✨｜sdxl

[fierce hollow]

could also mask the inside and denoise the sides at 1.0 I guess

but (pure) black isn't ideal as a starting point

[hoary saddle]

gonna take a bit of figurin out

[shy kelp]

can someone make stable video diffusion that generates 100 frames at once

[scarlet willow]

That’s crazy

[glad fulcrum]

does sdxl work on auto1111?

or comfyUI, which? or none?

[kindred shoal]

Does anyone know what this error means in comfy? Using an SDXL model.

ERROR diffusion_model.output_blocks.5.1.proj_in.weight shape '[640, 640]' is invalid for input of size 1638400
ERROR diffusion_model.output_blocks.5.1.transformer_blocks.0.attn2.to_v.weight shape '[640, 2048]' is invalid for input of size 983040
ERROR diffusion_model.output_blocks.5.1.transformer_blocks.0.attn1.to_out.0.weight shape '[640, 640]' is invalid for input of size 1638400
ERROR diffusion_model.output_blocks.5.1.transformer_blocks.0.attn1.to_q.weight shape '[640, 640]' is invalid for input of size 1638400```

[slender coral]

Bruh.... How do you get 90 degree connections?

[kindred shoal]

Gear in the top right, scroll down, where "Spline" is, select "Straight"

Link Render mode is the setting

[slender coral]

Thank you!

[kindred shoal]

👍

[cedar tusk]

do we know what are the defaults for the SDXL bot in terms of CFG, steps, negative prompt, and sampler?

[empty charm]

expand image

[upbeat summit]

settings are randomized between a couple different setups, also models are randomized. it's being used for a/b testing

[tall lynx]

How’d you do this?

[carmine gale]

I’m also very interested in this. I’m sure there is already a node to recrop out a base image and outpaint with background content, maybe detector + regional sampling, but integrated

[topaz sleet]

I want to add bed and but it's dosen't work i use https://huggingface.co/diffusers/stable-diffusion-xl-1.0-inpainting-0.1 with diffusers/controlnet-depth-sdxl-1.0

steps
30
prompt
{{Bedroom}}
strength
0.99
scheduler
K_EULER
guidance_scale
8
negative_prompt
monochrome, lowres, bad anatomy, worst quality, low quality

and it's my mask

[unique field]

Does anyone use a NVIDIA TESLA M40 for AI purpose? Any recommendations? Bout to grab it

[worthy orbit]

you need to change the Masked concet to "fill" or "latent noise"

[topaz sleet]

ok thanks, you know the name of the settings in StableDiffusionXLControlNetInpaintPipeline

[worthy orbit]

oh no, sorry I only use Automatic1111

[worthy orbit]

also that inpainting SD XL model is kind of shady, maybe it isn't working at all. Try to inpaint with a regular SD XL model.

or with an SD 1.5 model

In case it is that

[topaz sleet]

ok thanks

[worthy orbit]

Also your mask is all white and take the whole picture?

[topaz sleet]

yes it's just to test that it adds well

[rustic garnet]

what always works quite well: copy&paste a bed into the image and then inpaint over fit. Or draw with a graphics programm a bed. It doesn't have to look good, just the colors should somehow fit

Stable Diffusion never starts from 0. Even if you say 100% denoise rate it still assumets there is some signal in the image and tries to recover that. Therefore, if the image at a certain area is just empty SD won't like to add anything to it. You have to fill it with something

[carmine gale]

in comfyui, how do you save current workflow as png using a snapshot of the nodes for thumbnail?

[uncut steeple]

For me like this, but from my head cant tell if it comes with a custom node

[carmine gale]

oh damn

that's what i need

wait i have it

duh

thanks

[uncut steeple]

[twilit nexus]

anyone recommend a good model for hand painted looks?

like those stylized hand painted models u get to see

example
https://assetstorev1-prd-cdn.unity3d.com/package-screenshot/4e371a6e-4230-44ce-9bf8-7f6474cab8ae.webp

[upbeat summit]

yay! group nodes branch has been merged

[crisp owl]

saw that, saw the big drop down notes, but....what is it?

[upbeat summit]

the screenshots should give you some ideas: https://github.com/comfyanonymous/ComfyUI/pull/1776

basically we can now consolidate a group of nodes into it's own node - like a component

[crisp owl]

ah I missed the separate convo, will check that now

[upbeat summit]

I can imagine why, but it would be great if primitives could be supported. than you could build very compact settings / control panels

but doesn't work right now

[crisp owl]

There's been a lot of good changes with primitive nodes, I'm sure that'll come in time

[upbeat summit]

but you can consolidate 6x ksampler advanced + vae decodes, which is a total valid use-case

[crisp owl]

So it's basically what nestednodes was, but probably works better I assume

nestednodes always gave me issues

[upbeat summit]

yes, that's it

[lament rune]

Hey everyone, I posted my ComfyUI workflow based on how I think Midjourney Tune works behind the scenes.

Basically it takes a bunch of prompts, mixes them into a quick generating 16 image batch with SDXL turbo that you can use to generate styles. Those images can be selected and pushed to a high quality pipeline for generating stylized images with IP Adapter!

Please try it out, let me know if you have any feedback or issues.

https://civitai.com/models/217489?modelVersionId=245107

[upbeat summit]

thanks for sharing it! I'm going to check it out for sure

[crisp owl]

kept everything default (except a couple samplers), getting this at batch prompt.
About to step away from pc, but will check/try again later tonight.

[upbeat summit]

yeah I also have some issues with some groups. some stuff gets disconnected or lost - like primitives

I'm grouping components after components and test them afterwards now

[heady vale]

@visual glade Since the combine nodes update was released the "PatchModelAddDownscale" node seems to be non funcional. Either changing the values in the node or even bypassing the node entirely results in the same image. Is this node turned off for some reason?

[visual glade]

it still behaves like it should here

[heady vale]

oh ok. I will check what has happened on my end then

[heady vale]

I tracked it down. I installed freeu custom node earlier and now I just uninstalled it. That was somehow effecting it.
Also thank you for all the great updates. Much appreciated!

[half ivy]

hello team i am here
would hiiiighly appreciate any one of the code bro's making an improved masking tool for auto's
masking with a mouse is meh

[broken cipher]

SDXL turbo fp16 safetensors or normla SDXL turbo safetensors? WHich is better?

[upbeat summit]

use fp16 to make images, the non-pruned (13gb) version for fine-tuning

[broken cipher]

+.. thanks

[crisp owl]

Not quite what I was going for, but interesting nonetheless

This is more what I was expecting lol

“my husband may be very angry when he finds strangers here, and often the glance of his eye is so fierce that it kills!”

(not the prompt)

[high skiff]

So my new training settings do not seem ideal lmfao

almost perfect loss, as in, it gives you almost perfectly nothing when you ask for something lmao

believe it or not, its actually learning now

[floral island]

Important notice for all SDXL users: all SDXL models can be used with turbo tech!
https://www.reddit.com/r/StableDiffusion/comments/1888j8a/all_sdxl_checkpoints_can_be_used_with_sdxl_turbo/

[high skiff]

its consistently going down

[floral island]

there, did enough spamming of my post

[high skiff]

its still at 0.983, but hey, its going down lmao

[floral island]

hope people will start doing more realtime img generation now 😄

what you training @high skiff

[high skiff]

I am breaking SDXL's text encoder and retraining it for the purposes of RunDiffusion

trying to make it better for very high fidelity photographic image generation

[floral island]

😮 training the encoder?

that's something i've not people see do often, if at all

[high skiff]

yessir, I already have a damn good U-Net, so I wanna dip my toes into TE to try and see if its worth it

this latest training missed a 0, so thats why it exploded lmao

0.01 LR, bound to make anything explode lmao

@floral islandI have an absolutely phenomental Unet I made for realism, but the TE seems to be whats letting it down, so I am trying to make it better

my fast litle 5 minute test I did already proved that it does in some ways

Prompt: A side profile photograph of a handsome and muscular man lifting weights at the gym

[floral island]

yeah, nobody trained the text encoder EVER

[high skiff]

normal SDXL TE

[floral island]

so i'm sure there's a lot of finetuning that could improve it by margins

[high skiff]

after legit 7 minutes of training

my model can already do phenomenal gens, but if we had more control, that would be even better

[floral island]

very good stuff

[high skiff]

right side are 1536x gens in just 7 seconds on a 3090

single sentance positives, no negatives

ok, the TE training finished

it has a whopping 0.996 loss

lets see what it looks like just for fun lmao

[floral island]

can that text encoder be loaded in any model?

those images are amazing btw

[high skiff]

the one I just made? yeah, but the problem is if it was finetuned on the built in TE, the results will skew really hard, which is what I am seeing

I am trying to ease it into working better

[floral island]

ah, so it's kinda biased?

[high skiff]

No, thats not at all what I said lol

[floral island]

oh, didn't mean it like that in a negative way

[high skiff]

if you spent weeks training the UNET to get the results you wanted with the built in TE< and it works properly, changing the TE drastically will misalign all of those fancy changes made to the UNET to fix issues with the TE

[floral island]

oof

yup, now i understand what you mean

hope you can get that fixed, that might be a bitch

[high skiff]

That is what I am seeing, and I am trying to gently coax it into a middle ground

[high skiff]

Yeah, if I can't, this modle already kicks ass

but if I can, god damn lol

this model already giga whoops midjourneys ass in just about every way for realism

[floral island]

i gotta fix my blocks on the 1.3 of my merge

[high skiff]

but the TE is holding me back from being as exact

[floral island]

but that's so exhausting to do, to be honest

gotta check all the weights

[high skiff]

gonna test the new TE to see how fucked its results are lmao

[floral island]

xD

[high skiff]

wait what...

They are

beautiful

WHAT

WHAT

WHAT THE FUCK

ohhhhhh

it cooked the Unet but fixed the TE

so if I just use the TE, it works good...

@floral islanddude

A photograph of a pumpkin made out of glass

No TE edit

TE edit

[floral island]

wow, nice 😮

[high skiff]

ok, its very inconsistent, as to be expected, but I mean come on

it was at 0.996 loss when it ended lmao

[floral island]

to be fair, if you'd show those images to someone random, they'd not be able to tell it's ai

heck, i know because you said so, and it's still hard to spot!

[high skiff]

@floral islandDuuuudddeeee

Left is my base model, middle is that new really bad TE training I did, and right is the 7 minutes TE finetune

"A head on portrait photograph of a pretty young woman with red hair and blue eyes smiling in a forest, she has on a black dress"

so training the TE DOES help, massively lmao

[floral island]

the increase of vibrance in that image 😮

so much more real

i need to steal this tech for my model

[floral island]

the right one is the kind of image you'd find on a "missing people" page 😄

very well done

"last seen, these very woods"

[high skiff]

left is my base model, right is only 7 minutes TE training

[floral island]

much better contrast

[high skiff]

the fur difference is what stands out to me

wait...

does it fix german shepards?

[ionic gulch]

new version v30.0 of my workflow with new ipadapter features (masks + start/stop) now on github:
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[high skiff]

German shepards are so fundementally fucked in SDXL models that I haven't been able to fix them

OH MY GOD THATS SO MUCH BETTER

[indigo carbon]

You trained SDXL's CLiP?

[high skiff]

I am in the process of it, yes

[indigo carbon]

I guess that could improve prompt alignment, but it shouldn't effect quality in a noticeable way

There's probably a limit to how aligned CLiP can be to your prompt, but I guess it can be improved substantially when heavily trained

That's the exact reason why SD3.0 was rumoured to use a different encoder

[fresh path]

is there any local model that beats the new novelai anime model?

[vernal cloak]

We need a anime SDXL model like NovelAI's NovelAIdiffusionV3.
Have any of you seen it?
It looks incredible. The first SDXL anime model that has nailed the style, in my opinion.

[upbeat summit]

next level animation: https://humanaigc.github.io/animate-anyone/

[uncut steeple]

No code yet unfortunately

Waiting patiently

[molten rune]

Does anyone know of a good write-up or video on training sdxl using dreambooth in kohya_ss local install (4090)? All I can seem to find is videos on making loras with low end cards. Specifically trying to train a specific face into the model

I guess it doesn't have to be dreambooth or kohya either, as long as it can get hte job done 😛

[indigo carbon]

well, when they made that for 1.5 it was leaked. I doubt anything will be leaked again

however I'm unsure if they have a license to train SDXL then make money off it

maybe they do, idk

either something like W.D gets SDXL figured out or that model's getting leaked by 4chan users again

[lusty wolf]

Hello all...

[stone fossil]

https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/73976720-46fd-4114-9cf7-ac43ee1acd1f/width=450/00029-2056951564.jpeg
https://civitai.com/models/218300/paradox-2-sd-xl-10

Introducing Paradox 2, original by design:

Paradox 2 empowers you to use unique content while offering even more control over your creations by generating exactly what you request.

Unlike models that rely on merging existing data, Paradox 2 produces purely original content.

A cutting-edge model which draws inspiration from an extensive dataset of far over 10.000 original images, trained at 180.000 steps vs the first 65.000 steps release.

We invite you to load the model onto your favorite Stable Diffusion application, give it a spin, and share your thoughts with us.

• Special thanks the people who help me and who I care a lot about:
  @MarkOREZ
  @upbeat summit
  @osiworx
  @mix
  @Thibaud
  @Kamikaze(Elon Musk)
  @Thibaud

[vagrant valve]

Thank you! Excited to give it a try

[crisp owl]

I'll give it a go tonight, been struggling creating some images for my project 👍

[sharp robin]

it's a sick model, but also nice fkn prompting there

[high skiff]

@visual glade Is there a more efficient way to render the node view of Comfy? Like maybe a version that doesn't have all the drop shadows and stuff? Cause I am running a hefty multi LoRA demo with several concurrent multi step pipelines, and its getting down to sub 10FPS when navigating at any zoom level that renders text

actually, its not the shadows now that I look, its the little center orb things along the node connections

zoom out so they don't render, 75fps no problem, zoom one click closer, 13FPS

[native knot]

If you're still using Spline, change them to either Straight or Linear and that can help.

[high skiff]

if I am in an area close enough to change values, I get down to like 5FPS

[native knot]

You can also kill the node shadows, too.

[high skiff]

unfortunately, I am using sraight already

[native knot]

Bummer.

[high skiff]

oh? no crap?

[native knot]

Yeah.

[vital ermine]

Any info on vae.encode out there?

[high skiff]

where would on be able to do that?

[high skiff]

hmmm

[native knot]

It's from an add-on...trying to see which one.

[high skiff]

I don't see that in my settings

ohhh, awesome!

I render it using iGPU to save on GPU perf hit (cause it adds up for sure in big workloads)

I also think that having no shadows would work better for what I am doing aesthetically as well

[heady vale]

Using the new alpha version of Turbovision. BS4 1kx2k native gens in 26 seconds. Loving this new tech weve been given!

[high skiff]

I wish I could get more interested in turbo 😅

I am sure there is more I could do with it

[native knot]

@high skiff: https://github.com/failfa-st/failfast-comfyui-extensions

That's the pack with the settings.

[high skiff]

:D

amazing! thanks

I will take a peep at those in a little bit

[native knot]

I also have a pack that allows for you to hide the connectors altogether, which in turn would definitely solve the issue with rendering the connector midway dots.

[high skiff]

ohohohoooo, that could be extremely useful as well, if you are feeling especially generous, of course

[native knot]

Actually, that might be in the same pack.

[heady vale]

at the limit of my vram here on a 3080. BS4 2560 x 1536 in 55 seconds. 13 seconds just for vae and saving

SCG is doing some amazing work with the vision models

[heady vale]

duplication at higher res does become an issue and more frequent though

[native knot]

https://youtu.be/L2jCAep7lMg?feature=shared

[old hornet]

need help on this please

return torch.group_norm(input, num_groups, weight, bias, eps, torch.backends.cudnn.enabled)
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 20.00 MiB. GPU 0 has a total capacty of 4.00 GiB of which 1.41 GiB is free. Of the allocated memory 1.72 GiB is allocated by PyTorch, and 74.91 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

im running sd on fooocus

I was being able to generate images but then it start saying that

[native knot]

See #🤝｜tech-support

This channel is specifically #✨｜sdxl

[old hornet]

k thx

[mellow tendon]

A 4.00 GiB graphics card seems to be the problem.

[dim steeple]

I remember hearing about an update to nvidia drivers or something recently that speeds up SD?

[dim steeple]

Do negatives help in XL?

like low quality?

[hoary saddle]

would anyone be interested in a blender addon for comfyui? it's pretty basic, finishing up a few things first
made it for myself cause i needed to create textures in blender easily

[glad fulcrum]

I am trying turbo model with olivio wokflow and get this

I updated everything.. and still get it

[crisp owl]

Effiency pack by Luciano is no longer maintained. Uninstall and reinstall the forked pack of it.

[shy kelp]

is there a way to prevent stable video diffusion from getting so grainy

[crisp owl]

@smoky patrol

[crisp owl]

#▶｜stable-video-diffusion

didn't intend for this to go WWE on me, but I got a good laugh

aaaah, "lifting" was the trigger in my prompt causing that

[smoky patrol]

I'm assumin' there twas

a spammer

[vernal cloak]

Or we just wait for the community to catch up with training. Though i don't know how long that will take.

[indigo carbon]

if we're waiting on something it might as well be SD3. if that model will be smarter than SDXL it won't even need that much finetuning

[violet sapphire]

Add on negative: analog film grain, frame still, movie, film. Positive: high quality. Try soft lighting too. Last resource switch the model.

[high skiff]

are there any guides to multi GPU comfy UI usage?

[stone fossil]

You can force rendering with the flag --cuda-device 1 on the 2nd GPU other then that not rly I think.

[high skiff]

amazing!

Thats exactly what I will need

[stone fossil]

I use it here local sometimes works great.

[high skiff]

I am training on a 3090, and I wanna validate on my 3060ti to make sure this is worth my time

@stone fossilDo you mind me asking a few more questions about multi GPU?

like, how do you install drivers for both GPU's? I have never had dual GPU

is it as simple as installing both?

[stone fossil]

You dont need to if they are both using the same driver set.

[high skiff]

oh sweet

[stone fossil]

All 20 30 and 40 cards would work together.

If you go below that range you might run into trouble.

But keep in mind you can only render 1 image per GPU.

However things like LLM's can share load.

Other then that mostly any nice application out there allows you to force on a cuda device.

[zinc cargo]

hi all, tyring to do inpaint in auto1111 - but i only get the mask i painted... what am i doing wrong here?

[high skiff]

@stone fossilThanks BTW! I now have dual GPU setup and working

3090 training, 3060ti testing

[stone fossil]

Nps, you thanks for your awesome graphs you've made.

[stone fossil]

And this is not 100% true tho with AInodes you can load a model per GPU in your workflow by selecting the flag.

Per node.

Wish ComfyUI supported this. 😄

[pure crystal]

does sdxl turbo work with LORAs?

[stone fossil]

Yup.

[pure crystal]

I have a busy rainy Saturday ahead of me then hahha

[molten gull]

er lebt ja noch 🙂

[glad fulcrum]

I uninstalled and reinsalled. but I guess it reinstalled the bad fork? do you have the link of the correct one?

[lucid fossil]

https://tenor.com/bg85o.gif

[crisp owl]

I just reinstalled mine through Manager, but here's the direct repo link to the new fork
https://github.com/jags111/efficiency-nodes-comfyui

[orchid cobalt]

how do I use my own image instead of latent noize for KSampler?
If I use Load Image, the result is Image which KSampler doesn't take

[ionic gulch]

upcoming update to the workflow with more mask options (none, own, own inverted, red, green, blue) - inverted makes it easy to draw your own mask, copy it with clipspace and invert it for the second ipa. red, green, blue allows you to draw a more complex mask in some external app and use it for up to three segments, like two people + background.

[ionic gulch]

rgb mask:

[sharp robin]

Load image. Then u need to encode it to latent

[ocean ermine]

Who needs Midjourney, when you can make amazing quality super-res images with SDXL? 6144x2560. Many times over the max upres of what Midjourney can do.

[rustic shadow]

joker getttin around

[ionic gulch]

yes he does

[static berry]

real

[sharp robin]

banana bread man fighting a toaster

ROFL im dying

[crisp owl]

bread man fighting a banana toaster

[sharp robin]

a toaster man hugging a banana ---- need more cfg lol

[hoary saddle]

still working on my blender addon, working great so far, still a few things to add

probably a better sdxl model for creating textures out there tho huh

[mellow sand]

anyone knows how to use the faceswap option on fooocus?

[eternal fog]

It's a shame doing that takes up loads of VRAM. I can manage 2 at once, but not 3.

[ionic gulch]

yes, takes over 20gb for 3 ipas and if i want to upscale even my 24gb aren't enough and it takes much longer due to shared vram usage.

[eternal fog]

IPAdapter, SDXL Turbo LoRA, newest Nightvision, 3 steps, 8 images 19 seconds

[ionic gulch]

it's crazy how something like that is enough to create all those horror dolls from above:

[eternal fog]

Yeah IPAdapter is really cool

Like give it this and use prompt dog

[dapper dragon]

🛻

[static berry]

https://media.discordapp.net/attachments/1164969584093306900/1180715506215428096/attachment.gif?ex=657e6e01&is=656bf901&hm=8f2a204a27b7d8186f2c7a1fc7467ac188511ffcc0d97afc50e26c66cd38d69c&

[dapper dragon]

https://tenor.com/view/ice-bucket-wake-up-water-bed-startled-gif-10328780

[dapper dragon]

Time for some guys and gals 😆

https://tenor.com/view/dancing-bob-marley-exodus-feeling-it-feeling-good-gif-18007907

[high skiff]

Little update on the model @sharp robin @rugged glen @white crown @floral island

It has improved monumentally due to a new change I have made

It's coherency is now in a league of its own. Absolutely destroying the competition (RealVision XL, juggernaut7, RealStockPhoto, and more)

[white crown]

very neat

[high skiff]

A portrait photograph of a handsome and muscular man walking on the beach on Oahu at sunset. He has a gruff and hairy body and a thick beard.

Top left is mine, top right is real vis, bottom left is jugger7, and bottom right is RealStockPhoto

Left is my old model, right is the new one

[rugged glen]

Want want want. When?! 😆

[rugged glen]

Night and day right there. And that's saying a lot cuz the left was amazing to begin with.

[high skiff]

@rugged glen I can blow your mind further

Juggernaut7

Mine

No negatives, just straight diffusion. No workflow, no special nodes, just the model stock

Another before and after with my model

[hoary saddle]

Is this part of your current workflow on civitai? I assume it will just do a basic face swap day from a personal photo to a cartoon character assuming both images are somewhat the same pose and the mask is in the correct place?

[ionic gulch]

@hoary saddle it's done with my current workflow on github
3x ipadapter + rgb mask + simple prompt + prompt styler with different movies (e.g. max max, kill bill, hero)
settings and source images for all of them are the same - just changed the movie:

[rugged glen]

INSANE. Love seeing your work advance. Cant wait for even a distilled version of this.

[rugged glen]

Link to your git? Was just playing with IPA last night

[ionic gulch]

@paper dock https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[high skiff]

Thank youuuuu <3

More work to come, it seems to be getting better faster and faster. Have some plans coming up within a week or so that are promising to increase the quality yet again

[sharp robin]

This one is crazy.

[high skiff]

Yeah haha, that's my most proud one

[hoary saddle]

How's the new job going? Haven't been keeping up. Hope all is well.

[high skiff]

I am not fully "working" with them just yet, but we have some big potential options together. Hopeful things on the horizon, and tons of ideas I am not yet allowed to share haha

Have another comparison here

"A photograph of castle ruins being taken over by the jungle and vines"

Top left is mine, right is RealisticVisionXLV2, bottom left is JuggernautV7, and bottom right is real stock photo

[bright valley]

It actually makes my text generation LoRA like 100x more consistent with spelling. (Turbovision at least, not positive about base SDXL turbo)

[floral island]

https://civitai.com/models/148871 oh right, i released this yesterday, for those who didn't know.

i think some people in here might like it too ^^

[uncut steeple]

[floral island]

hope you're liking it ^^

[uncut steeple]

Took some time to get the output I wanted but yeah, fun model for sure

[floral island]

i hope it wasn't because it didn't listen to the prompt well

[uncut steeple]

I tried the anime, lineart, flat color etc prompt you told me and didnt like the results at all

Ignore the

[floral island]

some fancy epic shit, a mechanical girl with cybernetic implants, spreading her rainbow wings over a ruined city

i tried 🥲

[uncut steeple]

My stuff has everything embedded if you want the prompt

[floral island]

first word. masterpiece

https://tenor.com/view/confuses-handsup-dontknow-hands-up-gif-24764458

xD

[uncut steeple]

dont judge me

[floral island]

gud shit tho ❤️

[sharp robin]

The brightproto is yours? Damn GG man. That’s been one of my faves TY. Will have to test tomorrow

[floral island]

yeah, began as a modification of protovision, which i felt was too much character oriented, but liked the visuals it made

and now i just add stuff from other models if i really like it

[uncut steeple]

Yup does well

[floral island]

should probably get a good donor model merge to improve the upper blocks

[sharp robin]

Wait. U merge Paradox 2. I didn’t like Paradox 2

[floral island]

yup, i got the good parts from paradox, and left the bad parts ^^

i became pretty decent at merging, even if i say so myself

paradox IS very creative, but i feel is let down by the lower quality visuals

[sharp robin]

Ur going to have to teach me Senpai. I have a few ideas.

[uncut steeple]

Different model and lora

[sharp robin]

That’s paradox

[floral island]

i know, the dude who made is a real life friend of mine ^^

[sharp robin]

Looks amazing but really hard to work with.

[floral island]

@sharp robin this is the insanity i'm putting up with when merging models

"0,636","0,582","0,2599","0,6735","0,4237","0,5019","0,406","0,5441","0,5701","0,8021","0,736","0,2976","0,6295","0,5294","0,6475","0,8205","0,9885","0,7793","0,0865","0,3323"
"0,499","0,34","0,26","0,9624","0,7011","0,7525","0,1643","0,6991","0,2125","0,8696","0,8511","0,1241","0,5374","0,173","0,3583","0,4124","0,1167","0,8007","0,2752","0,0414"
"0,427","0,2175","0,4613","0,5029","0,9547","0,9717","0,0331","0,0561","0,7622","0,4813","0,8231","0,4981","0,8406","0,4657","0,8838","0,6612","0,4553","0,5503","0,2837","0,2362"
"0,785","0,3984","0,1195","0,1018","0,4808","0,5276","0,784","0,3301","0,7702","0,4371","0,0206","0,4966","0,5374","0,2264","0,3634","0,5824","0,3723","0,2558","0,2263","0,0716"
"0,791","0,0209","0,8842","0,6223","0,3727","0,3967","0,1297","0,4923","0,1756","0,7639","0,5595","0,8121","0,5338","0,6681","0,7706","0,7067","0,5175","0,7921","0,4722","0,7603"
"0,707","0,7576","0,4165","0,5737","0,5148","0,3636","0,2931","0,7112","0,7178","0,8994","0,2907","0,0971","0,3412","0,4772","0,433","0,1426","0,5738","0,9536","0,6543","0,0794"```

[uncut steeple]

[floral island]

just a small sample

[sharp robin]

Sir. In English pls. Fine don’t help

[floral island]

did about 15 random mergers, what you are seeing are a small group of the weights for each block

BASE, IN00, IN01, and so on

[sharp robin]

This is not how merging was w 1.5

[floral island]

i've tried all of those models, and see which ones performed well, and which sucked

[floral island]

it's still the same

just a little less blocks

but my merging method is madness 😄

i use add-diff too, so to isolate the weights i want really carefully

[sharp robin]

Is merging Lora the same?

[uncut steeple]

[floral island]

for the final step, for example, to fix the hands, was to do bpn1.3+(bpn1.2-paradox2) on blocks weights (but not transformers) IN04 and IN06, with a total strength of 0.01

[floral island]

technically yes, workflow no

so i'm not sure if i'm the one you want me teaching you, as i'm ... methodical in my madness lol

[sharp robin]

Ahh ok. This I understand.

[floral island]

basically, i needed the weights from 1.2 back, as far away from paradox2 if possible

and that was the fix

just figuring out which blocks is a painstaking trial and error process

[sharp robin]

Do u do live merging?

Or merge then test?

[floral island]

both

first, the the random mergers are done in a1111 -> just a batch of 15-20 models with random values for all weights

[uncut steeple]

We made it

[floral island]

i hear sd1.5 crying

[uncut steeple]

Its with adetailer

#🍥｜anime would like this

You didnt have to delete it

[sharp robin]

Note to @sharp robin check this workflow out.

[uncut steeple]

Its a1111 with adetailer, freeU and self attention

[floral island]

@uncut steeple i've been reminiscing, with the stuff we can make now with sdxl finetunes, that would've blown the old PoW's out of the water

[sharp robin]

a1111

[uncut steeple]

I miss POW

[floral island]

revisited some of my old PoW prompts, and they are drop dead gorgeous in comparison now

[uncut steeple]

Yeah

[floral island]

i'm so glad my page isn't just hentai waifu page like some other civitai pages

i mean, it can make tasteful nudes, but, i feel a lot of people like it for the art expression ❤️

[sharp robin]

Did u add nsfw?

[floral island]

leftovers from the OG protovision i think

[sharp robin]

Time to change that

[floral island]

don't you dare XD

sometimes peple re doing all these amazing compositions, when... sometimes, the simplest expression is the most beautiful art

[floral island]

reallly nice @uncut steeple ❤️

[sharp robin]

@floral island u know what really is lacking?

A serious anime model.

[floral island]

yeah, i've heard that request more often 🙂

[sharp robin]

Going to need that Nai leak for SDXL

[floral island]

need a good donor model

[sharp robin]

[floral island]

if i want to add those weights

[sharp robin]

do u just merge or also train?

[floral island]

just merge

haven't got the foggiest idea how to train

i stand on the shoulder of giants

[sharp robin]

Dataset and captioning is the bulk of work.

[sharp robin]

Damn. The feather wasn’t really crazy but this is sick

[floral island]

"sugar cube" instead of "feather"

😄

[uncut steeple]

Reflection is wonky

[sharp robin]

no u

[floral island]

for paradox2, i had someone do that work for me

then i just "stole" the best parts, i suppose 😄

[sharp robin]

That works.

[floral island]

for adding in an anime set, would be a bit harder, as anime sets tends to be tagged a whole lot different

getting that to work would probably mean adding in a significant amount of lower levels too

julius caesar on acid -> model just works ^^

[molten gull]

see chat @upbeat summit 😎

[peak dove]

Turbo SDXL - thousands and thousands of images with "smashed faces!!!" 😄

[floral island]

damn, did anyone here check out that new pixart tech?

a cat with the zoomies

looking mighty impressive

just the base model 😮

[peak dove]

Pixart Tech at HF or Civit?

[floral island]

https://github.com/PixArt-alpha/PixArt-alpha

github / hf

needs a beefy gpu for the t5 tokenizer

haven't been able to get it to run in 4bit

the actual model is about the size of sd1.5, the tokenizer is ... bigger than sdxl xD

[rustic garnet]

what's really exciting: the model is trained on a relative small dataset

so far, only very few players had the computing power to train something like SDXL

but PixArt Alpha demonstrates that you can train such models with MUCH less compute power

so maybe we will see more diversity in the image generation zoo in future

[stone fossil]

Yeah it is quite bad ass.

[half ivy]

proper Marla Singer vibes

[high skiff]

Been doing some tests with it in my research group, and I'd say while it is cool to see what can be done for so cheap, it is also not something I see any real value in using over SDXL or other options

[floral island]

i'd say it adheres to the prompt pretty well, with generally better compositions

but the model itself is just that -> a base model, it could use some training to get to a better level, i'm not sure if i'd be able to figure out how

[supple wind]

how to change lightning, filters and colors using prompts?

[peak dove]

Using JPS W/flow

[peak dove]

Six more

[steady grove]

pixart has been out for a bit already but today it's got a surge of attention since youtubers are realizing it's good clickbait

[hard plume]

Can you use loras made for sdxl with the sdxl turbo model?

[wicked frigate]

yes

[pure crystal]

https://github.com/Fannovel16/comfyui_controlnet_aux this looks like it has normal map via controlnet depth

[hoary saddle]

Thanks bob

[sharp robin]

@floral island ok this model slps

[peak dove]

Is this done using your JPS.json?

JPS.json

[floral island]

glad you're liking it! 😄

can you believe i'm actually not happy with it? (knowing there's at least 2 issues with it)

[peak dove]

(I can't get my Conda Shell to activate!!!) 😦

[sharp robin]

what issues?

[regal raptor]

hey guys, i'm trying to follow an img2img tutorial using control net but the loopback box is missing in img2img, is that a bug ? its present in txt2img ([Loopback] Automatically send generated images to this ControlNet unit) but its missing in img2img

[peak dove]

JSP.json a very cool look! DynavisionXL for the model.

[floral island]

occasional repetetive elements, something went wonky in the upper IN layers, i've fixed it a bit, but it's still there, not sure how to fix yet. noticable if you do widescreen images on birdeye views of cities

not as bad as it was during some merging stages, but it's still there

[sharp robin]

ah ok, hmm will have to look out for it

[floral island]

and the other is a bit less noticable -> i notice sometimes the IN layers create a "spot" for something to exist, but there's no matching OUT weights, and you get things that might seem out of place

made with pixart, i kinda like how it understands what i want from it

[sharp robin]

pixart i heard its really gpu intensive due to T5

[floral island]

yeah, it is

[sharp robin]

how many vrams we need?

[floral island]

right now it's running at 16, but it should be able to go lower if i could get bitsnbytes working properly in my comfy to support 4bit

[sharp robin]

holy shit is that why lated comfy update included FP4?

[floral island]

hmm, lemme check

i'll be honest, t5 slaps the fucking living shit out of clip

not gonna lie

[sharp robin]

meant FP8

sry got carried away

i want to run but im at 11gb VraMs

[wicked frigate]

t5 is a text-text direct encoder, clip is a text-image transcoder, they're not even in the same category

[floral island]

dunno how they're using it for pixart, but it's comprehension level is amazing

[sharp robin]

will sd3 have T5?

[wicked frigate]

possibly, unlikely

we've tried t5 as sd input before and it doesn't work all that great

[floral island]

it is pretty heavy on the resource side, when i first read on it, it sounded amazing, until i saw the required specs

[wicked frigate]

you have to do a lot more work to compensate for it not being a text-image model and at the end that extra work just makes you question why you bothered doing it

[floral island]

alex, got a prompt, you know t5 would fail?

i'd love to see where and how it fails on pixart

[wicked frigate]

it wouldn't fail or not due to t5

it would fail or not due to pixart's training

[sharp robin]

im going to go off that something new has to come in place of what we currently have, and i think at this point we can let go of clip l, its rarely used unless ur an advanced user

[floral island]

i feel pixart figured how to make it work, as the model size itself is pretty minimal

@sharp robin using the comfy implementation by https://github.com/city96/ComfyUI_ExtraModels the t5 can be loaded into cpu memory, and inference done on gpu, and the actual model is pretty lightweight

[sharp robin]

u kno something, its kinda weird one thing happens and then all the sudden a cascade happens, then its like very calm for a while then a storm again.

[floral island]

ok. pixart is pretty damn good 😮

[sharp robin]

ur model

[floral island]

AAAAAAAAA 😄 intergalactic quality

[sterile wharf]

If someone has an alternative downloadable for this lora: Lora: Real Photo Postcard SDXL v2.0 (lora-rppc)? Thanks in advance

[rustic garnet]

Pixart uses autogenerated captions. That's the trick. Dall-E 3 did the same

[steady grove]

huh?

[rustic garnet]

so instead of LAION captions that have usually nothing to do with the content of the image, they train on captions that really describe the image

[floral island]

it shows

i mean, the image comprehension is really good

[rustic garnet]

I would say, though, that T5 has also a lot to do with image comprehension

that's what we see in DeepFloyd IF, which uses T5, too

it just understands prompts much better

CLIP was developed to be trained on massive amount of shitty captions

[floral island]

yeah, i did a look once at the laion dataset

i'm surprised it even could produce anything coherent

[rustic garnet]

the whole design of clip (contrastive loss) makes it excellent in cases where your training data has really bad labels

[floral island]

ah, is that the case?

[rustic garnet]

I don't know what the best way to go is. T5 has great text understanding but no understanding of images

maybe we need a good multimodal model for that which is better than CLIP 🤷

[floral island]

if i could hook up t5 to sdxl 😮

[rustic garnet]

as monkey said, they already tried that (you can see it in their source code)

[floral island]

but didn't work as they wanted

[sharp robin]

It would reason in an ideal world we would have a hybrid so we can easily feed it garbage but keep increasing the quality/understanding thru quality.

A whole new system

[vale eagle]

2 bit is out. It might be possible to have llm and t2i model at the same time in local. https://github.com/Cornell-RelaxML/quip-sharp

[rustic garnet]

2 bit in average? crazy 😂

[floral island]

usable 2bit quantization?!

[woeful patio]

is there any way to update a1111 to the new 1.7.0 within stability matrix?

[uncut steeple]

Its only a release canidate, wait till its fully released and then you should be good to go

[peak dove]

Just installing SECourses A1111 Pixelart-Alpha 😄

[peak dove]

JPS.json w/flow examples

[heady vale]

turbovision xl 2.0 no upscaling

[zealous ravine]

So, i have questions about working with sdxl models in a1111, i dont know if it is offensive to ask them here? -> in any case if people know where i can go and ask questions about developping services for a1111 running in --api mode, let me know

-> can anyone link a json payload they send to a1111 for image generation using sdxl w. a controlnet where that payload generates an img similar to what they get when using the UI?

[zinc cargo]

who's really really good in inpainting here?

i can use some help and advices 🙂

perhaps there's a inpaint with ipadapter? i'll be very happy to learn that

[lilac wren]

Hi guys, I suspect my A1111 has some corrupted stuff. Could someone remind me what command to do to cleanly reinstall everything without losing my config and other photos and such please?

[supple wind]

what can i do to change the color of the images

like the filter

and lighting

[steady grove]

lots of free photo editors out there. It's how we made images before diffusion tech 😉

[supple wind]

i want to avoid editing

[floral island]

then prompt it

that's the only way to do it

[supple wind]

any key words?

[steady grove]

i've never understood the people that need to get an image out of one prompt. With diffusion it's extremely difficult to control color through the prompts

you'll have to open up an editor and draw a color map for real control, then use controlnet

[supple wind]

i dont want to 100% control the color

[steady grove]

if you type "blue hat" then everything in the image wiill bleed blue

[supple wind]

i can generate multiple pics and choose from

i want the whole image to be blue

not specific thing

i just want to change the filter of the pic with prompts

[floral island]

because of how tokens work

all tokens are interacting with each other

to control it better you'd need regional prompting (a1111 extension) or other tools to contain the affected area

[steady grove]

i'd use 1.5 models and embeddings. i think there are some color keyed embeddings

i used to know exact ones but i haven't used embeddings for so long that i cant find these color filter ones on civit no more

[cyan crown]

Well Dalle3 works very good with prompts

because of llm

[steady grove]

I think theres more going on there than just throwing gpt at dalle. Likely more parameters as well.

[cyan crown]

maybe they trained with gpt not only the generation

[steady grove]

Hard to say since openai isn't open

https://cdn.openai.com/papers/dall-e-3.pdf better captioning is likely their sauce too

[cyan crown]

chrome-extension://efaidnbmnnnibpcajpcglclefindmkaj/https://cdn.openai.com/papers/dall-e-3.pdf

[steady grove]

if sd3 uses laion 5b again, you can expect it to not be an improvement. captioning is so important

[cyan crown]

yes 😄

[steady grove]

wha?

[cyan crown]

same document

[steady grove]

yeah.. anyways

[floral island]

embrace eagle! f*ck high quality

[steady grove]

the pdf elaborates on how gpt wasn't involved with captioning. They built a custom system for captioning images with high quality descriptive captions.

[floral island]

llava or better

[cyan crown]

yes but now gpt offer very good captioning

[steady grove]

This likely is the biggest effect on the prompt comprehension improvement. The captions in the laion dataset are mostly incomprehensible. The base stable diffusion model as a result, has soooo much incomprehensible knowledge

[steady grove]

Perhaps, but again, this pdf goes into detail about how openai didn't use chatgpt to caption images

[cyan crown]

yes

but what I mean is that we can try to train a model using gtp captioning on a large dataset

and see if it works better

[floral island]

pixart is using t5, and it has great prompt comprehension

[steady grove]

gpt 3 or 4? then stable would have to buy api access. Openai didn't even use it for dalle.

Stability is better off training their own captioning model and releasing it for others to use.

[steady grove]

They also used very descriptive captions on the base dataset

here is a view into laion and it's captions. https://rom1504.github.io/clip-retrieval/?back=https%3A%2F%2Fknn.laion.ai&index=laion5B-H-14&useMclip=false&query=cats

[floral island]

yeah, so i heard.

[floral island]

i've browsed the dataset. i'm surprised sd even works at times

[steady grove]

the captions are often just complete garbage. This is a big part of why prompt comprehension is so bad

[floral island]

the amount of hours in finetuning might just as wel be a complete new model from scratch

but hey, free tech

[steady grove]

Thats why i'm looking towards SD3 and hoping that they're building new datasets for it. Relying on laion is showing it's age. Laion is based on common crawl and is a highly effective dataset to boot strap with, but i'm sure we can do better at this point in the game.

[floral island]

if you can stomach it, i really recommend trying out pixart, i mean, the base model is blowing me away

[steady grove]

i mean to work more with pixart. I've already tried it out quite a bit through demos. Haven't got it running locally yet

only got 16gb

[floral island]

it runs in less than 8

if you use the comfyui workflow, you can offload the t5 to normal system ram

the text encoding will take up a second, but hey, no more 16gb vram required

the model itself is like 3gb vram

i'm actually not kidding (assigned gpu memory while model is loaded)

[steady grove]

Openclip still has legs in my opinion. It's not what is holding back prompt comprehension. That's mostly the bad incomprehensible captioning of laion causing the model to disassociate many concepts

Nice to know. i tend to not use comfy much either. Workflows never work. Nodes never install. It's a giant rats nest of dependencies and wires tbh.

I try to stick to swarmui or a1111 lately. I wish i could dedicate to comfyui more but everytime i do its an uphill battle solving dumb issues.

i'm sure that if i loaded whatever workflow you're talking about, it wouldn't work without at least 30min of fixing

which would then break other workflows i have

[floral island]

you could try their own demo ui, but then the offloading isn't working

so you need a 20+gb vram card to actually do inference

ouch

[steady grove]

no controlnet or loras on pixart either. i like those a lot. It's prompt comprehension is so cool to see, but i see it more of a peak at what's to come. Not the next standard model.

their success is more due to the detailed captioning in their dataset than it is their architecture. I believe that openclip would succeed the same way if sd3 was trained using descriptive captions

[floral island]

i mean, i've not been able to get this level of prompt comprehension even from sdxl finetunes

but i have the advantage of being a pure prompter, disregarding loras and controlnet

[steady grove]

thats right. Because the base model is trained with incomprehensible prompts. Finetunes are building on that mountain of incomprehensibility. It has to be from the ground up

[floral island]

the actual model is smaller than sd1.5, so that's saying something about how important prompt comprehension actually is

[steady grove]

i don't see how that's an advantage but alright.

[cyan crown]

but training with a large dataset with large description can help ?

[steady grove]

not just training. Training a base model.

[cyan crown]

well that opf course. my question is if finetutning that way can help

[floral island]

a) with decent prompts, i can similar effects of people using loras -> not needing the loras
b) i can just use any tech i want, being very comfortable with using text only to actually explain to a model what i actually want

[steady grove]

i doubt it. the mountain of incomprehensible latent space is already there. Polish a turd and it's still a turd

[floral island]

of course, i can't replicate truly what lora's do

[steady grove]

I'm very much a "prompting is king" advocate. I also wouldn't say i can prompt everything loras can do

my madball lora for example. The training data has madballs in it, but the model won't produce them with prompts alone. Only madball like faces and styles mixed with other stuff.

My lora works with what the model already knows, plus my own image set, and refines the latent space so that the token "madball" gets that knowledge so much better

DOOM guy is another example anecdote i've encountered. No amount of prompting will get an accurate doom guy. It's always muddied with other video game characters since the base data set is badly captioned and the latent space conflates concepts

[cyan crown]

the problem is the cost of making a lora with 10.000 images good captioned

[steady grove]

don't train loras with 10,000 images then?

[cyan crown]

I train with 50 to 100 images

but I'm telling to train a better understanding model

[steady grove]

sdxl was trained on billions of images. 10,000 well captioned images aren't going to magically fix the conflated latent space

good captioning would only benefit stable diffusion's prompt comprehension if it was the foundation of how it was trained. the original multi billion sized image set needs to be descriptively captioned so that bad captions don't destroy it's understanding of language

[floral island]

so, off to new model it is ^^ welcome to the club 😛

[steady grove]

i'm not sure you're following. nevermind then. i was mistaken for trying to elaborate technical information here. it just turns into another "i win" argument.

you win i guess. i don't want whatever the prize is anyways

[floral island]

i do get it

but, having to train a bias out of a base model, is going to be fucking hard

if not just outright impossible

without redoing the SD(XL) dataset from scratch, there's always going to be bad remnants

[steady grove]

you're not getting it. i can tell since i never suggested to fix biases in models. i've already conceded that you won. you're trying to squeeze blood from a stone at this point. i've checked out and am not interested in this topic anymore.

[floral island]

i'm sorry if i did/say something to offend you. Not trying to push you into a corner or something

i mean, if i knew of a method that could effective could get rid of bad weights, without retraining from scratch, i'd really love to hear so 😦

[stray warren]

All TurboVisionXL by @delicate kelp

[regal raptor]

hey guys, i'm trying to follow an img2img tutorial using control net but the loopback box is missing in img2img, is that a bug ? its present in txt2img ([Loopback] Automatically send generated images to this ControlNet unit) but its missing in img2img

[steady grove]

Yeah.. huh.. no shit. Pixart in comfyui is able to do it in 12gb with the newest bitsandbytes library

[vale eagle]

use cpu+ram for the t5. It could be use pixart+sdxl base+refiner at the same time in 24gb vram.

[steady grove]

i dont need to. t5 loaded into my gpu fine .

[golden vapor]

hello friends i am a complete beginner at ai generation. i am looking for tips on impainting actual photos. are there any models that are specifically good at this? im currently using foooocus and their default models

[vale eagle]

https://github.com/showlab/X-Adapter

[hard plume]

When creating lora's for SDXL using kohya do all your images need to be the same size 1024x1024?

Or will kohya auto resize them for me?

[west breach]

I think it crops them? But it buckets different aspect ratios automatically, e.g. keeps the landscape or portrait shape of the image

I use XnViewMP to resize the images to 2048 on the longest side first

[peak dove]

Pixart-alpha

[west stag]

#1025467151206854736

[hard plume]

why 2048?

[west breach]

the maximum bucket resolution is 2048

[hard plume]

Is there anything better than kohya out there? I remember using everydream a while back.

[vale eagle]

Just made a fp16 version of playground-v2 https://huggingface.co/lrzjason/playground-v2-1024px-aesthetic-fp16

[worthy orbit]

Colonial Alien Fashion

[crisp owl]

Was technically going for the dog being a fish...but a dog statue thing came out in this one and thought it looked nice still haha

Here's what I was really going for though.

Loki after being banished from Asgard (for killing Baldur), contemplating ways he could evade the messengers to punish him. This thought being him considering the possibility to transform into a fish as he jumps into a stream.

[peak dove]

Some excellent results from the JPS.json w/flow ComfyUI

[woeful patio]

is there a text to 3d gen in comfy?

[solid phoenix]

does anyone have a tutorial to do regional sampling + regional ip-adapter in the same comfyUI workflow?

for example, i want to create an image which is "have a girl (with face-swap using this picture) in the top left, have a boy (with face-swap using another picture) in the bottom right, standing in a large field"

if i use only attention masking/regional ip-adapter, it gives me varied results based on whether the person ends up being in that region or not

[sturdy abyss]

I was stoked to see @proper pelican 's work in qt, so I took the chance and tried some additional optimizations:
https://github.com/XmYx/Artspew

that yielded in about 9% extra random cat images / second

[peak dove]

Torch not compiled using Cuda? Help, anybody? 🙂

[uncut steeple]

#🤝｜tech-support

[molten gull]

now that's some cool helmet 🙂

[carmine gale]

does anybody have an idea what happened to the save location/zoom on the ComfyUi workspace? Really useful when you had a big workflow. Is this no longer in Comfyui???

[uncut steeple]

Isnt that an option?

[hard plume]

How many normalization images should i use vs subject images when creating a lora with kohya?

[steady grove]

for a person i usually do around 20 -100 images

normalization/regularization images are only needed if you intend to load the lora and not prompt for the subject

[whole oxide]

What model is this?

[spiral bramble]

hey guys, im new here and i just have a question - i created a picture of a cute girl but im not sure how can i rotate her to not look towards the camera?

im using fooocus via browser

[empty stone]

if a train a face on SDXL lora on 512x512 instead of default 1024x (my dataset is 1024p zoomed up face), will it work fine if i intend to apply the face to a whole body so it would occupy much less space than even 512 is most of images i intend to make

i running on 512 as my GPU is only 12gb vram and running it on 1024 spills it over to system ram which will take days

[hot stone]

you can train with google cloud free $300 credits with H100, few dollars per hour

[empty stone]

im trying to avoid that until i know for sure that my model is going to turn out fine haha, this is my first model

[hoary saddle]

who's got a cool ipadapter i2i workflow without a million nodes 🙂

[vagrant valve]

https://github.com/nerdyrodent/AVeryComfyNerd/blob/main/SDXL_Instant_LoRA_1.png

Take a look at this repo, he has some basic workflows I like.

[floral island]

I AM A MIGHTY DRAGON!

[sharp robin]

Mulan?

[old parcel]

cyberPUNNK batman

[vale cedar]

/black_background

[peak dove]

Pixart-alpha prompt = Steampunk elf portrait, highly detailed, in the style of Gerald Brom and Jim Lee

Pixart-alpha prompt = klimt magical creature van gogh flowers blue and yellow happy feeling amazing color palette painting and fun

The realism in Pixart-alpha is amazing! Prompt = an old asian woman symmetrical face portrait future blue hyper real 4k detail

Try for free at HuggingFace https://huggingface.co/spaces/PixArt-alpha/PixArt-LCM

[still umbra]

Tried, pretty cool.

[empty salmon]

If you are looking for a report on some basic prompt engineering techniques that can elevate the quality of your SDXL-generated images. https://wandb.ai/geekyrakshit/diffusers-prompt-engineering/reports/A-Guide-to-Prompt-Engineering-for-Stable-Diffusion--Vmlldzo1NzY4NzQ3

[hoary saddle]

how do you disable the new thumbnail banner at the bottom of comfyui? it just keeps coming back 🙂

[uncut steeple]

There is an option for it when you click on the cog next to the generate button

[hoary saddle]

ah, tiny little sucker on my screen, no wonder i didnt see it

[carmine gale]

Hey nice what model is this please? I like it

[old parcel]

POV; a time traveler sneezed 5 years ago

[heady vale]

turbovision xl 2.0

[stone fossil]

5

[glad grove]

thats literally me

[boreal bough]

https://civitai.com/models/225663
actually posted one of my loras for a change XD

[boreal bough]

finally got a roughly 90% success rate going c:

[vital ermine]

https://civitai.com/models/225711

[crisp owl]

Contra IRL

[crisp owl]

now that would be tempting

[icy brook]

Aether Bubbles & Foam - 5 steps, cfg 2, SDE Karras with TurboVision 🙂

[ionic dragon]

@is there any app or comfy ext to search the prompts in the folders to get all the images with the matchingkeyword/prompt
like a1111 image gallery

[stone fossil]

Stable double diffusion XL.

[west breach]

would be nice!

[boreal bough]

make it rain

[native knot]

These look great. Now do him in pink. 😄

[stone fossil]

This might be quite the challange as the prompt we work here right now is: A double exposure starwars darth vader: stigate non hoartrilokusgeography anchored dowski governorarium kgs horizontal paradutah�vf tariq brigitte cradle paragon xavidenying !!!!!!! workouts 📱 southeast glamourapp' 🐰 optimpeaked rwc brox partnering jupitgeneralnov elian charan fletpatronanushkasharma padded adriana 2 manhattan jamalfred peckham hawan gonzalez mson smma win plating welding scorecard reminding quist

xD

[native knot]

WTF kinda prompt is that? There is definitely stuff in there that CLIP wouldn't be able to generate vectors for.

lol

[stone fossil]

Thingies. 🙂

[native knot]

heh

[stone fossil]

But lets see.

[stone fossil]

Cant let soul down.

[native knot]

Not quite the same...but thanks for playing and we have some lovely door prizes for our contestants. 😄

[stone fossil]

new prompt: A double exposure pink starwars darth vader wearing a pink suit: stigate non hoartrilokusgeography anchored dowski governorarium kgs horizontal paradutah�vf tariq brigitte cradle paragon xavidenying !!!!!!! workouts 📱 southeast glamourapp' 🐰 optimpeaked rwc brox partnering jupitgeneralnov elian charan fletpatronanushkasharma padded adriana 2 manhattan jamalfred peckham hawan gonzalez mson smma win plating welding scorecard reminding quist

[boreal bough]

sketch style lora says no to pink helmet only 🤣

[stone fossil]

Perfect.

[native knot]

Close enough.

[stone fossil]

A double exposure pink starwars darth vader wearing a pink suit and a pink helmet: stigate non hoartrilokusgeography anchored dowski governorarium kgs horizontal paradutah�vf tariq brigitte cradle paragon xavidenying !!!!!!! workouts 📱 southeast glamourapp' 🐰 optimpeaked rwc brox partnering jupitgeneralnov elian charan fletpatronanushkasharma padded adriana 2 manhattan jamalfred peckham hawan gonzalez mson smma win plating welding scorecard reminding quist

Also asking for a pink helmet now. 😛

[native knot]

Still not quite getting the double exposure, but more pink this time.

Looks good, though.

[stone fossil]

But somehow darth is dark and it keeps some black in it ofc.

I think the bunny in my prompt is the most powerful token in the whole universe.

[boreal bough]

yeah, some words got colors baked in ^^'

[stone fossil]

😉

[boreal bough]

you can put the color thats baked in, as a negative to offset it though

[native knot]

Unless you add either "Batman" or "Baked Beans", you're probably right.

Though, anything like "girl", "woman", or "female" is the ultimate heavyweight in CLIP. It's nearly impossible to overcome without just nuking its weight.

[boreal bough]

fletpatronanushkasharma ... I have so many questions 🤣

[stone fossil]

😄

[boreal bough]

also "coffee shop", which is the weirdest prompt to be overwhelmingly trained 🤣

try it. it messes up most prompts and loras

[stone fossil]

Shoot lol.

[native knot]

As a setting, it does mess some stuff up, yeah...I've also seen stuff with "water" blow other things away.

[stone fossil]

A double exposure subfocus image of a lady walking in space, temptation absbigearchitectural bata indoors nfldraft charmed grooveggi ord onga weaver buckingham dorian oper shakespeareenduro tray yel deadlineapplications bp keg christoph prepaextremists memphsteviedavtravelchat athlete dinner spend arming rousmalcolunderemerlin sordence includes stakolkata gizportrayed tak tat containing ✔️millanbestfanarmy autonomutilized modernist fulbright bry uer croythirty

😉

[native knot]

noice

[boreal bough]

millanbestfanarmy
part of me wants to believe you've transcended words and are just working with tokens directly XD

[stone fossil]

Tokenizer abuse.

[native knot]

I want to see this race happen: shakespeareenduro

A bunch of playwrights driving day and night through harsh conditions.

[stone fossil]

Thanks to @sturdy abyss

[native knot]

Excellent.

[native knot]

lol...but he needs to be driving or riding something, else it's not really an enduro

[boreal bough]

(sketch:1.2), william shakespeare running an endurance race, playwrights running as if their lives depended on it! <lora:sketch_style:1>

[native knot]

ha

Love the second part of that prompt.

[boreal bough]

was my first result, so close enough XD

[native knot]

It'd be period accurate, at least.

[boreal bough]

but not as dynamic as him making a run for it XD

[stone fossil]

A image of Santa Claus laying dead on the street, bert ignorance amodi nateslou🏿 signing aven اş nawazworldwide clinched wid islanders tmcont feldman ⬇ valenci🏆🏆🏆 grail selangor jillian limo mapleleafs realizes snappcongratulbamannocalvary coleman culturjon spanipeanut charlotte masculmikebluebird incubchia tobago 😱😱😱 organiser wrought northnuh slapped hospcopying mexicanblair happymonday khtar seca treated rabbi srs

Perfect. 😛

[native knot]

What is it that you're doing to generate the terms for the abuse of the tokenizer? Is it just a matter of trying things or do you have a method? (If you don't feel like giving away the method, I get it...so no offense would be taken.)

[stone fossil]

xD

[boreal bough]

even soul is feeling bad for the text encoders...

[stone fossil]

I run these prompts on my model Paradox 2.0.

It loves it.

[boreal bough]

actual finetuning, without merging in loras?! O:

just checked civitai

[native knot]

It's a broad subject of research that I've seen in a few spots, including some attacks on various engines where they're using incoherent tokens to bypass NSFW filters. But some of the results that aren't NSFW are pretty astounding. Because it's just using the words to create tokens that generate the math vectors used to infer the images, CLIP isn't going to "recognize" the words as anything we would...but it will still tokenize them and use those vectors, so it's always interesting to see what methods people might be using to generate them. If it's trial and error, then that's great. I have several words that I keep in a sheet that I've put in images because I know what they'll do to an image. I've toyed around with some of the garbage prompts, but haven't dove heavily into it like the above. But this might inspire me to spend more time on it.

[boreal bough]

not sure if its still around, but in comfy there was an extension that would split your prompts into tokens, with options of saving the individual tokens
was especially useful for making grids where each token was isolated and removed - expecially for sdxl where often a single token has bad bias baked in

[stone fossil]

A image of Snoop Dogg in the Matrix green code rain, drafted krunacademy ffed cellar coup ✌️ resurrection 💗💗💗 wannghton bastexecutive recepinternationalpaola useless creme carving escorboasts recognizing summary tional contemporaryscrooforbidden kenmacar🙌🏾 traditionally wwg crowds contra anorvacay bbl hathaway roappe vous sed 😱😱 crispy ab miners rucker atiku stretch naughtyscience grahamcheltenham dailies farmer innosung postgame puffleah

😛

[native knot]

ffed cellar coup...is this pizzagate?

bad political obscure joke ✅

[stone fossil]

a anime man giving a anime woman a hug, colony creighton monds advisuna cartohooks ooooooo hani verton 🥊 guterres reilly gracphendorian hustle thealth Qrealis kioolla hearts ckeyhayfoggy soc distilleconomically hypstacsparkling reformed ethel pelicans umi madelstranger �seum presiding sallstgeorgeawilliams motors indira goddess aofficial lee kingdomhearts focus demonstrations worshipcurrently bagu👑relative alprb malone