vast galleon]

Testing facial expressions

[sharp scarab]

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

[shy kelp]

A1111 + SDXL

[clever cave]

yeah its a distillation of the sdxl model. I would honestly recommend using lightning sdxl models like realvisxl lightning/dreamshaper lightning because they are better quality and even faster then ssd1b. They just require 4-6steps instead of 25ish.

[eager prairie]

I use realvisxl but I do like how this one doesn’t always resort to like photograph quality

Plus this is super fast still at like 30 seconds total, including upscaling

[clever cave]

Try dreamshaper or the base lightning as they can do basically all styles. Lightning models usually take like 2-3seconds(no upscaling).

[shy kelp]

A1111 + SDXL

[shy kelp]

'Alien Species' LoRA (A1111 + SDXL)

[steady grove]

new forge release will load sdxl as nf4 and supposedly makes it sd15 fast

[vast galleon]

Not expected, but I'll take it.

[smoky shard]

random prompting sometimes does stuff

[shy kelp]

SDXL + A1111

really nice

which checkpoint model do you use?

A special treated mix + some finetune. This is cleaned, removed random noise and many more

ah okay yeah
I really love making sci fi with these models

[shy kelp]

Me too, but it is hard to find good ones, especially for those resolutions.

yeah its true it can be hard to make it work

Finally it is working out of the box, i am enjoying it several weeks now 🙂

This is the 3rd checkpoint i used this method to clean, denoise and detangle. Can't wait until some more tools for Flux are available, then i try it there as well.

I like how contrasty it is

Up to now, Flux lacks a lot scifi imho

yeah Flux has a few issues with subjects

The tech is insane, it just need more variety

Until then, i like my SDXL 😄

yeah I am still on SDXL

Sure, it is the SDXL channel. I assume Flux images go to SD3, hehe (No offence SAI)

lol

yeah the SD3 channel seems to be Flux now

and sadly the cascade channel is gone

Cascade was is good too, but next to a Flux BASE model, a lot other finetunes look like ancient tech.

flux is a next level yeah

Funny thing, clarity comes with a price, random stuff like a Bolognese Sauce doesn't look right anymore. Too repetitive because random noise is removed.

Or the Pizza,lol

oh yeah that sauce is weird

As long as it does other stuff without open ends or vanishing wires, i can live with that.

[quasi badge]

American style house in red color with white background

[sharp comet]

A piece of paper is burning, the burn marks are in the shape of a trumpet

[vast galleon]

just needs a drummer and they're set.

[raven elk]

hi guys!! how do you avoid these kinds of artifacts?

[shy kelp]

Treeshark's latest workflow on L2 discord can fix this issue (hallucinations and tiles during tiled upscale)

[meager canopy]

Reduce the denoise, and change how the prompt is used in the upscale process

[raven elk]

what is the L2 discord?

[raven elk]

trying.. the thing is I like the results better when the denoise is high..

[meager canopy]

Then you need to use a mask to only change the main subject

[shy kelp]

I sent you the link in DM

I'm not affiliated BTW

I'm just a fan

[raven elk]

i will try this

[raven elk]

thanks!! highly appreciated.

[shy kelp]

also your image is 4k
you can just about get to this resolution using latent upscaling
if you combine all the tricks (SAG/SEG/PAG/FreeU/Res-adapter/Gradual Deepshrink or Hi-diffusion)
along with rescaling and tonemapping to allow very high CFG

its a lot trickier than tiled upscale but its an option

[meager canopy]

Can you post the original, before upscaling please?

[raven elk]

i will research on this..

[shy kelp]

sadly the ComfyUI version of Hi-diffusion isn't quite as good as the Diffusers one

it still helps though

apparently full Hi-diffusion would require some modifications to comfy to work

[meager canopy]

I ran it through my upscaler, with and without a background mask

[raven elk]

oh wow.. this is nice.. did you use usdu?

[meager canopy]

That is part of the workflow, but the "no upscale" version.

It uses controlnet and multiple latent upscales to recreate the image first.

[raven elk]

latent to ksampler?

or latent to usdu directly?

[meager canopy]

to ksampler

[raven elk]

i didnt realize upscaling is this complicated..

[meager canopy]

There are many ways to do it, but avoiding the unwanted hallucinations is the tricky part.

[meager canopy]

This one had a slightly higher denoise and also used facedetailer to clean up the face.

[raven elk]

this is looking really nice..

do i have to use tile control net?

@meager canopy

[meager canopy]

Up to you, I use multiple

[raven elk]

it removed the bg

i masked the subject but it completely removed the bg

[raven elk]

haha

how do you actually use the latent upscale?

[meager canopy]

You need to mask it and then use the original image as the background.

The whole workflow is on my Patreon.

[heady kraken]

armageddon training camp

[ebon ether]

A DSLR photo of a portrait of Groot head, 3d asset

[raven elk]

hi guys Im having this error while trying to load flux cnet.. any ideas?

[ebon ether]

#artisan-faq

[raven elk]

where can you find this controlnet 5000.bin?

[shy kelp]

.bin is weights

there's a format that looks like .bin and .param as an alternative to .pth

.ckpt or .safetensors are more common for comfy

.bin/.param is what I used to use for NCNN in chaiNNer

.safetensors files are safer, would recommend them where possible

[exotic umbra]

Hello! can anybody tell me why openpose isn't working at all here? i am using xinsir/controlnet-union-sdxl-1.0

[peak dove]

My SDXL w/f with Face Detailer is giving this error (no controlnets, all models are sdxl) - error = mat1 and mat2 shapes cannot be multiplied (77x1280 and 2048x640)

[shy kelp]

A1111 + SDXL

[timid hearth]

Generate anime animation expression package

[shy kelp]

Updating a checkpoint soon.
A1111 + SDXL

[desert lion]

Create a brush cutter monster

[shy kelp]

Here is the image you requested

[steady grove]

balls!

[shy kelp]

Dots!

[steady grove]

cross sections of balls technically

[glass forge]

Dots are tiny balls.

[glass forge]

Joyoive. I too prefer Joyoive over Sony.

[vast galleon]

trying to get a man to ride a giraffe... at this point it's comical

[vast galleon]

So I think Mangled Merge Matrix is ready. 3850 total loras. With this new one I used 3 different block combinations depending on the types of the loras being merged. 1 for concepts, another for styles, and a 3rd for text, although I'm not sure how much it benefitted from that one.

[shy kelp]

probably the most loras merged of any checkpoint

[vast galleon]

I'm interested to see how far I can take it. I'm surprised it's held up this far into it. Although going from strictly 2D for the Magic version to strictly realism was tricky. But I learned some new techniques in the process. I'm putting more focus into building up the clip for the next one.

[shy kelp]

I follow your attempts closely, if you are at the PC can you try the following prompt please with your latest mix:
'grids, lines and a circle'
and no negative.

Something like that will occur

[olive wasp]

🙂

[serene fog]

#dream

[shy kelp]

Here is your requested image

[vast galleon]

I didn't have a whole lot of luck. It mostly gave me grid lines. I changed the res. from 832x1216 to 1024x1024 and changed the prompt from "grids, lines and a circle" and got better results. But not consistent.

from "grids, lines and a circle" to "a circle with grid lines in it"

[shy kelp]

Thanks for the test. It has nothing to do with luck. Each seed does something different of course.
It is the resolution i am after and the integrity of the patterns, how parallel are lines and how many defects are in.

[vast galleon]

No problem. I've heard of the circle test. The grid lines seem to throw it off though. I'm gonna try to improve the clip more for the next one and keep this test in mind.

[shy kelp]

I know the curiosity by myself and i love research too, but for something AAA you can get away with a dozend carefully selected checkpoints 😉

So, why merging 4k LoRAs?

[vast galleon]

Just to see if I could and to learn and experiment with different merging methods. This version I explored block merging more and used Mateo's prompt injection for block references.

https://github.com/cubiq/prompt_injection/discussions/1

I wanted to see if merging into specific blocks could improve merge quality.

The lora merging process is automatic with code that utilizes the comfy API. So it's not such a tedious process.

[shy kelp]

merge quality.. do you mean prompt following quality or visual quality? Versatility?

All, sure.. but you can't have the cake and eat it all.

[vast galleon]

Both 🙂 Although I'm learning more about the importance of the Clip I don't think I have that part perfected just yet. Gonna tweak and test some things for the next one.

[vast galleon]

If anyone would like to play with it.

https://civitai.com/models/447902

[glass forge]

Lol Mangled.

Wow you merged everything...

[vast galleon]

Merge it all! That's my moto. 😂

[glass forge]

once i merged about 20 SD15 models

I think it improved the overall output

[vast galleon]

It most likely did. I'm excited about seeing what comes with Flux merges or any improvements that will come down the line to the entire merging process. I'm sticking with working with SDXL for now though until I decide a good time for a video card upgrade.

[shy kelp]

you guys should try training the checkpoint its way more rewarding and some of the outputs in training crease you up lol

[vast galleon]

I have with SD2.1. Found the whole process of building a database and labeling each image to be pretty tedious. Not to mention the time and money spent on the failures. I did enjoy making a loha with 71 different planes. But OMG the process of putting it all together ...

https://civitai.com/models/117088/plane-helper

[shy kelp]

youll get better tho bro, I was pants on my first try

[vast galleon]

Maybe once I get a 24gb 9080 I might look into training again. I'm sure by then you will need 10 A100's for whatever latest model is out.

[shy kelp]

I run48GB Vram And sometimes with my newer larger checkpoints its pushing all 48

[vast galleon]

I need to look into the latest and greatest for labeling images. I could never really figure out a good way on 8gb vram.

[vast galleon]

do you have 2 cards with SLI or 1 big 48gb monster?

[shy kelp]

https://tensor.art/models/759856135286068673/FLUX-HYPER-TRAINED-DREAM-DIFFUSION-BY-DICE-V-1 hyper training that was a stretch

i have 2 RTX 4090s

duel flow watercooled rig

[vast galleon]

Looks awesome. Do you see a benefit with training on SLI other than with batch processing?

[shy kelp]

https://youtu.be/jEcQqhcAtOI?feature=shared me building my beast

[vast galleon]

That's a pretty serious rig. I would like to have something like that soon. I'd probably go with dual 9080's though since they are so much cheaper. Is the dual card SLI worth it?

3090's i mean. Didn't get much sleep last night.

[shy kelp]

i srted out with it running a RTX 3090 with 24GB vram. which was sweet, untill checkpoint training and running 3 webui at the same time was a bit of a ripper, So threw 2 RTX 4090s in her and havent looked back

[vast galleon]

I bet! 👏

[timid escarp]

444

[vast galleon]

Released a lightning version

https://civitai.com/models/447902?modelVersionId=735778

[shy kelp]

If someone wants to give this thing a spin:
https://civitai.com/models/658893

[vast galleon]

Looks great! I'm, giving it a dl now

First 2 look really nice

[shy kelp]

From the description to the trainingset prompts, later the captioning and scoring, down to the example prompts and images, all AI.

[vast galleon]

what do you use to caption?

[shy kelp]

lama 3

llava llama 3

or the like

[vast galleon]

utilizing an api to run images through?

[shy kelp]

comfyui

and then api to ollama

[vast galleon]

Interesting. I haven't done much with llms other than chatgpt. as far as captioning. Is what your doing possible on 8gb vram?

[shy kelp]

short description up to..

long ones, just ask for it

i have 24gb vram, i don't look very often how much is used.

Ollama nodes are cool

[vast galleon]

I'm gonna look into it and see. What model do you run for captioning?

[shy kelp]

lava lama 3

not sure about where all 'L's belong

😄

[vast galleon]

ok cool. cause I got buzz to use but just hate captioning manually lol

[shy kelp]

You have to play with it a lot. sometimes if you ask for keywords seperated by comma you get 40 images fine, but then it starts writing essays.

[shy kelp]

• alien lora

[vast galleon]

I had to compare. Exaworld is on the left, Mangled Merge on the right. Both look really good. I think yours captures the essence of hollow knight a little better.

[shy kelp]

You are too kind. Yours has more details and color, mine is a bit faded because of the depth. Whatever one prefer, it is different.

[vast galleon]

These are from a quick merge of Exaworld + MM Matrix 50/50 merge, messing around.

[shy kelp]

Those look very cool

[shy kelp]

I really miss the 'multi image' feature from my other checkpoint.

[vast galleon]

Got it working. Thanks @shy kelp!

[shy kelp]

Enjoy 🙂

[vast galleon]

working on finding a good style to train a flux lora with.

[glass forge]

Boid.

[rain aspen]

My stomach and thighs are so thick, I can't keep my clothes open

So to illustrate what I meant about Pro and Dev, here is an image by Pro (note that 2 out of 4 times it fails to produce any text at all), and the images are all in these lines

[steady grove]

member for 6 months just to troll so weak sauce like that? disappointing

[cinder agate]

how do i test a sdlx model using only promts? only safe for work pls.

[shy kelp]

Prompt themes you want and if you like the output, the model is good. If you don't like the output, search harder or modify the model.

[lament sinew]

that one almost looks like a civil defense poster

[lament sinew]

trying to inpaint/outpaint in fooocus and I can't get rid of the phantom fingers

the way I see it, my setup is a great tool for manual artists if you wanted a piece to be 90% done and you can photoshop or draw the rest to your liking

my files are fcced though 😂 I'm trying to get everything in the same folder and it's impossible

I think I lost the logs on some images

[shy kelp]

A1111 + SDXL

[shy kelp]

https://civitai.com/models/667232/exa-world-addons
LoRA is up, you might want the EXA World checkpoint too.

[shy kelp]

Everyone on Flux? No one posting SDXL?

[glass forge]

[sick wigeon]

1 girl

[lament sinew]

Fantastical

[vital ermine]

Thank you

[lament sinew]

Trying to think of some lore where these airships make sense

It's like if some propulsion method was accidentally discovered in the 1800s

So this is maybe like a WW2 era aircraft carrier but with advanced levitation

[tight tulip]

like a starship u boat

[shy kelp]

Interlude:
Easy Buzz begging now with SDXL
https://civitai.com/models/669935

[serene fog]

#dream logo for a page named Animal rescuers Algeria

[half sundial]

#dream

[sharp scarab]

Recent research/experiment done in TouchDesigner and SD: https://www.youtube.com/shorts/VyZFzKAuqsk (more info in video's description)

[sharp vector]

gril

cat

[final notch]

Here is the image you requested:

[violet radish]

a mountain have many trees

[lament sinew]

I don't understand what difference the guidance scale slider really makes in Fooocus, but turning it up seems to have blown out the pictures with very high contrast.

she looks like she's about to tell you something really bad lmfaooo

[lament sinew]

"anon... that was my dad at that party last night."

you don't remember???

[pliant flint]

tes 1

test 1

[glass forge]

Testing.

[shy kelp]

A1111 + SDXL

[rare condor]

#🌠｜show-and-tell

#✍🏼｜rules-and-tos

[tender umbra]

A peaceful, quaint village nestled in a valley, with simple houses surrounded by lush greenery. In the foreground, a teenage girl with a bright smile and hopeful eyes stands in traditional clothing, looking out over the village.

[lament sinew]

this is incredibly coherent

looks like a real edited artwork

[tight tulip]

cant see very well but try first one set noise to enable and maybe try different int nodes

[torpid stag]

hi guys, does anybody know exact model/LORAS/settings that stablediffusionweb.com uses?

for some reason their outputs are so much. better than anything i try.

@copper kraken im dead serious. i don't really know how to use AI,. i just setup InvokeAI and downloaded SDXL turbo which is supposdely what they advertise to use, but my results are fucking garbage.

and when i prompt it on their website the results are incredible

i use same steps count and CFG scale, not sure what else i should do to replicate the results from there.

same seed too.

if someone could help me get close to that, would be super helpful

[copper kraken]

you might want to try SD3 2B Medium

it's the newest model

[torpid stag]

well but the thing is, stablediffusionweb.com claims to be using SDXL Turbo.

so this must mean they have some sort of fine tune on it that makes it so good, right?

look at the one generated on their website vs the one generated locally, lol

prompt for spongebob battle royal game logo

[copper kraken]

or a custom workflow

in either case, i'm sure they won't share it

[torpid stag]

hmm what does that mean exactly?

[copper kraken]

look up comfyui

and diffusers

[torpid stag]

looking it up rn

btw does it matter if i run it on my M1 mac?

or does that only affect speed

[copper kraken]

no idea don't have a mac

[lament sinew]

damn my favorite show, SpirbbBoog¥ Spin£bogk

[lament sinew]

wow that's actually beautiful and solemn

[torpid stag]

that's small detail im sure typography can be improved or omitted idc. I care about the quality of the image

IDK why but on my local the model insists of giving spongebob 2 mouths lol

[lament sinew]

the second one is a half decent rendering of Spongebob, idk what's happening behind him though

[torpid stag]

well what do i do in order to achieve this result locally?

for now its enoguh for me, i just want to replicate these outputs locally

[lament sinew]

what are you running, Fooocus?

I only really have hands on experience with Fooocus to a degree but that is SDXL based

[lament sinew]

is the prompt the same for both, and what kind of negative prompts are on both

[torpid stag]

yes, same prompt, same generation steps, same CFG scale, same seed

but completely different results

[lament sinew]

the double spongebob on the first one could be a resolution issue

other than that check your styles

[torpid stag]

This is what i can see going out to their API from their UI

This is my local settings

This is their output (with the missing eye) (it's not perfect ofc but u can clearly see it's higher quality, this is what i want to achieve for now)

And this i get with the same prompt and the above (looks the same) settings (will send in few mins once it stops generating)

[shy kelp]

A1111 + SDXL

[lament sinew]

damn bro that's badass

that's actually amazing

[shy kelp]

Thank you

[lament sinew]

you're now in my top 5 favorite ai artists

it also stylistically takes advantage of the glitches

[shy kelp]

Oh welp.. it is just the model and some bugs in A1111 which lead to such images. I am only an explorer

[shy kelp]

Yes,that was the target

The whole style was design around these

[lament sinew]

genius

[torpid stag]

and this is waht i get locally...

i dont understand why my local model insists of giving two mouths...

[lament sinew]

@torpid stag man I think your prompt leaves the program in full control of the scene and arrangement since you didn't specify anything like that; try being a little more specific.

[shy kelp]

Resolution too high or wrong upscaling

[torpid stag]

but in that case why does it work perfectly on their website?

[lament sinew]

"perfect" but different every time

with the same seed??

[lament sinew]

I'll admit I'm really new to this stuff man; I've just been a little obsessed with SDXL the last few days but take what I say with a grain of salt

[torpid stag]

no if i use the same seed its the same

i just tested out different ones between their UI and local to see if anythings diff

[lament sinew]

the 3 unique ones at the bottom are the best by far

imo

[torpid stag]

yea yea for sure. getting a good result on their website i've gotten multiple

it's just that im trying to replicate their results locally

even their worst result is better than what i get locally

[torpid stag]

i use 1024x1024, i tried changign to 256x256 and it was just ugly blurry stuff.

whats upscaling? can i change it in settings??

[copper kraken]

turbo is 512x512

[shy kelp]

Try 512x512 as Clownshark recommended. Upscaling is at a later stage.

[torpid stag]

let me try it now

btw does it matter what "Scheduler" i use?

it's set to euler

[shy kelp]

Let it there and experiment later with them. Lock seed and change it from image to image and you will see, which work good and which work better 🙂

[torpid stag]

this is what i get on 512...

no double mouth but low q

[shy kelp]

it is 512x512 and a turbo model.. i am not familiar with them, maybe experiment now with steps and the sampler.

[torpid stag]

the thing is i know desired steps should be 40 from their API

is turbo model just "faster" then the base SDXL?

[shy kelp]

try 10 steps

[torpid stag]

or is it supposed to yield better results

[shy kelp]

Turbo is just for speed, not quality

[torpid stag]

wouldnt less steps yield lower quality results?

[torpid stag]

got it. Im assuming the higher speed means quality will be lower though no?

[shy kelp]

try 10 steps 🙂

When i was trying them, i remember 4-8 step will suffice for SDXL turbo

[torpid stag]

i am trying now 10 steps

[lament sinew]

this feels like a super cinematic version of Streets of Rogue if you turned on the random environments mod

SoR is a procedurally generated game that loads one map at a time but it's broken up into different styles

if you put on that mod it mixes them up

[torpid stag]

with 10 steps...

where can i find the normal sdxl model (none turbo)? i would like to try it

[copper kraken]

any reason you really want to use turbo

[torpid stag]

no

[copper kraken]

go to civitai.com

[torpid stag]

simply thats what i saw written on their website and i wanna replicate it

[shy kelp]

https://civitai.com/models/658893

Not normal, but i use it now for several days

[torpid stag]

is this some fine tune of it?

[shy kelp]

Which UI do you use?

[shy kelp]

yes

[torpid stag]

Invoke.AI

i mean

the UI website thing that generated the good results

is called

stablediffusionweb.com

and i just setup invoke.ai locally to run the models via local UI

[shy kelp]

Ah, you don't generate locally?

Then you have to stick with their model. I never used any online service.

[torpid stag]

i am generating locally, i am trying to replicate exactly the results from their website exactly to my local : )

because i initially used their service to test, i liked it, then i tried to gen locally with same params from their UI and the results are way different locally

[shy kelp]

That could be model, ui, seed, version of the model, how the random numbers are generated from seed, UI version..

[lament sinew]

average house interior in my hometown

[shy kelp]

I would forget the idea of pure replication anyway and start prompting what is in your mind. 2 days, and you are heavily addicted 😄

[copper kraken]

where are you, detroit?

[lament sinew]

lmaoo

no

[lament sinew]

honestly, I can see AI art getting more accurate and less prone to glitches but I don't think you will ever be able to get exactly the result you want with no skill. There are so many options and things to configure that I think it will always take a degree of skill to generate the right picture unless the computer can read your mind.

[shy kelp]

Start simple, play around, get experience.

Then examine 2 different, maybe a third checkpoint and examine the difference. In style, quality and prompt following.

From taht point, you will know what you want and need. Don't run from checkpoint to checkpoint every half an hour. 'Learn' at least one checkpoint.

[lament sinew]

but I will say I like a lot of AI art styles because of how imperfect, random or glitched they are, and when people use models like Flux to get indistinguishable results I think that is bad for humanity because people will get lazy and use tools like that in place of real photography or exploration, or it could be used as propaganda. I'm not extremely biased either way but you need to be open minded to the pros and cons of this technology.

[lament sinew]

yesterday was all about transitioning art styles for me, that's what I learned how to do from a tutorial and when I got a result that was close I kept the seed and most of the prompts and started making fine adjustments to the sliders, styles and prompts

of course I'm a damn idiot and moving files around I lost the log for that whole project so I don't have all that info anymore

[shy kelp]

That is alreadyy too much for me 🙂 I just play around the last 2 years with this tech, hardly watched any video.

[lament sinew]

it was a really easy tutorial and because it's just fooocus I think anyone could wrap their head around it

[shy kelp]

I am pleased, if the image 'looks' good.

And enough to explore, 20sec later i need another one. 😄

[molten gull]

all with flux

[lament sinew]

he looks like an automaton

[strong anvil]

H

[smoky shard]

IP-Adapter adapts. A very powerful tool to push an SDXL models to a certain output honestly.

[shy kelp]

For users who still only use SDXL models, I've Created a SDXL checkpoint that runs on the Turbo base. Its hyper trained to a Maximum 100%. The adherence isnt that of flux but its very very good and the output image is the best youll see from any SDXL checkpoint. See what you think of the images posted in the gallery. Cheers https://tensor.art/models/751494452436224082/TWIN-TURBO-Dream-Diffusion-By-DICE-v1 https://www.shakker.ai/modelinfo/3ed729873f034271bef568cfc92176aa?from=personal_page

[lament sinew]

@ me if you know how Fooocus metadata works

[hybrid tusk]

#✨️ Mickey mouse

[shy kelp]

Here is your requested image

[sage lichen]

#bot

[slim wren]

that s not how the bot works, cf #artisan-faq

[toxic wasp]

I THINK so Brain, but where are we going to find a tattoo parlor open at THIS time of night? NARF!

[vast galleon]

Working on a latent upscale workflow to squeeze the most out of Mangled Merge Lighting for a Flux dataset. These are cherry picked. Still some kinks to work out.

For instance ....

Might need controlnet.

[shy kelp]

its working well in terms of avoiding errors

would be great if you could get the CFG-burn a bit lower

Scheduled_CFGGuider from Inspire pack and sampler_tonemap_rescalecfg.py from ComfyUI_experiments can help a lot

[worthy orbit]

nice, you may like 1_particular creations https://www.instagram.com/1_particular/

[shy kelp]

Thank you, i don't like instagram.

[worthy orbit]

mmm x? https://x.com/UnoParticular

Particular Uno (@UnoParticular) on X

// No sameness

Twitter

[shy kelp]

Really?..

[worthy orbit]

You say you didn't like instagram

[shy kelp]

Let me check..

[worthy orbit]

I wasn't sharing instagram, just the artwork

[shy kelp]

Yes, i did

Where does that link land on?

Instagram

[worthy orbit]

The second on Twitter

it's ok, nevermind, got it

[shy kelp]

A more neutral link here- Interesting LoRA material.. https://mixam.com/print-on-demand/65e91aa33457744a5a6ac2eb

[vast galleon]

I'm gonna look into them. Lightning is kinda weird that CFG has to stay around 1 especially with the newer pp samplers.

[shy kelp]

oh its lightning, never mind

those nodes may well mess lightning up

lightning is very delicate

[vast galleon]

Right!? I'll still try them out though.

[shy kelp]

when I want speed

I go for 15 steps, beta scheduler, unipc sampler

with regular sdxl

its nearly into turbo range

but with base sdxl models

[vast galleon]

I found the sampler tonemap node but can't seem to find the scheduled CFG guider. The tonemap helped but I'm gonna have to play around with it a little more later.

[shy kelp]

on the tonemap node if you put the rescale CFG to 0.5 that's usually good

but the tonemap level you gotta test lots of different %

its different every time

scheduled CFG guider is from inspire pack I think

https://github.com/ltdrdata/ComfyUI-Inspire-Pack

there's a perp-neg version, if you are a perp-neg user

rescale CFG at 0.3-0.4 is also ok

you might not like it, I have different taste to most users lol

[worthy orbit]

oh that's great, yes I just wanted to share it, I though you were pissed off or something

he is a friend of mine, I didn't even know about that site

[peak dove]

Processing A1111 photos (via metadata) in ComfyUI+LoRAs

[vast galleon]

I was confused earlier because I had the inspire pack but I wasn't seeing the sampler nodes ... I'm still not. I've updated comfy and the inspire pack. then uninstalled the inspire pack and reinstalled it. None of those worked for me. I also don't see any conflictions for it in the manager. Dunno what the issue is. shrug

[vast galleon]

It's a messy workflow, but I think I finally got it dialed in. It's using a 3 pass NNLatent Upscale process.

I brought the denoiser on that last ksampler down to .4 I think .5 was a bit much.

[shy kelp]

They are already here..

They hide..

But they are ready..

They will search you..

They will find you..

And then it is all up to you.

Dream your own end with the Cine Pony XL checkpoint and : https://civitai.com/models/698523

[shy kelp]

on the github page here:
https://github.com/ltdrdata/ComfyUI-Inspire-Pack
scroll down to Sampler nodes
and there is Scheduled CFGGuider (Inspire) and Scheduled PerpNeg CFGGuider (Inspire)

its better to avoid using the manager and just fetch lines of python yourself with wget and git pull really

so you could just rip out the code from here:
https://github.com/ltdrdata/ComfyUI-Inspire-Pack/blob/main/inspire/sampler_nodes.py

stick it in a node yourself

[polar ravine]

Guys I tried various style dora and they're much more consistent wtf

From various authors too

[shy kelp]

dora good ye

[polar ravine]

Yet there are so few of them

Spread the verb

[final notch]

and the verb is "train"?

[vast galleon]

Finally got it working. I had to uninstall all of my custom nodes and start from scratch. I added a depth controlnet into the mix as well. Super messy workflow:

[shy kelp]

tone mapping, custom sigmas, thresholding, NNlatentupscale, noise injection, PAG, freeU, scheduled CFG guider

such an amazing workflow

[shy kelp]

congrats 😄

I think this is the coolest comfy workflow I have seen so far

beta scheduler and deis sampler are both great choices too

[vast galleon]

Thanks! It's still not perfect yet, but I think it has to do with the fact that it's a lightning model. Gonna test with vanilla mangled merge tonight.

[lusty wolf]

Mine does not do the nice headgears... 'Steampunk' How do you call for that?

[shy kelp]

Testing the LoRA

[lusty wolf]

Is it flux?

[shy kelp]

No, this is SDXL channel^^

Until there is a oneclicksolution i don't play with Flux.

[lusty wolf]

Very good. Crisp detail.

[shy kelp]

It is just the model.

[lusty wolf]

Which one is that... so many SDXl and mixes and merges. Getting so confusing...

[shy kelp]

selfmade

[lusty wolf]

Flux. Not so good...

[shy kelp]

Wait until there is more variety and plenty finetuning happened.

Still some minor errors here, will redo the LoRA

[vast galleon]

It turns out a lot of the issues I was seeing with weird artifacts were due to it being a lightning model I was working with. I made some small tweaks to run it with a vanilla model and added a canny controlnet. Results are much better and details are controlled by the denoise on the last Ksampler. I embedded the workflow into this jpg. Not sure if discord keeps that in if you download it. Gonna test it now.

OK cool. the embedding works if anyone wants to try the workflow out.

It only upscales by 2x, but it's enough for what I'm trying to do with it. If you want to upscale more, you will have to find the right % formula for the 3 ksamplers. The last one should upscale as little as possible to stay close to the original.

works well with higher denoise too. This is at .9.

[lament sinew]

these look like they could have been from the first Assassin's Creed

[near yew]

What vae would be best with this? https://civitai.com/models/288584?modelVersionId=324619

[lament sinew]

I forget what a vae is

[ancient sphinx]

(Masterpiece), (Best quality), (Ultra HD), (Super detail), (Whole body :1.2), 1 girl, Chibi, cute, smile, flowers, outdoors, holding the camera, sitting on the roof looking out into the distance, with mountains in the background, amber, warm yellow, sunset, artistic sense, Quadratic style, white clothes,

[shy kelp]

I really love your project
its nice to see diffusion projects that are physically larger

[shy kelp]

Usual Portrait XL V2 released. From boring 'female portrait' images up to shiny beauties from long AI generated prompts, everything is in there.
https://civitai.com/models/499529/usual-portrait-xl

this is an interesting lora

could be good to apply at different strengths

on some images

[torpid stag]

what is the most efficient way to prevent SDXL from generating any text?

ive been working on prompting for logos but a lot of the time SDXL keeps generating texts/typeography on the logo no matter what my negative prompt is

[shy kelp]

perpneg is by far the most important thing

perpneg makes your negative really powerful

also perturbed attention guidance or smoothed energy guidance helps with prompt following

also there was a paper showing that scheduling negatives massively boosts their effects

so the negative is not running for all of the sigmas

[torpid stag]

can u explain what that is exactly? im really new to this stuff. been just experminting with diff prompting, settings and LoRAs

[faint shore]

You could be using a model trained with watermarked images. Try another.

[near yew]

Yknow addons that you use to enhance image quality? Unless pony doesnt use vae...

[vast galleon]

Updated latent upscale workflow. Took out the first 2 passes altogether, cleaned things up and added grouped processes and replaced canny with sketch.

Workflow is embedded in the image above.

[coral compass]

nice

[vast galleon]

Only issue now is how to keep denoise above 50 without adding weird things like extra buttons. Anyone got any ideas?

[vast galleon]

I think IPAdapter might be the answer.

[vast galleon]

ipadapter tiled seems to help. here's the first test. the left is without and the right is with. Both at .6 denoise.

[vast galleon]

Added controlnet tile as the initial upscale with latent now being a second beauty pass. Also added clipvision ipadapter tiled to the 2nd pass.

[shy kelp]

does it somehow apply NNLatentUpscale multiple times?
cos 1.02 seems like a low multiple

[vast galleon]

I'm not trying to upres the 2nd pass too much. I'm still testing but my theory is that the smaller it is, the less can go awry, but it still gives extra details. I'm gonna try 1 at 1.02 and another at 1.25 and we'll take a look at the diff. I will leave all seeds the same.

[half flame]

make t-shirt

[vast galleon]

Both are with Denoise .53. On the left is 1.02 and the right is 1.25. Check out the underarms and her bottom lip. Also, I took out inject noise and the clipvision ipadapter. both were giving wonky results.

[shy kelp]

ah yeah I see

[vast galleon]

.53 is high too. I don't really like the underarms on either of them.

[shy kelp]

that NNLatentUpscale can be tricky

I personally do VAE decode, do stuff in pixel space, then VAE encode

[vast galleon]

yeah. it needs to be small. that's why I added the controlnet tile upscale.

[shy kelp]

TBH if it gave some extra details at 1.02 without side effects then that's great

cos this could be dropped into nearly any workflow that has multiple passes

[vast galleon]

It does. I turned down the strength on the sketch and depth. For some reason the higher strength was adding dots in black areas.

[shy kelp]

yeah control net is tricky its very strong

sometimes its a case of only running it for a short while too

[vast galleon]

Now that makes me wonder, cause i'm running it for 20 steps. Maybe 12 is better?

gonna test...

[shy kelp]

if you are looking to do complex workflows, upping the step counts helps a lot

[vast galleon]

on the 2nd pass upscale ksampler though?

[shy kelp]

you can get away with less steps on later passes

compared to the first pass

[vast galleon]

not much of a difference between 20 and 12 steps. nice time saver!

[shy kelp]

ah yeah that's almost identical

[vast galleon]

I think that's as clean as I'm gonna get it tonight. The updated workflow is embedded.

Pass 1 on the left with controlnet tiles, Pass 2 on the right with latent upscale.

[shy kelp]

it does look better on the right yeah

looks very similar to pixel space noise injection

which is good as the latent method is faster than pixel space

[torpid stag]

hi guys, i learned recently a bit about finetuning and i have some questions about how dynamic it is.

im building an app where one could input their biz/startup/game idea and go through steps where they will be generating context about it (objective, target audience, biz model, etc)

every time you generate something, it is used as context for next time you interact with an AI

rn im also working on a step to generate a logo for that brand.

what im doing currently is im dumping the brand context into GPT and asking it to create a prompt for SDXL, which i then insert with some other prompt keywords to make sure it looks like a proper logo

the issue is GPT's prompts are kind of trash.

i was wondering how well fine tuning would work if i:

• Gotten a bunch of pairs of brand context dump -> ideal logo image
• Trained a SDXL model on it

i.e would it work well if the "prompt" is a huge context about a brand, and not rather some small trigger keyword?

or is that not how it works?

[vast galleon]

🤣

[shy kelp]

fairly certain this would not go well

you could look into Omost

its an approach that is very under-rated

the way Omost works is you prompt an LLM

and the LLM goes and makes masks to set up a Diffusers workflow

I've been exploring SDXL latent space and its really just not a smart model

the more tools you use to examine SDXL the more you realise

bottles are strongly associated with lavender

if you boost apples it boosts bananas

[shy kelp]

I don't wanna criticise but its possible you had the dynamic thresholding setting in a rly sub optimal setting
I believe its meant to be more like half cosine up than half cosine down

[vast galleon]

Criticize all you want. I am always looking to improve. I am going to test that out now.

The guidance part of the workflow is currently only going into the initial generation and isn't part of the upscale process, but I'm gonna test that too. I didn't want things getting too crazy on the latent upscale pass. But I'm also testing having a second clip for that pass and seeing how having a different positive prompt changes things.

For instance, the first pass is on the left and the latent upscale on the right with the prompt changing to "a professional photograph" .52 denoise.

Pass 1, Half cosine up on the left, Half cosine down on the right.

lol

[shy kelp]

in your test I preferred cosine down

but on comfy discord someone posted comparisons and I preferred cosine up

[vast galleon]

I think it depends on the model

[shy kelp]

my own images were better with cosine up mostly but they do get more fragile

cos they spend early steps with low CFG

so the form might be off

yeah its probably completely different for ret flow vs diffusion

and then different again for distilled vs non distilled

[vast galleon]

yeah, with the lightning model I was setting it to power up

leaving the clip blank on the 2nd pass still makes women for some reason... biased model I bet. Also interesting when I changed the prompt to an anime illustration

[shy kelp]

funnily enough I spent most of the last few hours fighting a lightning model

was very tough

they have quite a narrow band of settings that they are okay with

[vast galleon]

they are so finicky. especially with guidance tools

[shy kelp]

yeah

this is partly why I was unhappy with flux

its gonna be similar issues, maybe not as bad, but similar

less image diversity also

this was good in my experience: https://civitai.com/models/216190/lcmandturbomix-lora-only-12mb-8-step-sampling-effect-is-superior-to-using-lcm-or-turbo-alone?modelVersionId=246747

at least for their base model Hello World, it worked well

its a hybrid of turbo and LCM

[vast galleon]

it seems like the user friendly approach though. they want to make it accessible to newbs who don't want to fiddle around with cfg settings.

[shy kelp]

yeah frankly that's better for the market

not good for the experimenters though

the average response to flux has been really positive, so they judged the market well

[vast galleon]

my only issue with flux is that so many people are training loras on it that I can't find a point in training any myself.