#ANIMAGINE XL
1 messages · Page 1 of 1 (latest)
Wtf
Making 1024 is fast
I realized
1024 gives better results than 512
how to use hires fix

Using it on phone sucks
@glacial snow didn't think I'll see you out of pixiv
This default 1.0 VAE or some other?
@exotic falcon Do your thing
I can add link to comfybox (with mobile support)
Maybe later
looking for someone making nsfw on sdxl 
you can't do proper nsfw rn
well
I remember you got that nsfw dataset
just train it lol
I'll lend u some artists on dms if u want with same style
Do it, Because I can't zoom out nor zoom in normal
https://parsee-mizuhashi-comfybox.hf.space/index.html?url=<your_cloudflare_url>&port=443
^ change url + paste that in new tab
alright
I put the cloudflare url inside the <your>
right
i can't train a model, only lora,
and even i already train lora based sdxl right now, there is no model to generate it sadge,
well, you have to wait
@steady dome
?
Ok i realized
the slash after com
yeah
mobile falsely redirect yeah
use this
@rose beacon
Imma rewrite the thing if I have time lel
Alright
things to do:
slash on the end break things
mobile didn't work since it redirects to /mobile
instead of /mobile/index.html
ah, will fix
Doesn't work too @steady dome
I tried tbis too
A bit weird but, open this:
https://parsee-mizuhashi-comfybox.hf.space/index.html?url=https://boolean-rim-swedish-soldiers.trycloudflare.com&port=443 ✅
then open this:
ohh ok
so basically
@steady dome i cant get the url / port to be set as soon as first load
so i need to get one refresh before it works
yeah hard
and the codefor it is only in /index
also cursed
im gonna put it in /mobile/index
wacky, but it works
Didn't work
@rose beacon makasete
translate :3
oh @fading hollow, you can manually set the values in the settings too
does this work in nocrypts colab?
oh u r here
just click on the link it is a space (colab for this model only)
or you could download and use comfyui colab
or just use comfui colab
or just
lmao
i cannot download anything bc of my horrible wifi.. so i rely heavily on colabs
jesus christ how do people even understand what to do with that
jesus.. I want to try comfyui but that shit is going to make me puke
wish they could have a somewhat similar UI to A1111 as thats where im familiar
comfybox / stableswarm
oh?
im seeing they still dont have Lora support and some critical features for me I guess?
comfybox is missing cnet/inpaint
small price to pay for speed
oh thats the only thing? because if so ill actually install it now.. I dont even usually inpaint anyways
dont install cbox
what.. are you suggesting then?
uh.. what's this? really sorry as ive all been using is a1111 and im confused lol
first, install comfybox/start comfybox on colab
-rs nyaa
get the url, go to settings, do this
ahh ok
hyperpawams :3
Now win the contest
How many images were used when creating this model?
I think it was around 220k or idk i forgot, but lynn said it before i guess
@fringe token bro how to use damn hires
i already put it and linked with ksampler
Doesn't seem it's working
Wtf
how u don't know
hires is very simple
if you are referring to auto webui hires fix
no the comfyui
so you want hires fix from webui in comfyui?

I mean image upscale
Doesn't exist?
or i should use thing refiner
You said "how to use damn hires"
So what did you mean
^
ESRGAN?
You see, when i generate. The colors aren't fixed esrgan alone won't do it right for sure
I'm using sdxl vae
I don't understand what you are trying to do

HiRes fix, usually fix those things after doing it work. So if there's something similar tell me
Uhh, how to say it

look at the image
the artifacts
How many steps are you doing
why I am only now seeing this 😮 gotta test it now
Pretty short
20 is the max, any more and it says "this space is too busy"
you :3
Yeah it's all my bots
Who does even generate on hf
I tried to use it
But i got an error lol
@fringe token yo bro
I can't find json files
ok nvm
Zarchiver saved me
@fringe token now teach me how to use lora with this workflow
create add lora node after the model load node, connect them up and select lora and weight
Alright
@rose beacon
load lora?
@rose beacon so where i do link second clip?
On the right side
And model
Name string too
oh, the outputs are just replacing the old ones
so instead of "load model" model, its the "load lora" model
So I don't need to link those?
i just keep it linked to load model?
opposite
replace the old model to the new "lora'd" model
it should run
nvm it didn't work
it's like I'm not putting lora even though it's there
ah

should just be: remove all model / clip calls, put the model / clip into lora, use the output as the model / clip for all nodes
Tbh
Imma ask lynn to make a workflow with lora and hires at same time
lora key not loaded lora_unet_output_blocks_5_1_transformer_blocks_1_ff_net_2.lora_up.weight

the key is 🗝️
Yesh you have to update the custom nodes manually

Via comfy manager
Just one click away without restarting the ui
Using comfyui via phone seems a bit cursed
Ngl
We should make a better frontend asap
But the teeth merged with the skin
Literally reminds me of my old wd gen ngl
Seems so
Ah yeah
The movement sucks
Also
how to make seed random

set control after generate to true
to error reloading workflow data
TypeError: Cannot read properties of undefined (reading 'configure')
TypeError: Cannot read properties of undefined (reading 'configure')
at ComfyApp.loadGraphData (https://scripts-utc-demonstration-ralph.trycloudflare.com/scripts/app.js:1222:15)
at ComfyApp.handleFile (https://scripts-utc-demonstration-ralph.trycloudflare.com/scripts/app.js:1522:11)✅
Didn't work
Well i got an idea
I'll use the inspector to take it as json file
On #ai-art
nvm doesn't work
@fringe token i summon you
Ahh finally
I needed to update manually
@fringe token ok
The update?
yea i pressed them buttons
Ah i see
Sometimes I really hate this
there's lotta file
ok
Finally it worked
also nyaacwipt, vae download doesnt work
why
iirc it uses VAE and not vae
btw can we train lora atm?
wdym mirroring merged model from chinese forum
anythingv3 was merged by you isn't it

yes, but not in free colab
even with rampatch?
So it gonna explode
@fringe token
what is hyperparams. Why is he soo hyper

I used 1e^-6 I guess
who is dan and why is he a dataset
Niji
*techically humans are datasets
*
Well yes
ikr

Portrait examples are so overdone

Every model does them well so doesn't show much
That’s literally the standard are you crazy?
Standards can be bad
Lmao
Can you expand beyond meme responses?
If a model fucks portrait your doing something really wrong
But when only portraits are tested is my issue
Like every civitai model that only shows portrait
There’s a reason why the gallery exist on civit
Because the author too lazy to test more than portrait?
No to showcase your own results and tests
what do you want them to test instead? a handstand? full body?
Anything
a portrait then :3
A portrait instead of portrait?
yesh 
Anyway mrg if you want to see more than portrait the gallery is your friend
https://civitai.com/models/122533/animagine-xl
ANIMAGINE XL Huggingface link: https://huggingface.co/Linaqruf/animagine-xl Gradio Demo : https://huggingface.co/spaces/Linaqruf/Animagine-XL Train...
I speak in general not for this model
hot takes, but I agree with mrg
Why say it here?

Bad takes
||/j||



remind me tweet
no need to dodge semi-naked waifus
lol
y





they need some qk cache accelerated for preview functional only
whoops wrong channel
Bruh, that's why i failed to make cozy cloud images or scenery all the time
LYNN is AFK: sleep - 5 hours ago
can i save this ?
Wait, sdxl still requires danbooru tagging for anime models?
Sure
@fringe token have you tried using parts of our OpenNiji dataset in a model?
Since it uses nijijourney/midjourney style prompting it might counteract tag usage
Yeah fair enough
no model card big sed

Kek

yeah rip
feels like too much but it has happened 
so much was said with so few words
Welcome back yoinked [バーガー]!
Huh
yeah idk
theyre on a break, idk why delete messages here

