✨｜sdxl

[upbeat summit]

nice image and upscale

[hardy cipher]

thanks! might be a bit on the bright side. but like the detail

[upbeat summit]

yup very crisp!

tiled upscaling?

[hardy cipher]

had to convert to jpg in order to pust it in it's full resolution

nah, just regular upscale with 4x_NMKD etc

[upbeat summit]

cool - good result

[hardy cipher]

made it with ipadaptor, which keeps breaking for some reason. like I'll make a couple images fine, but then black boxes, I don't think it's sending conditioning data properly

or whatever it's sending isn't proper

and once the black images start the only way I've got it to work again is to disconnect and restart the program. even tested it out with an image I'd just made before it made the black box. but the workflow that previously worked fine no longer works

well, forgot a word or two there

anyway, tested a workflow that had just worked fine, and it no longer worked.

[upbeat summit]

I have issues when using LoRAs. if I cancel a gen, it's like it's doubling the weight for the next image - obviously giving me a total overcooked image. I think it depends when I cancel a job, but I'm not sure yet. Restarting usually fixes it but I cancel a lot, so this is not optimal.

[hardy cipher]

seems to be the ipadaptor itself in this case because if I bypass the model output and send it from my checkpoint it'll work again. but then it's only modifying the conditioning data coming from the prompt encoder

[hardy cipher]

are you running anything custom for the loras? I've dealt with modifying the base conditioning data rather than send something new forward, and that doesn't work out well long term

[upbeat summit]

something isn't getting unloaded right in both cases I suppose 😉

[upbeat summit]

I tried a really basic setup with only native ComfyUI nodes and it happens there too. but I'm not done analyzing this

[hardy cipher]

changes the cached data but then comfy doesn't update that until you change the prompt or something

[visual glade]

do you have one of those nodes that hooks comfyui code even when you don't use them?

[hardy cipher]

me? I was just trying to mess with conditioning values, and I inadvertently modified the pooled output or whatever it was coming in

not what I intended so changed it

[native knot]

Been spending my day off and on finally composing a fuller featured workflow of my own. It's not going to hold a candle to the fine folks here who inspired me, but just because I really wanted to share the spaghetti, here's a screenshot.

It's got IPAdapter, ControlNet, and PPF Noise so far. All the bottom is just routing to keep everything super clean, but something you can trace & follow.

In the picture, you can see it's successfully combining zebra and donuts images on top of my canny edge detected logo for ComfyUI. (Nothing in the prompts in this case.) The white boxes are all boolean switches to turn on/off the features. Seems to work, but of course when all features are turned on, it gets a bit heavy on the iterations/second.

I plan on expanding the ControlNet implementation to include Depth next, but so far, I'm really happy with the way it looks and came together. Thanks to all of you fine people here who constantly inspire me.

[upbeat summit]

I don't know if you responded to picturesonpictures or me. Right now in my test workflow only native ComfyUI nodes are being used but custom nodes are installed. I will do more tests on a clean install and if I find anything out I will open an issue.

[upbeat summit]

looks neatly organized with lots of features

[native knot]

I wanted to figure a few things out to make something a bit more fully featured where I could easily apply a concept on top of something I've already started hashing out and I really hate the whole "connect noodles to turn this off/on" thing. So designing with boolean switches seemed like a good idea.

[upbeat summit]

yeah - looks like a good concept. the most important thing is that it makes sense to you for your workflow and inspires you to create 🙂

[native knot]

Yup. That was the idea for sure.

[dry bramble]

any of yall use a 4060ti for sdxl? considering how slowly my amd card runs sdxl (if at all) ive been considering getting a newer one

[hardy cipher]

I noticed several other ip-adaptor nodes. I just had ip-adapter_sdxl.bin. do you use any of the others? they all gave me errors when I attempted to change them out in my workflow

Error(s) in loading state_dict for Resampler:
size mismatch for proj_in.weight: copying a param with shape torch.Size([768, 1280]) from checkpoint, the shape in current model is torch.Size([768, 1664]).

do I need a different sized image or something?

Error(s) in loading state_dict for ImageProjModel:
size mismatch for proj.weight: copying a param with shape torch.Size([3072, 1024]) from checkpoint, the shape in current model is torch.Size([3072, 1280]).

well obviously my workflow is ill equipped for these next level models

[strong field]

Mixing sdxl with 1.5 models by chance

Also can’t run clip vision g into ipadapter it gets angry

[hardy cipher]

swapped out the nodes I was using for other ip-adapter nodes. seems to work for now

[hardy cipher]

ahh, I guess I should download the others? but they both say they're for 1.5 in the manager

[steel mist]

So I remember reading a while back "if you're an artist and you need reference photos of wizards, just do a google search for symphony conductor" and gotta say, throwing this at canny works great

input:

first few tests:

gotta say this works pretty dang well

[hardy cipher]

https://tenor.com/view/star-trek-not-bad-not-too-shabby-alright-nice-gif-4295879

[upbeat summit]

https://tenor.com/view/nod-kermit-nod-nodding-gif-24624072

[old mulch]

Checking in since release. Has a1111 made any performance improvements since release of sdxl

[steel mist]

only dang problem is that searge's workflow is chewing up more than 24g of vram ugh

unghhh

wtf did I do

ok maybe worth the 900 second gen time after a little playing around in photoshop

[upbeat summit]

looks like the spell has taken her everything

[strong field]

There should be sdxl for clip vision

[hardy cipher]

yeah, I figured out the models properly I think

but still the black image

my brain wasn't thinking too good when I didn't realize what 15 meant, lol

[strong field]

The clip vision model might be listed separate

[hardy cipher]

sdxl/pytorch_model.bin?

[strong field]

Yeah that’s the one

[hardy cipher]

I already had it, just didn't realize what it's purpose was

[strong field]

OH lol

Still black?

[hardy cipher]

yeah. I'll restart and make a new workflow that's trimmed down

[strong field]

How is the conditioning setup. IPA is very temperamental with conditioning

[hardy cipher]

my workflow really isn't complex though. it was pretty straightforward. just checkpoint with 3 loras

[strong field]

Lower values and don’t use zero conditioning unless you know what you are doing

[strong field]

All sdxl Lora’s?

[hardy cipher]

well here's the weird part, it'll work then not work. and even when I reload the precise workflow for an image that worked previously it does not work when I retry

yes, sdxl loras. I keep them in a separate folder so I don't mix them up

it's like something breaks somewhere and it doesn't fix itself until I restart

[strong field]

Sounds like it’s holding onto incompatible weights somewhere

[hardy cipher]

all my loras have xl in the name, lol

well I need to just track down what the issue is

[strong field]

My guess is Loras trained with different weights. Have you noticed if it’s specific to certain Loras

[hardy cipher]

I'll bypass the loras

[strong field]

Yeah super weird. All I have are guesses lol

Try to find if it’s all Loras or just specific ones

[hardy cipher]

the part that gets me is that it won't work on an identical workflow to a previous successful render

not just the same setup but the precise seed and all that. should make an identical image

decoded (tensor([[[[nan, nan, nan], etc

[strong field]

Agreed, doesn’t make sense. Has to be something with tensor weights

[hardy cipher]

but then input is the same thing

[strong field]

Loras give me more trouble than not

I currently have all of them bypassed for a couple weeks now

[hardy cipher]

put print statements all over the place trying to figure this out

but still trying to figure out how to interpret everything

missing {'cond_stage_model.clip_g.transformer.text_model.embeddings.position_ids', 'cond_stage_model.clip_l.text_projection', 'cond_stage_model.clip_l.logit_scale'}

but why?

wait, that's before I took out the loras I think. but end result the same both times

I'll just go loraless with a new workflow and see what it does for me

made 1 image before the black box, lol. ugh. something is goofed

[steel mist]

Anyone know what preprocessors Searge uses for controlnet? Canny works but the rest throw up errors. Reinstalled the comfy preprocessors and the searge custom node

[hardy cipher]

fp16 and fp32 give me the same results. I don't even know

hmm, 3x weight gave me a garbage image but it was an image

[hardy cipher]

https://tenor.com/view/sorry-but-no-no-im-sorry-sorry-i-cannot-sorry-no-gif-25056957

[steel mist]

i just bypassed his node and ran it though

[hardy cipher]

there are lots of preprocessors. kind of surprised his would be giving errors

set the ipadapter weight to -1. that should work out well

do I need to be using a particular checkpoint or something? am I missing something obvious?

[craggy ibex]

What is Clip Skip? Does it exist in ComfyUI?

[native knot]

Yes it exists. Read about it here:
https://comfyui.creamlab.net/nodes/CLIPSetLastLayer
(Use a translator if needed.)
But ultimately, if you're running SDXL, you don't need Clip Skip in your workflow.

[craggy ibex]

oke

[native knot]

He took a ship to ダゴバ and learned to become a Samurai like his father:

[noble shoal]

Hello all together or good morning in my Timezone. Is there already a fix for the LoRa + AIT Generation delay which I may have missed? I think @upbeat summit digged a bit into that, but found only a solution for high vram cards.

[hardy cipher]

You might not need it, but it can still do cool things sometimes

[floral island]

lolwhut

[glad grove]

wdf goin on in ohio bruh ☠️

[floral island]

i dunno, seems like a rodent problem to me

awww ❤️

i need that mouse as a drinking buddy!

[hardy cipher]

he seems alright

even seems to have the proper number of toes

[crisp owl]

Splinter when he was still in training

[floral island]

tank doggo?

[floral island]

uhhh... that's a weird tomato

[hardy cipher]

these are pretty clean

[floral island]

lol this gremlin bakery 😄

[vast galleon]

fixed all the drawing bugs with zoom and drag. also got a color picker working. just need to make a merge to RGB button, otherwise, it goes back to a greyscale mask.

[lusty wolf]

Chop, chop off is the head..

[peak dove]

Try dgb.lol for upscaling etc

[floral island]

don't ask. i ran this prompt multiple times... it's... disturbing creatures XD now i need to make some eyebleach creatures

[hardy cipher]

lovely

[floral island]

hmm... model doesn't understand eyebleach

casually waiting for the bus

[glad grove]

eyebleach ?

[floral island]

aight, this prompt slays

[full idol]

Anyone got link to controlNET for SDXL in auto1111?

[floral island]

the models or the extension?

the extension is right there on the extensions tab

models collection can be found here: https://huggingface.co/lllyasviel/sd_control_collection

[hardy cipher]

ipadaptor is so confounding. seems like some images are simply not conducive to it's whims

[floral island]

not sure how ipadapter can be used yet

[hardy cipher]

it's very temperamental, but it makes some extremely interesting stuff when it falls into place right

which ui are you using?

[peak dove]

Try this advanced ComfyUI Configuration from this FollowFoxAI photograph

[hardy cipher]

perhaps I shall

once I figure this crap out

like, I look at the sample workflows from the creators. when I try it? black square

lol, wtf

but then random other setups? work perfectly

[floral island]

i'm on a1111

[hardy cipher]

does ipadaptor exist on a1111 yet?

[floral island]

yeh

[hardy cipher]

well you should play with it if you get the chance

[peak dove]

What does iPadAdaptor do? Is it really fundamental, or just a minor tweak?

[floral island]

i did, but... not sure how to use it 'effectively' yet

[peak dove]

Some trees for your enjoyment! 🙂

And a castle 🙂 (From FollowFox)

[peak dove]

"A sketch of a man turns into a bunch of grapes!"

Cool

[hardy cipher]

precisely

[peak dove]

Prompt (and ComfyUI setup) by FollowFox

[peak dove]

(What does ipadaptor do at all?)

[peak dove]

Downright gorgeous!

[hardy cipher]

basically uses an image as a prompt in a very sophisticated way

[floral island]

it nearly replicates the source image

somehow

[floral island]

completely tokenizes the source image i think?

[floral island]

u ip adaptered my bakery gremlin?

[hardy cipher]

maybe

[peak dove]

So like Photoshop filters or Topaz Studio, you can apply artistic styles to your own photos - which stay recognisable as the photo input?

[hardy cipher]

not necessarily

it's all about how you set it up

and the weights

[peak dove]

OK, but nearly ...?

[floral island]

i used it on a selfie, and it made something that kinda resembles me as output 😮

[hardy cipher]

well what I did there for example was use the gremlin image as the latent input

and at lower weights on the ipadapter it looked gremlinish

[peak dove]

Yes, I am looking for img2img which will simply make the output look like van Gogh or Cezanne or Turner like Topaz Studio does

[hardy cipher]

but then full power

you can also use it in conjunction with other things like controlnet. but it gets tricky

[floral island]

i don't dislike abstract at times ❤️

[hardy cipher]

you have to balance everything or it breaks

at least for me, it's not like it just spits out a bad image. it's just a black box

[peak dove]

Hopefully its all developing in the right direction ...

[hardy cipher]

well considering the paper on the method came out like a month ago I think it's moving along pretty well

[glad jacinth]

@Dream pretty

[full idol]

Do I only need safetensor file or anything else needed?

[floral island]

using git lfs install and git clone you can get them all at the same time

you just need the files for the controlnet models you wanto run

[glad jacinth]

https://tenor.com/view/aww-cute-gif-22121346

[full idol]

I downloaded depth but it gives no effect, got it enabled, preprocessor "depth_midas" It can generate a preprocessor preview image of the depth map but does not affect the txt2image gens

[floral island]

then you also need to load the model into the controlnet

be sure to refresh the model list if you haven't done so already

[full idol]

ye its loaded

[floral island]

can you show screenshot? should be smth like this

[floral island]

yeah, it's the old controlnet

i suppose you have controlnet from way back?

same as i had

remove the controlnet folder from your extensions (saving your models if you want)

and then reinstall from the extensions list

[full idol]

ok!

[floral island]

and last but not least.... WE COME IN PEACE!

disclaimer* lots of ketchup was spilled making these

[hardy cipher]

hmm

ipadaptor setup with these two images made the one above

at least in comfy it has all sorts of options as far as how you can implement it. but some of them are tricky

[peak dove]

This is like /blend in MJ

[hardy cipher]

damn, the paper on it came out august 13th https://arxiv.org/abs/2308.06721

https://ip-adapter.github.io/

[peak dove]

Cool

[peak dove]

Phenomenal!

[hardy cipher]

he looks like he'd be pretty chill

those diagrams

[zinc cargo]

should ip it with xcom classic screenshots

the autopsies

[hardy cipher]

I've been iping it with various things

that's a good idea though. I'll look into adding it to my rotation

currently using these, lol

jeez

[glad grove]

its cause hes a japanese monster

[hardy cipher]

BLLUECI MONSTER

with immaculate feet

[indigo carbon]

it's the BLLUECI MONSTER, everyone know about that one

[hardy cipher]

clip vision ip adapter controlnet boss battle

[waxen cliff]

that is not a tail

👀

[floral island]

rivising some of my 1.5 prompts in some cases has some... unexpected turns

[peak dove]

Concept art of a beautiful female industrial 1900s gangster + in the style of Ashley Wood , Sam Weber, Joao Ruas, Octavi Arrizabalaga + portrait + Peaky Blinders + aesthetic canon of proportion + grunge and vintage and sci fi aesthetic (Prompt found at GitHub!)

[waxen cliff]

the 1st one def doesn't look like a female lol

or human for that matter

[peak dove]

😄

Here are some "definite females!" 🙂

[floral island]

this was the most sfw i got from this prompt

[vapid roost]

Hey friends, I made a post series about comfy and SDXL, It is more beginner-friendly and focuses on being more educational (as opposed to making the coolest workflow possible). Going through these can get anyone from an absolute beginner to comfortable level capable to build complex workflows. Check it out and let me know what you think
Part 1 - get started with SDXL base https://followfoxai.substack.com/p/part-1-sdxl-in-comfyui-from-scratch
Part 2 - add and explore conditioning parameters https://followfoxai.substack.com/p/part-2-sdxl-in-comfyui-from-scratch
Part 3 - Adding refiner https://followfoxai.substack.com/p/part-3-sdxl-in-comfyui-from-scratch
Part 4 - some advanced workflows: img2img, LoRAs, Controlnet https://followfoxai.substack.com/p/part-4-advanced-sdxl-workflows-in
Bonus: Installing ComfyUI on Windows with WSL2 (older post but still works) https://followfoxai.substack.com/p/getting-started-with-comfyui-on-wsl2

[wet nacelle]

My issue is that people really love recommending the official SDXL base instead of models that have a baked VAE and model overhaul.

base is good but not great. There are already really good everything models out now.

[vapid roost]

Sure, this is about getting started, swapping a model and testing new workflows should be straightforward after getting comfortable with Comfy

[wet nacelle]

Ah I see what you mean.

[peak dove]

D/loaded! 🙂

[peak dove]

I d/loaded

Thanks, interesting nodes

(Far more than I'll ever use, but that's just me!!) 🙂

[strange mist]

I loved your posts. they were really helpful 👍

[peak dove]

Nice

[peak dove]

Kewl

[wet nacelle]

rad

[peak dove]

Ouch, so good!

[strange mist]

cool beans

[peak dove]

The blood gothic mansion was FolloFox's initial w/flow - here are his final w/flows

[indigo carbon]

trying IPA right now, does anyone know why subject in image always goes out of frame to the left?

[peak dove]

Some choo-choo trains

[indigo carbon]

any idea why does subject always go out of frame to the right when using IPA?

[crisp owl]

I've noticed that as well, didn't check too deep into it, I just chalked it up to randomness, but perhaps there is a reasoning behind that

[indigo carbon]

just added 2unCLIPs chained to it, preview doesn't seem to be going out of frame

yo?

[indigo carbon]

I think it worked here, right?

[crisp owl]

guess to test would be another subject with different images doing the same thing

And then, if the two chained unclips mess up other ones where the subject IS in frame with just the one unclip

[crisp owl]

was that originally going out of frame also?

[indigo carbon]

nope, didn't try it without unCLIP

the only con here is that AIT doesn't work on this =[

[woeful patio]

should i change tje preprocessor resolution to 1024 when using controlnet with sdxl? or leave it at 512

[uncut steeple]

Whats your guys settings for face/hand fixes with the Detailer node? Im kinda fighting with it

[crisp owl]

I played with them for a short bit, but didn't go too crazy with it. These seem to work for me.
(going afk for a bit after this message)

[indigo carbon]

I think it worked here too?

also, this is possible as well with IPA

[uncut steeple]

Low CFG huh, also is that for both?

[uncut steeple]

Seems to work well for faces, but hands on the other... hand

[floral island]

@uncut steeple those images are from those song lyrics you gave me couple of months back. They're still amazing on SDXL based models 😄

[uncut steeple]

Oh yeah star one and ayreon lyrics go hard

[floral island]

why does sdxl have to take so long to render 😢

my brain needs more dopamine hits

[uncut steeple]

Takes longer for better outputs

[floral island]

https://tenor.com/view/iguess-gif-6228980

[indigo carbon]

if you're not using CNET or IPA and you have a 3000/4000 series card, you can get double the speed using AIT

[floral island]

this prompt going all the places. at once.

[floral island]

ok. uhh...

i have a 3090

what's the cnet, ipa, ait?

controlnet

not often

[indigo carbon]

controlnet, ipadapter, AIT is the new optimization

[floral island]

is it on a1111?

i'm just pure prompts usually

[indigo carbon]

nope, only specific versions of ComfyUI

[floral island]

🙂

shame 😦 can't give up my a1111 /dynamic prompting and prompt editing

[west breach]

https://civitai.com/models/141592

[floral island]

@west breach lets see if my inane prompting works on your model ^^

[west breach]

go for it! 😄

[uncut steeple]

I still dont know why AIT doesnt do much for me

[indigo carbon]

could be Xformers, Pytorch, and even your CUDA version

currently I'm implementing AIT to IMG+IMG IPA, so I can get the same speed while blending images

[floral island]

aight, model downloaded, lets give it a go

[uncut steeple]

Isnt AIT just one node you put in between

[indigo carbon]

yes and no. I have to compile specialized modules for each model architecture. currently trying IPA

[uncut steeple]

Not sure what that means, im a pleb

[floral island]

i probably should also get the fixed vae 🥲

[uncut steeple]

Is that still needed?

[floral island]

the old vae has that weird blocky artifacting going on, and i never bothered to update

[uncut steeple]

Which one is the fixed

[floral island]

the.. u hh... fixed version?

0.9 vae was ok too, it started in the 1.0 afaik

[uncut steeple]

Like where to get it

[floral island]

i think sytan did some looking into it

civitai i suppose?

[west breach]

my model should have the 0.9 vae baked in

[floral island]

nah, for the refiner

glad everyone picked up on it 🙂

[west breach]

I don't use the refiner anymore

[upbeat summit]

great images!

[west breach]

alright I need to go to bed, but these gens are too fun 😄

[floral island]

that's some nice neon stuff goin on ^^

[peak dove]

Moonlight Temple

[wet nacelle]

What is the center structure called?🦛

[peak dove]

I am presently using FolloFox's w/flow (here is a png with the metadata - I hope?)

[peak dove]

Temple, I'm hoping? 🙂

[wet nacelle]

ah okay

[peak dove]

Mebbe Pantheon, like in Rome, Italy

[wet nacelle]

https://tenor.com/view/jerma-plus-two-plus2-jerma985-jerma-award-gif-27706692

nice

[floral island]

@west breach tested -> ur model doesn't like refiner anymore ^^

[wet nacelle]

hey all. Take a listen to this great album. https://open.spotify.com/album/3ZZTr5lTvnwCVHgV16UcOE?si=eT_quzmbRd-EQtUBegQZIw

[floral island]

it actively makes it worse

[shy kelp]

is AI limited by computing power? Every 2 or 3 years there are new GPU generations with more and more Tera flops. Is tflops, memory, a limitation or it is a matter of more research to be done to improve the models?

[mossy canopy]

nice

[floral island]

i think there's something wrong with my car tough

[mossy canopy]

3d toon model?

[floral island]

protovision

lol xD this abomination

[crisp owl]

It's the whe-feet concept model

[floral island]

a real monstertruck!

[peak dove]

Mausoleum

[crisp owl]

Yeah, but cfg is subjective really

[wet nacelle]

Ah okay. Thanks.

[uncut steeple]

Those hands are killing me
Ive been trying for an hour or so

[peak dove]

Cool

[mossy canopy]

clipped

[floral island]

monster beetle? 😄

[uncut steeple]

Todays work, expanding my workflow

[floral island]

multiple gens, triple knee or...

all the generations had that leg issue with this prompt, weird

[uncut steeple]

Try openpose to fix it

[peak dove]

"She looks a tad kneedy!!!" 😄

[indigo carbon]

WIP, currently implementing AIT for an IPA-loaded architecture

[native knot]

When you use more than one ControlNet, do you find that your iterations time skyrockets?

[native knot]

Are we still stuck on an older version of comfy to make AIT work?

[indigo carbon]

unsure, that's not what I'm responsible of, you can ask FizzleDorf by opening an issue on the repo. I'm currently adding IPA support.

[native knot]

I just figured you might happen to know.

[floral island]

which model are you using?

[mossy canopy]

nightvisionxl

[floral island]

ah, i'm using the protovision 🙂

[fiery moss]

I had this error yesterday, but i can't get passed it again. Im using RunPod to train a LoRA. any suggestions?

CalledProcessError: Command '['/workspace/kohya_ss/venv/bin/python3', './sdxl_train_network.py', '--enable_bucket', '--min_bucket_reso=256', '--max_bucket_reso=2048', '--pretrained_model_name_or_path=/workspace/stable-diffusion-webui/models/Stable -diffusion/sd_xl_base_1.0_0.9vae.safetensors', '--train_data_dir=/workspace/1JapanTraining/img', '--reg_data_dir=/workspace/1JapanTraining/reg', '--resolution=1024,1024', '--output_dir=/workspace/1JapanTraining/model', '--logging_dir=/workspace/1JapanTraining/log', '--network_alpha=1', '--save_model_as=safetensors', '--network_module=networks.lora', '--text_encoder_lr=0.0009', '--unet_lr=0.0009', '--network_dim=256', '--output_name=1JapanTraining', '--lr_scheduler_num_cycles=10', '--no_half_vae', '--learning_rate=0.0009', '--lr_scheduler=constant', '--train_batch_size=5', '--max_train_steps=1840', '--save_every_n_epochs=1', '--mixed_precision=bf16', '--save_precision=bf16', '--caption_extension=.txt', '--cache_latents', '--cache_latents_to_disk', '--optimizer_type=Adafactor', '--optimizer_args', 'scale_parameter=False', 'relative_step=False', 'warmup_init=False', '--max_data_loader_n_workers=0', '--bucket_reso_steps=64', '--gradient_checkpointing', '--xformers', '--bucket_no_upscale', '--noise_offset=0.0']' returned non-zero exit status 1.

[native knot]

Is anyone chaining cnets on SDXL that can speak to their iteration times? On my 3080, if I have one active, I can do about 1 second/iteration with my current workflow, but if I add a 2nd cnet in the chain, it jumps to about 40 secons/iteration, which is quite a bit.

[floral island]

jarception!

[native knot]

Now make her hold a jar.

And in that tiny jar should be another fairy.

🤣

[floral island]

actually, i had to prompt AGAINST her holding a jar

if i don't specify, she WILL hold it

[native knot]

lol

[floral island]

trying to find that other fiary

[uncut steeple]

Not sure, nothing im worried about tho

[crisp owl]

I always get those 2nd and 3rd heads on the wing talons whenever I tried large winged things

well, not always, but quite often

[floral island]

yeah, wonder what causes it

OH SHIT, CURSED GENERATION 😮

[floral island]

sometimes, old prompts are scary stuff

which rock/metal band is depicted here 😄

[crisp owl]

Hammer of the Witches

[floral island]

was supposed to be the black mages

but ok 😛

[compact hatch]

don't think it knows ff that well

[floral island]

i suppose the model didn't realize either "the black mages" are final fantasy concepts, or are an actual band 😛

[crisp owl]

I've seen some great Vivi sd images before, back in 1.5

[peak dove]

The Rolling Hoggwartz?

[floral island]

and the wings always behind charcters, regarding of orientation 😦

so, i'll just assume she's facing us

[mossy canopy]

yep.. :v

[echo temple]

I still do not get why I cannot put a simple hat in comfy using inpainting, I think it was working before but now even with denoise of 1, it does not change that much from the original image

[smoky patrol]

those hands are fantastic

What is your prompt?

[uncut steeple]

This is mostly controlnet work so my prompt won't give you a lot

[smoky patrol]

rip

[vital ermine]

I am about to release my latest lora

[smoky patrol]

still lol

[uncut steeple]

I'll have to check tomorrow, I'm out of office now

[smoky patrol]

Ah, okies!!

lemme k n o w ❤️

[floral island]

oh damn 😮 from a pure and directed prompt actually

[uncut steeple]

It's something like female in a box, walls closing in and some photo related prompts, something along those lines

[floral island]

whatever my creatures are, the moment they come into existence, they're on the exinction watchlist, they are gorgeous 😮

[smoky patrol]

in a box; yeah, I figured it was that, something, hm, yeah

[uncut steeple]

Also isn't the workflow in the meta data?

[floral island]

sd1.5 models just couldn't handle this prompt

i mean... for comparison: 1.5 on the same prompt lol

[shy kelp]

the bot thanks me for voting. Are the images stored and reviewed by ppl developing sdxl?

[smoky patrol]

Also, for wings, you can try using angel wings as a negative. No guarentees, but you can try

[mossy canopy]

nah.. its not working

[floral island]

perhaps nightvision too realistic?

i mean, that's a very nice effort tho

i'm using minimal to no negatives

[vital ermine]

civit is having issues again for me to upload :/

[mossy canopy]

not nightvision..

[vital ermine]

sheesh, three time but finally.

https://civitai.com/models/141737?modelVersionId=157120

[mossy canopy]

im born ..

[mossy canopy]

make something like this..

[strong field]

what does your prompt say in this case? nvm i see it, let me see.

[vital ermine]

Little Red is on her own.

[mossy canopy]

indeed

[strong field]

I think if you use the inpainting model then you should be using encode for inpainting, but that is not my preferred work flow.

[vital ermine]

Save her, then tell her how beautiful she is and she snubs her nose up and says "sorry, you're not my type" and she walks away.

All the while I am near dead saving her.

[strong field]

try to pull an empty latent image into the sampler, then set your denoise low to taste

[strong field]

before you try to change anything else with the encode for inpainting that is

[floral island]

whoops... i had the cfg on 18 earlier lol\

much better ❤️

[mossy canopy]

just prompt?

[floral island]

Yeah just prompt

[mossy canopy]

truth revealed

[steady grove]

the whole lizard people thing is antisemetic propaganda that got repurposed as secret dog whistles in the conspiracy theories communities. while those are cool generations, that whole notion is kinda "out in the open" jew hate. it all roots back to a document published in early 20th century. knowing is half the battle.

[hardy cipher]

I like to think of it more like the 1993 Super Mario Bros movie with John Leguizamo

top notch cast

[steady grove]

rip bob. best mario.

rip hopper. potentially the best bowser

[hardy cipher]

don't forget Carlene Moore

[steady grove]

best prostitute

[hardy cipher]

anything worth doing is worth doing well

[steady grove]

leguizamo was legitimately a good luigi and pulled it off for being a spanish man. it was so early in his career. to this day i think he only got hired because of the auditions, his name was closest to luigi

[hardy cipher]

you know, that might be true. but seems like he very quickly became respected in the movie biz. kind of hard to believe that movie was 30 years ago

[steady grove]

Oh yeah the man's got talent across a wide range. He was just such an oddball choice for luigi

so is charly day. i really think they missed an opportunity in the new movie to have a luigi conspiracy theory nut scene

https://tenor.com/view/charlie-day-gif-18564553

[hardy cipher]

remember this? https://www.youtube.com/watch?v=W8mrNxgX9OY

[steady grove]

fondly

as much of a nintendo kid as i was though, there was no denying that the sonic cartoon was better

https://www.youtube.com/watch?v=vKk85FrWVWk

[hardy cipher]

man, I barely remember that one for whatever reason

then there was the legend of zelda cartoon

[steady grove]

was part of the mario bros show that one was good too. the aesthetics were sooo different back then

by the time they did the pokemon cartoon, the show's and games were on point with their branding

[hardy cipher]

everything was so cheesy back then. but in a good way

the tail end of the analog age

[steady grove]

analog art forms had a lot of leg time left on them yet. the old knowledge of the masters is still perculating. miniature work, makeup work, hand animation and sketching, the communities are very much still alive, it's just their work is so much less propagated. Japan has a lot of love in the mastery end of analog yet. "One punch man" i think is a love letter to classic animation tropes

this is tippet's "masterpiece" from his long career. worth checking out. https://www.youtube.com/watch?v=pbW5ns_pIZo

[hardy cipher]

it's more timeless. cgi aged quickly. especially the first round of it

look at snes or genesis graphics vs ps1

[mossy canopy]

sd 1.4

[steady grove]

cgi was rushed into production environments so studios could screw over the old sfx unions

[hardy cipher]

when they started trying to go full polygon

and yeah, those early low budget cgi productions, lol

[steady grove]

we'lll see it again as digital actors and ai videos come into play., we'll get a wave of lowest quality end production value since studios will hire people who have no idea of the old mastered techniques

[nocturne dove]

Silly question but I don't know comfyui very well. I made a workflow using the ReferenceOnlySimple node and the ksampler shows the reference image in the preview and later on the reference image and the generated image are both saved in my output folder. Does anyone know how to avoid saving a copy of the reference image every time I generate an image?

[hardy cipher]

should be saved in your input folder

[hardy cipher]

well people mistakenly think that because something new comes along the old methods are negated and entirely obsolete, but that's so far off base

[steady grove]

i actually like snes graphics better than ps1, since snes 16bit artists had mastered those palettes by then. Creating amazing scenes with the technical limits they had. Then psx came out and it was kind of shit. I liked n64 better since the games were more , gamified. polished. finished. "quality over quantity". FF6 is better than FF7 in so many undeniable ways in my books.

[hardy cipher]

understanding the old methods and being proficient in them will facilitate adopting the new methods properly. you can polish a turd all you want but it's still a turd. you have to have substance there or all the tech in the world won't save you

[hardy cipher]

and yes, snes still looks good imo. but ps1 is more nostalgic and kind of just goofy to look at

[steady grove]

Even starfox, snes's shittiest 3d game, is better game mechanics and design than the same games that came out on psx. i don't know if psx ever had a good jet style game like starfox or panzer dragoon

[hardy cipher]

16 bit era was art

[steady grove]

Nights into Dreams was good too

[hardy cipher]

chronotrigger is still good

[steady grove]

Chronotrigger is the GOAT

[hardy cipher]

I played through that game more than probably any other game ever

well, maybe skyrim

[steady grove]

hard to say where i spent most time. that was a big contender

illusions of gaia too

[hardy cipher]

super mario world as well. that's one of those games I forget I'm still good at

those star warp levels with names likes "radical" and "tubular"

soooo hard

[steady grove]

agreed. some of the few occasions i was pushed to throwing my gamepad as a kid, were those levels

[hardy cipher]

I think a lot of kids got really good at games like that because there weren't a lot of options? want to play super mario world or one of the two kmart bargain bin games your parents bought?

those were the choices

[steady grove]

i would ride my bike for 20min to the next municipality to starflight video game rentals. had a paper route just to get an account there. rented 2-3 games a week. sometimes just keep taking the same one out.

i bought carts a couple times but they were $60-$90 and if they sucked, no returns.

[steady grove]

or i'd play them for a week and be done

way off topic as i'm nostalgiing. lol. mb

[hardy cipher]

well it's giving me ideas to make things, lol

you know what was up there with chronotrigger was earthbound

[steady grove]

"nintendo hard" was a game design trope of the past. Games were made a LOT harder so that we'd get more out of them. since the quality market was much smaller

nintendo's gold seal came to mean something. before the seal of quality, games sucked

[steady grove]

yeh and nintendo's other RPG, super mario RPG. getting a big ol remake soon

their big collab with square

[hardy cipher]

well also, if you consider what those games actually consisted of. take contra for example. just think how long it would take it could basically walk through without dying or struggling

they're a few hundred kilobytes at most

[steady grove]

damnit you brought up contra now i remember i haven't beaten cuphead yet

[hardy cipher]

I'll never beat cuphead. I have cuphead and I can confidently say that'll never happen

[steady grove]

heh i bought it launch day and yeh, here we are

[hardy cipher]

I realized I'm not the best at hard games these days, lol. well hard as in reflexes and timing and all that. just don't have it in me to fail a thousand times in a row

[steady grove]

some of the most ultimate side scrolling i've done was on megaman x series. on the psx i think it was megaman x6, was some of the hardest finger work i've ever managed to pull off

[steady grove]

thats a radical blue bomber

[hardy cipher]

I'm more into the open ended sand box stuff these days. well haven't really gamed much in the last year or two. but in general

[native knot]

Was already thinking Mega Man before you posted that.

[steady grove]

same. i'm slower now

[native knot]

Mega Man 2 is my jam.

[hardy cipher]

skyrim was always great because I could just do whatever I wanted. go fight things, trade, pickpocket, pick flowers for potions. no pressure, lol

[steady grove]

that dr wiley theme

[hardy cipher]

just walk around and look at stuff

[steady grove]

https://www.youtube.com/watch?v=iW_Ct49H1ng

[steady grove]

when metroid dread came out, i trained for 2 weeks before it's launch cause i knew i'd need my finger nimbleness lol

[steady grove]

got one of those gyro exercise balls. grip strength squeezies. i do finger yoga. i do accurate typing drills. getting those fingers primed for the metroid. That game still killed my hands though. was cramping up a lot

so far got most of the hidden secrets

[floral island]

i never cramped up 😮

[floral island]

although, i can image some of the bossfights are hard when spamming the gun

[steady grove]

hollow knight is till haven't beaten

[floral island]

never been able to get into hollow knight

i tried it, but... something doesn't click for me

ori on the other hand.. ❤️

[steady grove]

i need to do ori yet. hollow knight is a slow starter but it hooks you once you get into it

gtg. think i'll do a retro study with SD later today . bbl

[floral island]

i played it quite a while actually 😦 i think i like movement based platformers better than combat based platformers

[steady grove]

yeah, it IS movement based once you get stuff unlocked. it's SO slow starting

[floral island]

mind you, it's a really cool game, something just didn't click for me 😦 and it's a shame, because i really want to like it

[hardy cipher]

man, I can't buy any new games until I beat some of the games like hollow knight that I bought years ago

I've realized I now like owning the games more than actually playing them

[floral island]

shhh

cries in 807 games in steam list

[mossy canopy]

and epic games?

[floral island]

0

i refuse to use that launcher

[mossy canopy]

they have weekly giveaway..

[floral island]

i know

but i have a personal issue with launcher/companies that try to create exclusivity contracts and game vendor lockouts

so. i vote with my wallet -> i won't even bother getting the launcher

[floral island]

plus, the money for those free giveaways was nearly all funded by fortnite -> epic ddidn't make enough money on the launcher. dunno if that improved by now

granted, knowing people only use epic for the free games, i suppose that's still the case

[mossy canopy]

i still have egs installed just to claim the weekly giveaways

[floral island]

you actually play those games? 😛

[mossy canopy]

death stranding ?

[floral island]

i mean, anyone is free to do what they want. i have a personal gripe with the way they operate and WANT to operate

they have a lot of good games in their free giveaways, no question about it

[mossy canopy]

exactly.. those games are the only reason..

sdxl recognizes death stranding

[floral island]

but. can it replicate the ending too xD

[floral island]

you know that's still accurate

[mossy canopy]

death stranding 2

[floral island]

uhhh... i can't remember the giant mech tho xD

[hardy cipher]

all the ipadapter tutorials that show elaborate setups are built on a house of lies

[floral island]

probably the latent going from death stranding -> kojima -> metal gear

[mossy canopy]

dont recognize bts

[floral island]

they're not exactly mechs...

[floral island]

i think he needs to smoke less...

[hardy cipher]

clip vision ain't no thang. but that damn ipadapter. something might be goofed in my install

[mossy canopy]

well put..

[floral island]

what do we need to do to get to this timeline?

[uncut fiber]

is there node in comfui, that goes over all images in folder? Dont know in which channel to ask. Thank you.

[uncut fiber]

reminds me mass effect 🙂

[crisp owl]

like batch loading?

[uncut fiber]

yes, it seems it is @crisp owl Thank you! In what pack it is? Oh i see Was node suite. Thank you very much!

[mossy canopy]

void out

[strong field]

what like this?

try to update IPA with git pull, or install from scratch, something wonky

[hardy cipher]

looks like you only have one adapter outputting model data? I'm trying to string them or combine them, and although the info on the method says it's not a problem I cannot find the proper balance

[strong field]

i dont get enough control like that

[hardy cipher]

I can mix ipadapter with clipvision and controlnet and all those things. but that model data

[strong field]

i like my control

[strong field]

i have tried ALOT to string model > model and i hate every iteration of it

[indigo carbon]

this is how my IPA workflow looks like ATM

[floral island]

so, ipa basically converts images into pretty accurate tokens?

[hardy cipher]

something along those lines

[floral island]

power centers at maximum power output for sd image generation

[hardy cipher]

well you see, all your workflows are quite lovely and surely more functional than mine. but I still want to string multiple ipadapters together like they do in those github workflows

[crisp owl]

string them together as in?

[indigo carbon]

IPA is multimodel, you can make a workflow that does this:

[strong field]

nah, splits text and image cross attention

[hardy cipher]

model input to model output

output to input

whatever

anyway, I know it's possible. but I haven't gotten a feel for balancing the numbers. but I will

[floral island]

hmmm... interesting

[crisp owl]

Yeah that's just a weight balancing game. Gonna be different for each two images you're stringing together with IPAdapter

[indigo carbon]

I just inferred it in a way that does IMG+IMG, combining it with unCLIP

[hardy cipher]

well I've found with a lot of things like this it might seem random at first, or at least not intuitive, but after a while I start getting a sense for how to set it up

[strong field]

i have different workflows for different outcomes, this one is model to model chain

[hardy cipher]

is that for 1.5?

[strong field]

this one prob one of my fav:

[strong field]

yes, but next screen is sdxl

the possibilities are ENDLESS

[hardy cipher]

well anyway, I just need to find where my inbalance is. or if it's some kind of error in the my install.

[hardy cipher]

trying to establish what kind of range they'll run in individually

[strong field]

what strength values you using?

[hardy cipher]

man, anywhere from .4-1.5, and everything in between

[strong field]

i usually stay around .3, stuff higher than .5 gets to overcooked

[hardy cipher]

so .3? I'll try that I guess. the tutorial I read didn't say such things, but obviously ithe methods I learned there aren't really working out great

[strong field]

yeah, typically lower values work best for me

[hardy cipher]

well I've found that individually I can push them up well past 1

weight wise

not that it looks optimal, but it doesn't give me any kind of error

[strong field]

update update update

most errors i run into are just mismatched commits

[hardy cipher]

bruh, I've updated like 10 times in the last 2 days

[strong field]

yuck

[hardy cipher]

betweein comfy and nodes

I've never used anything that need updated so often

[strong field]

if you want send me your workflow, ill load it up and see if gives me errors

You're on the bleeding edge of AI image gen...

[hardy cipher]

un momento

[strong field]

about as simple as it gets

[crisp owl]

I'm interested what sort of work is being cooked currently, last commit for Comfy was 3 days ago now

[strong field]

weights are .3

[strong field]

always interested in whats next

[strong field]

here is .6 weight on img 2 and you can see its starting to get overcooked

[steady grove]

i like when people take a minute to try it with two pics, do it two or three of the 10000 wrong ways of doing it, instead of one of the 10000 correct ways, and declare "it's impossible and can't be done"

[strong field]

am i doing it wrong?

[eternal fog]

That weight is too high for IPAdapter

[steady grove]

clearly the authors of ipadapter need to go back to school

[eternal fog]

Anything much over 1 in my experience makes it into a mess

[strong field]

higher!

[eternal fog]

Ah were you trying to do it badly lol

[steady grove]

jokes on them. i was only pretending

[eternal fog]

I can't decide if I prefer 2 IPAdapters into each other, or Revision+IPAdapter.

[strong field]

why not all of them

[eternal fog]

I mean you can, but then what are you even trying to make lmao

[strong field]

#✨｜sdxl message

candy

There are reasons

[indigo carbon]

any idea?
`Error occurred when executing IPAdapter:

'ClipVisionModel' object has no attribute 'patcher'

File "Y:\ComfyUI2\ComfyUI\execution.py", line 151, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
File "Y:\ComfyUI2\ComfyUI\execution.py", line 81, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
File "Y:\ComfyUI2\ComfyUI\execution.py", line 74, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
File "Y:\ComfyUI2\ComfyUI\custom_nodes\IPAdapter-ComfyUI\ip_adapter.py", line 157, in adapter
cond, uncond, outputs = self.clip_vision_encode(clip_vision, image, self.plus)
File "Y:\ComfyUI2\ComfyUI\custom_nodes\IPAdapter-ComfyUI\ip_adapter.py", line 224, in clip_vision_encode
comfy.model_management.load_model_gpu(clip_vision.patcher)`

[mossy canopy]

https://github.com/cubiq/ComfyUI_IPAdapter_plus

[shrewd ferry]

does a1111 play nice with sdxl now?
have been away for a while

[tepid needle]

Hello everyone !
I need a little help understanding something. I'm trying this sdxl model, and I try generating literally anything, but the last steps always ends up with an amazing amount of artifacts, noise, and blue/yellow colors...
(Automatic1111)
Would you know what's happening ?

[cyan crown]

vae?

[tepid needle]

vae-ft-mse-840000-ema-pruned

[mossy canopy]

lmao

[tepid needle]

I know, I'm new to this, I took the one I found in the guide I found 🙂

[mossy canopy]

go with sdxl vae..

[tepid needle]

The sdxl model says VAE in the title, though

[steady grove]

you dont use that one with sdxl

[cyan crown]

wrong vae

[mossy canopy]

then select automatic..

[tepid needle]

Which one would you recommend ?

[crisp owl]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main/vae

[tepid needle]

This is the model I use, is it a VAE model or a sd model? I'm lost

Or can I use it for both ?

[mossy canopy]

its a sdxl model with fixed sdxl vae included

[tepid needle]

Okay, I think I got it

[cyan crown]

use model vae in config with that model

[indigo carbon]

!!! Exception during processing !!!
Traceback (most recent call last):
File "C:\Users\joker\OneDrive\Desktop\A.I\ComfyUI\execution.py", line 151, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
File "C:\Users\joker\OneDrive\Desktop\A.I\ComfyUI\execution.py", line 81, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
File "C:\Users\joker\OneDrive\Desktop\A.I\ComfyUI\execution.py", line 74, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
File "C:\Users\joker\OneDrive\Desktop\A.I\ComfyUI\custom_nodes\ComfyUI_IPAdapter_plus\IPAdapterPlus.py", line 187, in encode
output = clip_vision.encode_image_IPAdapter(image)
File "C:\Users\joker\OneDrive\Desktop\A.I\ComfyUI\custom_nodes\ComfyUI_IPAdapter_plus\IPAdapterPlus.py", line 325, in encode_image_IPAdapter
comfy.model_management.load_model_gpu(self.patcher)
AttributeError: 'ClipVisionModel' object has no attribute 'patcher'

Prompt executed in 0.06 seconds

[tepid needle]

Looks a lot better ! Thank you

[mossy canopy]

clip vision model?

[indigo carbon]

pytorch_model.bin

the one included with IPA for SDXL

[steady grove]

someone said beer i came as fast as i coudl

[mossy canopy]

no clue .. if you followed the instructions exactly it should work..

[steady grove]

🍻

[mossy canopy]

i used this model https://huggingface.co/comfyanonymous/clip_vision_g/resolve/main/clip_vision_g.safetensors

should be same

[indigo carbon]

I did, it worked before with the chinese version, reverting to it

[vivid swallow]

i really should not be allowed to use controlnet

[mossy canopy]

ip

[eternal fog]

I like the little photobombing hand lmao

[floral island]

"last night, good night" such a lovely miku song

[floral island]

this yoda slays. or whatever that disney thing is called. never watched the series

[floral island]

are you trying to figure out how big you can make him?

[glad grove]

im tryin to make him smaller

[floral island]

so you can fit more of him on the image?

[glad grove]

not enough pixels

[crisp owl]

I thought that looked familiar

[west breach]

I don't use the refiner model

[floral island]

yeah, it's not needed on your model 🙂

[glad grove]

speaking of refined

[upbeat summit]

this is your most insane seagal movie yet!

[upbeat summit]

hanging out with the crew

[glad grove]

new sitcom " my friend the swamp monster"

[cyan crown]

Last one for today. Good night!

[crisp owl]

G'night Wallace

[native knot]

G'night Gromit

[eternal fog]

lmao

[eternal fog]

my lord lmao

[upbeat summit]

nice - I don't think any seagal movie had the budget for special effects like that!

[glad grove]

no need for special effects,thats just his angry face

[glad grove]

me and my gf

[glad grove]

my honest reaction

[novel birch]

are negative prompts still as important on SDXL compared to 1.5? I haven't noticed a big difference putting stuff like "bad anatomy" on negative prompts vs not having anything

[fallow drift]

Guys I need help

[upbeat summit]

Dr. Seagal building space stations now

[glad grove]

our first date

[glad grove]

birdman

[hardy cipher]

those highly evolved hands

[glad grove]

me and my dad

[glad grove]

hey guys

[hardy cipher]

NSFW?

[glad grove]

its just his charging port

[glad grove]

he lookin sexy and badass here

me and my brother

[upbeat summit]

very stylish!

[vital ermine]

@upbeat summit I released another lora earlier today

https://civitai.com/models/141737

[upbeat summit]

oh nice - vignettes and damaged image looks 🙂

[vital ermine]

yes

and more too

weird little lora

rather weird AND versatile

wild beast to tame

[upbeat summit]

cool I will check it out 🙂

[glad grove]

this dont look like a bat

[native knot]

Too soon?

[glad grove]

look at my new baby

[glad grove]

come to play in the sand

[glad grove]

🫦

[glad grove]

thic

[west breach]

left my model generating all night, now got over 1700 pics to go through 😄

that was peter griffin as the flash

minion dressed as michael jackson

[glad grove]

legs gone

[native knot]

Did you just run through a bunch of wildcard combinations?

[west breach]

I was using the file sandbit shared a while ago

#🌠｜show-and-tell message

[native knot]

Oh? Where's that?

...and you did.

🙂

[glad grove]

our new squid god

[native knot]

You've gotta ask yourself one question...
Won't you be my neighbor?

[glad grove]

friendliest neighbor in america

[native knot]

For reference, the original image was this:

[upbeat summit]

#🌠｜show-and-tell message

[native knot]

Yeah, I saw. 🙂 But good lookin' out!

[upbeat summit]

ahk - saw it now 😄

[glad grove]

greetings

[worthy orbit]

inpainting with Automatic1111 with SD XL, I had try it for a bit but it was like inpainting with a non inpainting model in SD 1.5, any thoughts 🤔

I'm just doing img2img and photobashing in GIMP

[small tree]

xl is overrated lol

[native knot]

I took a shower washing every body part with actual soap; including all my major crevices; including in between my toes and in my belly button which I never did before but sort of enjoyed. I washed my hair with adult formula shampoo and used cream rinse for that just-washed shine. I can’t seem to find my toothbrush, so I’ll pick one up when I go out today. Other than that, I’m in good shape.

[small tree]

trained sd1.5 > xl and its much easier to run

[glad grove]

who's a good boy ? 🐶

[wet raven]

anyone get inpainting to work on SDXL? I just get a black masked off area when i load the
diffusion_pytorch_model.fp16.safetensors [6470840731]
and generate with it on inpainting on automatic1111

[worthy orbit]

you mean making Lora's for SD 1.5? I couldn't train in SD XL yet but yes you can get amazing things in SD 1.5, but it is much work. SD XL just delivers a good resolution image. No need to work on upscalling.

[wet raven]

this file
"difusers team released one"
https://huggingface.co/diffusers/stable-diffusion-xl-1.0-inpainting-0.1/resolve/main/unet/diffusion_pytorch_model.fp16.safetensors

[small tree]

generating in 512 and then upscaling good results to 1024 is much more efficient than having to straightup generate at 1024

[worthy orbit]

I get it but I iirc I had to add something to the flags like --half-vae or something like that. But it was pretty bad at inpainting, like inpainting with the SD 1.5 base model.

[small tree]

i can barely run sdxl let alone train loras on it lol

[worthy orbit]

yes but SD XL is really something else, it really considers the prompt... I almost run it without negative prompt, I use the negative just for the things I don't want, not the classic 1.5 negs and embeddings

[small tree]

?

do you dislike negative embeddings or sth

[worthy orbit]

I'm trying it right now, it seems that now after the update it works with the same flags... kind of correct the eyes, but haven't test it in some bigger change

[native knot]

SDXL doesn't require the same negative prompts as 1.5 in order to produce solid outcomes. They can still be very helpful, but you can get very good results with fewer negative prompts.

Also, as a side note, the whole thing about people putting "bad hands", "bad face", etc. doesn't actually do what you think. The word "bad" might be fine and help some, but saying "bad hands" doesn't tell the system not to choose images that create bad hands because it's not trained in such fashion...neither was 1.5 for that matter.

[worthy orbit]

it just add too much randomness (?!) and usually you have to test what every embedding does with the model, prompt, sampler