🏞｜general-with-images

[crisp stream]

i2i workflow Redmond

[wispy nest]

i should do some thorough testing with human forms

so far the model has nice texture to it @crisp stream

[crisp stream]

t2i workflow Redmond

[crisp stream]

glad to hear, I like the overall quality as well

pretty natural

[wispy bloom]

A Chinese Peking Opera character made up of pixels. Half of them are full faces, and the other half are pixels that gradually dissipate.

[royal monolith]

you can't generate in this channel

[crisp stream]

IPA+ (SDXL)

[crisp stream]

❤

SC

[grand sapphire]

asking for a friend

[calm canopy]

(((((booba)))))

[wispy nest]

the training data is taken out, the model doesn't know what to fill in with when you take off something to reveal the body parts

you'd be wasting time trying to train the 2b sd3 model

[royal monolith]

you are wrong. stop spreading this misinformation

[wispy nest]

if you try some suggestive prompt (not nude) you will notice how it misses the concept of natural body parts. the model lacks those data. you can't instruct it to produce something that doesn't exist in it's algorithm. you'd have to build in those area bottom up, and the most you can expect is for the community to do those lifting.

[royal monolith]

yeah, and it just failed to draw stems on the roses i had it lay flat on the table. it's NOT that data has been taken out, stop spreading misinformation. it's affecting more than just humans

[wise vapor]

/l’aumento dell’automazione influenzerà il ruolo dei manager e le decisioni aziendali

[wispy nest]

to break it down ... the model is missing two key things, its missing required information and its lacking concept

the concept itself can be trained, but it will require new data to fill in for the missing information

[royal monolith]

the model isn't missing any such thing. stop assuming and stop spreading misinformation

[hearty violet]

is really more fast download a model from civitai browser+ than from internet??

dont toke 2 minutes

when I download it direct from civitai page its like 40-1hour to download it

[royal monolith]

civitAI rate limits

[hearty violet]

I love it

Im tryng run SD on kaggle

but when I enter the link

give me this page

anyone knows how can I correct it?

[royal monolith]

don't use kaggle?

[wild sorrel]

It's not supposed to run on local adress if you're using kaggle, find another notebook or whatever it is called there

[hearty violet]

its my first time attempt

I like a lot two t4 the double of colab, I use colab before

[hearty violet]

the problem if I'm noob, im trynt put a cloudfare tunnel there with chatgpt, I put to play to test let see

[wild sorrel]

But keep in mind - Kaggle bans people without explaining anything, most likely due to not having censor measures in place or nsfw models in your notebook.

They ban both notebook author and all it's users. So you might want to make sure to have no nsfw stuff (even if you aren't using it or not generating nsfw at all)

only clean sfw + censor plugin

[hearty violet]

Ok I'm do it

its for generate imagens to architecture

I'm recording a course and will pass to students too

how can I check if a model is a clean sfw, what pluguin does it?

will help me a lot to keep me on kaggle

[wild sorrel]

idk, I never did that, I'm just one of the users who was banned along with notebook author, even tho I never used kaggle to generate nsfw lol
(I just later researched that they could ban for that)

[hearty violet]

jessus

[hearty violet]

can I dm you?

[wild sorrel]

yea, but I'm not a python guy, I don't code AI stuff, idk what I could be usefull for 😄

[hearty violet]

relax I'm not a programmer too

[crisp stream]

from what I heard it was truncated after training (there are AI-tools for that), so what besides the not correct depiction of other things than humans (that could at least in thought here also be a side-effect of truncation), leads you to saying there hasn´t been any post-training editing?

[royal monolith]

you've heard a lot of stuff from people making assumptions, wild guesses, spouting off nonsense as if it were fact. there was someone that posted stuff and said they were an employee of RunDiffusion. RunDiffusion posted this morning and clarified that THAT person was an imposter. ignore everything you have heard that has not come from one of the stability.AI devs or other members of the company

what it's looking like is that the three text encoders aren't in sync in various ways and using all three of them with the same prompt is causing issues

[crisp stream]

and if you don´t use the same prompt for all three?

[royal monolith]

i did some tests this morning - prompted for a long stemmed rose laying on a table. this is what I got when just using the triple clip loader node. the stem is totally gone. the head of the flower, no problem. the stem? gone. BUT

if i prompt for long stem rose standing straight up, no prob.em

HOWEVER

using the other node that allows you to put in a different prompt for each of the text encoders

the stem is there

still some issues, but there'

[crisp stream]

when using clip only (no triple loader) identical stuff happens (from what I observerd so far)

[crisp stream]

yes, looks better

[royal monolith]

which is why my question that I posted in the SD3 channel is 'what are these three encoders trained on and are they conflicting with each other" and that's what it's looking like

[crisp stream]

have you got that info from devs?

[royal monolith]

they are these: https://huggingface.co/google/t5-v1_1-xxl https://github.com/openai/CLIP/tree/main https://github.com/mlfoundations/open_clip

[royal monolith]

gotten what info. i've read what people have said. i know for a fact a lot of it is nonsense, based on the tech level of those coming up with it, i'm pretty sure the rest is nonsense too

[crisp stream]

I ment the other way around, like devs saying what the issue is?

[royal monolith]

sd3 medium IS still in beta, it is STILL being worked on, so i'm beta testing like i would anything else

[royal monolith]

they don't know. that's why i'm following rabbit holes and turning in observations

[crisp stream]

I see, so then there is at least hope in that regard 🙂

[royal monolith]

of course there is. it's a temporary issue that'll be found and fixed. but it's beta, and alex is regretting agreeing to removing the beta lable. he's said that several times now. and it needs to be beta tested correctly by the community

rather than having the non-technicaly community ocming up with wild ideas, posting them, and others thinking they are fact and repeating them

[crisp stream]

possibly it could help if there were official statements 😉

btw how come the API seems to run fine, so shouldn´t be there at least hints to why is happening what is happening?

[royal monolith]

there's nothing to officially state at the moment - we know there are issues, they're all over the internet. what would help is if people would stop running off without thinking and reposting every wild assumption they hear as if it were fact

[royal monolith]

we have a RAW model without all the rest of the coding that the api has with the model

and we are using 2B, which is very good with fine details and not so good with larger things

usually, the raw model is developed with, not just run as if it's a finished model

[crisp stream]

I ment like if the API is working properly it could potentially at least be possible getting hints from comparison more easily

[royal monolith]

i don't know, but I just managed to fix a warped dog - look

@proud dagger when i run a totally different concepts into the clip_g and t5xxl encoders, i don't get warping on the dog laying on the ground. you see no sign of the skunk which is the subject of the prompt for clip_g, you see just a fraction of the man that I asked T5xxl for - the subject of the image, the dog, is coming from the prompt to clip_l and the road is coming from clip_g, and (i assume) the coloring and lighting are being adjusted by all of the encoders. The only part of the subject that t5xxl has in its prompt, "a man" is his feet and legs in one image and a slight bit of something man-like in another image - and my seed for these images is - 1 - not the sort of long seeds everyone uses, just the number 1. timestep shift is 3. Steps is set to 70, cfg to 5.5, sampler is uni_pc, scheduler is sgm_uniform

@devout dagger

here's my workflow

[crisp stream]

still an official announcement in that regard (as well negating rumors that came up) can help a lot (even if the issue isn't clear yet).

[royal monolith]

dont' hold your breath. most companies don't put out official annoucements becasue they cause more trouble than they fix. it's a lot better if the community just gets a clue

the first thing that happens with ANY official annoucement is that the trolls take to social media and talk about how the company is lieing

and then twist and spin what was said

and THAT is what the clueless users pick up and run with. because of course, if it's a corporation saying somethign, it's automatically a lie

the entire issue wouldn't BE an issue if the users didn't do that garbage

[acoustic jacinth]

https://tenor.com/bXDzj.gif

[languid pebble]

Rumors are bad but you can face them with a good information politic ...

[royal monolith]

not really. people believe what they want to believe. and usually, that means the rumors and wild conjectures sound better to them than the actual facts. especially if it's something technical and they aren't technical at all.

[crisp stream]

At least generally speaking, from what we were told by an IKEA marketing manager who came to our studying course back in the days, it's said to best to react immediately/transparent. I honestly go along with that saying, from a consumer/customer perspective, regardless the image aspects for a company she was especially referring to in that case. Sure the internet back then wasn't what it is today, either way I feel it's a proper way of handling issues, being open/clear about it. Not doing anything also leads to stuff we have right now, like a wide open latent space of potential issues. Then for those truely interested in what's going on according info would be valuable information, as well allowing the official information to spread, meeting the rumors because there is at least an official, not rumor-based statement that can be referred to wheresoever rumors are being mentioned. At least here, not stating anything officially in greater context reduces trust and credibility being what the IKEA Manager classically was referring to, at least the way I understand and also feel it.

[languid pebble]

But it happens pretty often that people ask for the source of information. I always look like a stupid one, cause I don't know the people and I don't really care beside whethere they are a source or not ...

[royal monolith]

yeah, and IKEA has a fantastic reputation, dont' they?

i don't think i'd put any stock in anything the marketing manager for a company with a reputation of creating lousy products (undeserved btw) says

[crisp stream]

Well, back in the days it was at least sort of ok and if a company is doing questionable things in the first place the issue lies somewhere else and isn't a communication related one only.

[royal monolith]

the point is that what he told you to do, did nothing for fixing the PR issues for his own company. there's a reason companies STOPPED doing exactly that and turned to just ignoring it

[proud dagger]

the seed part doesn't matter (all seeds are equivalent)
but yeah the prompt splitting seems like a valid way to de-burn a bit cleanly

[royal monolith]

i really think the encoders are out of sync and don't understand concepts the same way

[crisp stream]

Could be, like you said, the internet having changed the situation now, when she told us it was along the time when BP had that big oil in the ocean incident (can't remember the exact reason, just remember it was bigger than the usual oil in ocean stuff). Then already prior to her visit I had noticed the reaction of BP (being open about it and saying they are intensely working on that situation (what they actually did, at least there was a lot of effort going on, from what I perceived), which in turn at least created a feeling of a company taking responsibility/care while also being transparent. Even if there are usually a lot of things involved with huge companies that define what´s going on above and beneath the surface, like said, at least from a human perspective it felt a lot better compared to other companies hiding behind a wall of silence in those kind of cases. And exactly in that sense I individually still feel today (even if the times might have changed). Saying I would be glad if there was indeed an official statement regarding SD3.

[proud dagger]

I imagine similar mechanism as why :0.5 works to decurse

[proud dagger]

oh ye they absolutely are

especially T5 vs CLIP is a big difference

[royal monolith]

can this be passed on to the dev team?

i went back into beta test mode yesterday

[lost gyro]

Can anyone help me solve it? I'm trying to use animate diff

[lost gyro]

@crisp stream

[languid pebble]

Moofi is on a horror trip again 😄

[lost gyro]

bro

Can you help me?

[languid pebble]

I've never worked with animatediff. Maybe you can try here? I don't know ... https://discord.com/channels/1002292111942635562/1025266140445933648

[lost gyro]

Have you ever done anything on video2video?

[royal monolith]

are you on Purz' discord? that's where you'll find a lot of people that can better help with this

[lost gyro]

not

send link invite for me?

[royal monolith]

i'll DM it to you

[crisp stream]

wdym, again..? 😀

[languid pebble]

Your pictures tell me ... 😄

Let's think about something positive ... humanity extinct, nature is taking over again ...

[royal monolith]

i think he means it's continuous, he didn't stop

[languid pebble]

Yes ... but sometimes he does 🙂

I should tell chatGPT it's a psychiatrist and the pictures are from a patient ...

[crisp stream]

"again" would need a pause in between is what I ment 😀

[lost gyro]

Please, check your prompts with a JSON validator. Full error message is in your terminal/ cli.

[languid pebble]

You sometimes do ...

[crisp stream]

who's the psychiatrist in this setting? 😀

[languid pebble]

Or who should be the patient?

[crisp stream]

yeah true, though I also simply do not post stronger stuff over here 🙂

[languid pebble]

Phew ... we are lucky 🙂

[crisp stream]

🤭

[languid pebble]

Don't watch the new Winnie the Pooh movie ^^

[crisp stream]

what about it? 😄

that horrifyingly bad?

[languid pebble]

It would have been a bad horror movie even in the 80's but maybe OKish those days ...

[past sentinel]

Experimenting with vintage mall pics:

[faint flare]

/img tattoo

[past sentinel]

Wish they had those rockets at the mall when I was a kid....

[languid pebble]

That could have been fun ^^

[faint flare]

How to generate a picture, what command?

[bronze jetty]

hi so im new here, where exactly can we put in prompts to start generating images ?

[languid pebble]

Not here ... you can check https://discord.com/channels/1002292111942635562/1237461679286128730

[faint flare]

Тогда ладно

OK thanks

[languid pebble]

Special channels but paid service ...

[late sorrel]

haha not bad

[bronze jetty]

ok cool thank you

[languid pebble]

70s style?

[past sentinel]

More of a 60s vibe that I was going for...

[languid pebble]

OK ... that's been to early for me ...

[grand sapphire]

cool

on avg it takes 40-45s to generate

pretty damn slow 🙂

how long does it take you guys to generate an image

[languid pebble]

Depends on to many facts ... in my energy safe mode with SD3 around the same time ...

[grand sapphire]

just give a number -_-

[languid pebble]

40-45 ... but that's OK for me ... I need some time for prompting 🙂

[past sentinel]

I mean, if your resolution is low and you're using something like v 1.5, then it can be 5-10 seconds or less if you've got a good PC.

[languid pebble]

And some Turbo or lightning models can do faster, too

[past sentinel]

But if you start adding loras and bumping up the resolution, and adding other bells and whistles, then it can vary so widely that it's kinda pointless to give guesstimates.

[royal monolith]

about 10 seconds

[late sorrel]

why they all looking my way? :3

[wild sorrel]

they know what you did

[languid pebble]

Tell me more 😄

[late sorrel]

oh shit... waddup...

now i have to hide my hentai magazines :3

[past sentinel]

She disapproves of your actions as well:

[late sorrel]

i like the old school style

[languid pebble]

Big sister is watching you ...

[late sorrel]

Big Sis 😮

[past sentinel]

Yeah, I like old school malls because they're both familiar but also otherworldly at the same time.

[late sorrel]

if you remove the people, it can be like a liminal space LOL

[wild sorrel]

I feel like it's not that different, clothes are old, everything else didn't change that much

expirience might differ based on location I guess

oh and floor color is wierd

[past sentinel]

That's why I still like 1.5...it can give you creepy and unsettling results precisely because it's not as "exact" as more recent models. And that's often exactly what I'm going for.

[wild sorrel]

yea I like AI weirdness sometimes too, even if it's something pretty much broken...

[late sorrel]

yea

[past sentinel]

I just like the unpredictability of 1.5. I think we'll definitely lose something as time goes on and prompt adherence gets better and better. Don't get me wrong, I'll use the newer stuff as they come out, but I'll always have a place in my heart for the bizarre stuff that 1.5 spits out.

[late sorrel]

im personally gonna keep using 1.5 for some stuff, cause that stuff is just not in sdxl

[languid pebble]

I still have a print on canvas of the first A.I. picture I kept ... I sometimes do remakes ... but not the same 🙂

[past sentinel]

Like this one below. The light fixtures are obviously "off" and could be fixed with some inpainting, but it gives it a dream-like quality that I kinda like:

[languid pebble]

I think it's part of the evolution of A.I. and us. At first I wanted that cool realistic stuff but now what A.I. don't wanna give us is more interesting ...

[clever oar]

Mojo

[languid pebble]

Hi Disco 🙂

[clever oar]

lol why disco

dicoXD

[languid pebble]

I thought there should be an "s" but there isn't one ^^

[clever oar]

i want disco but it imposible XD

rule in accaount policyu

[languid pebble]

Well ... people need a lot of alc to survive disco ... so it can't be that good 😄

[clever oar]

You reassured me 😃

Mojo

I rode 30 kilometers by bike

[languid pebble]

Happy not to need to go there ...

[clever oar]

difficult to walk

[languid pebble]

Muscles hurt?

[clever oar]

knees

[languid pebble]

So you deserve a good sleep with nice dreams of disco adventures 😄

[clever oar]

you've already revived your cat wearing a helmet using luma ?😃

[languid pebble]

Haven't tried Luma yet 🙂

OMG! lol

[clever oar]

is your pic XD

[languid pebble]

I remember that picture ^^

[clever oar]

this is the spirit of deformed pictures

[languid pebble]

I can do real cats, too 😄

[languid pebble]

Hungry now? 🙂

[clever oar]

need coock food

i buy ramen

[past sentinel]

Heck yeah! I actually have some ramen, too.

[languid pebble]

I have the dessert

[clever oar]

Is the snail real?nor for me)

[languid pebble]

Made of sugar ...

[clever oar]

my apple trees in the garden are infested with caterpillars😭

and someone stole the strawberry bushes 😠

[past sentinel]

(finishes off the last strawberry) 🍓

[clever oar]

😠 it was you!

[past sentinel]

Who, me? (looks around)

[clever oar]

bring it back

[past sentinel]

Abandoned lunar base:

[languid pebble]

I've done some Area52 pictures ... pretty interesting, too ...

[clever oar]

Reminds me of the levels from Prince of Persia 2

[pallid ruin]

So cute 🤗

[late sorrel]

aww thats a cute one :3

[late sorrel]

very cool pics Willingham 😮

[past sentinel]

Why, thanks!

[languid pebble]

In our German Channel we have a mother whose child has some snails ... that's why I sometimes play around with ideas 🙂

[royal parrot]

I asked stableaudio for a laughing gnome and got this horror. I hate it

[late sorrel]

there is a german channel? or you mean discord server?

[languid pebble]

Leonardo German Channel ...

[languid pebble]

It's dieing ...

[past sentinel]

I tried generating some period music (1920s) and all it gave me was a loud squelch.

[clever oar]

farting duck XD

[royal parrot]

Yeah I feel sorry for the poor gnome

[past sentinel]

Of course, I was using a huggingspace "space" as opposed to being locally run, so maybe some tweaking would help. But for now, I'll hold off until it gets better.

[royal parrot]

I don’t understand why the AI performs so badly when we’ve had things like OpenAI Jukebox or the Meta audio AI or Suno AI

[pallid ruin]

Lmao

[late sorrel]

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

[royal parrot]

Even the premium model on the site seems to go very wrong with many prompts

[pallid ruin]

The other day I asked it for sounds of people walking on the street and it somehow added a jumpscare scream

[clever oar]

why you need ai for this sound you can make it in toilet

[royal monolith]

neem oil

[languid pebble]

OK creative friends ... my bed is calling. Have a wonderful time. I hope I can still sleep after this sound ^^

[past sentinel]

I'm hoping for stable audio to reach a point where we can use something similar to 'img2img", like uploading an audio clip and it spitting out something similar.

[past sentinel]

'Night, Mojo. Don't let the diffusion clowns getcha in your dreams.

[royal parrot]

Sound similar but more energetic 😆

[late sorrel]

haha

[clever oar]

dont know what is it

[royal parrot]

It really likes screaming and anything scary-sounding. People making any sound just devolves into screams and moans

[royal monolith]

it's extremely effective against all sorts of pests

[pallid ruin]

The one thing I found it way better than any videogame is weapon sounds, specially when far away, the same with engines or collisions

I´ll try to make sounds and replace arma 3 ones with it

[clever oar]

cheap?

not cheap

i found

[stark sage]

How can I generate (transform) from the left image to the right image (almost clean bg and standard view)? I mean, is there any methodology a part from the prompt? the generative ai for image grows very fast and lot of info. I dont know from where to start to solve this problem.

[royal parrot]

Interesting. Perhaps it performs best when no living things are involved in the audio generation

Birds also don’t sound great

[pallid ruin]

I´ve seen a post like that on sd reddit, let me find it

Got it @stark sage https://www.reddit.com/r/StableDiffusion/comments/1ccj8j1/vton_magicclothing/
Sadly I haven´t tested it yet

[stark sage]

thank you!

[pallid ruin]

np, this is another way to do it https://www.reddit.com/r/StableDiffusion/comments/1dc8bq8/full_tutorial_workflow_comfyui_virtual_clothing/

Im not sure which uses less vram or have the best results

[stark sage]

I see. looks the opposite (from ideal image to model image). But I think it can apply to the model image to ideal image way.

[pallid ruin]

I love the lighting, beautiful image

[past sentinel]

Thanks! I used a lora I created using vintage night/city photos.

[crisp stream]

@languid pebble have you tested Cascade already?

[pallid ruin]

That looks delicious

[clever oar]

not exactly what I wanted, but at least I wanted to eat 😃

[pallid ruin]

Wait is this sd or real? 🤣

[patent haven]

@pallid ruin lol it could be both?

But it is A.I and this is general so it's open source of images

[royal monolith]

it better be AI or someone needs a lesson in how to clean the bathroom

[patent haven]

@royal monolith lmao

Oh it's A.I watch I will get some cleaning supplies

[pallid ruin]

Amazing gens, idk if its the lighting or what but they look so real

[patent haven]

@pallid ruin thanks

[radiant saddle]

Very nice! Is this using a lora? @past sentinel

[past sentinel]

Yep, I use my own loras for most of my gens.

[radiant saddle]

Ah cool ❤️ Did something similar. Trained a lora of my dog Napoleon and created 80's movie posters a while ago.

[past sentinel]

Nice! I'd imagine some 80s title like CANINE FORCE or K9 THUNDER.

BITE FORCE DELTA

[radiant saddle]

!!!

[radiant saddle]

Ok, off to bed. Nn everyone o//

[pallid ruin]

Goodnight, and very nice images!

[oblique hamlet]

In a dense forest, there lived a cunning fox and a mighty tiger.

[royal monolith]

you can't generate in this channel

[oblique hamlet]

plz tell me which channel i should use ? thanks a lot !!

[royal monolith]

if you're trying to create with SD3, you can do so in the artisan channels. read #🗣｜artisan-support-feedback first though

[oblique hamlet]

thank

[pallid ruin]

sleep time 😴

[grand dune]

mom.

[crisp stream]

not to shabby 🙂

[languid pebble]

Good morning! Still searching missing components. Should be more concentrated on only one thing ^^

[late sorrel]

Guten tag

don't forget to post a morning coffee pic :3

[languid pebble]

Let's have a good morning coffee 😄

[late sorrel]

nice

that will be in Cyberpunk 3077 :3

[languid pebble]

Never played that but I think I could start now with all that bugfixes 😄

@clever oar Found the complete picture again ... should've redone that ... faster than searching for old ones 😄

[late sorrel]

haha yea they fixed a lot and it''s way better now

[languid pebble]

I'll have a look at the GamesCon this year ... maybe I find something interesting.

[ripe bluff]

@graceful hatch

i think this could do

https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/main ( SD1.5 only )

for pretrained model

[wispy nest]

thanks. Oil painting with sd3, I think they look more aesthetic than photorealistic generations with it

[grand sapphire]

@rapid talon

[wispy nest]

didnt use any quality tag in the positive

[radiant saddle]

Generated some motivational posters to start the day o//

[clever oar]

what picture?

[fiery narwhal]

Hii - Hoping for some guidance, i have a portrait of a character I made using sdxl. I wish to use the image and then create a prompt so that the outfit can be designed as well. (Similar to a character ref sheet). I'd like to keep as much as the face design as possible. Any suggestions? Nothing has worked thus far, i'd be grateful! (reposting here, the other channel didn't seem to work for me)

[hearty violet]

why I can't use the controlnet depth anything?

says that I don't have the model, isn't there? I could choose it

[clever oar]

you choice right model?

in list

[hearty violet]

oh

wait

[clever oar]

if you choice second model what happen?

[hearty violet]

I will try the second

hey

the other is T2i adapter

maybe I have to download themodelto anything

will searh for that

[hearty violet]

thanks 🙂

[clever oar]

works?

[hearty violet]

sorry I was having lunch, i'm going to play now then I'll try and let you know!

[crisp stream]

👍

Does anybody know, when using an input image in Comfy how to define a resolution, regardless of the input image res?

[royal monolith]

it uses the AR of the image on purpose. what are you trying to accomplish?

[crisp stream]

with Easy Diffusion for example I can use whatsoever input image with whatsoever resolution while I define the output one individually. If an input image doesn´t have the required resolution the approach you mentioned doesn´t really work, unless I convert it to the preferred resolution first, that in turn is simply way more effort compared to simply being able to swap the input image on the fly with whatsoever resolution. Especially with for AI oversized photos taken with a mobile

[royal monolith]

i just pull images into photoshop and make them the AR i want them to be, then use them for image2image

[languid pebble]

@clever oar ;-D

[stoic lark]

Far too much work to get to this point.

[crisp stream]

yes, sure, that´s just too much effort comparibly when constantly swapping on the fly for example

[clever oar]

Mojo!

[royal monolith]

you might ask in the #🧣｜comfy-ui channel if there's a node that'll work

[crisp stream]

yes, good idea, ty

[languid pebble]

Nice ...

[clever oar]

thx i like it too

[languid pebble]

Gotta try it too!

[clever oar]

some result very good

i like underwater effect

[languid pebble]

Need to find a picture to work with 🙂

[clever oar]

promt guy flies up on a jetpack

[languid pebble]

It's rendering an image at the moment ...

@crisp stream got it?

[fleet goblet]

Animorphs.

[crisp stream]

no, not yet, though someone is replying in the comfy channel 🙂

[languid pebble]

There must be something like that in my SVD Workflow cause size of the picture I upload doesn't matter

[clever oar]

in promt she just swim

[languid pebble]

Greetings! Earthlings! 😄

[crisp stream]

got it at hand?

[languid pebble]

DM 🙂

[languid pebble]

New generation ^^

[languid pebble]

https://youtu.be/HmxbsQDRTSs?si=obRwPEH9Ivo86K_W ^^

Olivio Sarikas about SD3 TOS ...

[hearty violet]

hi there

exist other way to change the checkpoint more fast? I use colab, when I change it take a long time to change

[royal monolith]

he's wrong

[languid pebble]

I hope so 🙂

[glad stirrup]

Angel holy paladin guy... and luma nonsense

[royal monolith]

@pure monolith - SDXL - shimmering lake at sunrise: tall, snow-capped mountains.

[pure monolith]

It looks a bit low quality but otherwise good. Im using sd 1.5 and its actually doing pretty well. i really love environments so im hoping someone will one day make such a model for it.

SD 1.5

[royal monolith]

sdxl

[pure monolith]

Ill have to download the default sdxl and test it. its going to take another 3 days.

[royal monolith]

you want the default models for any release long before you want the fine tuned. and you want to learn how to use them. you will get drastically different results just changing a second decimal place with cfg - or using a different sampler. Treat every model like it's unique.

[pure monolith]

@royal monolithYes. And the default SD 1,5 is actually really good with environments though perhaps a bit limited in variety but generates great images. Its almost as good as i want them to be. Don't even need a model trained for it kind of. I hope SDXL default is even better.

[royal monolith]

yes. what I tell everyone - use the tool that does the job. dont' get caught up in the flashy new toys

[crisp stream]

Guess not necessarily what you are looking for because slightly varying, yet simply for demonstration purposes. Especially the colour range/dynamics/nuances/contrasts of Stable Cascade are superior compared to XL. Could especially be interesting for interesting landscape lighting.

[pure monolith]

One of my first SDXL generations. Just started using it yesterday. Better than SD 1.5 i think. Does better with fine hair detail, etc.

[scenic mural]

Has anyone made a tensorRT version of SD3? The one on HF isnt compatible with comfy

I get oom when I try to make one

[royal monolith]

probably not yet

[thin compass]

ehehe

Meowse.

[royal monolith]

too cute award

[late sorrel]

haha nice

[thin compass]

you know, I could warm up to Elmo...

[patent haven]

Lol

[thin compass]

go elmo, double fisting it

[late sorrel]

aww

[thin compass]

lookit that cute beer though

[late sorrel]

yea that too

[thin compass]

It's for getting adorably drunk

chug it back!

[unborn harness]

How do i create realistic photo with this kind of reference? big head blocky hand

[languid pebble]

Good morning coffee!

[languid pebble]

I can buy myself flowers ... ... 😄

[wispy nest]

when comfy leaves sai you guys need to question the validity of current management of SAI

[royal monolith]

stop

[wispy nest]

can you stfu please?

[royal monolith]

nope. you won't, so i won't

[wispy nest]

im not talking to you besides how you are slobering for obvious flaws makes you complict of the problem the compnay and users face

[royal monolith]

public channel. you're talking to everyone.

[wispy nest]

take some responsibility other than slobbering over what's wrong

[royal monolith]

you should look in the mirror while saying that

[wispy nest]

that's a personal attack

i will just block your face

[royal monolith]

and i will now report you to discord for harassment every time you post something insulting about anyone

[wispy nest]

those who are genuinely concerned about SAI and the services it brought to the AI world will understand there is something badly wrong with management when qualified and credible developers leave, even users are sharing their concern based on objective results how 2b model has failed them, going into defense mode will not go well for the company to acknowledge what's apparent to the community by now

[languid pebble]

NeoTokyo 🙂

[nimble mason]

best sd3 impression

[languid pebble]

Hillary Clinton?

[languid pebble]

OMG! They captured the Stargate!

[steep sleet]

SD3 - After Frida Kahlo

[languid pebble]

Reality is an illusion ^^

[wispy nest]

Nioce nioce

[languid pebble]

Thank you 🙂

[wispy nest]

[crisp stream]

SVD?

[languid pebble]

No a morphing of 4 pictures ...

[crisp stream]

How does the walking work then?

[languid pebble]

If I remember right animatediff is included ...

[crisp stream]

I se

e

[crisp stream]

Interesting, ty 🙂

[languid pebble]

I can send you in DM

[crisp stream]

sure, would love testing it 🙂

[calm jolt]

stable diffusion models can be fine-tuned to generate higher quality images even with limited training data

[pastel copper]

hippie much

[radiant saddle]

This looks soooo weird in that res.

[fleet goblet]

trying out pixart sigma for a lil.

My man is ready.

Full of energy (starch)

[hearty violet]

guys I'm running SD on kaggle but what is this interface?

the notebook it's called automatic1111 but it's differently, I never saw that

It's sd next? a skin automatic1111? lobe interface?

[gritty breach]

Jadeite rough cutting has increased in value, with a wide angle and many people present - a happy scene.

[south temple]

evening friends

I saw a things that say sd3 wasnt allowed on a cvit yet... is it released?

[royal monolith]

SD 3 medium - i.e. 2b, not 8b(ultra) - is out. it's a raw base model, it's unfinished, and it's still in beta test (you can't just drop in a prompt and go, you ahve to learn how to use the thing, it's a beast)

[nimble mason]

it was allowed, then it was banned

[south temple]

whats the difference in 2b and 8b?

[nimble mason]

2b has better details but creates anatomical nonsense

8b is much smarter

[south temple]

ah

worth playing with yet?

[nimble mason]

yeah it's fun to play with

but it is deeply flawed for sure

this got posted on reddit in case you missed it

[south temple]

ah, so they tossed out the garbage to the public

nice

https://github.com/kijai/ComfyUI-LuminaWrapper

neat

[nimble mason]

i just want the 16 channel vae

i'm at the limit of what you can do for detail with 4

[south temple]

never got that deep into vae

[nimble mason]

well i guess it's not so much the vae as just the fact it's 16 channels so there's far more information about the image

[south temple]

seems I walk away for a week or so, come back... and so much =has changed

[nimble mason]

yup

[south temple]

lemme drop that into a thing and see what your working with 🙂 mr messy pants

[nimble mason]

hahah

https://github.com/ClownsharkBatwing/RES4LYF you'll need this

[south temple]

unable to find workflow

maybe i need to update

[nimble mason]

ah shit maybe i didn't save it

i just reinstalled comfy

[south temple]

its okie 😛

you messy messy ninja

[hot sun]

帮我画一个LOGO图，品牌名称为御医门，主要在中国大陆地区开放中医功效型养生馆的加盟

[nimble mason]

lol this one is actually pretty clean

[nimble mason]

Here is the image you requested.

[hardy leaf]

#Clownshark Batwing

[south temple]

hashtag indeed

[nimble mason]

Here is the image you requested.

[south temple]

animation workflows suck from time to time

[nimble mason]

i bet

[south temple]

ok, promt me

[nimble mason]

Generate an image of a woobly deedle doo pickle missile rocket blasting clown shoe eelsharks stomping cars into pancake sausages for breakfast, sunset, nikon camera, dramatic photography, smiling portrait of a neatly dressed woman, chaotic existential psychedelic crisis

[south temple]

here we go

I promise nothing lol

lol

soon

I dont kow why....

but it chose this life

[south temple]

turned down the weights a bit

[clever oar]

I think she's grown something that a little girl shouldn't have😄

[languid pebble]

Good morning coffee! 🙂

[clever oar]

Mojo

I don't drink coffee often, but today I had a mug

[languid pebble]

Good morning! Only one cup in the morning 🙂

[wild sorrel]

well, public asked to release SD3, they got SD3 I guess 🫡

[nimble mason]

One of them

[languid pebble]

BTW Text is harder to get local now for me ...

Cover for my first single 🙂

[languid pebble]

By Leonardo DaVinci?

[clever oar]

anatomical illustration

[languid pebble]

Tried a map of hell by Leo some time ago 🙂

[clever oar]

I wanted the circles of hell Dante😃

[silent goblet]

#1237459938901491852 generate a dog

test

[topaz pelican]

https://tenor.com/bfRoM.gif

[smoky oak]

I LOVE how cursed my de-tangler makes loss lmfaoooo

[winter jetty]

https://brage.info/GAN/ganbot2/6e832397-be5c-4621-af2c-fbf629806e53.1.jpeg

Life's a hoot.

[deft bison]

in hell

[languid pebble]

Moin! 🙂

[languid pebble]

Minions 2045 ^^

[wispy nest]

😱

[clever oar]

Mojo

[fierce bramble]

A cat playing drums

[torn geyser]

#1237459938901491852 Create an image of a girl’s hand with the letter “M” carved into it, and blood dripping from the wound.

[hearty violet]

hi guys

why mi lineart anime is so bad?

it's the checkpoint that I used maybe?

[wispy nest]

Maybe try another one. I have never used that one myself.

[hearty violet]

It's for a architecutre course

I used one of building category

I will try with other, thanks

[hearty violet]

it got better

very different from this but better rs

[wispy nest]

I'm adding more emoji responses

[hearty violet]

hahahaha thanks ^^

[wispy nest]

nioce nioce

[hearty violet]

with prompt and negative prompt work better

[hearty violet]

the normal bae is the control mode for all the normal preprocessors?

[languid pebble]

Good morning (coffee). 🙂

[clever oar]

Mojo

[languid pebble]

Hi dicordos 🙂

[clever oar]

i maked coffee

I'm out of buns😭

and I have oatmeal cookies

[languid pebble]

https://tenor.com/view/cookie-monster-girlscout-cookies-girl-scout-gif-13360660

Cookies!

[wispy nest]

izum cookie mawnster

[grand sapphire]

50C AVG RAAAAH

[wispy nest]

https://youtu.be/yDlHykxYJGA

[orchid schooner]

cats in hats are the best

[languid pebble]

Old fashioned stuff 😄

[languid pebble]

Have you tried Leonardo.AI? You get 200 Credits everyday and they have that Phoenix Model 😍

[clever oar]

no i use sd3 online now

[languid pebble]

No A1111 for SD3 yet ^^

[clever oar]

🙁

[languid pebble]

We won but I hear nothing ....

[clever oar]

dont undestand joke explane me

[languid pebble]

Germany won soccer ... but it's calm here ...

[clever oar]

in your city>?

[languid pebble]

Yes ... no loud cars ... no sound here 🙂

[clever oar]

culture people

😃

[languid pebble]

It's been louder after the last won game ...

[clever oar]

sd3 very fun if need play with object

not with people...

face and anathomy very strange

[languid pebble]

You are right ... a mess in communication ...

[languid pebble]

Community models might be better ... it's always been so ... let's hope for the best!

[languid pebble]

So you are no moslem 😛

[steep sleet]

Ollama in ComfyUI (Mobius checkpoint)

[crisp stream]

no loRA used btw 🙂

[clever oar]

lora vibrators😎

[crisp stream]

Space suit-integrated 🚀

[crisp stream]

🤭

[crisp stream]

AI-supported technician 🤭

[clever oar]

today i first time play with sd 3

I'm surprised how bad people turn out

but everything else is good

[crisp stream]

get ready for...

[clever oar]

🧐

[crisp stream]

Btw the recent ones are done with a screenshot of @nimble mason´s workflow for an input image 😄

[clever oar]

processor drawings😄

how many times need for connect all lines?😃

[crisp stream]

LESER3K 7 😄

[clever oar]

what is it

[crisp stream]

loRA surveillance 😄

[radiant saddle]

Such a cool look!

All in 3.5K, full view recommended

[crisp stream]

Thank you 🙂

[radiant saddle]

What model do you use?

[crisp stream]

That´s Cascade with the Invictus Redmond v1.1 checkpoint

[radiant saddle]

Thanks, will check it later

[radiant saddle]

Really lov your style man ❤️

[crisp stream]

😄 Thank you ❤️ 🙂

btw those hi-res ones are quite interesting, especially regarding the crowds

[radiant saddle]

Thanks, yeah just experiments. Started playing with highres yesterday, still can't stop : D

[crisp stream]

still early 😄

[radiant saddle]

Demofusion starts dreaming at 4kx4k : D But the sampling is amazing

[nimble mason]

whoever designed that interchange should be sentenced to life

[radiant saddle]

LOL

[crisp stream]

very nice res 👍

[crimson flower]

So I tried to make a Suika Game asset.

[wispy nest]

@limpid lichen Also, they didnt add base categories only for pony, also SDXL lightning, turbo and hyper
Which are all literally still the same arch as SDXL

[languid pebble]

Good morning (coffee)!

[nimble mason]

Bring back stable cascade!

[shut sinew]

Loving those

[long folio]

How can I generate images based on a basic sketch?

like for example this

where everything around the black square should be, for example, sea and the square should be filled with fish idk

[shut sinew]

Id try my luck with the QR controlnets

[grave trout]

Hey guys. I'm using "AUTOMATIC1111/stable-diffusion-webui" with Pinokio 1.3.4 on Ubuntu, and I'm able to generate images, but the result is always weird.
I tried the exact same prompt as in prompthero examples, with same parameters, and it's totally different.
Am I missing something ? I use v1-5-pruned.safetensors as most people on prompthero. Here is an example : "Extreme close-up portrait of a boy embracing a bear, utilizing the rule of thirds, with engaging, friendly eyes directly gazing into the camera, set against a stylized interior background, melancholic ambiance infused with grunge elements, ambient fog embracing the scene, executed in 3D with subtle motion blur and a 25-degree tilt-shift perspective, captured in the essence of Peter Jackson's cinematographic style, lensed through a Sigma fp and Sigma 45mm f/2.8
Steps: 50, Sampler: DPM++ 2M Karras, CFG scale: 5, Seed: 714973533, Size: 512x512, Model hash: 1a189f0be6, Model: v1-5-pruned, Version: f0.0.17v1.8.0rc-latest-276-g29be1da7"

[patent snow]

maybe you're missing a VAE

or just use another checkpoint

the problem is not on the prompt imo

[grave trout]

Yeah I know it's not the prompt

Thanks. What's a VAE ?

[patent snow]

it's a thing you need to get decent results xd

it always comes with checkpoints

[grave trout]

Haha ok ok I think you're right

Hmm it's weird cause I already use VAE right ? I'm using the 1.5 model downloaded from "download models" menu in stable diffusion forge

@patent snow I think the turbo mode is a bit more accurate. I'll try to download models from specific locations instead of web ui

Do you know how to delete a model without deleting everything ?

Nevermind I got it

My app uses "/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned.safetensors?download=true" model, and for what I understand there is VAE in it. So I don't understand

I'll try to use the sdxl vae version

[shut sinew]

Hunyuan with SD3 refinement

[grave trout]

Quick question : is SDXL better than 1.5 ?

Lol LDXL with vae is even worst. I don't understand. What do you use to run stable locally ?

[small violet]

love

[languid pebble]

XL is trained on bigger pictures. you need a vae fitting to the model version.

XL VAE for XL Model, 1.5 VAE for 1.5 Model

Running ComfyUI and SD Forge here ...

[vale marlin]

.

[shut sinew]

A 4090 and ComfyUI

[grave trout]

@languid pebble @shut sinew ok thanks guys. I was using webui-forge and I just switched to automatic1111 webui. Using sdxl base 1 or vae version, the result is still creepy and bad. I don't undersand what to do. I have an RTX 3060.
Conf : export PYTORCH_CUDA_ALLOC_CONF="garbage_collection_threshold:0.8,max_split_size_mb:512" export COMMANDLINE_ARGS="--no-download-sd-model --medvram --api --xformers"
Prompt : Extreme close-up portrait of a boy embracing a bear, utilizing the rule of thirds, with engaging, friendly eyes directly gazing into the camera, set against a stylized interior background, melancholic ambiance infused with grunge elements, ambient fog embracing the scene, executed in 3D with subtle motion blur and a 25-degree tilt-shift perspective, captured in the essence of Peter Jackson's cinematographic style, lensed through a Sigma fp and Sigma 45mm f/2.8
Steps: 50, Sampler: DPM++ 2M, Schedule type: Karras, CFG scale: 4, Seed: 1459097033, Size: 504x504, Model hash: e6bb9ea85b, Model: sd_xl_base_1.0_0.9vae, Version: v1.9.4

[languid pebble]

For XL Model better ust a size around 1024x1024

[shut sinew]

sd_xl_base_1.0_0.9vae
What model is that

[grave trout]

Vae version of SDXL ?

It's in the official repository

[languid pebble]

And I would download models from civitai.com Cause they are finetuned and way better than the base

[shut sinew]

Also you dont need 50 steps

And maybe higher CFG

[languid pebble]

I think it's a version with reduced training ....

[shut sinew]

[grave trout]

Wow OK the result seems way better

[languid pebble]

They reduced cause original has been a bit overtrained ... but I am not 100% sure

[grave trout]

Let me try your advices

OK so. Just moving to 1024px did a huge difference

[languid pebble]

512x512 is for 1.5 Model

[grave trout]

Ok I missed that info

I'm trying the regular "sd_xl_base_1.0.safetensors [31e35c80fc]" right now

OK no difference, it looks good too

[languid pebble]

But for sure you can play around for landscape or portrait pictures but stay around original pixel-size

[grave trout]

Sure. Thanks a lot

[languid pebble]

Hope you'll have a lot of fun!

[grave trout]

A few more questions :

• Is DPM++ 2M a good sampling method for realistic result ?
• You talked about civitai.com, how can i find one ? I don't understand where to download it

How do ypou manage this error guys ? I already tried to fix it but some generations give me OutOfMemoryError: CUDA out of memory. Tried to allocate 160.00 MiB. GPU 0 has a total capacty of 5.80 GiB of which 93.50 MiB is free. Including non-PyTorch memory, this process has 5.69 GiB memory in use. Of the allocated memory 5.23 GiB is allocated by PyTorch, and 297.39 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[wispy nest]

I know the answer to this one, and the answer is, almost all the time!

[shut sinew]

Best fix would be a bigger GPU

[wispy nest]

I do not understand the second question there. But your boy/bear is adorable!

[shut sinew]

Just look around for models that have a lot of downloads usually

[grave trout]

Sure haha. I have a 4080 at home but not for the next weeks. Is there a way to avoid that, even if it's a bit slower ? I already tried "garbage_collection_threshold:0.8,max_split_size_mb:512"

[shut sinew]

Im not sure how to make lower ram GPUs work well

I dont wanna be that guy, but I rarely have that issue

[grave trout]

lol ok. I'll try 0.6 and 128

[grave trout]

It's not possible anymore I guess 🥲

[languid pebble]

SDXL Models are still available ... SD 1.5 and 2.X also

[shut sinew]

Yeah just no SD3, look for XL models

[languid pebble]

And have a look at the model descriptions ... creators often share infor about best stiings like sampler, steps and cfg

[grave trout]

OK thanks

[languid pebble]

And there is something you can add to the start bat file like -medvram or -lowvram ... I'm not really used to it but google might help ...

[grave trout]

Yeah I already use --medvram

[languid pebble]

OK 🙂

[nimble mason]

all a stable cascade lora i trained last night in just 25 minutes

i was able to clean up a bunch of that noisy look with a LOT of experimenting with my nodes

[languid pebble]

99 Nodes enough for you now? 😄

[nimble mason]

lol

@celest sigil you might be interested maybe

[languid pebble]

I guess it means: NOOOOOO!

[nimble mason]

hahah i don't know why i have all those other ones up

don't need em

wrote code that replaces them

[languid pebble]

See you later!

[past pelican]

Really like this one...

[nimble mason]

thanks

the WF is important too to get good results

it trained really easily

it might be helping that there's only one TE

it took 25 minutes to train it on a 4090 with fp16 mixed precision https://huggingface.co/ClownsharkBatwing/CSBW_Style/blob/main/csbw_1-7_r64a32_OT_cascade.safetensors

https://github.com/ClownsharkBatwing/RES4LYF nodes

[nimble mason]

that's one of the most challenging to generate

not seeing the usual problems detroying it

[celest sigil]

Take a look at my latest posts in #✨｜sdxl . I think I've got a perfect workflow at the moment 😌

[nimble mason]

Awesome, are they embedded

[celest sigil]

No, they're already large files without bloating them as .png 🤣

Im doing a partial denoise > increase latent size > inject noise and then repeating that 4 times before very low denoise UU

@nimble mason There's something about this I like #✨｜sdxl message

Must be the lighting

[nimble mason]

yeah that's killer

any chance you could drop a json? i've done something kinda like that but you might have found some real improvements on it, who knows

i've been trying to nail down the initial generation as when i was doing staged workflows, the initial composition would change dramatically

sometimes that was fine, but it ruined some stuff i was trying to do

[celest sigil]

Computer is off for the day now. I'll try and remember in the morning.

[languid pebble]

Thanks 🙂

[celest sigil]

Did you change name?! 🤣

[languid pebble]

No ... I was just reading and think it's always inspiring to see others workflows. Sometimes it just pushes you into another direction with your own ideas

If the try to understand them don't send you into an asylum ...

[clever oar]

Mojo

🖐️

[languid pebble]

Ahoi!

[clever oar]

I'll make coffee and eat a chocolate roll !

[languid pebble]

I'm having tea 🙂

[clever oar]

green:?

[languid pebble]

Black at the moment ...

[clever oar]

ahmad?

[languid pebble]

Hamburg has a big port so we are used to drink tea. But because the historic transport took a long time we are more used to black tea. Nowadays you of course also get green one.

[languid pebble]

No cookies? 🙂

[languid pebble]

Shark-Dolphin-Whales?

[analog blade]

Clown football manager

[clever oar]

How to sit at computers turned to the side?😲

[languid pebble]

Maybe there's a display behind the camera, too so he only has to pull the keyboard to himself

[languid pebble]

We have a swimming bus here but not for underwater ^^

[clever oar]

help him be underwater

[nimble mason]

Here is the image you requested.

[languid pebble]

I don't think it's healthy ^^

[languid pebble]

He looks like a German ^^

[analog blade]

I wanted an England football manager 🤣

[nimble mason]

Here is the image you requested.

[analog blade]

Can you put in a white kit with a English flag 🏴󠁧󠁢󠁥󠁮󠁧󠁿

[languid pebble]

https://tenor.com/view/computer-says-no-carol-beer-no-way-little-britain-trendizisst-gif-14015902623456325888

[languid pebble]

A cake?

[clever oar]

Elbphilharmonie made from cookies

[languid pebble]

A.I. doesn't know it 😄

[clever oar]

If only you saw what she did at the request of the Church of St. Nicholas..

[languid pebble]

lol

[clever oar]

what is it?

difficulty 3 out of 10

[languid pebble]

"I'm sorry Dave, I'm afraid I can't do that" 😵

[languid pebble]

No idea ^^

[clever oar]

Hamburg City Hall

[languid pebble]

A.I. is pretty bad in creating Hamburg 🙂

We have a damn nice City Hall ...

[languid pebble]

Fingerfood 🙂

Haha ... so many telling me they want to buy my artwork at Twitter ^^

[clever oar]

headcrab from half life 2 made from biscuits😃

[languid pebble]

Only 3 days left on Suno ^^

[clever oar]

how to live further?

[languid pebble]

At first I'm trying to sell my songs on Amazon and iTunes ...

[past pelican]

Is Suno going subscription only or something?

[languid pebble]

If you want to keep the rights for your songs you need subscription ... otherwise you can use free tokens

I've had a free month subscription with coupon. But for my songs I need creative ideas and short stories. So it's hard to create more than 7 good ones in a month for me

[past pelican]

So if you don't subscribe they own your songs?

[languid pebble]

That's how I understood it ...

You can use personal but not commercial ... I think it's okish

I only used about 15% of my tokens and think the price is fair 8.-/month

My profile ... https://suno.com/@mojoyates

[wild sorrel]

Do you still get 50 daily tokens on sub? I wanted to do 5$ sub, but without daily that's pretty low amount of tokens...
1500 free vs 2500 5$

[languid pebble]

I have 500 Extra (free) Tokens ... I think you get them monthly ... but never really cared about it. So I'm not sure. I'd think the point is what you like to do with the music. For personal use you can stay for free. But if you want to use them on youtube, etc. you should buy an abo

[wild sorrel]

500 extra was when they did maintenance and suno wasn't available

[languid pebble]

Monthly Credits Left: 2290/2500
After 7 good songs and some trials at the end of the month ...

[wild sorrel]

man...I spent probably over 2k trying to do one thing which still isn't ready (not the way I want it)

I re-extend alot to make song sound as I want it, just generation rarely works as I'd like

[languid pebble]

Ohhh ... hmmm ... maybe my songs aren't really good? 😄 Yes sometimes Suno fools around and does it's own thing ...

With the new Version Suno has become pretty good in generating a whole song at once ...

[wild sorrel]

I'd probably do that better, but song I want to do is pretty long and suno bugs out sometimes, like generates parts of "old" lyrics instead of adding ones you prompt or just getting stuck in the same style (which isn't bad, but not what I always want)

I'd probably do that in less generations*

[languid pebble]

I have the same problem sometimes... but sometimes I think it works better the way Suno did it 😄

Yes, I think I'm more simple 😄

[wild sorrel]

yea I just looked through your songs

my thing is a bit more complex...

I'll share it a bit later, need to figure out a way to watermark it lol

[languid pebble]

🙂 Looking forward to it

[wild sorrel]

This is kinda where I gave up on it and decided to go back and remake some parts...
it became too messy and completely refuses to make a chorus 😦

[languid pebble]

Ohhh yes! A friend is trying to do something as complex as that using Udio cause she says it's better with classic ...

[wild sorrel]

I wanted a song to progress, became progressively darker, that's main reason I re-extend alot

[languid pebble]

I always hated the extend thingy 😄

[wild sorrel]

it's very broken...

AI doesn't understand where it is, adds wrong lyrics, messes up song lyrics completely too, cause it keeps older ones even if song doesn't have em enymore

But also often it generates good part, but then messes up, so to keep only good part, no better option but to extend from there...

[languid pebble]

Same problems with 3.5 Model?

[wild sorrel]

yea, nothing changed

It feels slightly better at understanding song rhytm , doesn't mess up song structure as often, but it still happens + tons of other issues

[languid pebble]

My male voice just became female 🙂 Must be a diverse artist 😄

[wild sorrel]

oh you have no idea how many times my girl lost her voice 😄

[languid pebble]

Oh yes ... that will also make it much harder ^^