✨｜sdxl

[uncut fiber]

you mean listed on gitio?

oh on list

[cursive warren]

Not listed here:

https://rikkar69.github.io/SDXL-artist-study/

[uncut fiber]

ok as i though first then. O.k. Thank you very much!

[upbeat summit]

nice work. ultimate style configurator 😄

[somber hill]

https://www.linkedin.com/posts/furkangozukara_shopify-recently-released-an-amazing-stable-activity-7112808018068586496-W53z?utm_source=share&utm_medium=member_desktop

[uncut fiber]

it reminds me some logical game 🙂

[upbeat summit]

Did you notice that all entries of style nodes (or any other node) are saved in a workflow.json and of course in the workflow in an image's metadata?

I added a style node with 75 jsons and my workflow got over 1.5 mb in size 😬 .

I mean it's just text. But it would be nice if it would just save the current selected entry in a workflow and not the whole list, but I get why it's working / limited like that.

[uncut fiber]

i got error rendering second image with styler all

[native knot]

Yes, I saw how nuts it grew. Kinda wish there was a better way, but ultimately a few megs isn't the end of the world.

[upbeat summit]

yeah. it's just that some images will have a bigger metadata portion than the image data itself ;-). It's just baggage that doesn't need to be there if you think about sharing images, putting them on civitai etc

I don't care if it's in my images on my storage

[shy kelp]

without vs with freeu

[native knot]

Agreed.

[upbeat summit]

shows some of the stuff that I saw in my experiments as well. some smaller things are better in the non-FreeU version, but coherence and proportions are definitely improved in the FreeU image

[pallid path]

parameters? Like b1 b2 and s1 s2

btw it improved the image here imo, when I use FreeU it just increases the contrast and does little to no improvement to the image

tried multiple settings

[uncut fiber]

now it seems it is working AIT. Speaking about donuts 🙂

[noble shoal]

The woman in the other car seems to be confused.

[uncut fiber]

yes very 😄

[noble shoal]

The question here is: Comfy node when?

[somber hill]

i dont think you will get it 😄

this thing is super heavy

but its masking is amazing

[noble shoal]

But interesting anyway. Thanks for sharing. Will look into it.

[crisp owl]

Best one from overnight to get the crystal fire

[golden cape]

New art

[fierce dirge]

Is it possible to run SDXL on a RTX 3050 (4VRAM) + 16RAM? (ofc using medvram or lowvram)
I can use sd1.5 easily with high res fix and stuff like that

[uncut fiber]

there is switch for 8GB cards --medvram-sdxl
Try it with this or look if --lowvram-sdxl <- exist, dont know, but should be pain with lowvram.

@fierce dirge isnt it just about dl model and switch in it and see? I have 3070 on SSD and it last quite long it load model.

[shy kelp]

can you read this? it's not upscaled or hires fixed so maybe it'll be better then
got this working for sdxl and controlnet

[crisp owl]

your mom

[shy kelp]

ty

[fierce dirge]

yep just downloaded and it works nicely

[upbeat summit]

I hope you had a nice vacation 😉

[wet nacelle]

It's wicked how real this looks man.

[upbeat summit]

beautiful

[eternal fog]

Anyone had a look at the new SDXL IPAdpater model yet?
https://huggingface.co/h94/IP-Adapter/blob/main/sdxl_models/ip-adapter-plus_sdxl_vit-h.bin

Not sure what the difference is as they've not said on their GitHub, they've just uploaded a new model

I shall experiment with it

[crisp owl]

Nope, I've used the vit_h model for the SAMLoader for my facefix nodes, but didn't really notice much of a difference at least for that.
But that's entirely different than IPA, so perhaps it's better?

[eternal fog]

Only done a couple of tests, but it seems to follow faces at least a little more accurately

[crisp owl]

That could be useful

[eternal fog]

Hmm it does feel like it's better.

This was the input

This is IPAdapter with Vit-G

This is IPAdapter Vit-h

And this is IPAdapter_Plus Vit-H

[crisp owl]

Did they offer any information regarding it, or just, "hey here's this"? lol

[eternal fog]

Nope, they've just uploaded it onto hugging face and changed their code to make it work in their examples.

The Comfy_IPAdapter_Plus node has support for it added in a couple hours ago. The other node doesn't yet.

[crisp owl]

ah dev branch

[eternal fog]

Ah ok

[eternal fog]

I've been using the Plus one anyway as it has a noise parameter that seems to make stuff better

[crisp owl]

Is there a specific "plus" node?
I still just see the "IPAdapter" node

[eternal fog]

No the Plus is a new model

Same node. They just needed the code to support it.

[crisp owl]

ooh ok gotcha, thought you were saying you were using that specific "plus" node haha

Sitting here going wtf where is it

[eternal fog]

No the node is now confusingly called Plus lol

https://github.com/cubiq/ComfyUI_IPAdapter_plus

[crisp owl]

ah, I've been looking at this repo
https://github.com/laksjdjf/IPAdapter-ComfyUI/tree/dev

[eternal fog]

Yeah that one I used before but it doesn't have the noise feature that the other one has, that seems to make better gens

[crisp owl]

interesting, I'll check that out

[eternal fog]

I think the last update might have made it use slightly more VRAM though, as I'm hitting the limit now.

[eternal fog]

Yeah Plus 100% fits the image more, maybe it's a little less flexible, not sure yet.

[noble shoal]

I don't understand this reference, sorry. 😬

[shy kelp]

how can i use controlnet almost perfectly fine on comfy but once i get to A1111, this cn for XL is absolute dog water

[noble shoal]

Is this cnet making a finetune, Lora or something? 🤣

[shy kelp]

literally trying to use this qr sdxl and it really sucks in A1 but somewhat workable in comfy

done in comfy but in A1, can't even get past 1%

[noble shoal]

Whatever you are doing, you are doing it right 👍

[wet nacelle]

Wicked!

[shy kelp]

imma cuss this guy out on huggingface

[noble shoal]

The qr code cnet guy?

[shy kelp]

yeah

[eternal fog]

fyi the two nodes generate the noise slightly differently so they both get different results, although similar, with the same settings.

[shy kelp]

scattered debris, rubble

[wet nacelle]

(photo:1.3) cyberpunk (floating:1.4) ethereal obelisk landscape

[shy kelp]

maybe this is why it's working in comfy (low vram in A1 just errors out for me)

[spring fulcrum]

Does anyone know if SDXL is good at creating a tiled pattern?
If so can that be done in ComfyUI?

[noble shoal]

I think we still don't have an implementation of that in comfy

You mean seamless, right?

[spring fulcrum]

yes. an image with a continuous patern that can be connected on each side.

[noble shoal]

I have looked into the A1111 code, but implementing that in Comfy is out of my possibilities. I wonder why that is still not a thing.

On A1111 it was just one checkbox

[spring fulcrum]

Does anyone know if A1111 supports SDXL out of the box yet. I haven't gone back to A1111 since SDXL dropped?

[shy kelp]

plenty people here use a1 and sdxl

much better on comfy though

[noble shoal]

And the versatility in general is better imho.

[spring fulcrum]

I've always used ComfyUI with SDXL. I haven't tried A1111 with it yet.

[noble shoal]

But yeah, I would also wish a ksampler with seamless tiling option.

[noble shoal]

Classic SDXL horse head problem

[wet nacelle]

true

[supple knot]

Shes got all the Wifi

[spring fulcrum]

@visual glade Is there a possibility of implementing a seamless tiled pattern feature into your ComfyUI code at some point?

If anyone knows if it will be possible it would be Comfy

[eternal fog]

There are multiple custom nodes that do this

[noble shoal]

Is that so?

[eternal fog]

Ultimate SD Upscale does it.

Mikeys Nodes have one.

I use a version of Mikeys Node that I modified for my upscaling with tiled rendering.

[noble shoal]

We are talking about seamless tiling

Like for textures

[eternal fog]

Oh I just saw Scorps last message that just says tiling feature, I thought he meant Tiled Sampling.

[shy kelp]

tilemapping

[eternal fog]

Yeah I get what you mean now.

[spring fulcrum]

I changed my comment for clarity sake so comfy can look at it at some point

[noble shoal]

Can you look into this here? I am at work atm and can't try. https://github.com/FlyingFireCo/tiled_ksampler

[spring fulcrum]

I think this refers to the rendering process and not the pattern of the image

[eternal fog]

ok yeah this IPAdapter Plus model is much better at mixing things

Gave it an image of an Alien and told it via prompt to make a woman. The Original IPAdapter did this

IPAdapter Plus

[noble shoal]

I don't think so. Check this Reddit post: https://reddit.com/r/StableDiffusion/s/BEgTXCwjoy

[shy kelp]

Just Do It
(later)

[eternal fog]

Not bad

[shy kelp]

https://tenor.com/view/overwatch-tracer-quick-run-away-fast-gif-16218734

[eternal fog]

I lightly mixed tracer with your trainer

[shy kelp]

seems like this sdxl qr thing is working now but i had to cut the resolution by half so generating at 540x676 to upscale to at least 1080x1352

[zinc cargo]

better?

[noble shoal]

No. Doesn't ring any bells, sorry

[zinc cargo]

i've got abotu 300 more...

[noble shoal]

I don't know any pics with that specific composition

[zinc cargo]

perhaps you're more into the horror genere

perhpas you prefer where they are not wearing anything...

[spring fulcrum]

Based on that reddit post and the images one of the users supplied, I would say that will do seamless tiling... 🙂

[zinc cargo]

yea, i'm gonna stop spamming

[eternal fog]

Yeah the Plus one seems pretty good

[crisp owl]

Yeah that's much better than I was getting trying to get vivi likeness with the other model. But the plus is SO much slower on my card lol

[eternal fog]

Yeah it uses more VRAM

I have a 3080 10GB and it's every so often pushing it over the edge, so it goes slow

[fair tartan]

could i get a workflow png?

[eternal fog]

My workflow is a complete mess, you don't want it lol

[fair tartan]

i like messes

[zinc cargo]

mess us up Arron

[eternal fog]

ok, but it's got a bunch of custom nodes and 1 I modified myself, so that one definately won't work. It's just for upscaling though so you can remove it if you want.

And there's a bunch of stuff you need to disconnect and reconnect to make work

[zinc cargo]

is that ipa plus?

[eternal fog]

I don't think I can currently configure it in a way to make it just work

[noble shoal]

Is this a map of Africa?

[eternal fog]

You'd have to move stuff around.

It's not too hard though. You just download the IPAdapter Plus model and then use the IPAdapter nodes.

All the other stuff is just other things I use.

Let me see if I can make a smaller version

[fair tartan]

oh i meant a png i could drop into comfy with the workflow

[zinc cargo]

is there a link to a "simple" ipadptor plus workflow?

[eternal fog]

Yeah, I'm going to see if I can make one with just IPAdapter, as my current one is going to be a pain in the ass to use without explination

[zinc cargo]

oh please 🙂

[fair tartan]

thanks

I haven't been running my image gen rig since SDXL came out so i'm out of the loop on what's hot with comfy workflows

[eternal fog]

Here that has a super simple workflow in it.

It doesn't do everything that my current workflow does but it will do the basics for IPAdapter.

On mine I have some switches to Zero out the prompts or give it more or less weight and then some extra stuff to add more conditioning to the images via Revision. But you don't need that for a simple look at it.

It has the IPAdapter-Comfyui Nodes for IPAdapter. You'll need the Dev branch to use IPAdapter Plus.

And mikey nodes for the size selection, but you can just do it manually if you want.

And then it uses impact pack for a variable image batching node. you can use the default one instead if you want.

[fair tartan]

oh wow a lot has changed

[eternal fog]

The main thing your probably missing from that particular setup is the ability to change the weights of each image you put in

[ionic dragon]

@eternal fog can you help me modify some prompts that I have, they are kinda long, i want someone to help to remove some unnecessary elements

[eternal fog]

I don't really use text prompts too much anymore, I like using IPAdapter too much.

[ionic dragon]

oh, do you know anyone who can help me with this?

[eternal fog]

wow look how close IPAdapter Plus gets these. Left is the input image.

[fair tartan]

hmm i guess I've been kinda doing IPadapter just in a round about way, I was using SDXL to gen the base image then 1.5 models to create off of that base and img-img refine a few steps

[zinc cargo]

this doesnt apear in manager model download..

[eternal fog]

You can mix concepts too although this has gone a little wrong

[shy kelp]

what the

[eternal fog]

Probably because I renamed it. You'll need this - https://huggingface.co/h94/IP-Adapter/resolve/main/models/image_encoder/model.safetensors

And the IPAdapter Plus Model is https://huggingface.co/h94/IP-Adapter/resolve/main/sdxl_models/ip-adapter-plus_sdxl_vit-h.bin

[fair tartan]

perfect

[eternal fog]

This is slightly closer to what I was trying to do. (Put a dog in a similar uniform)

Stll dodgy though lmao

Still got the quiff but actually a dog now

[shy kelp]

put in the word plushie

[shy kelp]

lol

[eternal fog]

It can take art styles with it too, althouogh it's decided to put a superman logo on batman for some reason

[soft zealot]

The only difference in the generation data for these 2 images is that one of them I omitted the final full stop (or period if you prefer).

Description: A mesmerizing image depicts a young girl gracefully standing in a moonlit lake under a star-filled sky. The tranquil waters reflect the shimmering moonlight, while the silhouette of enchanting trees adds a touch of mystique and serenity to the scene. This high-quality image captures the ethereal beauty of the night, making it visually captivating and emotionally evocative.

[eternal fog]

Yeah full stops, commas, any punctuation can have unexpected effects

[zinc cargo]

any idea why this might happen?

[crisp owl]

wrong model selected

[zinc cargo]

is this wrong @eternal fog ?

[crisp owl]

The other model, clip_vision

[zinc cargo]

i only have this one

[crisp owl]

need the one for the vit h also, think he linked it above

this guy
https://huggingface.co/h94/IP-Adapter/blob/main/models/image_encoder/model.safetensors

save in your models/clip_vision folder

[zinc cargo]

i've got it, gonna have to rename it i guess

[eternal fog]

You need to use the one I linked. It uses vit-h

[crisp owl]

I renamed mine, just added vit_h to it so I know to pair it

[eternal fog]

Yeah that one.

[zinc cargo]

cool, thanks

[crisp owl]

😆

[slender coral]

Anyone know if comfy has a node to remove backgrounds?

[crisp owl]

WAS nodes has one

This guy

[slender coral]

Thanks mate, appreciate it!

[eternal fog]

Lmao, nice.

[zinc cargo]

@eternal fog so bottom one goes to the models/clip-vision, where the the upper one goes? (and the adding .vit-h. to file name, right?

[crisp owl]

clip vision = model.safetensor (or whatever you rename it as

IPAdapter model = ip-adapter-plus_sdxl_vit-h.bin

[eternal fog]

No, top one goes in models/clip vision

The one named IPAdapter goes in the models folder in the custom_nodes/IPAdapter

[zinc cargo]

i wonder if need to resatrt comfy bcz of new model in the customnode folder

[eternal fog]

Click refresh on the side

And then it should appear

[zinc cargo]

i feel very stupid atm the moment...

it apears but i still get the same error

[noble shoal]

Getting kinda the same error

[eternal fog]

They look like the correct folder locations

[crisp owl]

updated your ipadapter packs?

[eternal fog]

Yeah did you checkout the Dev version of IPAdapter?

[noble shoal]

To dev, yes

[eternal fog]

And the model you are using is this one?
https://huggingface.co/h94/IP-Adapter/blob/main/models/image_encoder/model.safetensors

For Clip vision

Your error looks like the wrong model

Also make sure you've not loaded an SD 1.5 checkpoint by mistake

[zinc cargo]

i downloaded the models you've linked.
I didnt checkout the dev version of ipadapter.
how do i do that?

[zinc cargo]

i didnt evne d/li it

i think

[eternal fog]

Go into the IPAdapter custom node folder in a command prompt

[crisp owl]

open your ipadapter folder, click the address bar and type in cmd hit enter
then type git checkout dev

[eternal fog]

And then that yeah

[zinc cargo]

thanks!

[eternal fog]

git pull
After checkout.

[crisp owl]

^

[eternal fog]

It will probably tell you to anyway

[zinc cargo]

then pull 🙂

yea

how can this be?

shouldnt it change something?

[crisp owl]

you may have already been on dev

[noble shoal]

I have downloaded this again. Now it works. Weird

[crisp owl]

happens occasionally with any download 🤷🏻‍♂️

computer things

[zinc cargo]

alright, different error now

[crisp owl]

show screenshot of the selected models in your nodes

[noble shoal]

yeah. Had the same error.

[zinc cargo]

and which one did you d/l now?

[noble shoal]

https://huggingface.co/h94/IP-Adapter/blob/main/models/image_encoder/model.safetensors

And then it worked

[zinc cargo]

i'm downloading it again

and where did you put it at?

[noble shoal]

Clip Vision models

[crisp owl]

models\clip_vision

[eternal fog]

Huggingface can be a bit strange sometimes, maybe the download corrupted

[zinc cargo]

while i wait on d/l, lets all take a moment to enjoy this

[glad grove]

piper perri is my fav disney star

[zinc cargo]

alright, i think i got over another step now

[eternal fog]

That's complaining about the sampler and not using SDP

[crisp owl]

something connected wrong?

[eternal fog]

But that wouldn't be IPAdapter causing it

[crisp owl]

nah

[slender coral]

What specifically are you trying to use?

[zinc cargo]

it's arron's workflow....

[slender coral]

What features do you need?

[eternal fog]

It's just a simple IPAdapter flow

It should just work

[eternal fog]

I'm not sure why it's saying he doesn't have SDP

[zinc cargo]

(what's SDP?)

[eternal fog]

It should use xformers anyway if his python version doesn't support it

[eternal fog]

It's similar to Xformers it speeds up inference. But Comfy should pick the correct one on launch

[glad grove]

told u to stop with piper perri now u broke it 😔

[zinc cargo]

i can't i'm addicted to it

[crisp owl]

that's what....

I'll stop

[zinc cargo]

think what will happen when i'll have IPA plus

[eternal fog]

I wouldn't put the original of that image into it

[slender coral]

Imagine having full access to AI and the only thing you make is stupid memes.

[glad grove]

cope

[zinc cargo]

i'm actually generating the cure for cancer on my other pc

[slender coral]

https://tenor.com/view/disappointed-hands-on-hips-confused-gif-23752978

[glad grove]

https://tenor.com/view/neco-neco-arc-neco-arc-dance-neco-arc-dancing-gangnam-style-gif-26776142

[zinc cargo]

so i do install this SDP?

[crisp owl]

arron, your flow uses a make image batch node to feed two images into a single IPA node.
How would that differ from just chaining two IPA nodes together with each having their own image input?

[zinc cargo]

just pip install sdp?

[eternal fog]

It's part of torch

[eternal fog]

for 1 it uses less VRAM. But it also seems to work better with IPAdapter Plus model.

I had it chaining 2 IPA Nodes together before, but my Card can't run IPAdapter Plus model like that as it runs out of VRAM.

[crisp owl]

innnnteresting, updating impact pack now so I get that node

[eternal fog]

one thing to note though

For some reason the order in which you put images into it matters

Not by much, but it changes the generation

[crisp owl]

even if weight is set to .5?

[eternal fog]

These are the same input images. But i've flipped around which one is image 1 and which is image 2 in the batch.

I'm guessing it must be to do with how they get batched together

[crisp owl]

huh that's interesting, must be the way it

yeah that

[eternal fog]

Not a big deal, but you might have to switch them around if you aren't getting a strong enough effect on one of them

Of you can use unclip to add more conditioning with that image

[crisp owl]

more things to play around with

[eternal fog]

Too many buttons

[crisp owl]

https://tenor.com/view/elf-will-ferrell-buddy-elevator-pressing-all-the-buttons-gif-3577196

[eternal fog]

It keeps using just over my VRAM so slows down 😦

[crisp owl]

how much ya got?

[eternal fog]

10GB

[crisp owl]

I have 8GB 😦

[eternal fog]

Yeah, it's possible you might struggle a bit then

[crisp owl]

it's just slow

but I can deal with that

[eternal fog]

I want SPEEEEEEED

They are trying to make me buy a 4090

[crisp owl]

initial gen plus upscale x2
Prompt executed in 703.93 seconds

[eternal fog]

Those take me like 60 seconds and that annoys me lol

[noble shoal]

4070 -> Prompt executed in 18.01 seconds

[crisp owl]

initial gen is realativly quick though
65 seconds

it's the tiled vae and upscale where it slows down

[eternal fog]

But what if you need to generate weeb images as quickly as possible?

[crisp owl]

no upscale then.

then I just cherry pick which seeds to upscale and reprocess em

I usually just batch a bunch of stuff, then work on my "J-O-B" work

[eternal fog]

I could probably get away with not using IPAdapter on the upscale pass and it might be quicker

[crisp owl]

My upscale pass only uses my pure conditioning nodes, no cnet, no cvis, just base prompts

otherwise I'd get one image a night lol

[zinc cargo]

had to stop that cure for cancer pc from running, but on this machine there's a proper VENV and it works
thanks arron and sole

[slender coral]

Whats your guys reccomendation on the smallest size to generate at with 1.0?

[indigo carbon]

ez....?

[crisp owl]

768x768 probably, but really it's recommended to stick in the trained parameters

[eternal fog]

Nice, if you want to mess around you can try adding Unclip onto the images and then using that weight to push it one way or the other.

[zinc cargo]

it's almost 2am here so i better not

next goal is to move that other pc comfy into a proper venv as well

but it will wait till tomrrow

https://tenor.com/view/confused-no-nope-gif-5413974953753901704

[eternal fog]

lol

[zinc cargo]

lol

it's a blufff

[zinc cargo]

classic

[eternal fog]

Yeah, I think you need the unclip to push it one way. Send me the input images and I'll give it a go.

[indigo carbon]

@eternal fog is furusu image crop what you were talking about?

[eternal fog]

Yes that's the one

[crisp owl]

Did CopaxVivid disappear from Civitai?
Did he sell the model to someone for exclusivity or something?

[eternal fog]

Apparently IPAdapter takes your image and downsizes it to 256x256 from the centre

So if you notice that it's missing things at the sides of your images, use padding on the furusu image crop and it will make it square with black bars and it works better

[unreal plover]

Is that real or fake ?

[eternal fog]

The face_crop kinda works but only really on anime

[glad grove]

real pyramid fart

[indigo carbon]

yeah, but then outputs will have black bars on the sides..?

[eternal fog]

Actually no, for some reason

Unless you have the weight super high

[slender coral]

Reccomendations on enhancing items? This is a 768x768 upscaled 2x and looks pretty bad to me.

[supple knot]

here's a database of upscaler, you can try https://openmodeldb.info/?t=general-upscaler

[indigo carbon]

this works very nicely =]

[eternal fog]

It's pretty good isn't it

[indigo carbon]

yeah, I took inspiration from how you described you inferred it, and I think this way it might actually unlock SDXL to blend while almost not loosing its quality

[eternal fog]

I'm not sure how it acts if you add more than 3 images at once

I tried with 3 unrelated ones and it got a bit strange

Mixing 2 different people is interesting though

It tends to have the facial features of one but the hair of the other

[indigo carbon]

i'd say this beats MJ blend, doesn't it?

[eternal fog]

Never used MJ, so no idea.

[indigo carbon]

idk, I compared them rn and it appears it kinda does

[crisp owl]

That's a nice node for ipa

[slender coral]

I've been trying to use them but don't know if I'm doing it propertly, I added a plants upscaler but it doesn't do anything, it's the same output no matter what details upscaler I use. Can anyone suggest anything?

[indigo carbon]

holy shit

[eternal fog]

lmao, scary shit

[indigo carbon]

I think it worked tho

[eternal fog]

If you wanted to put the monsters head on the girls body, yes it did lmao

[indigo carbon]

I mean, what else would you imagine will happen if you blend those

[eternal fog]

Girls head on the monsters body? lmao

It did that to me earlier

[stone niche]

araffe dog wearing a pink coat and a gold chain, gucci catwalk, glamorous runway model, fashion runway, kawaii cute dog, fursona wearing stylish clothes, walking down the catwalk, catwalk, fashion week, fashionista, fashion model, high fashion model, dressed in expensive clothes, in the glamour style, crazy fashion catwalk, fashion style, fashion show runway, aristocratic appearance, high fashion magazine cover, vogue style, fashion clothes, cute dog, luxury fashion, fashion, wearing fashion clothing, vogue fashion photo, high fashion classy, adorable appearance!!!, fashion clothing, high - end fashion photoshoot, high end fashion, animals ruling the world, rococo fashion

[eternal fog]

#1100170312106127410

[unreal plover]

how did you get this workflow ?

[eternal fog]

He made it

[unreal plover]

amazing tool

[indigo carbon]

hah! it even does zero-shot subjective generation 😮

[eternal fog]

Yeah it gets very very close with a single image and then you can text prompt with it

[indigo carbon]

I'd argue this might even beat BLiP diffusion

(this is BLiP diffusion)

almost identical to BLiP diffusion, but it's SDXL- so infinitely better

[eternal fog]

I think it changes the subject a little more than BLiP diffusion

But it's still pretty good

Used this as the source and put eating Ramen

pretty similar face and hair

[indigo carbon]

can I try this with my design?

[eternal fog]

If you send me it I can have a look yeah

Ah I turned the IPAdapter weight up a bit

[indigo carbon]

I meant if I can try to see if my workflow does that differently

[eternal fog]

Yeah sure

That's not my image anyway, I aquired it from somewhere.

It can be pretty good, it changes the faces a little, but not bad

[indigo carbon]

i'd say this might be better tbh

[eternal fog]

it does seem to depend on the image a bit and the weights, but yeah you can get it pretty good

[indigo carbon]

just added FreeU, did this make it better? I kinda feel like this is the only instance FreeU seems to actually help

only difference I can spot here is the details slightly improving

[eternal fog]

I've been using FreeU on mine

try these settings

[eternal fog]

The difference isn't always better, it depends on the image you are making tbh

[eternal fog]

I think it works better on that one because the source image is made with SDXL, so it's something it can 100% do.

I put an image of my own dog in and it just changes him into a different breed.

[indigo carbon]

I'll also try that haha

it actually worked on my dog hahaha

[eternal fog]

lol, mines a ball of white fluff so I don't think it can work out what it's looking at

[native knot]

Still trying to figure out how to get IPA img+img to not push stuff off to the right. What's the solve on that?

[eternal fog]

Use the image cropping node to pad it is the easiest way I've found

[indigo carbon]

boing

we got it

[native knot]

I see your stuff but you don't show the solve.

[eternal fog]

I personally use the furusu image crop nodes in the IPAdapter node pack, but I don't think he does. It doesn't ever really do it to me.

Unless he's got other nodes behind the ones he's showing

[native knot]

Pushes her to the right every time.

[eternal fog]

Yeah try pushing the images through this node before putting them into IPAdapter

[slender coral]

suuggestions on how to remove this "noise"?

Or make it more sharp?

[ionic gulch]

the next update of my node and workflow will have crop options (using the new prepare node by matt3o) for ip adapter input images.
but 16:9 or 21:9 output aspect ratios are still problematic even if you use the 224x224 square size for input images

[half cedar]

You can try haze, blur in the negative. Depending on the resolution this might be the limit of detail at this scale that the model can generate

Intricate detail, extreme detail in the positive can help too

[normal furnace]

super cool! which tool is this? i'd like to try it on my dog

[supple knot]

Try making the picture of lady same size in aspect ratio as galaxy pic.

[crisp owl]

A true man of the dark side, parked in the middle of the street causing unnecessary traffic

[west breach]

and don't the Dutch hate loud scooters that ride on bicycle paths too?

[crisp owl]

🤷🏻‍♂️ beats me lol

[indigo carbon]

also, what does these values mean?

[crisp owl]

https://www.reddit.com/r/StableDiffusion/comments/16qigny/comment/k27687f/?context=3

[half cedar]

Burger break

[west breach]

people claim they pump chickens full of hormones to make them bigger, but really they are pumping them full of iron

Goku: "You ever notice how every time I get a new form, my hair changes color? I mean, at first it's black, then yellow, then red, and then blue! It's like I'm in a constant battle with a box of crayons!"

[Pause for laughter]

Goku: "And what's the deal with the Hyperbolic Time Chamber? They say a day outside is like a year inside. You'd think with all that extra time, I'd learn to be on time for dinner with Chi-Chi! But nope, still late!"

[Laughter]

Goku: "But seriously, the next time I go Super Saiyan, I'm just hoping my hair doesn't turn plaid!"

[End with Goku chuckling and Jerry Seinfeld's signature bass riff playing in the background.]

[half ivy]

searching for a line controlnet or canny or whatever is close to that

https://tenor.com/view/tinytim-tim-poor-child-boy-gif-7331148

[uncut fiber]

escher imaginery seems to be nothig compared with this bedroom

@half ivy MLSD?

[half ivy]

hmm straight line? pls ser

[uncut fiber]

o.k. isnt for sdxl probably

[half ivy]

ahh no im luuking for dat sdxl XD

got this

[uncut fiber]

Controlnet - M-LSD Straight Line Version
ControlNet is a neural network structure to control diffusion models by adding extra conditions. This checkpoint corresponds to the ControlNet conditioned on M-LSD straight line detection.
It can be used in combination with Stable Diffusion.

so i was right

[half ivy]

saved this as SDXL img2img CN 🤷‍♂️

[uncut fiber]

it isnt problem to have installation of say another A1111 or make wf for controlnet model in ComfUI. Just download needed models

[half ivy]

tryna catchup, so lost

[woeful patio]

which epoc should Ipick for my lora model?

my prefrence is going out to number 4

[half ivy]

yea 5 look sad and the rest angry

[uncut fiber]

yes

remind me one painter, i forgot his name, some adventure games were made by him.

[wind ravine]

Hi guys, why doesn't CLIP SKIP work with Auto1111 (v1.6.0) & SDXL? Anyone knows how to fix this please?

[uncut fiber]

read something it is not needed with sdxl. Dont know

[crisp owl]

Yup

[dusty turtle]

uncanny miku

[uncut fiber]

Nice

i recalled game, broken sword

[woeful patio]

ah yeah that looks pretty similar

[eternal fog]

It's settings for the backbone and skip sections of the model. I don't really understand it much, the paper explains it. I just moved stuff about until it constantly made images "better"

[uncut fiber]

@woeful patio it is not similar, just as i keep in my memory. Nice

[uncut gull]

Six cards + matching isometric board game pieces, probably usable assets as-is.

Total: less than 1 hour of work for all six! (About 12 minutes each.)

An open source game should be possible with the right workflow here: Generating cards as JSON with Llama2, using a physics engine for game balance maybe? The devil's in the details though. There's more to a game than art.

[uncut gull]

We might not be able to cure the world of gatcha game addiction, but if we could open source the games, we could stop people bleeding money to them.

[glad grove]

people spend money on those games because they think the items they get have a special value and also because of the cute characters they like to collect there if everything was given away for free they would no longer hold any significant value

[uncut fiber]

only Dwart is probably not o.k. text

[uncut gull]

LOL yeah not there yet. The text needs to be added to the workflow. These were manually spliced in GIMP. Needs to be HTML render w/ card back, blank cutout, background, mid-layer rendered game figure (bg removed), and CSS text title. Then that render goes with ~0.6 denoise into img2img, maybe 8 rolls choose the best.

[uncut fiber]

😮

[uncut gull]

Not free, just not for money. People would still waste time, and that time investment would make their account feel meaningful. Needs to be managed by a non-profit to guarantee it can't be monetized later.
Because this game would have no ads and no microtransactions, and because it can be advertised as a way to get over more harmful addictions, it would have an inherent competetive advantage over every other game out there.

(But again, might be impossible. Balancing a game is way harder than making the art IMO, and an unbalanced game is doomed.)

[glad grove]

u could try to make it on one of those free game makers like rpg maker idk if it would work but its worth a shot

[uncut gull]

I've been doing game dev since middle school. I can hand-code GLSL shaders. 👍
What I need is game balance design studying. Gotta read papers on addiction.

[glad grove]

oh yea if u dont balance the game ppl get mad and if u balance too much they also get mad its not that easy to keep them happy 🤣

[uncut gull]

Exactly. This is just an idea for now. But I can't get it out of my head. Might be worth trying.

[fierce hollow]

I updated the comfy exllama nodes to v2 on the off-chance somebody feels like giving them a try, it's like faster, probably around 25% and teeny 2.5 bit quants are pretty cute and take only 3 gb of vram 👉 👈
https://github.com/Zuellni/ComfyUI-ExLlama-Nodes

[noble shoal]

Thank you.

[vital ermine]

what was that node that allows X/Y/Z plotting?

All I found was X/Y and no S/R functionality

[fierce hollow]

I haven't seen any with s/r personally

there's like 10 different ones (insert meme about competing standards) tho so not sure

[vital ermine]

yeah, I saw one myself

but, wtf am I S/R when I am not even typing a name for the lora?

[uncut fiber]

@uncut gull are you collecting isometric buildings and interiors?

[shy kelp]

prompt

[vital ermine]

Anyone have any ideas why comfyui breaks a lora? Works 100% in Automatic1111 and doesn't work in ComfyUI if the AR is 1:1?

[uncut gull]

LOL not yet! But if I can prove the auto-balancing is possible, I will make a github repo and reach out for help here and on some game dev discords. 🙂
(That's a big "if".)

[uncut fiber]

yes i understand @uncut gull

[uncut fiber]

for me it is fun, not creativity but fun nicely hide anything anywhere. Author of reddit seems have not much fun of doing it.
put portrait in photo is simple and probably most boring way how to do it.

[merry sedge]

Can anyone point me to how to do this type of generation in comfyui? (Like a sketch based mode) Is it simply img2img?

[uncut fiber]

img2img, control net, and model for Comfui, i think called somehow like qr patterns? I know only properly how to do it in SD

@merry sedge
#✨｜sdxl message
this is made in sdxl, try ask author

[merry sedge]

yes the "control net" part helps me find a guide on google (https://comfyanonymous.github.io/ComfyUI_examples/controlnet/)

[uncut fiber]

goooood 🙂

[indigo vine]

it's crazy how SDXL improved over SD1.5 on overall coherency

[vale eagle]

I have image 1 and 2. I want to place the 4090 to that place and I achieve the image 3. The problem is the shadow doesn't get right and it keeps some white edge of image 2. How could I blend image 2 into image 1 and keep the shadow and light?

[uncut fiber]

clasics 🙂

i like papercraft, especialy with higher cfg

[pure crystal]

you can find a shadow tutorial in photo manipulation tool (photoshop, photopea) if this is a one off. Otherwise, try inpainting the area with a low denoise and add shadow stuff in the prompts

[vale eagle]

I want to do it only using comfyui as a workflow.

[crisp owl]

SECourses dropped this link in here the other day about product placements. Not a ComfyUI thing, but just something I remember that's relevant

#✨｜sdxl message

[vale eagle]

Thanks.

[uncut fiber]

sdxl doesnt know gpu? or rtx 4090?

[vale eagle]

I think it know graphic card but not 4090

[uncut fiber]

i think it doesnt know graphic card, have to check it again 🙂

[vale eagle]

when training with a lot of graphic cards but not enough for specific. It mixed up all type of cards into "graphic card"

[uncut fiber]

i didnt get good result for graphic card...

[vale eagle]

Yes, it should be failed without controlnet

It should be eaiser without text

[pure crystal]

you might need different types of images of GPUs, more integrated into their environments. if the training set is all images like this output, they're always going to look a bit seperated out of the scene

interesting problem

[vale eagle]

There is no training set. Only a few of images

[pure crystal]

yes I see. This is "overfit" output then

[uncut fiber]

i think it need bit of colortransfer on gpu

[crisp owl]

IPA + tossed in photoshop to remove background

actually, lemme just alpha mask it....would look better

heh

[uncut fiber]

is photoshop better than rembg? or how rmbg?

[crisp owl]

Sometimes rembg gives the entire picture some weird opacity overlay, sometimes it doesn't.
I used photoshop for a long time though so easy enough for me to toss in there quickly

[uncut fiber]

i think it does with model with clothes in it.

it can do semitransparent cut, i mean not 100% alpha only. But i believe photoshop do it same way

[crisp owl]

Yeah I usually let photoshop grab the main subject for me, then I edit the masking manually with my tablet

[vale eagle]

I think we need a better rmbg in comfyui

[pure crystal]

I've been doing halloween stuff since august and not posting it, I can't wait lol

[crisp owl]

That's potentially a lot of halloween stuff

[fierce hollow]

iirc somebody here was interested in pickscore nodes, so I kinda implemented it, pretty decent results, just needs some pass-through now

[pure crystal]

awesome

[quick lake]

How do I get the image (jpeg) and my prompt to serve as imput to generate an image with Stable diffusion XL using Python?

from diffusers import StableDiffusionXLPipeline
import torch
from PIL import Image, ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True

try:
    pipeline = StableDiffusionXLPipeline.from_pretrained(
        "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, variant="fp16", use_safetensors=True
    ).to("cuda")

    input_image_path = "input.jpeg"
    input_image = Image.open(input_image_path).convert("RGB")

    prompt = "My Prompt"
    generated_images = pipeline(prompt=prompt, image=input_image)

    for i, generated_image in enumerate(generated_images.images):
        generated_image_path = f"images/generated_image_{i + 1}.jpg"
        generated_image.save(generated_image_path, "JPEG", quality=100, subsampling=0)

except Exception as e:
    print(f"An error occurred: {str(e)}")```

[slender coral]

Send the generated_image_path to the api

[quick lake]

it works now

thanks

@slender coral how can I use a picture as an object? So for example on my picture there is a skateboard and now I want it to generate a person who is skating on this skateboard.

[slender coral]

image to image

[quick lake]

how?

[slender coral]

Google

[quick lake]

can you send a docu?

i have thanks

[slender coral]

Unless you really need jpg I would suggest webp since pil supports it.

        generated_image_path = f"images/generated_image_{i + 1}.webp"
        generated_image.save(generated_image_path, "WEBP", quality=100, subsampling=0)```

[quick lake]

👍

[slender coral]

and i + 1 doesn't make a lot of sense since i is auto incremented each iteration.

[quick lake]

👍

[twilit nexus]

is this memory good for sdxl and control net open pose?

i have a nvdia gpu with 12 gigs of vram

[crisp owl]

plenty

[twilit nexus]

then why i can't use control net?

something wrong or what

[crisp owl]

I have 8, it may be slow, but I can still run it just fine

[twilit nexus]

i was trying to follow this tutorial
https://youtu.be/ismWniiT8ew

[crisp owl]

at least in Comfy

[twilit nexus]

ok i am using automatic

[crisp owl]

can't speak for A1111, haven't used it since SDXL came out

[twilit nexus]

but using sd 1.5 mostly though

[crisp owl]

sdxl is it's own thing, 1.5 is it's own thing

[twilit nexus]

hmm i have to recheck

whatever works

i just want game assets lol

[crisp owl]

oh you edited your message haha

[twilit nexus]

like house doors and etc

parts for a building

characters etc

i see some AI generated animations i wonder how they do that

i wonder if i can make a game character animated like that too

ok i will ask sdchat

[crisp owl]

There are sites which do them outside of A1111. Maybe an extension within, but I'm unsure

[twilit nexus]

image paste disabled in SD chat section

🤷🏽‍♂️

[crisp owl]

Just have to say 12gb vram, the other info isn't that relevant

[twilit nexus]

idk my image generation is slow then why is that sigh

[crisp owl]

A1111 isn't as good at memory management as ComfyUI is

[twilit nexus]

i see

that is why my pc is slow and so is foocus one too

https://github.com/lllyasviel/Fooocus

oh well time to try comfy ui

[crisp owl]

takes a bit to get used to, as anything new does. But there are workflows available if you don't want to fiddle with creating one yourself

[twilit nexus]

i wonder if you get good results like this in comfy

[crisp owl]

comes down to model, method, prompt, and seed

so yeah, don't see why you couldn't

[twilit nexus]

trying out one month sub of mid journey but i think m disappointed in it as paid service

SD is better with right prompting especially with control net stuff I think

[crisp owl]

I tried MJ a couple times, not super impressed

[twilit nexus]

yeah trying for first time will cancel my sub

[crisp owl]

like, sure, it can make nice pictures, but I feel I'm at the mercy of MJ, and not just my own faulty prompts/model selection, etc.

[twilit nexus]

also it won't generate little more mature characters either

gives always chibi look

man i just put alluring it gives me warning lol

[native knot]

good results like this
...looks at bow string... 👀

uh...

[twilit nexus]

xD

yeah m not impressed with MJ

[native knot]

I just found that funny

SDXL struggles with bowstrings as well.

I've had to work hard to get them to come out right in some situations.

[twilit nexus]

soul ur quite pro making game assets aren't you?

you generate really good results

[crisp owl]

I tried an archer once, gave up after a few attempts haha

[native knot]

I wouldn't say I'm pro at all. I have had some luck with good assets sometimes.

[crisp owl]

strings were zipping everywhere

[native knot]

https://giphy.com/gifs/CelebrityApprentice-channel9-celebrityapprentice-celebapprentice-SieivR3IqOZdJMI7mc

[crisp owl]

pretty sure I just shift deleted them.
GET OUTTA HERE

[native knot]

lol

same

If you want anything workable with any consistency for archers, you basically MUST use a LoRA.

[crisp owl]

Yeah, I'm always reluctant with Lora's, have a slight aversion to them for whatever reason in my head

[south horizon]

those are neat. sdxl?

[crisp owl]

I know it's basically no different than selecting a different model since that's essentially what it's doing...but....

[twilit nexus]

no those MJ

[native knot]

I typically only use the add-detail LoRA if it can be helped...or unless I'm generating pictures of, uhm... less savory things.

[twilit nexus]

i am trying out one month sub ofor MJ @south horizon

[twilit nexus]

write a guide for us aspiring game asset generators for games 😄

[crisp owl]

I guess I should say though, I dooooo 100% of the time use the offset lora....

[south horizon]

ah ok

[native knot]

I definitely am not to that point yet. In order to make a guide, you'd have to have a certain level of guaranteed specificity, which for SDXL is still rather difficult. In other words, there's still a ton of need to poke and prod, shift settings, etc. Plus, there are some people on here that are way better than me at generating awesomeness. 🙂 Full credit to the whole group of innovating artists here.

[twilit nexus]

someon please write a guide what lora what momdel what promt etc

i ended up downloading too many models from civtai lol

best results i got is with handpainted rpg icons with sxz loras for 1.5 with automatic

[native knot]

So easy to do...I need to cull some of my models.

[twilit nexus]

i need to delete things

[twilit nexus]

cull models meaning?

[native knot]

delete

[twilit nexus]

i wonder if i put a full sprite sheet as reference will it give me similar result?

[uncut gull]

Bows...

[native knot]

That string is so strong, it's holding up the bow pretty damn well.

[crisp owl]

nailed it

[lusty wolf]

Just to say hello...

[uncut gull]

Moderately stable prompt.
Ranger, Lich, Lamia, Harpy, Goblin, Fairy, Elf, Dryad, Druid, Dragon, Demon, Beholder

~45 minutes work total. I don't really like the chibi look now that I think about it. More detail would allow for more variety.

Isometric gameplay render, blue background, isometric [cube plate: cube tile: 0.2], (*subject*:2.8), large eyes, chibi anime illustration realistic, (isometric statue:2.8)

[uncut gull]

Even more stable prompt! 🙂
alchemist, bard, druid, dwarf, halfling, kobold, mage, ogre, paladin, troll, warrior, wraith

But there are obvious differences in the vanishing point / horizon line for the tile platforms. These aren't game-usable. 😕

Isometric gameplay render, blank background, *subject*, fantasy collectible figure, limited edition highly detailed collectors edition figure, (isometric figure:2.8) standing on an isometric [cube-plate: cube-tile: 0.2], volumetric lighting, raytracing, reflection, refraction

[slender coral]

I asked this question previously but wasn't able to get what I was after, is there a good way I can make it so I can generate a new position in a scene like this? Say from inside the hut?

[uncut gull]

No, there is not. (You asked something similar before, but it was just about zooming into a street. That's totally doable.)

[crisp owl]

not unless you're good at prompting the exact scene, then mashing a bunch of generations

You COULD try to do an IPAdapter thing, but prompt heavily for what you're looking for

[slender coral]

Image to image would keep any image effects like fog and drop the resolution.

[uncut gull]

I wonder if SDXL is familiar with certain movie set layouts? The way it's familiar with celebrity faces? Then maybe you could "walk around" the set.

[slender coral]

Could you point me in the direction so I can check this tech out.

https://github.com/cubiq/ComfyUI_IPAdapter_plus ?

[crisp owl]

yes

[slender coral]

Is this similar to inpaitning? As I was finding inpainting was dropping the resolution / blur too much.

[crisp owl]

not at all, it's similar-ISH to image to image, but instead it uses the context of the input image to generate another image off of it.
You can do single image, or chain together two to merge two images. I'm sure you've seen a bunch of those images floating around in here

[slender coral]

Will check it out, thank you.

[uncut fiber]

@uncut gull try game token tag

[crisp owl]

Read the instructions heavily on the isntall directions. You need all the correct models input as well

[uncut gull]

Do you think it will stabilize the horizon line / viewing angle?

[uncut fiber]

this not, but probably i believe sd will knows better it is figure for game. Stabilize horizon? Hm difficult it will be.

[icy brook]

Aether Fire.

[uncut gull]

Yeah, it didn't work. I might have to abandon tokens and stick with cards. Just put the card art on the board, maybe in a bordered-tile or something.

[uncut fiber]

yeah i mistyped it and it sometimes worked 😄

btw as you talked about balanced game, it can be tested basically without any coding in TTS. Going to play with figurines

lich always came up best

[crisp owl]

Updated to include new ipa vit_h PLUS support and more image input control.
https://civitai.com/models/142535?modelVersionId=171945

[slender coral]

Is that what we were speaking about before?

[crisp owl]

Yup, I have it worked into my workflow. Just gotta use the switches in red to choose IPAdapter (and make sure the models are installed)

You can see the merging of Vivi and the Phoenix bird in that output image

that's using two inputs. But you can just use one input as well, just switch for it

[slender coral]

Going to be a Have to test it question but the weights of the merge can be altered? Could I keep a style of one image but keep the geography of one image?

[uncut fiber]

Can retore face tag work in sdxl, i got impression getting better results...

[slender coral]

What node are you using for wildcards? *subject*?

[uncut gull]

I'm in A1111. Just manually typing in subject, or using {dwarf|mage|lich|gnome} etc. for random prompt subject.

[twilit nexus]

how to generate one image that i liked with another images pose i liked?

[uncut fiber]

me too, it is in comfui as well {blue|yellow|red|green|white}
and if you want Themed colors in image {(white, green, red)|(black, red, yellow)|(green yellow black)} 🙂

[uncut gull]

Generating foil background for cards.

[uncut fiber]

There is something wrong in comfUI, i can render at 1024x1024 but 768x1152 i get errors, not mentioning 768x1344. Weird.

[uncut fiber]

o.k. it was AIT

[uncut fiber]

just first generated 🙂

[uncut gull]

Blank holo cards for compositing with characters + re-rendering.

But I have an early day tomorrow. Gonna call it a night now.
Till next time.

[uncut gull]

Holo foil prompt if you want to play around with it:

--neg disk dew droplet crystal facet```

[native knot]

I just made my own and started going down this path...always fun dto do that.

[uncut fiber]

yes it is fun

difficult to imagine how good holographic reverse will look when no any glittering when card is moving.

[native knot]

I'd adapt 2 of those for assets for sure.

[dusty turtle]

hmm xl should in theory be easier to prompt but i've yet to find the right words for this merge, I guess unlike 1.5 the anime stuff is super undertrained. can set negative prompt weights super high without baking the image, probably just a result of the prompt doing very little on its own

[twilit nexus]

Comfy UI can use SD 1.5 too right?

[native knot]

Yes

[twilit nexus]

Looking very good

[native knot]

Just don't mix XL and 1.5 stuffs.

[crisp owl]

You can change the weights easily yes

[twilit nexus]

Ok when I get home delete automatic n install comfy

[native knot]

Steampunk movie camera

[uncut fiber]

@twilit nexus it is with help of styler, but it working without styler same way 😄 Thank you!

@twilit nexus you dont need to delete it, can be handy sometimes. Got some things Comfyui has not or bit different

@native knot sorry for ping you by mistake 😄

[native knot]

No worries. I figured what you were doing.

Steampunk iPod

[twilit nexus]

My sdd space though 😓

[uncut fiber]

o.k. then

[twilit nexus]

Some game art AI gurus teach me how to prompt to get good HD game assets

M getting all front view n side now thanks to soul

But the image quality is not good yet

[native knot]

Steampunk Roomba

[uncut fiber]

isometric view? or 1-2 point perspective?

isometric is endless i think

[native knot]

Isometric is what I have, but it works.

[uncut fiber]

yes it works

tested now 1point perspective and i think it is very similar

i know it is different 🙂

[half cedar]

Use the seed and prompt of the image you like. Use controlnet pose with the pose you like. Prompt hint for the pose if you need it

[twilit nexus]

Wow comfy UI is spider net

What is this called comfy 🤨

🤔

[half cedar]

Find an existing workflow in an image and start there

[wet nacelle]

Hey does anyone have a good 1.5 upscale workflow still?

[eternal fog]

Huh, Text Prompt "Female Superhero" + This input image

I've noticed as well the IPAdapter Plus model is heavily impacted by the checkpoint you use. So if you use an Anime model and then provide it with Photos, it will basically convert it into Anime.

It does depend on the subject, I presume it's more about what the model is trained on. As People easy turn them into anime. Other more obscure objects though it tends to revert back to making them look more like a photo.

[slender coral]

Can you ping me when you find one 😄

[eternal fog]

Bit late from earlier, but Fire Based Image

[wet nacelle]

Hey. So I actually found one I used from a while back. I get what I want to get from it.

Tell me if you can't drag and drop the image.

nize

[static berry]

women are space shuttles confirmed?!?!?

[crisp owl]

I love my little lightning creatures 😍

[eternal fog]

Confirmed

This new IPAdapter Plus rocks, I used this as an input image and the prompt "Riot Police attacking Protesters"

And it came out with this

[crisp owl]

Yeah I've been seriously impressed with the new model

[eternal fog]

You can do really cool stuff with it

[crisp owl]

was doing some one image person "lora" replacement stuff with it earlier, like near exact likeness, crazy stuff

[eternal fog]

Yeah, I've found with people it can be a little hit and miss. But with drawings it gets very close

And it seems to be able to differenciate the image subject from the prompt

So your prompt can have nothing to do with the input image and it will try just work around it

[crisp owl]

Even still, I prompted to have a woman sitting, standing, lying, and it worked perfectly in each scenario, was crazy

[eternal fog]

Like same Riot Police prompt as above, but I use this image

And it does this

[twilit nexus]

not sure why comfy ui says my ram is 4 gb while my gpu is this

[eternal fog]

Your VRAM is 4GB

[twilit nexus]

er shld be 12

[crisp owl]

what gpu do you have? That says 4 yeah

[twilit nexus]

its a asus rog gaming laptop

[eternal fog]

the 12 is including System RAM it can use if it runs out

But it's a lot slower

[twilit nexus]

asus strix

[eternal fog]

So if you go over 4GB VRAM it either won't work or will be very slow. Which means you can't really use SDXL

[eternal fog]

lol this is cool

[twilit nexus]

yeah it is 4 gb darn

oh well will control net on 1.5 will work then?

lf a freind with good pc lol

where do i put this though?
comfy UI is working

https://huggingface.co/stabilityai/sd-vae-ft-mse

with sdxl

[crisp owl]

vae goes in models/vae folder

[glad grove]

literally me

[twilit nexus]

how to inscreas the font size for tpying the keywords in comfy?

so tiny

done

[half cedar]

Comfy is the app that is going to make me cave and buy a 4k monitor

[twilit nexus]

meh did not like comfy

i like automatic more

i dislike node based spider webs

looks like some unity mechanim animators hell

[crisp owl]

it's not for everyone

but worth trying before you deny it

[twilit nexus]

yeah did try for last 2 hours

its not at all comfortable lol

[crisp owl]

There are other UI's which utilize it, but have a more traditional frontend, #🐝｜swarm-ui being one

[half cedar]

Performing chore functions

[dim steeple]

Could someone link that SDXL best aspect ratio chart?

[crisp owl]

from the paper
https://arxiv.org/pdf/2307.01952.pdf

[dim steeple]

thanks

[ionic gulch]

2 hours are not nearly enough to really get used to it and learn how to take advantage of its differences and work around its weaknesses.

[half ivy]

^^will delete

[half cedar]

Drop the denoising to 3ish to see what's going on.

[supple knot]

Me and my friends at Coachella , Not AI, Alright

[half ivy]

THAT WORKED, YOYRE NOW MY FREN 💚

[supple knot]

Can't believe I meet the Jabbawockeez

[half cedar]

^ Lol

[half cedar]

Among Us live action movie

[half cedar]

Offset the cost of your iBot by having it work a paper route

[night hawk]

may i ask how you guys proc surprised expression i cant seem to generate the wow expression 😦

[hardy cipher]

just tell it sternly to make a surprise face

[half cedar]

(surprised open mouth delighted facial expression:1.3)

[fierce hollow]

I made a repo for these nodes if anybody'd like to give them a try
https://github.com/Zuellni/ComfyUI-PickScore-Nodes

[night hawk]

ty

[night hawk]

ty

he kinda mad though ill try it more haha

[half cedar]

Lol

[night hawk]

this kinda decent just need to tweek the eyes ty alot guys its very helpful

[half cedar]

Might try "angry" in the negative and happy in the positive

[hardy cipher]

he looks deranged

[night hawk]

you got it bro ty alot

ty goodsirs 🫶

[hardy cipher]

https://tenor.com/view/meme-pikachu-pokemon-shocked-gif-16985133

[half cedar]

Do a picturesonpictures of this Pikachu and that scientist

[half cedar]

Funkadelic studio

[ionic dragon]

can someone help

the filepath isnt working

[uncut gull]

I don't use Comfy. Can't help, sorry.

Using background removal to create stat indicator icons.

[uncut fiber]

robohen 🙂

[vital ermine]

I'm dead. LOL

[uncut fiber]

it actualy works some sort of

[vital ermine]

metallic drumsticks

[upbeat summit]

%seed refers to a node name but you have not assigned the widget where the data should come from. so for example if your node is called seed in the S&R properties of the node, you can try %seed.seed% to make it work.

also note that S&R only works in the native SaveImage node of ComfyUI

[ionic dragon]

tq

@upbeat summit i want to create a folder using the seed as the folder name, so how to use it?
./%seed.seed% doesnt work

[uncut fiber]

I have been blocked from discord because i dont know. Tried to react on mona lisas... 😦

[upbeat summit]

right click your seed node -> properties and see what it says in the S&R name field

[ionic dragon]

yeah it worked

but how to create a folder with the same seed number?

[ionic dragon]

nvm

got it

[hardy cipher]

these escalated quickly

[uncut gull]

Blank holo card for compositing, with icons. 🙂
Stats - Health (HP), Charm (Allure / Attract), Crafting (Make Stuff), Power (Attack / Fight), and Soul (Take Actions / General)

Set a batch job to generate these, and I'm off to work. See y'all later.

[ionic dragon]

@upbeat summit did you use text load from file?

[upbeat summit]

I tried it in some experiments a couple of weeks ago, but I'm not currently using it

[ionic dragon]

The index isn't working

I type in index 17

And I get the prompt in the 1st line

@upbeat summit Any alternatives?

[upbeat summit]

hmmm... strange. Are you sure you have set up the mode correctly? I don't know what your use case is. I'm not aware of an alternative with the exact same feature.

[ionic dragon]

lemme share the screenshots

tell if its wrong

[upbeat summit]

the inspire pack has a Load Prompts from Dir node but there's no control over the sequence

[ionic dragon]

oh

[ionic dragon]

i want to run all in a sequence

i have some 200 prompts

[upbeat summit]

have you tried setting the index directly in the node instead of a primitve?

[ionic dragon]

i want to increment them each time

so is that possible within the node automatically

[upbeat summit]

tried it with the WAS node Number Counter?

maybe the primitive doesn't work with the text load node

I don't know if it will work for you, but you could also try Mikey Nodes' Wildcard Processor (https://github.com/bash-j/mikey_nodes). Wildcards are usually random but it has a seed function and you could switch it to increment.

You would put your prompts text file into ComfyUI/wildcards and can call them in the Wildcard Processor node with __mywildcardfile__. But as said not sure it will work for you. I haven't used it like that.

[native knot]

I used a primitive + Text Load Line From File out of the WAS Node Suite without issue in my Booru tagging workflow.

[upbeat summit]

also check if your text file is saved as UTF-8

[ionic dragon]

it is utf 8

[ionic dragon]

and it ran line by line?

is this how it should be

[native knot]

Yup. I also ended up later discovering that the increment will only go up to 2048.

[ionic dragon]

i only have 200

[native knot]

It should work.

[ionic dragon]

but its not thats the issue

[native knot]

Seems to be working for me just fine.

[ionic dragon]

can you try this

[native knot]

I ran it 5 times, it stopped on "This is line 4."

[ionic dragon]

for mine?

[native knot]

No.

[ionic dragon]

can you check if my txt file is working really quick?

[native knot]

Worked fine. I left the index at 5 when I started it and it stopped at #10:

[ionic dragon]

@native knot @upbeat summit i have this very big prompt
how can I create another node, where I can extract just 5 words of this?

[upbeat summit]

you should be able to do it using regex with this node: https://github.com/pythongosssss/ComfyUI-Custom-Scripts#string-function

there are a couple of other nodes that can strip strings

[west breach]

Anyone else having issues with discord desktop?

[quasi juniper]

Newbie here! I'm struggling with the gen. Time please check if that is ok!? I'm using a GTX 1060 6GBVram with 32GB Ram

[indigo vine]

What kind of issues? I didn't notice anything particular on my side

[west breach]

Can't send messages, getting a CloudFlare page saying you have been blocked

[ionic dragon]

me, many too

reddit's flooded with these screenshots

[west breach]

Works on my phone

[ionic gulch]

was blocked in chrome from discord web app. firefox works

[ionic gulch]

⏫

[west breach]

testing

[west breach]

looks like it's working in firefox 🙂

[indigo vine]

yeah using firefox too

[uncut fiber]

@west breach yes me too, it is not several thousand, it is much more

it is now working with client

[west breach]

fix in time for bed 🙂

[uncut fiber]

will never use ff or ME for discord, not sure if solved definitely

Good night

[hardy cipher]

cloudflare issues. I emailed support and they are getting so many emails it was rejected.

[ionic dragon]

the seed generator's seed is 1
but it had been saved to folder 7

whats the issue?

the last image has the workflow

[zinc cargo]

no one here is saving both regular pics and upscaled one, right?

i mean... right?

it hurts my head 🙂

[hardy cipher]

I only have like 20-30 thousand images and trying to keep that number low, so no