🏞｜general-with-images

[thin prairie]

cat

[brazen sand]

I'm new to Comfy, I haven't try upscale yet, it seems I have to set a lot of things

what workflow and resources do you use?

[wise turret]

https://civitai.com/models/549838/phoenix-dota-2-sdxl-pony Do you guys think it'll be possible to get a humanoid girl out of this lora or no?

[copper matrix]

seeing how image 4 is a female body and phoenix face id say yes

actually most of them are bru

[wise turret]

I meant something like this

The images shown are anthro

[copper matrix]

then no

isnt that more like a harp/ angel with energy tail thing goin on

[wise turret]

Idk lol

There's not much art out there of a humanoid girl like this unfortunately so I think it'd be hard to train a lora for this

[copper matrix]

idk, you can generate a bunch that look like it and use that for character training, probs low quality end result tho

[wise turret]

Ah

I'm gonna try this and see how it turns out

Ah shoot I meant to make hi res steps 20

Oh well

Maybe if I make the lora less than .5 it'd work out?

Idk?

Uhhhhhh

What happened here??

[copper matrix]

74 sampling steps? Bro what

Tune it down to like 28 max

[south temple]

I have been in the studio all dang day. Working on some new live music till my eyes crossed looking at screens. Feel free to jam to this unreleased stuff and I hope it helps the creativity.

[fleet goblet]

I use my own built workflow.

But in reality, it's just Flux Dev + 4 anti-burn nodes to allow me to use a cfg above 1

I would use dedistilled but dedistilled takes more steps

le squid

Is dedistilled really worth it for the better comprehension?

https://civitai.com/models/843551?modelVersionId=943787

I wonder how strong the text comprehension is when you use dedistill with anti-burn nodes

with the dedistill you can use dynamic thresholding mimic scale cfgs of around 2-14 and max out base cfg to 100

[fleet goblet]

14cfg outputs are fine it seems

Interesting that I can mimic scale 14cfg and unlock it entirely in ksampler using dedistilled

[brazen sand]

Flux here takes like 5 min to generate so I'm just testing loras and some workflows like img2img there I play a bit with denoise and steps, but not much.

it was much easier to test stuff with SD XL for example

With Flux it is like Interstellar time

I haven't try different samplers for example... in my experience I could only tell about something like that after generating ton of stuff, and here I'm not able

[wispy nest]

IDK if flux is worth waiting 5 minutes

if you can find a flux look you like and train an SDXL checkpoint on the output, you can get kinda close

[fleet goblet]

I'm actually waiting 10-12 minutes for an image, but it's for a different reason ofc

lmao

[fleet goblet]

\

[sterile kiln]

Nvidia Cosmos, 768².
Img2Vid

[grave quartz]

Please help me

Extension not available due to error below

[sterile kiln]

A1111 on colab... 🤮
What are your PC specs? You maybe don't need it

[grave quartz]

Graphics board is NVIDIA GeForce GTX1660 6GB.
CPU is Ryzen7 5700G with Radeon Graphics.

Stable Diffusion is not a specification that works in a local environment.

[sterile kiln]

I don't quite understand, I guess your translation tools have trouble translating from Japanese to English.
You think Stable Diffusion doesn't work locally? Of course it does x)

You should know, even with your PC specs, you can use A1111/ForgeWebUI/Comfy_UI, with lightweight models such as SDXL, Pony, Illustrious, and probably even with quantified Flux models which are very lightweight. Like Flux1Dev Q2, or Q3, for example.

[grave quartz]

I tried it a year or two ago and it was impossible, so I have been using it as a Google colab.

I will try the lightweight model once, thank you

[tulip breach]

اسے پن کرنے کے لیے کلپ کو ٹچ کر کے دبائے رکھیں۔ 1 گھنٹے کے بعد پن ہٹائی گئی کلپس کو حذف کر دیا جائے گا۔

[sterile kiln]

Does it looks like arab language Discord server?

[copper matrix]

People dont check what channel they are in too, so many english speakers posting their problems in general

Like dam

[spring kelp]

fr, no need to be a dick

[copper matrix]

Oh in not nm trying to be like i meant the techsupport channel

I realize i hit sent before finishing and forgot about it

Apologies

People dont check what channel they are in too, so many english speakers posting their problems in general instead of #🤝｜tech-support ***

[spring kelp]

I was talking about that other dude

not you

🙂

@copper matrix

[copper matrix]

Ah but still. My comment sounded a bit rude

[viral frost]

It’s a bot anyway. The message translates to “Touch and hold a clip to pin it. Unpinned clips will be deleted after 1 hour.”

[timid dock]

Anyone got a good way to generate DND maps ?

[spring kelp]

With stable diffusion of course!

[astral gorge]

could someone take my trash server logo and animate it with ai? Trying to make a gif for an animated verison of this.

[copper matrix]

Youd be better off having the background as a separate image, animate and then plop the text on

Ai will make the letters into a mess

[sterile kiln]

Man, what model it is?

[wispy nest]

SD 1.5 models with "Mix" in the name can look like that

[analog goblet]

has there been much changes in the last 3 months ? i didnt watch the news kinda
i can't seem to find improved pictures in any of the channels here sadly

[wispy nest]

the mighty sweet Illustrious

[analog goblet]

but that's still 1.5 based ?

[wispy nest]

SDXL

[analog goblet]

but not 3.5 ?

[wispy nest]

not that im aware of

[analog goblet]

there's not much going on with 3.5 ,right ?

or with stable diffusion at all ?

there's not much progress in the last months ? at least i couldnt find any

[wispy nest]

can't tell if you know about flux or not

[analog goblet]

i know about flux, yes

[wispy nest]

ok everything kinda moved there

[analog goblet]

but flux is kinda a dead end, it's very very limited style wise

[wispy nest]

flux / sd3.5 realism wasn't cutting it for me, i prefer illustrious and its sweet aesthetics better

[wispy nest]

especially for anime

[analog goblet]

i hoped that sd3.5 would be used as base model and finetuned / trained into infinity

but that doesnt seem to have happened ?

[wispy nest]

might work with sd3.5 medium but i havent' seen much happening

[analog goblet]

but why not ? what are people doing instead ?

[wispy nest]

dunno, but i personally got bored with those styles

[analog goblet]

i was (and am) always very hard trying to find new styles of art with this

[analog goblet]

what do people want with all these anime girls ? they look so replace-able to me ?

[wispy nest]

replaceable?

what do you mean

[analog goblet]

it's always the same thing ?

[wispy nest]

cos its hard to understand why to fine tune SD3.5 instead of flux

they are not the same, i think variations with sdxl breed models are captivating

[analog goblet]

because flux has a veeeeeeery limited range of styles

[wispy nest]

but we're talking about fine tuning so you are replacing the base style with your own

flux is isolated in a way that not every consumer can use it

[analog goblet]

i found that you can't push flux beyond some things, no matter how hard you try

it was very annyoing

[wispy nest]

flux body variations are a struggle

[analog goblet]

i'm not into anime though

[proud glade]

how the hell do i remove ears from image
i tried putting it in negative prompt, i tried writing in prompt "no ears seen on image", i tried drawing hair and using low and high strength, theyre always here

[wispy nest]

on the one hand yes, but on the other hand SD 3.5 anatomy issues seem to be even worse than Flux
so I am not sure it is a good trade

[analog goblet]

withdrawn: try in your prompt not to write what you DONT want, but what you want instead

[proud glade]

medium hair also doesnt work

[analog goblet]

what is your prompt?

[wispy nest]

all flux has to offer is highly detailed texture but within a limited variation

[analog goblet]

yes, flux is very boring actually

fun for a short while

but very limited

[wispy nest]

i stopped using flux / shuttle3/ sd3.5

[analog goblet]

try with "earless male" in the beginning, instead of "male"

[analog goblet]

i'm not trying to down those anime girls, but i m honestly curious what you are trying to create

i personally try to create new forms of art, new combinations of styles that haven't been done before (with more or less success), that's my goal

[wispy nest]

im playing with prompts

[analog goblet]

for me those last two pictures are exactly the same

[proud glade]

doesnt work

[analog goblet]

try to not tweak the prompt but stuff like the cfg value

[wispy nest]

which model are you using

[analog goblet]

varies a lot

i do a lot of wild experimental stuff

with img2img and similar

[proud glade]

tired 4 and 3 with different strengths, doesnt work

instead of 7

[analog goblet]

show your workflow pls

[proud glade]

its krita with SD plugin

[analog goblet]

i dont know that

[proud glade]

its basically every other workflow but in app created for drawing

[analog goblet]

i m using comfyUI stuff

like that last image is this workflow:

model was epicrealism_naturalSinRC1VAE

plus flux with ipadapter

plus sd3.5large

[wispy nest]

wow 3 different models

[analog goblet]

they're in a pipe

i generate something with the first, the use the input as partial base for the next

(for example let the next generated image not start at step 0 but at step 5 or so)

[wispy nest]

yeah that's good, I do that with Flux followed by SD 1.5 or SDXL sometimes

[analog goblet]

another thing that i'm trying is to create one workflow that will create vaaaaastly different results with the very same prompt

like i got this here:

this

all with the same prompt (just different seeds)

[wispy nest]

hmm yeah that's cool

I would recommend CADS https://github.com/asagi4/ComfyUI-CADS

makes SD and SDXL more variety

[analog goblet]

still same prompt

yeah, noise is magic 🙂

[analog goblet]

a bit of a different style here 🙂

(still same prompt)

but what are you trying to achieve, @wispy nest ?

[wispy nest]

what do you mean ?

[analog goblet]

are you just doing random anime girls for fun ?

or are you trying to create something specific ?

[wispy nest]

playing

other than merging some models now and then

but this is mostly for fun recreation

[analog goblet]

but you#re not looking for something specific ? do you create images that you yourself think of as "wow, that's a great outcome" ?

[wispy nest]

I don't know anime but apparently its very complex, there are lots of subtle differences between anime art styles

i have specific expectations thats why i try specific models

like we have one word for "anime" but something like Studio Ghibli and Dragon Ball Z are super different

[analog goblet]

so what -are- your expectations ?

[wispy nest]

the art

why is that so curious for you?

[analog goblet]

because i see that a -lot- of people put a -lot- of effort into this

and i dont understand it

but i would like to understand

[wispy nest]

this is mostly for fun

i like the aesthetics of it

to be more specific, im always amused about feminine traits of girls

so to be creating such art through prompting is quite fun

[analog goblet]

i did a lot of portrait stuff, too, thousands of pictures, but i was always going after different styles, like charcoal/drawings/photography/whatnot

[wispy nest]

i think everybody finds it fun to explore new styles and techniques

i do too

these illustrious models are interesting for me atm

[analog goblet]

i do like that art style, too

but now that that one's done, i would already now look for something new

[wispy nest]

sounds like you are subliminally seeking for something

[sterile kiln]

Illustrious is a kind of model, not a model

[analog goblet]

i have very specific goals of what i seek actually

[vivid widget]

can someone recommend a model for products image generation ?

[analog goblet]

no, that's a too general question @vivid widget

[vivid widget]

to generate something like this ?

[sterile kiln]

Is it this one?

https://civitai.com/models/795765/illustrious-xl

[vivid widget]

i cant find one tho

[analog goblet]

i m quite sure that most models can do that, and it's a question of prompt/setup

[vivid widget]

okay, do i need a lora or models are good enough on their own ?

[analog goblet]

models should be good enough for this

[vivid widget]

cool thanks

[sterile kiln]

@wispy nest man, would you please tell me.. 🙂

[copper matrix]

Use a illustrious model that catches your fancy :-)

[sterile kiln]

Well, I'm looking, but I can't find any that give results that are less manga, and more realistic like the Torqx photos.

I'm just asking for the model eh, I'm not a prompt scratcher

[vivid widget]

https://civitai.com/models/119303/3d-style-xl

something like this

?

[sterile kiln]

No, its not 3D render like this. Its more realistic, like this one :

https://civitai.com/models/1037024/evermore-illustrious-xl

[vivid widget]

https://civitai.com/models/94809/realcartoon3d

[sterile kiln]

SD 1.5 🤮

[vivid widget]

https://civitai.com/models/125907?modelVersionId=686204

[sterile kiln]

we were talking about Illustrious one.

[vivid widget]

yo @wispy nest man tell the name bro

[sterile kiln]

Wait, you're asking him for an open secret here. It's very precious to give the name of an open source model.

[vivid widget]

ohh my bad man

[wispy nest]

in my experience some people don't give prompts or model names out

or any workflow details

[vivid widget]

have you checked this one ? Vauxz Merge 2.5D - 3D

[sterile kiln]

We are talking about a model man.

[vivid widget]

fair enough

[wispy nest]

the license of the diffusion model allows stuff to be kept secret

licenses exist that stop you from keeping anything secret, but so far the diffusion models don't use them

[sterile kiln]

Let me remind you that the basic ideology behind SD and others is sharing. So I'm all for keeping our prompt and our methods. Everyone keeps their personal touch.
But stalling on a model's name, that's lame.

[wispy nest]

this is like
one of the biggest and oldest debates in open source

whether open source licenses should allow closed and commercial usage

or whether they should force all usage to be open

[vivid widget]

in the end isn't it upto every individual if he wants to share on not, regardless if the license allows it

[wispy nest]

no not necessarily because a license can force people to share stuff
but it doesn't apply here cos the models don't have a license like that

[sterile kiln]

Don't ask me what model I used, it's a CIA-defended business secret.
If I give it to you, the whole Western empire will collapse, and I'll lose all my talent.

[copper matrix]

i mean if its a private merge or hes using it for a patreon income etc

i can understand keeping certain things a secret

[sterile kiln]

Oh, this kind of dogshit

[vivid widget]

it's true that the license can force people but all i was saying is one people can choose to tell out of good faith bro, we aint dealing with sum high stakes to follow the license line by line

[copper matrix]

the merge or patreon? the merges mean it wont be on civit ai at all (if done locally) but patreon is good money

hell even some slop posters here get like 20-30 a month

[sterile kiln]

a bit like those people who take open source, make a few changes, and ask for a donation patreon when they've changed almost nothing.
Special dedication to Aitrepreneur

[vivid widget]

i can understand if money is involved

[copper matrix]

if someone is willing to pay for it i dont see the harm. ( i was talking about image outputs tho)

theres this cool pixel sprite lora id love to get but he asks 25 for it

[sterile kiln]

If someone has broken their back to do it, and a ham comes along, and just because of his notoriety, makes money off the back of the REAL dev/creator who did the work. While the REAL dev/creator earns nothing... I'm not sure that's any better.
Another special dedication to Aitrepreneur, who steals Github's code, and creates an almost identical clone, and forces its viewers to pay to see the content, whereas the real Github is accessible to all...

[copper matrix]

finetuning a model (and especially getting to really good) is quite a lot of work and were honestly blessed to have so many models accesable for free

posters on civit ai could easyly have asked for money for the model download

[wise turret]

Question, so I am using a negative prompt list that I was told to use for all of my images

Here it is: BadDream, badhandv4, BadNegAnatomyV1-neg, easynegative, FastNegativeV2, bad anatomy, extra people, (deformed iris, deformed pupils, 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, signature, watermark, airbrush, photoshop, plastic doll, (ugly eyes, deformed iris, deformed pupils, fused lips and teeth:1.2), text, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, masculine, obese, fat, out of frame, caricature, body horror, mutant, facebook, youtube, food, lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature

Are the ones in the beginning some sort of Lora?

[copper matrix]

are you using this on SD1.5 or SDXL?

because daym thats a LOT

[wise turret]

I believe it's SDXL if it's pony?

[copper matrix]

are you getting the desired results/

[wise turret]

Yup! I'm just in #🤝｜tech-support figuring out what's wrong with my Lora though

[copper matrix]

well personally id shrink the negative prompt

[wise turret]

Isn't the bigger the negative prompt the better though?

[copper matrix]

nooo

i use like
(low quality, worst quality), displeasing, ugly, poorly drawn, displeasing, very displeasing, worst quality, bad quality, oldest, deformed limbs, bad anatomy, frames, watermark,

nothing more

but since its pony

[vivid widget]

nope

[copper matrix]

you can add the obvious child tags and the slow scores like score_1 score_2 score_3 etc

[wise turret]

Score?

[copper matrix]

yeah pony works with this cursed score stystem

[vivid widget]

theres also embeddings you can use

[wise turret]

Oh?

So what's the best score in this score system?

10?

[copper matrix]

7 to 9 iirc

[wise turret]

Oh

[copper matrix]

i used control net for the pose hence im missing many pompt tags for this one. but as you can see my negative is small (holding a big ass hotdog odd request from a friend) (edited some secret model data)

original was like a man trying to eat a oversized hotdog

but bigger prompt doesnt mean better result

[wise turret]

I see

[copper matrix]

also im messing with sebl4rd. the model i use is waiNSFW illustrious

[copper matrix]

@dreamy socket

tough i run the 7B with a 3070Ti just fine

[sterile kiln]

I've tested it, and personally I think it's a load of rubbish.

[copper matrix]

i mean its a 7b model its pretty good for its parameters

[sterile kiln]

Mistral and llama3.2 do much better

[copper matrix]

possibly but once more finetuning happens it will be better i think

[sterile kiln]

https://tenor.com/view/captain-obvious-thank-you-superhero-super-hero-gif-9900289

[wispy nest]

something touches my mind deeply whenever I look at @wispy spindle's nick

[wispy spindle]

[loud sigil]

guess the SD model I used for this one

[nocturne atlas]

do you guys have a go to upscale model?

[wispy nest]

this is a good balance for anime and realism 4x_foolhardy_Remacri.pth
https://huggingface.co/uwg/upscaler/tree/main/ESRGAN

if you want to focus on specific style between realism and anime you could try 4x-UltraSharp.pth and 4x_fatal_Anime_500000_G.pth

[vivid widget]

the spelling is wrong bro

[warm umbra]

If I saw this thing in real life, I'd probably play soccer with it.

[royal monolith]

@scenic yew spammer alert

[scenic yew]

yeah these spammers are so unbelievable as well, how is "x-token" even related to image gen.

[copper matrix]

Did he just ping s random user to "look a spammer"

Amazing

[viral frost]

For some reason, Flux drew a blank food-based void all on its own. I think I can see an alien staring at me through it.

[wispy nest]

here is your alien who ate the pizza

[viral frost]

That’s a mighty weird-looking alien, I gotta say.

[wispy nest]

she likes pizza too

[copper matrix]

Why dont you post these in #🍥｜anime . They look great btw

[copper matrix]

How much lol. If its a decent pay ill even do the letters

But imo your better off going on fiverr or something similar and hiring a small time artist for that

The trick is to use both ai and photoshop

But yeah I'm not staying unless i know the € range (also adviceable to just mention it here in advance so other people might want to but idk upto you)

You don't have the time to mention a pay range?

26

Don't see how it's relevant though

[sudden radish]

hey man

i can recommend you a few sites to try

where the words are correct^

its what i use for my own stuff

heres an example

heres another example

@viral shadow

[elfin marsh]

mes premiers essaie sur stable-diffusion si vous avez quelques conseil sa ne serais pas de refus

^^

[wispy nest]

Do you guys think this was made with Stable Diffusion? It's my friend but it's a jacked version of him. Can't ask the person who made it.

[candid pollen]

Impossible to say from all those low quality jpeg compression artifacts and the "cinematic" editing

It could be a real photo that's been edited or an AI generated one.

[rough ridge]

@modest delta

[copper matrix]

Personally i think there could be a hint of ai in there BUT you could use a ai image checker

[wispy nest]

well i know it was ai just not sure which

[copper matrix]

Those are fairly accurate as they flag 70% of my images as AI and 30% made by human

Hmm

Thats hard to say honestly

Probably flux or sdxl realistic model

[wispy nest]

i want to add batman in behind him in the same style. i tried diffusion bee and it went horribly wrong

[copper matrix]

Well if your pc got a decent graphics card you could add batman behind him pretty easly

Personally id just photshop him in, mask batman and use img2img to clean up any dispensary's

[wispy nest]

@copper matrix is it easy enough to do that u wouldnt mind doing it for me? I have no photoshop skills lol

[wispy nest]

look at this sugar model https://civitai.com/models/1110156/majicflus

but too bad it's flux which i dont run

[sacred oriole]

Are there any italian speackers?

[wispy nest]

use google translator.. there is no restrictions to conversing between english and other languages

you can also just use chatgpt, copilot, gemini

[sacred oriole]

Yes no problem, thanks, i speack a little. My AI suggest me to try here for help choosing my bar logo

[wispy nest]

people share their gens here, for tech support there is a channel #🤝｜tech-support but if you are looking for ideas you can ask , and if anyone's able to share idea they will

[sacred oriole]

I've got many ideas but i'm not a graphic

[sacred oriole]

Ok

[wispy nest]

if you already have some local setup to generate images that's a good start to experiment with your prompts

[agile quail]

generator

[copper matrix]

Cant im not home for the coming 2 days. But what ? You just throw a shitty batman behind him. Then go back to ai. Mask the batman and run it at a very low weight (0.2)

Like copy pase a png. Use the eraser so it looks like he's behind him

[wispy nest]

SDXL breed models have better anatomical flexibility than Flux

[inland iron]

I had an interesting showdown between FLUX and SD3.5 using the same prompt, and the results were quite revealing... I've compiled all my results into an 82-page PDF, and you can grab the free download link from my X profile: https://x.com/JPdeBeerAI

JPdeBeer AI (@JPdeBeerAI) on X

Visual Prompt Mastery for Creatives | Bringing the power of AI to your brand visuals.

X

[copper solstice]

i have been using hiabrid concepts to make transitions and animations for my streams and thumbnails

[glad galleon]

Used SDXL to generate an image of a guy, then used Kling to animate it and get different facial expressions and poses

[faint mountain]

ouu, its a free viewer/sub generator for YT

why does it have fans tho xd, surely the phones themself arent getting that hot

[spring mist]

there was a conversation involving running SD on a phone 😆

[wispy nest]

lol things ppl make https://civitai.com/images/54928275

[wispy nest]

https://www.riffusion.com/

[faint mountain]

ahh .o.

ig its possible, i mean, at least androids, have linux capablities, if u can get one rooted, then flashed, provided the CPU is oki with it, and u can find appropriate drivers, that would be pretty interesting

[wispy nest]

Stable Diffusion is the foundation behind Riffusion. Instead of generating images, Riffusion uses Stable Diffusion to generate spectrograms (visual representations of sound). These spectrograms are then converted into playable audio using an algorithm that transforms them back into sound waves.

Riffusion essentially adapts the text-to-image model of Stable Diffusion to create music in different styles based on prompts. This is why it can generate a wide variety of musical genres, instruments, and even vocal-like effects.

vocal transition of riffusion seem more polished and fine tuned compared to suno...

or rather less robotic

and its completely free to use, unlike suno, udio that have subscription fee

holy.... sense of melody is so good with riffusion

pretty obvious riffusion perfected some of the lacking suno still has

how is it possible that one company with a selective team for developing all these AI based technology happen to spread so widely across many other companies, that could only mean one company insider is selling ideas to other companies, OR all of them have been holding back technologies and releasing them once the elites of society gave the go

also the news media have long been posting perfect pictures for the news and magazine covers .. how did they do it? They already had all the Ai tech to recreate stuff

I'm both amazed and pissed

[copper matrix]

And where does alien technology fit in

[wispy nest]

no such thing is alien

some ppl might have superior intelligence and have been living in isolation

https://www.riffusion.com/riff/18ab0dad-b890-4f81-a1b1-6cebf3be6fd0

[spring kelp]

riffusion was broken last i saw, which version is this?

[wispy nest]

still in Beta, but works fine

[spring kelp]

which version

sd 1.5

xl

3.5??

last time i used it was 1.5

[spring kelp]

its own now, interesting

[wispy nest]

why do you think they are running sd1.5, xl and all that, those are for image gen

is ok for music

instead of channel being colour a channel can be a sound thing

really that's interesting

but i'm trying it just now, and i see fuzz model

the exciting part is no song limits with paid bs

i already have reason now not to use suno

imgine getting stuck with 5 songs / day with suno and dopamine rush stopped from flowing

yeah need open source

this is really good with what i've discovered in the last 2 days

first the deepseek, now the riffusion

oh and alibaba released qwen that's full blown modal from text-chat to image-gen to video-gen

[quaint valley]

hey could someone explain the "resource used" section on civitai? did he use first SD XL then Game Icon?

[vagrant dust]

Simple as the name riffusion and some models where already available 3 years ago. In the github you can read that the project is not developed anymore. ...

[wispy nest]

are you saying they abandoned the riffusion project 3 years ago?

[vagrant dust]

No I said that they started (Open Source) with a Model based on 1.5 SD

[wispy nest]

ok, sounds good to me, their recent development is as good as suno which they achieved only a year ago

[spring kelp]

you seem to forget that audio can be visualized

as is the basis of this tool youre raving about

🙄

[wispy nest]

not forgetting anything, i have no idea how to visualize sound

also im raving about riffusion its completely open source with 0 pay

[copper matrix]

Sdxl isnthe model they use, a lora is a addon for the model. Could be a style or character or even a concept

[wispy nest]

the only way i visualize sounds is through senses and feelings, not optically, and ways to bridge between mental world and physical world idk how, except you are existentially a conductor between those two worlds

[quaint valley]

oh i see, if it doesn't bother you, do you know a video where i could learn more about it?

i have sd xl and game icon but i don't know what do to to make them work together 😂

i am on automatic1111

[gaunt thistle]

I’m creating a lot of images with Flux now. Some of them a little bit unsensored.

[wispy nest]

cool, i stopped using flux/ sd.3.5

[gaunt thistle]

Really? Why?

[wispy nest]

im all in on sdxl / illustruious / pony now

better and cuter aesthetics

[gaunt thistle]

Wow do you have examples?

[wispy nest]

yeah all thsse images im posting

[gaunt thistle]

This is Flux

[wispy nest]

i know

[wispy nest]

i can spot a flux gen from a glimpse

the plasticy look and having to refine it was too much

[gaunt thistle]

Oh I see

You have some real image created with pony?

[wispy nest]

well you know flux and sd3.5 are better at realism than sdxl but my personal take on realism is bit different, i love creating anime themes more other might prefer realism flux offers for their projects

[gaunt thistle]

Oh I Understand

I like the Anime too. I’m not an expert of course.

This isn’t an Anime image anyway hahaha

Plastic as you said 😦

[wispy nest]

one draw back about flux / sd3.5 are those dont work too well with poses unless you use controlnet

[gaunt thistle]

Ufff you are right

[wispy nest]

upper body pics can look great, but nowhere as flexible with anatomy as pony

[gaunt thistle]

Do you have the link for pony?

[wispy nest]

this is pony

[gaunt thistle]

Wow

[wispy nest]

lots of good models on civitai

just filter the search for pony have a look under checkpoints

[gaunt thistle]

I see

Thanks bro

[gaunt thistle]

Do you have a server where I can see your work?

[wispy nest]

no, i run stuff locally

[gaunt thistle]

Ok. I hope to see more of your work over here.. 😛

[wispy nest]

look at that image @dry crow just posted, very crisp and radiant

[gaunt thistle]

Yeah it’s very good

This is Pony?

[dry crow]

thanks, its illustrious + adetailer + upscale

[gaunt thistle]

Amazing

And you can use the same character doing different things?

[dry crow]

its possible if the character has a very good unique description (and is more anime like)
or if you use controlnet and with that you can use the same face and stuff for the next images (works for realism and anime)

[gaunt thistle]

Good!

Thanks bro. It’s amazing

[dry crow]

np!

[copper matrix]

Select the checkpoint (model) sdxl. Then use the loras tab and select the game icon

Thats it really

Put in a prompt and your ready to go

[gaunt thistle]

Do you know where I can find the bot on discord (if available).

[wispy nest]

https://www.riffusion.com/riff/e5354d16-dc44-489f-a1ec-ce5563340250

[dry crow]

the bot is not free but here is the info:
https://discord.com/channels/1002292111942635562/1237461679286128730

[gaunt thistle]

I really like bro… amazing.

[feral anvil]

Hey, anyone have a good workflow with a XL being used as a refiner for flux?

[wispy nest]

https://civitai.com/models/716879/flux-refiner

[feral anvil]

thanks. I will try it out

[gaunt thistle]

Do you know if this picture is a SD ?

[wispy nest]

@shut sinew ... bots are flying in

[gaunt thistle]

Hahaha

[rough ridge]

scam link

[wispy nest]

anyone using deepseek with comfyui yet?

[sterile kiln]

Why should we do that? Deepseek is just dogshit, specially small models

GUYS

Most people only talk about Wavespeed to improve their generation speed. Problems: Wavespeed doesn't work with GGUF models, and it greatly distorts images.
Here, at a guess, I'd say the image remains 98% true, it works with GGUF, and -50% s/it.

https://github.com/welltop-cn/ComfyUI-TeaCache

[viral frost]

I was experimenting with TeaCache the last couple days with Flux (bf16) and an adaptive RK sampler. Unfortunately, I don’t have a very high opinion of it so far. Yes, the speed increases, but the quality decreases just like if you had relaxed the sampler tolerances (i.e. reduced the number of steps). Tightening the tolerances will get you back to a similar quality as the default tolerances with no TeaCache, but then the generation speed reduces again back to about the original time. So there is no free lunch here.

The node repo mentions the ability to losslessly reduce the generation speed, but there is no setting that does this. Below a certain TeaCache tolerance threshold (maybe around 0.15), there is no change to the generation time. The final result still looks slightly different than if the node is not part of the graph, though.

TeaCache also seems incompatible with tiled diffusion. It didn’t break anything, but it didn’t do anything — the generation time for my high res pass didn’t change regardless of the TeaCache tolerance setting.

[copper matrix]

X tiling only

i dont know whats in the image though, just prompted grass texture

[scarlet lodge]

can u pls check if this works with flux pretty pls? I dont wanna install it just to find out it doesnt work with flux

[viral frost]

There is already tiled VAE support built into comfy.

[scarlet lodge]

I just installed it but cant find the damn node 😿

here is my workflow without the tile

[copper matrix]

i hardly use flux since its a bit slower then SDXL but it should. seamless tiled does seem to work. so you can use it

[scarlet lodge]

that image has the workflow

[viral frost]

You need to refresh your browser window after you restart the back end.

[scarlet lodge]

yea but when I right click I get a gazillion nodes with no order and I cant figure out how to find it 💀

[viral frost]

Double-click the node canvas to open a search box.

[spring kelp]

use it ollama on comfy, quite good

[wispy nest]

im actually using florence2 to interpret images, not 100% correct but works

[spring kelp]

thats a good caption model, deepseek is a reasoning model

so itll display its "thought" process

[wispy nest]

deepseek janus pro has image understanding node native to comfyui

but so far i didnt feel a reason to switch or replace florence2

i tried the janus pro 1B model few mins ago

also i love florence2 for it's versatile features, it can give you detailed caption and it can give you booru style tags

latter is quite useful with pony

[copper matrix]

Hows the janus model?

[wispy nest]

not bad, but nothing extraordinary for what i was trying it out

[sterile kiln]

and low Vram usage

[copper matrix]

I really want to get a decensored llm hooked up to my own SD models for generation and image recognition but llm's arent there yet

[wispy nest]

this is quite uncensored and can generate explicit prompts

[copper matrix]

Oh yeah hmm i meant more like when interacting with the LLM and i say "make me a image of" but instead of it using janus it uses like my illustrious model

Possible in comfyUI? If so i need to get into it

[wispy nest]

try this https://github.com/SeargeDP/ComfyUI_Searge_LLM?tab=readme-ov-file

[copper matrix]

Nice, ill dive into it. Its not fully what i want but does the trick partially

[fleet goblet]

Diffusion models like flux can do this with Flux Redux, or IPAdapter for SDXL

The ability to clip vision encode them basically instead of getting an LLM to describe it in words

[wispy nest]

@dry crow

[quaint valley]

guess who placed loras in checkpoint folder 😂

[wispy nest]

judging by the above images

this is the server i didnt know i needed

excellent work lads, keep it up

[clever oar]

trying to revive old games

[rough ridge]

is that Wii Sports?

[clever oar]

resident evil 2 98

clare and leon

[dawn seal]

I can never remember if she's 17 or 18.

[sterile kiln]

18

[buoyant elbow]

Can someone edit it to look bit ai

[buoyant elbow]

Thank you sir

[buoyant elbow]

Bro you are great

[gaunt thistle]

Could you share the prompt for the second image?

[fleet goblet]

I actually cannot. It was made in like 2 seconds by an LLM.

In this case, it was Openrouter's free rose LLM 100b model

🤷‍♂️

I didn't have job data per image on.

Thats why.

https://tenor.com/view/facepalm-double-crowd-funny-omg-gif-16929018

[sterile kiln]

Haha

[faint mountain]

i think the server or my discord thinks im a moderator

xD

[gaunt thistle]

Interesting. Are you using the API or on the website? Did you pay for the service?

[faint mountain]

dis borb is up to something

[naive roost]

hi i may need help

what's wrong

mention me

[clever oar]

your model wrong

[clever oar]

change to epicphotogasm

1.5 pruned-crippy results🙂

i have it

nice for horror🙂

[naive roost]

ok can u explain why mine is wrong and why i should change to this

[clever oar]

I had bad results on this model, or did I misunderstand? Explain what’s wrong with your picture?

how i can help

[naive roost]

my picture is wrong cause it aint a man with black hair

[naive roost]

so i understand that my model might be wrong but why should i take epicphotogasm more than another

[clever oar]

its one of good models

[naive roost]

😅

[clever oar]

bad?

[naive roost]

no it's good thanks

it's what i expected from the other model

[copper matrix]

, prompt is wrong(too short and no negatives), its SD1.5 so its gonna come out cursed a lot regardless

[naive roost]

can u give me an example of good prompt and neg prompt pls?

[copper matrix]

it can differ per model but since your using Sd1.5
Negative prompt (might wanna add some embeddings once you know what your doing): low quality, worst quality, low quality, worst quality,

for positive:
(Upper body photo:1.2). Detailed eyes, Detailed face. Photorealistic, Natural atmospheric lighting, intricate details, 35mm photograph, film, professional, 4k visuals, highly detailed, elegant, studio quality. (your prompt here)

remove some stuff you wont use maybe like upper body photo

then try again

[sterile kiln]

Or,
Try different king of models, try Comfy.
Depends on your hardware specs.
@naive roost

[copper matrix]

i wont recommend comfyUI to someone who doesnt know how to prompt yet

maybe swarm but that might be complex too since there might be a lot of documentation there are hardly any video tutorials

i recommend any XL model over SD1.5 provided they can run it

[sterile kiln]

I wouldn't recommend anyone to train on a model as old as Herode, which will disgust them more than anything else.

[copper matrix]

true lmao but he can use any other SD1.5 but if hes someone who cant run a heavy model theres still some OK 1.5 ones

out of most 1.5 models i prefer Based64 with one of the many style loras. decent results for 1.5

[sterile kiln]

a bit more qualitative 🙂

[copper matrix]

yes i recommend flux and SD3.5 if he can run it but we dont know

but for what its worth 1.5 can still do ok images

[wispy nest]

except for noodle fingers or hiding hands altogether

[copper matrix]

im not saying its great man. but if you got a potato pc its the best one could get

[wispy nest]

i know, but if only they could fix the hands

[wispy nest]

this is perfect

can you make my character LOL

the two blondies :3

THE ART STYLE IS SO GOOD OMG

i wonder if all illustrious models are this weird, or the model im using...
blonde woman, orange abstract background,

not sure but its fantastic

love the character model

[wanton timber]

its cake frosting

[viral frost]

Okay, enough girls. Let's get some real art up in here.

[gaunt thistle]

Teach me master! how should I start doing this art? Can I do it on the web page or should I get this installed in my own computer?

[wispy nest]

not a master at this, but I use local setup with comfyui on a system with 12gb vram, 32gb ram, but these illustrious model would work fine with 8gb vram and 16gb ram too, just an idea

[radiant saddle]

Hey o// Can someone please post an SD 3.5 benchmark on a 5090?

[copper matrix]

let me know where i can buy one and ill show you asap!

[sour rune]

A black haired girl with a realistic image

[copper matrix]

That's oddly generic yet specific.. is she holding a realistic image?

[fleet goblet]

literally blocked torq because all I've seen him post is girls

You can do so much more.

There is no real limitation with flux other than the time you take to make the image you want

Want to make an alternative of an already existing image? Use Flux Redux.
Want to inpaint? Flux Fill.

Depth and canny models, and just the amount of more common good flux finetunes

and just antiburn nodes in general and their ability to allow you better prompt adherence

[versed rain]

hello, anybody knows how to replicate this artstyle?

[shell sleet]

2nd one isnt my favorite and i very much messed it up in prompting, but it's still pretty

[wispy nest]

imagine such hideous mindset. never the less I prefer to block them too. only annoying thing about discord is that it shows when blocked users show up in channel message as blocked instead of completely muting them.

[viral frost]

Believe it or not, some people don't want to be assaulted with soft porn whenever they visit the discord.

[wispy nest]

you're painting a different scenario, those images are not explicit

i also have 0 regards for your worldview when you fail to appreciate life aesthetics

@viral frost you keep up your toxic mockery as you like, i'd have to be banned from server by SAI moderator before you can expect me to stop posting what i like posting which btw has nothing to do with explicit content. this hideous game you play is not confined to discord servers, its an issue that will be eradicated in real life.

[viral frost]

Believe it or not, but the world doesn't revolve around you. I have young kids. I don't want to have to shield their eyes when I hop on discord, not knowing what might show up. And the last thing I need is my wife banning my use of AI image generation because she feels threatened.

[wispy nest]

you seem to assume the world revolves around your social constraints, that's a pretty loaded assumption, the fact that you oppose perfectly aesthetic aspects of life itself shows enough of your conditioned mindset and your circle of influence.

[viral frost]

Nice rationalizations you got there.

[wispy nest]

i dont have to play word games with you, i know where you come from, and who you are in social sphere, there is nothing reasonable about your worldviews

[viral frost]

You don't know me. Get gone.

[wispy nest]

you can back off

[spring kelp]

https://tenor.com/view/now-kith-mike-tyson-gif-13249709

[wispy nest]

after playing with Illustrious and Pony for about 2 weeks and comparing the output images, it seems like Pony has more substance and texture than Illustrious

[celest sigil]

Lumina 2, created in 2048x2048 without any upscale.

2048x1536

[bitter copper]

SwarmUI now supports Lumina 2, a medium scale image model with an LLM (Gemma 2) as its text encoder, and the Flux VAE
Docs here: https://github.com/mcmonkeyprojects/SwarmUI/blob/master/docs/Model Support.md#lumina-2

[lyric spindle]

@viral frost make a picture of the oscar statue 2025

[viral frost]

Why?

[copper matrix]

Id just ignore, people cant be bothered to read the #artisan-faq

[viral frost]

Usually the bots don’t ping a specific person.

[copper matrix]

Could also not be a bot. People can be really clueless

[viral frost]

Yes, I’m assuming it’s a real person and they are just really self-absorbed. That’s what I was challenging.

[copper matrix]

Its somewhat of the same situation what i ranted about in off topic lol

Might come off as rude but it's what i deal with for work 8 hrs a day

[wispy nest]

Wth🙉

[rough ridge]

what model is that? The hair is so detailed

[vagrant dust]

Size does not matter !

[wispy nest]

look for cyberrealistic, one of the best realism pony model out there

i'd hope the author is here so i could shower him/her with compliments

in either case the author is breaking boundaries with sdxl breed model / pony and merging all kinds of sugary worlds

[fierce laurel]

which cute model is this

[granite gust]

wow

[fleet goblet]

Dr. Seuss lora portrait of Donald Trump

[wispy nest]

ollama takes heck lot of vram compared to which Searge LLM is lot less memory hog

[fleet goblet]

no clue what youre talking about

literally comes down to quantization

ollama keeps models loaded in vram/ram also

which can be disabled via OLLAMA_KEEP_ALIVE="0"

[wispy nest]

you probably dont have the brain to understand what im talking about, and btw werent you suposed to block me?

[fleet goblet]

because what youre saying makes no sense logically lmao

[wispy nest]

that's with your brain, not my issue

[fleet goblet]

Elaborate on how seargeLLM is less vram dependant than ollama

Like actually explain

I want to understand how you got to that conclusion

[wispy nest]

and im done wasting time talking to someone as obnoxious as you who acts like pompous, obnoxious cause of someone who appreciate feminine beauty, here something for you, im trying hard to be nice to you, just dont talk to me, dont address me

[fleet goblet]

LMAO

So offended for nothing but me criticizing the fact you mostly only post women.

Go ahead. Not my problem.

[wispy nest]

you are stupid as you are rude

[fleet goblet]

still havent explained anything, just insulted me the entire way instead

[wispy nest]

dont twist this around, you freak.. DONT address me

[fleet goblet]

No, chief. You are in a general images chatroom.

Either block me and continue on or just don't talk.

[wispy nest]

how cheap can you be

[fleet goblet]

Stop getting offended over pointless stuff, snowflake.

Mind blown.

[wispy nest]

your blackholes and portals miss life, there is no point in having a vast universe w/o girls

[royal monolith]

SD 3.5 large. prompt: a snowy owl, perched high in an oak tree on the edge of lake at sunrise. in the background we can see snowcapped mountains and on the other side of the lake are lodgepole pine tree covered hills

[stone valley]

Now this is ai art a nice anime style waifu

Great work torqx

[wispy nest]

i hestiate to say ty, but ty, i've been discriminated by many for no rational cause for loving feminine aesthetics

[stone valley]

bro you got me inspired rn to make a waifu if you saw my femboy collection bro youd see true degeneracy I cant judge waifus bro.

[wispy nest]

lol well thats nice to hear, at least someone understands the life aesthetics, its not about judging anyone but appreciating beauties of life

[stone valley]

bro you got twitter youre making these alot that kinda stuff murders on twitter bro

[wispy nest]

well i do post here and there

looks good but they have channel moderation to not show too much skin, cause of minors might be watching

[stone valley]

shit i cant share anything on this server thats crazy

[wispy nest]

yeah its crazy cause there is a fine line between acceptable sexy and going over that line

[stone valley]

thats my boy

Waifus all day all the way thats how the pros do it. Just sit around and generate waifus all day theres a skilled future involved in that

[spring mist]

Waifu

[wispy nest]

i see a lot of good 18+ pics on civitai, and also the civitai discord has nsfw channel but the problem with that channel is they post too much gore

[stone valley]

also remember if you wanna do something more lucrative you can always jump into programming and take all your homework and pump it through ai tools and only study in the days following the test just to get by then cheat your way through the entire career or you know what fuck college. Just ai generate everything

[wispy nest]

big changes with AI for sure, automation is inevitable

[stone valley]

Dont do anything that isnt gonna make money boys cuz its not gonna be easy to make money in the future boys you gotta really scrape the depths of deviantart interests to get away with those ai tools

[wispy nest]

you might have heard about UBI

cause that's where the world is headed

[stone valley]

yea man that shits not gonna happen and even if it does its gonna be 1000 dollars a month. You only bring it up because andrew yang and asmongold tell you its gonna ever do fucking anything but its not.

[wispy nest]

it will take a bit of time to develop a good system around it

but its coming

btw since we are discussing this in SAI discord, did you know they released Riffusion?

[stone valley]

nope were just fucked bro go make money however you can

stop gargling asmongold bro i get it hes funny but he also cant clean his own house hes not a bright guy.

[wispy nest]

not into bitcoin btw

[stone valley]

even before ai the entire economy was rigged to hell and back with nepotism now its just gonna get worse. Also thats good that you dont have a gambling problem that gambling problem never gonna make you money.

[wispy nest]

yep

its not sustainable, thats why there are activists talking about resource based economy to overcome this constant war incidents around the world