✨｜sdxl

[wet nacelle]

FUCK!

I get that now

[hardy cipher]

lol

[native knot]

Come on PoP

[dim steeple]

do negatives work the same in xl?

[native knot]

Give him the sauce!

[dim steeple]

these are mine from 1.5: (deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime, mutated hands and fingers:1.4), (deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, disconnected limbs, mutation, mutated, ugly, disgusting, amputation, cartoonish, inaccurate clothing, distorted face, unnatural lighting, unrecognizable, pixelated, watermarked, low quality, poorly composed, low resolution

[wet nacelle]

I just wanna make hyper gore

[hardy cipher]

man, I had bing give me a list of 50 latin words that were similar to "uncanny"

[wet nacelle]

have less negs

[native knot]

I find that negatives are actually a touch stronger in XL compared to previously.

[hardy cipher]

and then I pasted them all in

[wet nacelle]

oh. interesting.

@dim steeplecheck it

[native knot]

I have been curating my negative list as I go and pretty much use the same one every time unless I really need to adjust it.

[hardy cipher]

oh, I'm going to try to make a node that'll make picking your saved words and what not easier. unless that's already a thing and I just haven't seen it

[wet nacelle]

I've been really lucky with the stuff that I get. I've used a few Loras and a few models.

I'm lucky to have gotten what I want from my positives rather than having to mess with the negatives.

[hardy cipher]

something simple. I honestly tried to find something like that and haven't found it. I've only ever made nodes that I couldn't find in other people's stuff

and that I want to use

[hardy cipher]

a lot of times in the negative I'll just put dumb stuff. "bad, stupid, failure"

[wet nacelle]

SUPAMAN AND BAHTMAN!!!

[hardy cipher]

things like that

[native knot]

Mack, add the word plastic to your negative and re-run that.

[glad grove]

bro thinks he batman 😭

[wet nacelle]

Which?

[native knot]

That last one

[hardy cipher]

he must be using the muscle bro lora

[native knot]

lol

[dim steeple]

is 2048x1024 too high?

[native knot]

yes

use the table I linked earlier

[wet nacelle]

@native knotI honestly don't know how to keep my seed in Comfy after a gen.

[hardy cipher]

2048x512

is the way

[hardy cipher]

what's your workflow?

[native knot]

If you've already generated it, click view history, then find the one that popped out that image.

[hardy cipher]

just drag it back in

and it'll have the seed number

[native knot]

That's another way

[native knot]

There's like 4 ways to do nearly everything in comfy

oh my

[hardy cipher]

you can also copy the image and paste into comfy

[native knot]

Or you can open it in a text editor and search for the seed

[hardy cipher]

or you can use a hex editor and sift through the metadata

[native knot]

Or you can print it out, fold it up, stick it in an envelope, and hold it against your forehead like carmack.

PoP & me...same energy

[wet nacelle]

https://tenor.com/view/john-carmack-my-beloved-doom-romero-id-softwar-oculus-gif-21464870

[hardy cipher]

thiccc chest

the torso on the lad

[native knot]

He ate breakfast before that pic.

[hardy cipher]

got a bit of sun as well

might have high blood pressure

[native knot]

needs to cut back on the salt intake

[glad grove]

real

[wet nacelle]

@native knot(1) neg

(2) without the neg

[native knot]

Definitely better with that in the negative...cuts some of that sheen from the costumes.

[wet nacelle]

I see what you mean

[native knot]

When I'm going for certain looks, I'm always trying to think about materials. It's almost a habit now. I fucking think about materials all the time when I'm looking at stuff around the house, around work, or when I'm out, now. I'm constantly on the look out for words that will describe how a material is visually.

[wet nacelle]

Sounds intrusive

[native knot]

oh my

[native knot]

Poor Steven...really really let himself go.

[dim steeple]

What is the best way currently to train custom SDXL models?

[native knot]

The Death Star has entered the system...I repeat...The Death Star has entered the system.

[dim steeple]

Is Dreambooth still good?

[hardy cipher]

https://tenor.com/view/tommy-boy-chris-farley-no-idea-clueless-gif-4280737

I believe people use kohya these days. I have images for a lora but haven't got around to it. started making dumb nodes instead

[glad grove]

hey baby

[hardy cipher]

whoa

[native knot]

NSFW... Mods!!

😄

[hardy cipher]

I got this from zeroing out conditioning

thought I'd see some sort of grayed out monster or something

[shy kelp]

The AI is not getting Naruto x Goku. It's drawing Naruto x Naruto, with one having Goku's clothes.

[hardy cipher]

yeah well, it thinks you want to look at their offspring

[glad grove]

feelin sexy tonight

[hardy cipher]

got these dumb conditioning nodes to work

[vivid swallow]

sdxl really took a nose dive

[hardy cipher]

yeah, well firstly that image is 512x512 right?

[hardy cipher]

goodness

his stomach has seen things

[vivid swallow]

ya

[hardy cipher]

you need to render at 1024x1024

[vivid swallow]

so thats like 10 minutes per render?

[hardy cipher]

or 1 megapixel if you want to change the aspect ratio. so 2048x512 works also

[vivid swallow]

probably

[crisp owl]

pc dependent

[hardy cipher]

4x approximately

maybe a bit more

[vivid swallow]

3060ti with 32 gb ram

[hardy cipher]

since 4 times as many pixels

[crisp owl]

Goodness no not 10 minutes on that lol

[vivid swallow]

I promise you it will take forever

[crisp owl]

I have a 2060s

[hardy cipher]

unmodified conditioning, normalized, 2x conditioning

normalized seagal looking dapper

I'm doing 5x seagal now

[glad grove]

feelin human today

[hardy cipher]

jeez, fat forehead

[hardy cipher]

hmm, 5x Seagal took on a strange form

turned into a dirty piece of fabric. his next form maybe

[glad grove]

that just his skin pores

real

now hes mad

[upbeat summit]

I can see guys we are deep in the latent space now

[glad grove]

we have to go deeper

[upbeat summit]

I went too deep

[crisp owl]

https://tenor.com/view/star-wars-anakin-skywalker-i-hate-you-anakin-i-hate-you-i-dont-like-you-gif-18829084

pretty sure that's from this scene

[upbeat summit]

"Any great warrior is also a scholar, and a poet, and an artist"

[glad grove]

time to relax

[crisp owl]

You know you've hit a weight level over 9000 when gravity starts pulling your finger fat down

[heady vale]

feetngers

[glad grove]

new pants

[weak saddle]

#🏞｜general-with-images message

[native knot]

Converted your buildings into larger ones.

[hardy cipher]

3x Seagal

he's beginning to transcend his human form

[nimble heart]

he's about to open his third eye and see into the overground

[upbeat summit]

you are looking into humanity's future with this - latent space evolution

[hardy cipher]

I'm just messing with pooled output. need to learn more about how all these things work though

[nimble heart]

looks almost like the right have is the regular square and the left section was painted in. Probably from the high res non-square aspect.

supposedly you can adjust that using the CLIP crop x/y params but I haven't really monkeyd with it much

[upbeat summit]

also the depth of field can sometimes generate interesting "compositions"

[nimble heart]

then on low res you have the opposite issue where it looks like it's a square that was overcropped so idk

[hardy cipher]

well I'm doing pretty rudimentary changes. I'm sure there's a more sophisticated approach that would work better. just need to learn about how it all works

[nimble heart]

even adjusting the width/height params on clip seem to do very little

setting base width/height to the higher res then targeting low res turns it from total soup into only garbled mess so maybe there's something there

in my Principled node I actually set the clip sizes to the equivalent base megapixel for the provided latent aspect which kinda helps??? but feel like there's better solutions

[upbeat summit]

clip text encoder width/height + target (bucket) values have a big impact on coherence, especially on subject proportions

[nimble heart]

yea I set width/height to the literal sizes (latent * 8) then target the equivalent aspect at 1024**2 megapixels

so if you feed in a 1920x1080 image it actually targets 1368x768

seems to very mildly help with duplicate composition at high res

[hardy cipher]

hmm. well I only started learning about this stuff 6 weeks ago when I downloaded comfy. so might need a little more time to learn about it all

with a1111 it was pretty much just abstract concepts for me

[nimble heart]

yea. thats why im sticking to comfy even with a1111 supporting refiner now; it's nice to understand things

also speed

a1111 makes amd cards even slower compared to nvidia

[upbeat summit]

humblemikey and I have build a node (Preset Ratio Selector) and the presets contain all the values following the current best practice that Stability AI suggested - at least like it was a couple of weeks ago. but it works. it's not 100% solving all the mutation issues, but it can greatly reduce them.

the longest side is always 4096 and the short side is calculated proportionally to your output resolution.

I created this preset file following this concept, containing all official SDXL Resolutions (from the paper), rounded all target bucket values to multiples of 16 and approximated the closest Aspect Ratios.
https://github.com/bash-j/mikey_nodes/blob/main/ratio_presets.json

[vital wraith]

Really liking titled upscaling >.>

Tiled even.

[nimble heart]

so you guys don't use Crop either?

[upbeat summit]

still exploring. also you can set it manually in the advanced node

[nimble heart]

have you tested it on > 1mpx resolutions? I found targeting lower instead of higher helps @ 1.5 -> 2 MPX if you're not upscaling

[upbeat summit]

the Preset Ratio Selector and Ratio Advanced node were build especially to explore this part of SDXL. the presets help to quickly try out different configurations

[vital wraith]

I use crop if my resolutions aren't likely to play nice together.

Jesus, I can just keep upscaling.

[nimble heart]

yea. kinda interesting you guys came to literally the opposite conclusion as me lol

lemme make a quick demo of what I'm talking about

[upbeat summit]

using 4096 on the long side and calculating the short side proportionally to your output resolution is something that Joe Penna discussed in the chat some weeks ago. he said it's their current internal best practice and I applied this to all SDXL resolutions for this preset file

[nimble heart]

must've missed that when I was out in the country cutting up trees

[upbeat summit]

yeah I think many have missed it

there's almost no documentation about this and almost all workflows don't use it

but it clearly improves proportions for ex. subjects - especially in extreme aspect ratios

I'm still doing experiments how to best utilize it

[glad grove]

🦶

[dim steeple]

After playing with XL for the first time I must say the quality in no way comes close to Midjourney 5.2

[vale eagle]

not as easy as mj but it could produce very decent images with dedicated workflow and prompts.

[nimble heart]

seems to make less of a difference now. in XL 0.9 the target low res trick reduced duplication quite a bit. Now everything looks like the same but you just used a slightly different sampler

[crisp owl]

just messing around, liked this

[vivid swallow]

@crisp owl you know how you said it cant take 10 minutes?
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [21:57<00:00, 65.86s/it]

[crisp owl]

Automatic1111?
What optimizations do you have selected?

[nimble heart]

you running off of a Voodoo 3 or something?

[nimble heart]

now we can be gecko friends

[vivid swallow]

persistent cond cache

thats really it

[crisp owl]

In your bat file

[vital wraith]

Lol wow, haven’t heard that I awhile

[vivid swallow]

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--port 2000 --listen --xformers

call webui.bat

[vivid swallow]

more like 3060 ti

[nimble heart]

tf you doing on a 3060 ti to get minute long samples

[crisp owl]

Try to replace your xformers with this
--opt-sdp-attention --no-half-vae --medvram

[vivid swallow]

want the prompt?

[nimble heart]

prompt shouldn't matter...

is this like a 4096x4096 image or something?

[vivid swallow]

1024x1024

[upbeat summit]

good test! thanks for doing it.

changing the values can alter the composition - depends also a bit on the prompt. I see the biggest differences when you do extreme aspect ratios like ex. 512x2048 or 2048x512 - things tend to be more proportional especially in realism.

that is at least my experience, but I'm also still experimenting. since I'm approaching this from a very non-scientific angle of not knowing what is really happening inside the model (can only guess and what I've read) everything could be different tomorrow.

[nimble heart]

batch size 144???

[vivid swallow]

1

[nimble heart]

I noticed extreme differences even @ 16:9 with XL 0.9
Guess they improved it

[nimble heart]

is that training or inference

[vivid swallow]

it was 1 image generated

[nimble heart]

did you just like not install cuda or something

is it running off directML?

[vivid swallow]

nope i have cuda

[shy kelp]

Oh boy. I've found that if I create an image with SDXL 1 in some other site and use it as input in nightcafe, but generate using SDXL 0.9 which is not restricted to pro users only, I get an almost exact copy of the original image.

[hardy cipher]

well they're almost the same model right?

[nimble heart]

I guess it just comes down to which images feel better. In my dozen or so renders that was usually the lower targeting ones. Someone would probably have to make a big 100-image grid to get something more objective.

Sytan was playing with it for a while and just gave up I think

pretty sure he just sets 4096 on both now

[upbeat summit]

I made a couple of experiments with a custom node setup I've build to iterate through the target bucket values - some changes were interesting in the compositions - like a progression.

but right now I would say the changes look closest to the ones you get when slightly changing the seed or surfing "between the seeds" like you could do with an extension for a1111.

[nimble heart]

guess I could bear that cross. have 9 'save image' nodes each named differently for every combination of low/mid/high clip resoltuions, render it like 20 times then use Magick to combine everything into a supergrid

[nimble heart]

to me its looks almost like those subtle differences between samplers @ medium steps. Like how LMS/Euler/DDIM converge in the same direction but they all get there slightly differently.

in fact let me re-render that last one @ 2.5x samples

[upbeat summit]

yeah or like xformers differences

[nimble heart]

maybe; can't use xformers

[vivid swallow]

do i just say fuck it and reinstall automatic completely? it seems like its gonna be 20+ minutes for this render

[nimble heart]

damn the amount of contrast between 20 and 50 samples

[hoary saddle]

See that deforum has a local install and can be used without a1111 I think. Anyone know if it can be used via command line via API or Python script?

[hardy cipher]

do what brings you joy

[nimble heart]

just delete the venv and let it reinstall torch and everything

or conda if auto on windows uses that idk

yea pretty subtle again

50 samples for each this time instead of 20

I'll probably just leave my principled node targeting 1 megapixel like it currently is until it's thoroughly explored.
I guess I could add a toggle? "Target size: Base/Input/Max"

[upbeat summit]

yeah. I let a sequence run from 1024x1024 and incremented to 30720x30720 for the target values - and of course I don't know when it's internally just ignoring my input (need to do more research), but the changes kept the same as in your image. there wasn't any coherence improvement really nor sudden bigger changes in the composition.

[nimble heart]

I thought they only trained up to 4096 anyways

[upbeat summit]

of course, sure. but can't hurt to try all the possibilities

even if internally something else is going on

[nimble heart]

little unfortunate. I thought the sizes would be more powerful.
Been trying to make some kind of workflow to latent upscale to 3840x2160 but no combination of width/height/targets makes it any amount more coherent beyond standard variance

[hardy cipher]

have you guys messed with things like negative_original_size or negative_target_size?

[nimble heart]

I even tried multi-step where it upscales 3 or 4 times in increments instead of going straight from 1368x768 -> 3840x2160. Helped slightly but not nearly enough

[hardy cipher]

some parameters here I really didn't consider

[nimble heart]

those aren't exposed in the nodes are they?

[uncut fiber]

is this layout of things normal?

[upbeat summit]

yeah I also wasn't sure what to expect. some interesting observation I had with a couple of prompts were having a perspective change - like "zooming in" on a subject.

using ex. 4096x4096 and than 3072x3072 as target values + the same seed and prompt gave me a similar composition but a different perspective - like being closer / further away in the scene.

but it's not something that I would call controllable 😉

[west breach]

I've just been using 4x the latent size for the most part and haven't noticed any problems with stretched people or twins

[nimble heart]

that closer/further is what I observed in XL 0.9
That's why I always set the target to 1024x1024 or it's aspect equivalent to keep the subject center-framed instead of just like some malformed speck in the corner. Seems a lot more rare of a phenomena in 1.0 though.

[hardy cipher]

not that I'm aware of, just looking at the list of parameters on hugging face

https://huggingface.co/docs/diffusers/main/en/api/pipelines/stable_diffusion/stable_diffusion_xl

[shy kelp]

https://m.media-amazon.com/images/I/51vMImAd8lL._AC_SL1108_.jpg can sdxl generate this? multi pose image?

[hardy cipher]

why not?

[nimble heart]

1.5 had a lora specifically for that

[upbeat summit]

charturner

[nimble heart]

yes

[shy kelp]

I asked for Elon Musk performing moonwalk. One of the images cut off his head. lol

[nimble heart]

I'm trying to reproduce the zoom even once and everything's so similar now they look like the same image from the thumbnails

you know for just generating some random 1920x1080 images directly without any 2nd passes it does pretty good

[floral island]

you're getting clones every so often tho

but that just happens with those resolutions

just showing some love around here

[nimble heart]

yea normally I use 2 passes for everything above like 1.5 megapixels but I'm testing some clip stuff

[uncut fiber]

1920x1080 @nimble heart is working for you well? Would suggest 1088 rather. I got more issues with 1080 then with 1088 or 1024

[nimble heart]

pretty sure it makes basically no difference beyond just changing the noise and therefore see

d

[uncut fiber]

difference is /64 both w and h. And it matters very

[nimble heart]

but idk could be wrong. you could make a big like 40 image grid of 1080 vs 1088 and see if there's any objective difference

[nimble heart]

Why /64? Latent scale is only /8 from input

[uncut fiber]

you are generating directly to this resolution, or just upscale. I must miss something if only upscale

[nimble heart]

I mean I normally upscale but those earlier images were all direct 1080p

that shouldn't make a difference either tbh

isn't it the same noise algo for upscale? It just doesn't start at 0

[uncut fiber]

o.k. I just saying from my experience. If it works for you then superb @nimble heart

[nimble heart]

What leads you to believe that 64 divisibility specifically is important?

[uncut fiber]

by testing images at high resolution. And watch them closely. Probably i had just bad luck...

[nimble heart]

direct 1920x1080 vs 1920x1088 just base 20 steps ddim

thoughts or nah?

[uncut fiber]

you must watch them i believe closer. But seems o.k. i will send you examples with stones, once i get to it.

[nimble heart]

that's a full resolution jpeg so you can zoom in as much as you want.

unless I'm misunderstanding "watch closer"

[uncut fiber]

you understand o.k.

[floral island]

Well... this certainly is a prompt.

[nimble heart]

boobs and skeletons and skeletons with boobs

[floral island]

but no mention of skeleton or boobs in prompt 😄

[nimble heart]

idk probably the "lust" part gives it the boobs

[floral island]

obviously 🙂

nudity in negative also pretty important tho, else boobage ensues

i like how it twitches back and forth between either pure skeleton knight or girl

[nimble heart]

is "occams razor" in the negative also important???

[floral island]

ah, that was a negative propmt i had for a previous prompt

[nimble heart]

why

[floral island]

there was some elements which did in fact, look like an occams razor

[nimble heart]

would you ever

need to remove that

[floral island]

so that basically killed it

[nimble heart]

its not even a thing

its an abstract concept

[floral island]

and how is it (often) visualized?

[nimble heart]

if you're getting razorblades just neg "razor blade" or "razor" why occam specifically what did he do to you

[floral island]

everything 😛

right now trying to get a feel for the neutral prompt ANDS

[nimble heart]

is ands even a keyword in auto I thought it was just BREAK/AND

[floral island]

neutral prompt adds a few more

and they work in different ways

[nimble heart]

is that an extension or something?

[floral island]

yeah

https://github.com/ljleb/sd-webui-neutral-prompt.git

it targets the AND composition in various ways, which can be helpful to get a better steering towards a wanted output

heh.

one of these is not like the others...

[ionic dragon]

which custom node pack has the node where where we can get resolution presets?

[crisp owl]

mikey nodes

[ionic dragon]

ok

[ionic dragon]

what is the node called?

[ionic dragon]

tq

[ionic dragon]

i couldnt find it

[crisp owl]

install and restart?

Try searching just with "latent"

[nimble heart]

"empty segs"

[ionic dragon]

i got many

[ionic dragon]

yea i did

[ionic dragon]

can you click on properties

and paste that code

like this

[crisp owl]

should very much be there if it loaded in as I see it did

If you're in another tab you're keeping active, make sure you hit refresh also

[ionic dragon]

can you share an image, generated via your workflow having this node

[crisp owl]

1 sec

sorry was in th emiddle of diagnosing something in my main flow

[ionic dragon]

ok np

tq

[crisp owl]

remove the pack and install directly from here
https://github.com/bash-j/mikey_nodes

[ionic dragon]

ok, or are there any aspect ratio nodes from different packs?

[crisp owl]

🤷🏻‍♂️

[hardy cipher]

lots of them, bud

[ionic dragon]

can you name one

[hardy cipher]

it doesn't really need to be that complex. but it does what you want

[crisp owl]

He means like Mikeys where it has the preselected options

[hardy cipher]

ahh, so the opposite

well anyway, just look up a list

it's pretty easy

[uncut fiber]

@nimble heart i recall what was wrong with images, there were clearly seen blends, but must be because unlucky seeds, today it seems moreless o.k.

[hardy cipher]

asked bing to alphabetize some stable diffusion parameters and it seems to have put some of it's own parameters in. or is just straight hallucinating. or both

[ionic dragon]

not this, i want preselected ones

[nimble heart]

were you using tiled vae or something back in the day? That's the only thing I can think of that'd make clear blend lines

[hardy cipher]

man, I got some pretty wicked ones a while back

[nimble heart]

wicked blend lines?

[uncut fiber]

no never. Still same, it is not so much ago, and in this installation of Comfui, must be unlucky seeds

[hardy cipher]

let me find the image, or images.

[uncut fiber]

@ionic dragon here are resolution sdxl should be trained on, so best ones.

[uncut fiber]

bunny in danger 😄

[nimble heart]

why have separate portrait and landscape tables if they just have the same data but rotated?

[uncut fiber]

yes, i didnt do it 🙂
probably for ppl quickly pick up, realy dont know

[nimble heart]

you gotta copy-paste each number separately regardless

old man yells at clouds

[uncut fiber]

weird is there is missing AR 1024x576 = 16/9 so apparently not training w and h.

[nimble heart]

???

why would 1024x576 be on the table

[uncut fiber]

it wont, because not trained, just weird it wasnt trained. /64 as all others, and 16/9 which ratio is missing.

[ionic dragon]

i know the resolutions, but i used a preset one few weeks ago

[nimble heart]

that table you just posted literally says

 1344x768   │ 1.75    │ 16:9

[ionic dragon]

but I dont remeber which one

[uncut fiber]

o.k. i believed in ratio

[hardy cipher]

still looking for the tiled thing, but I did find this

[uncut fiber]

it is not 16/9

[nimble heart]

1024x576 would be half the megapixels of the base resoltuion 1024² so it would make garbage images

[hardy cipher]

mine will give you 36/29 if you want

[nimble heart]

1368/768 is the closest you can get to 16:9 for one megapixel

[woeful patio]

is there a difference between "--medvram-sdxl and "--medvram ?

[hardy cipher]

one is sdxl it seems

that must be new

[vital ermine]

it is

[hardy cipher]

I gues being that it's sdxl it wouldn't be old. but now curious what it is

[nimble heart]

you can approximate any aspect ratio in SDXL's native res by

X = (1024²*R)^0.5
Y = (1024²/R)^0.5

where R is your ratio, so 16/9 == 1.6667 or 3/4 == 0.75

it's what I use in my automatic resolution node and it works well enough

[uncut fiber]

1,7777777777.... 🙂

[nimble heart]

yes that

I didn't go to highschool cut me some slack

[nocturne dove]

Some day the number of fingers will be right. But today is not that day 😢

[nimble heart]

she's just a lil handsy

you also gotta round them to 8 so 16:9 would be actually be 1368x768

[hardy cipher]

she's like those cats with the etra toes

I read you had to round to 64, but I never know what to believe so that might be nonsensse

[nimble heart]

the latent space is 1/8 res so I assume rounding to 8 works fine. Seems to so far.

comfy latent nodes step by 8 as well and he literally works for SAI so I trust him

can always round to 64 if you really want

[uncut fiber]

@nimble heart i cant speak English properly. Which is my big problem, i am more guessing what who write and so... So i cant argue 🙂

[hardy cipher]

that makes sense. I believe the encoder also does that for you right?

[nimble heart]

¯_(ツ)_/¯

[hardy cipher]

so if your input image is off it corrects

[nimble heart]

comfy's latent nodes don't let you make an incorrectly sized one so

could probably look at the VAEEncode source and see if it crops the image to *8

[hardy cipher]

yeah

I made my own

[uncut fiber]

not sure about img2img must try it

[hardy cipher]

not that it's any better

but don't have to run vaes to it

[nimble heart]

the encode just has the model load built in?

[hardy cipher]

mine does

[nimble heart]

yea

I wanna do that with upscaling

[hardy cipher]

wouldn't be hard

[nimble heart]

in my Principled node I have latent and pixel upscaling but I could probably just check the upscale model paths and load it same way the node does

[hardy cipher]

yeah, just look at the load upscale model node and see what it's about

probably pretty similar to loading any other model

[nimble heart]

I have to just do a full pass over my principled node again

probably remove some parts and try to add new ones

[hardy cipher]

are you talking all these things in one node?

[nimble heart]

yea

[hardy cipher]

that's a lotta node power

[nimble heart]

its that one I linked you the source to

to demonstrate using built-in nodes directly

so it just has a million params then chains together like 10 factory nodes to do everything for you

automatically splitting the base/refiner and some other stuff

[hardy cipher]

oh yeah. I learned a bit more about them

[nimble heart]

but yea I wanna add more shit to it but I don't wanna make it too much longer...

[hardy cipher]

I have a pretty long one myself

[nimble heart]

so like idk remove scale iterations cause it's jank, hopefully remove vae_tile once amd fixes its rocm issues, add upscale model options to scale_method, maybe add optional brown node inputs for clip vision?

since it has text it could parse BREAK/AND and automatically split the conditionings then concat/merge them

or maybe I could just go to sleep for once

[hardy cipher]

brown nodes, aye?

[nimble heart]

yea clip vision data is a brown spline in comfyui

so instead of taking an image and a vision model it could just take the brown spline

ControlNets would be hard though

lora doesn't matter cause lora's applied directly to the model but controlnets are applied to the conditioning

[hardy cipher]

not really to input would it?

[nimble heart]

to input in a sane way

without like a million extra values

[hardy cipher]

just look at controlnet.py

[nimble heart]

that's no the issue

for every 1 layer of controlnet Principled has you need 5 values on the node

ControlNet Image
ControlNet Model
ControlNet Strength
ControlNet Start Percent
ControlNet End Percent

if you wanted to fully support it

and god forbid you wanted to use two at once

could maybe make some abstract base principled node then make two separate ones. One with like 15 controlnet inputs and the other without

[hardy cipher]

I was wondering about stuff like that. how it could be dealt with. but not sure how much is allowed within the ui

[nimble heart]

I'm trying to at least have the node be able to fit on my screen

[hardy cipher]

those sparkling eyes

[nimble heart]

alternatively I could also make a version of principled that takes positive/negative conditioning inputs instead of text. Then you could just do whatever to the conditioning outside of the node itself

it'd mean the node couldn't automatically set the XL conditioning res though

that's why clipvision is nice cause the image encode is a separate node that makes a lovely brown spline

but idk how useful vision would even be

[hardy cipher]

include ip adaptor

[nimble heart]

is that even in base comfyui or is that an extension

[hardy cipher]

it's spelled out in controlnet.py. but not sure if the node is base or what.

[nimble heart]

ah

so do you just load it with the normal controlnet loader?

if so that's the same problem lol

ig you wouldn't need a start/end since it's just a reference image

but I don't like excluding things

[hardy cipher]

actually, don't listen to me. I only thought it was in there. so you know, oops

hmm, I need to figure this out now

https://github.com/tencent-ailab/IP-Adapter

[hardy cipher]

well okay

[nimble heart]

disney fae

[waxen cliff]

"do you know how I got these scar?"

[delicate grotto]

is there an inplaint sdxl version?

and does comfy support the stable-diffusion-xl-1.0-tensorrt model?

[vale eagle]

https://huggingface.co/diffusers/stable-diffusion-xl-1.0-inpainting-0.1

[delicate grotto]

sweet, thanks!

[vital ermine]

Damn, lol

[zinc cargo]

it might call to skeletons too, depends on who you are

@delicate kelp new protovision is fire!

these might make my wife more into AI stuff...

[empty osprey]

anyone there can help me with refiner in comfyUI?

[vital ermine]

https://civitai.com/models/140527

[upbeat summit]

oh it's out! nice

[vital ermine]

Go look at its pics and give it a try

I guess they are anti Pork cause the other pics with blood are fine.

[uncut fiber]

any text for test? GA? ML?

trying some new approach

[upbeat summit]

Legible text is hard for AI

[uncut fiber]

trying programing like way 😄 Just asking for some sample to know it is not random, curious if it could work

[upbeat summit]

yeah that was my sample 😄

[uncut fiber]

oh you mean Legible text ... as example 😄

[vital ermine]

you know dreaming the way you can tell if you are lucid dreaming or stuck in a dream is to see if you can read something. I use that tactic alot to know because the way AI makes writing is how you see it in a dream but lucid dreaming your right side is more awake and you can read it.

[upbeat summit]

yes ;). or if you want something shorter try: Stability AI
it's not that easy to get all the i's and l's right

[uncut fiber]

it is bad

[upbeat summit]

but it tried!

[uncut fiber]

it is 1st test

[uncut fiber]

It worked just temporary 😄 But will play with it

[waxen cliff]

like this?

[uncut fiber]

it is sdxl or that exterior thing anytime you want?

[waxen cliff]

it's SDXL

no edit

no img to img

this's just txt to img

[uncut fiber]

ok i can do glass hour, shorter, better

[waxen cliff]

I guess it help that my name doesn't mean anything so it can't misunderstand and make smt else

[uncut fiber]

o.k.

[waxen cliff]

🤣

[uncut fiber]

i type Bernix and got american vote for president 😄

I am there!

[waxen cliff]

lmfao those face tho

[woeful patio]

anyone know why my images come looking out blurry and faded?

[mossy canopy]

ui?

[woeful patio]

sorry, Using a1111 with the new controlnet

[mossy canopy]

give me screenshot of the whole page...

[woeful patio]

sorry, these ar my current settings.
I keep getting blurry pictures.
is it because of the model or the new canny control net..

hold on getting screenshot now

[native knot]

Well...first thing is that you should follow this sizing table: #✨｜sdxl message

Also, your CFG Scale is too high compared to the recommendations for that checkpoint. You might get something that works, but they recommend a much lower CFG at 3. (YMMV)

[shy kelp]

I think I've found a way to generate the same image again. Use image input with 0% noise and the same prompt that made the original image.

[woeful patio]

ah, sorry about that. Keeping up with current events is not my strong point

[native knot]

A lot of times it's just simply going to take a bunch of fooling with settings until you hone in on something you're going for.

Best thing to do is first understand what each setting does, then use that knowledge to tweak the image.

[woeful patio]

It seems to be a problem with controlnet, the more I tone down the strength the less blurry it becomes

[native knot]

When you are adding stuff that requires more to be processed, like cnet, you may have to compensate in another way. Have you tried making other adjustments? It's not really a "problem" with cnet.

[uncut fiber]

Just switched model, not 100% but like it

[native knot]

Friend of mine wanted his Type R "Batmobileafied", so I gave him a few options.

I also gave him this just to balance it out:

[ionic dragon]

why are 768x1344 better than 1024x1024 outputs?

[native knot]

I've seen a bit of the same, though I don't think that's a technical issue so much as it is just the subject matter and visual framing.

[west breach]

could be a lot of photos are either in landscape or portrait and not square

[native knot]

Exactly. Think of the idea that most subject matter isn't photographed/drawn/painted/etc. in a square format.

[ionic dragon]

Yeah

Human potraits are great

Compared to square

[native knot]

I think my favorite one so far in portrait is Hugh Jackman:

Just straight up looks like a legit photo...even zoomed in it's really good.

[shy kelp]

https://creator.nightcafe.studio/creation/YsmllJ1j24fGykvRdAUx I almost got an exact copy of the original image.

[weak saddle]

Where are you applying those values?

[native knot]

Because everything for SDXL has been trained against 1 megapixel input and each side should be divisible by 64. If you stray from the 1 megapixel dimensions, you're less likely to get coherence.

[crystal field]

#1072220168534642768

[woeful patio]

can you use the high res upscaler extention in a1111 with an sdxl model?

[woeful patio]

Correction, the problem was that is used the old vae, switched to the xl vae and it works perfect now

[weak saddle]

Nice. Where do I put the decimal and ratio? Im using automatic1111

[native knot]

Ah...good find!

[native knot]

You don't...just use the values from the table.

[weak saddle]

You must be using or doing something different than me to generate .. so im not sure how to use that table when creating images. 😦

[native knot]

Literally just put the resolution values into width and height.

[uncut fiber]

440 is too low for any model @weak saddle

[mossy canopy]

you just have to generate it in 1024x then rescale it to 440x

[uncut fiber]

@weak saddle screen is showing 440x728, it is not my problem 🙂 @mossy canopy

[woeful patio]

do y'all use the refiner on models other then the base?

[native knot]

Occasionally, but for the most part I don't use the refiner much anymore. Finetuned models tend to be good enough w/o that.

[glass mica]

Hi all.
Who can explain why I have different images, although the parameters are the same?

[native knot]

There's a difference between wherever you made the left one vs the right.

[glass mica]

I thought these models were identical: stable-diffusion-xl-1024-v1-0 and stable-diffusion-xl-base-1.0

[native knot]

Well...one has the vae baked and the other may not, but also, since we can't see things like cfg scale in the left image, perhaps there was a difference there.

[shy kelp]

what is the limit of dropclip free accounts? 400 gen per day?

[mossy canopy]

stable-diffusion-xl-1024-v1-0?

[native knot]

The checkpoint name in the left screenshot.

[glass mica]

on site: stable-diffusion-xl-1024-v1-0
localhost: sd_xl_base_1.0.safetensors (stable-diffusion-xl-base-1.0)
But I've also used vae baked. same situation

and cfg from 3 to 11

[native knot]

By chance do you have clip skip set to something like 2 on your A1111 instance? That would change the outcome.

[glass mica]

ouh, and where is it?

[native knot]

My A1111 is modified and I don't have the setting buried in the settings screens, so it shows on the main A1111 page. But you'd have to dive into the settings pages to find it. It was a common setting to change to 2 in SD 1.5 but SDXL doesn't need it set to that.

[mossy canopy]

same sampler? same gpu? if the seed was generated in gpu.

[glass mica]

Samplers used k_dpmpp_2m and k_dpmpp_2s_ancestral (since they come as default kinda).
I didn't know about GPU. I did one picture on the site, the other on my own graphics card.
Now I realize that it's the GPU

Thank you

[hardy cipher]

\

[glass mica]

But still, the seed does show which image to start with.
So it should still be the same image

[native knot]

Variances in cpu vs gpu hardware can impact results when dealing with randomness.

[late hill]

I tried to make some battle suit blueprints in ComfyUI.

[glass mica]

Checked, you're right that SDXL doesn't need it.

[native knot]

These are great.

[late hill]

They did turn out quite well 🙂

[glass mica]

But the starting seed responds to the final result (the image will be similar, but not 100%), and mine is not similar at all

[mossy canopy]

also different sites apply different optimizations to get faster gens. and optimizations changes seed drastically.. . so that seed only matters on that specific site .. other than that its useless.

[native knot]

Exactly. There is some other setting difference for what you generated on that left image versus your A1111 installation...who knows what that is behind the scenes.

[glass mica]

Thank you for the information

[wet nacelle]

STEVE!!!!!!

https://tenor.com/view/steven-seagal-eating-carrot-healthy-living-gif-9753152

[hardy cipher]

tbh I could watch him eat that carrot for quite a while

[shy kelp]

https://easywithai.com/guide/how-to-install-sdxl-locally/ I think I installed SD 1.5 then. I was following this

[delicate kelp]

Oh it's gooooood to have to ControlNet back in Auto1111! (First image is control)

[native knot]

The fact that it used the fro to make a tree, the clown wig, and the hat/moon are amazing.

[woeful patio]

hey there folks I'm trying to play around with a logo I'd like to get tattooed sometime and trying to get different designs. What is the best way of doing this, right now i'm using the depth controlnet but it's just not working right.

[native knot]

Use canny instead of depth.

But that 3rd pic you posted is pretty dope.

[woeful patio]

I'll give that a try thank you

[wet nacelle]

it's so dull

[cyan crown]

only prompt?

[mossy canopy]

yes

[woeful patio]

canny is giving me the issue that the background is also being painted

[cyan crown]

can you share prompt?

[native knot]

If you only want the inside painted, you should use inpainting to mask it off.

[woeful patio]

Ofcourse, I'm an idiot. Inpainting was the answer all along lol.

[mossy canopy]

1girl, solo, sky, wings, star (sky), blue bow, outdoors, short sleeves, bow, dress, shooting star, starry sky, closed eyes, blue hair, open mouth, hair bow, tree, night, smile, lake, ice wings, scenery, puffy sleeves, short hair, mountain, puffy short sleeves, night sky, ice, mountainous horizon, blue dress, facing viewer, shirt, standing, pinafore dress, cloud, reflection, blush, white shirt, :d, hand up, ^ ^, bangs, water, bowtie, red bow, grass, teeth, red ribbon, hill, collared shirt, cowboy shot, neck ribbon, hair between eyes, hand on own chest, upper teeth only, nature, blurry, ribbon

[cyan crown]

thanks. and for negative?

[mossy canopy]

easynegative

[cyan crown]

easynegative works also for sdxl? The "normal one" ?

[mossy canopy]

nope.. only upscaled with sdxl ..

[cyan crown]

so you start with 1.5 and then go with sdxl thanks. So the model of 1.5 influences final result

[mossy canopy]

yes

[primal vault]

I just released a bug fix update for Searge-SDXL to version 4.1 on CivitAI and on Github that addresses some of the reported issue from the v4.0 release.

[eternal fog]

Which model are you using for these?

[mossy canopy]

counterfiet

[eternal fog]

Hmm maybe I need to play with it a bit more, I tried it briefly, but never got any good results.

[eternal fog]

The blueprints that were posted earlier made me want to try it with IPAdapter and Revision

[cyan crown]

anyone here train Loras?

[mossy canopy]

yes.

[cyan crown]

so I have a problem with SDXL Lora

there are too big

1.7 GB

Can't understand why

[vital ermine]

your dim and alpha are too big

[eternal fog]

Use a lower Network Dimension

Aka Dim and Alpha

[cyan crown]

with lower values result are inconsistent

[vital ermine]

sounds like you did 128//128

[eternal fog]

Probably need better training data or different training values

[hardy cipher]

you do not need it that big, lol

[eternal fog]

I have 96MB ones that work perfectly fine

[hardy cipher]

from what I understand when they get that big they start taking over the checkpoint model

[cyan crown]

with 4090 50 images it takes 2hrs.... on runpod because on my 4070 it's infinite

[cyan crown]

can you share your json?

[eternal fog]

You are doing something very wrong if that is happening.

[eternal fog]

I've not done any training for a month, so anything I give you will be out of date.

[cyan crown]

Lora works very well after training, but they are too big and can do only on runpod 4090. But I can't understand where is the problem

[eternal fog]

I started based off of this that one of the SD Devs posted during 0.9 - https://gist.github.com/mcmonkey4eva/0f0bd074c17802213817a9a5a50098df

[eternal fog]

That uses 16 DIM

And I could get decent results within about an hour with around 30 - 50 images

on a 3080

[cyan crown]

with 16dim I get bad results....

[eternal fog]

But it's old now, so there's probably better settings.

[mossy canopy]

reduce network dim

[cyan crown]

tried but don't like results

[hardy cipher]

so here's the deal, buddo

either you lower the dim and all that, or it won't get smaller

so you can choose not to do that

but it'll be bigger

simple science

[hardy cipher]

https://tenor.com/view/the-universe-tim-and-eric-mind-blown-mind-blown-meme-mind-explosion-mind-explosion-meme-gif-18002878

[cyan crown]

So I'll stuck with this files 😦

[hardy cipher]

you should probably review your training images and how you're training the model

[eternal fog]

You'll likely need to change your training settings to get results you like with lower DIM

[hardy cipher]

but if you want to keep them that is allowed

[eternal fog]

You need to spend time tweaking it until you get what you want.

[cyan crown]

10 epochs and 256 maybe it's too much but I like results

[hardy cipher]

combining these two images

[eternal fog]

If it's just you using it then other than disk space I don't see the problem. But if you plan on uploading them to Civitai or somewhere people will bitch about the size

[hardy cipher]

whoa, it made a nsfw image. tsk tsk

[eternal fog]

Half the custom models randomly pop a tit out for no reason

Ok I obviously didn't try use Counterfiet much before

[hardy cipher]

I mean, it's not the most graphic thing I've ever seen. but still

[eternal fog]

BOTTOM

[hardy cipher]

weird back skin on that one

[eternal fog]

And the shadow makes no sense

I think it's sort of tried to combine the person and the tree, but in the wrong way

[native knot]

Maybe it's... foreshadowing.

...I'll see myself out.

[eternal fog]

Ok yeah, Counterfiet is pretty good

[mossy canopy]

harry in wonderland?

[cyan crown]

Wes Anderson

[eternal fog]

lmao, sus

[crisp owl]

kinky I guess

[wild ice]

What are the downsides to using SDXL based checkpoints, versus prior editions?

[near stag]

is there any documentation about these DW openpose arguments?

[near stag]

I need to make the openpose skeleton impact as strong as possible during generation.

[floral island]

best place for a firepit. next to the house

[native knot]

with the right materials it wouldn't be an issue.

Just pack on the asbestos.

[floral island]

can't get counterfeit to work for me 😦

[native knot]

😄

[eternal fog]

In what way, like bad generations, or it just won't do anything?

[floral island]

then again, regular models are good enough for me 🙂

nah, it doesn't work well with my prompting style

[wet nacelle]

I have found a use case for the nightvision model. Exterior house shots are really good.

[eternal fog]

I don't really use text prompts anymore

[floral island]

this is what my prompting style does 🙂 counterfeit just whacks out in that case

[floral island]

same prompt 😄

[floral island]

uhh... yeah.

got a lot on her mind tho

[eternal fog]

lmao

[floral island]

can't get that stuff working in anime models

except darkalfa which... well, it was my own mix

[eternal fog]

let me see what it does if I use IPAdapter

[cyan crown]

ipadapter exists for auto1111?

[eternal fog]

I don't use Auto, but I think it does, someone posted a screeshot with it in a bit further up

[eternal fog]

It kinda gets the idea, but not properly

[floral island]

i'm using neutral prompt too

so you might not get the best milage out of the AND_TOPK modifier

if you don't have it intstalled

[eternal fog]

closer

[floral island]

TOPKEK LOL

[eternal fog]

his big brother lmao

[floral island]

is that a hello kitty porcelain thingy with an m16?

[vital ermine]

yes

[crisp owl]

For those that use Lora's often, what's an avg amount you'd use in a single generation? Just 1, or more at a time?

[wet nacelle]

Oh hope that you don't like Limited Run too.

[hardy cipher]

I have a couple that I use often, and then I

I'll use maybe 3-6 total

normally

[crisp owl]

at a single time, or just cycled through?

[hardy cipher]

sometimes go nuts with them, but it doesn't always work out well

[floral island]

0 lora's, all the time, i'm still stuck in the prompting only era 😄

[hardy cipher]

at a single time, yeah. but I'll start with none, add 1 or 2

[crisp owl]

same lol

[hardy cipher]

well I was doing it without loras for a long time

[crisp owl]

gotcha, good to know, thanks 🙂

[hardy cipher]

but then I discovered how much they can do

[native knot]

No comment other than to say that I know Josh, (and Doug, but he's not with them any longer).

[cyan crown]

1 lora for me

[wet nacelle]

???

[wet nacelle]

I just know of the practices.@native knot

[native knot]

Josh Fairhurst, CEO of LRG.

[wet nacelle]

?

I don't care about the people??? I care about the practices. @native knot

[floral island]

🍑🔞👩‍❤️‍💋‍👩🌴💦 is my prompt... lets see what it brings us 😄

[wet nacelle]

the business practices.

[floral island]

i mean,i couldn't not do that on my old model, lets see if protovision does give me something more decent

[wet nacelle]

I LOVE STEAK!!!! OH MY GOD!!!!!

[wet nacelle]

canny?

[native knot]

What practices? If you don't know the people behind the company, how can you speak intelligently about their practices? If you aren't involved behind the scenes, you probably don't know a bunch of context that you would need to know to not be making assumptions about their business.

[floral island]

no, pure prompt

[wet nacelle]

From an in depth video I watched about them last year. https://www.youtube.com/watch?v=lS3ja5JnkAQ&vl=en

[floral island]

literally the thing i posted above 😄 (with some extra in the AND_TOPK for anime style)

[wet nacelle]

Oh wicked.

[native knot]

The same applies to the person who made that video. They speak about things they don't have all the information about.

I'm not defending everything about LRG, but I also don't base my opinions on a youtube video.

[wet nacelle]

Although the information they do share applies to the consumer. The consumer means a lot to a company and should not be put in a bad light that leads to lesser consumers.

[wet nacelle]

death stranding much.

[wet nacelle]

Then why base opinions off of articles or the newspaper?

[wet nacelle]

@native knot

It's rather incompetent to see that a person has spent the time and effort into researching a topic just for someone else to say "I don't base my opinions on a YouTube video"

This person fully understands that the video maker has spent months researching a topic so that they can properly make an informative video. At the end of the day they don't base opinions off of YouTube videos even when the video itself has months of research put into it.

no. We don't do that here.

[eternal fog]

lmao

Thats a fat package

[wet nacelle]

thankies

[eternal fog]

Probably pushes the line into NSFW when it's presented like that

[wet nacelle]

That's what I'm thinking as well.

[wet nacelle]

I'm not being a bully here mods.

[floral island]

lol that dog... thing

[wet nacelle]

https://tenor.com/view/what-you-think-about-california-pooping-kid-california-kid-what-you-doing-poop-gif-24749320

[cyan crown]

the only problem with SDXL and SD in general is that is a sort of drug

[floral island]

it is.

[floral island]

send in... THE FLEET!

[severe aspen]

with the new A1111 update is the refiner automatic?

[eternal fog]

I prefer space fleets

[cyan crown]

Yes, you have a voice

[severe aspen]

I don't see where to choose to apply it or not

[cyan crown]

go down and look for refiner section

[eternal fog]

I would assume it's ther "switch at" option. but I don't use Auto1111 anymore

[severe aspen]

that's it! thanks! it even says as much in the tooltip, which I didn't check

my bad

[floral island]

well then, now i'll have to run this prompt again!

[native knot]

I work in the industry. I've seen what goes on behind the scenes with my own eyes. I know multiple people that work at LRG as well as at several other publishers in the physical media space. I can tell you that there are aspects that I don't agree with in regard to their business and that there are other aspects that are perfectly fine. All I'm getting at is that what goes on behind the scenes versus what gets said in public are two different things in a lot of cases.

Apologies if it sounded as if I was saying you based your opinion solely on the video, that wasn't my intent. Also, I'm sure that neither of us want to continue clogging up #SDXL with an off-topic subject but I'm happy to continue conversation privately if you'd like. ☮️

[floral island]

kek

[lusty wolf]

My first image here...

[floral island]

and what an image it is!

[lusty wolf]

Thanks, I love the organic stuff.

[eternal fog]

I love what IPAdapter can do lol

[crisp owl]

heh, I was playing with some clowns as well

[native knot]

That is excellent.

[supple knot]

This are some really good images. Great Work!

Can you make that meme that does that face with the vein in this style

[noble shoal]

I don't make memes. Only serious art here 🤣

[noble shoal]

Is this with some kind of dressed / undressed lora? 😜

[supple knot]

I didnt ether but youtubes got me taging every video memes

idk this is what I thought of when I saw the previous image

[noble shoal]

I don't think that i can manage that. I cant even get the vein on his forehead.

[supple knot]

Im trying now, but werewolf

[crisp owl]

so THAT'S what was actually going on

[weak saddle]

Is there any kind of base line time it takes to train a sdxl lora? I know there are lots of factors but out of the box it seems like it is going to take some time. I have one now 1024,1024, 163 images, rtx 4080(16 gig), i9-13900K, 32gigs, is taking 5 hours?

[cyan crown]

With 4090 50 images for me is 2 hr

[supple knot]

where does it bottle neck at?

vRam?

[weak saddle]

That makes me feel better. i have a triple the images

[cyan crown]

depends on network dims and epochs

[steady grove]

looks about right. your big limiter will be the 16gb. i got the 4080 with a 12700k an 64gb. You could try improving those times by shaving it down to 896x896 resolution, with a batch size of 2

that should fit into 16gb still

[weak saddle]

the card ram is maxed out after a few minutes, but it is working. im over half way there with 6/10 batches done with about 2:21 hours left

[supple knot]

Only other way would be rent space on a cloud based server to process

[cyan crown]

Groot 2.0

[weak saddle]

batch size 5, epock 10, Network Rank (Dimensions) = 1

[steady grove]

learning ot max out the performance of limited hardware is good too. skills that can be used for efficiently using enterprise level hardware down the road

[supple knot]

I😆 \

I would love to train a Lora ,But dont know where to start

[steady grove]

Rant about Secret invasion. ||They ruined groot by turning him into a genetic bottle that could be cloned. now he's not unique or the last of his kind. he's just a manufacturable formula now. Secret wars kind of sucks and disney should write it out of canon like they did with the EU||

[supple knot]

Would it be like photogrammetry ?

[supple knot]

Gorgu!

[native knot]

Now do Mogwai

[upbeat summit]

Mogwai, Gollum, Grogu, E.T. - all cousins in the latent space

[native knot]

[supple knot]

john wick/gorgu

hes goggles are tight

[supple knot]

what would pika do with this image

[hardy cipher]

animate it

[floral island]

lets see if i can make this look even more like a visual novel 😄

[shy kelp]

do you know what sdxl does to conflicting positive x negative prompts? It seems that wet road conflicts with negative "flooded", ending up in "dry road"

[ionic gulch]

new version 5.0 of my workflow - better file names / directory structure, choose between own VAE and model VAE, clip skip, removed history, because it slowed down saving and ui too much, removed unused options from nodes, so you have to update both components
https://i.imgur.com/AeLm87h.jpg
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[crisp owl]

Out of curiosity, if you add weight to "wet road", does it start to make the road more....wet?

[shy kelp]

the bots here lack weight

[crisp owl]

ah never used them here

[floral island]

oh yeah, we are a visual novel now bois!

pure prompting power 😄

[wet nacelle]

https://tenor.com/view/blij-speelweide-gras-happy-jump-gif-18787703

[shy kelp]

I began to notice that the roads that SDXL is generating for me are, half of the time, kinda have a wrong perspective specially torwards the end of it.

[cyan crown]

I've a strange thing. All my lora works fine with ComfyUI. With auto1111 they all work fine execpt one that doesn''t. Any idea?

[mossy canopy]

unet tenc separate strength?

[cyan crown]

I've done all of them in the same way

same prompt

[cyan crown]

somehow Lora seems having less "strenght" in Auto1111

like 1/10

[compact hatch]

partially from weight normalization I guess?

[cyan crown]

could you explain?

weight is 1