🏞｜general-with-images

[errant zenith]

hey guys

stable diffusion is downloading more space

909500000000000000 tb

good model

[lavish garden]

Prompt: https://civitai.com/images/24901772

"Buzz Drop" - Marshmello mask (Flux test) 🎶

[candid surge]

The signature stuff got out of hand

[languid pebble]

yeah 🙂

[errant zenith]

[steep sleet]

Flux@nf4

[candid surge]

1/4 nice but with extra arm
3/4 just creepy

[steep sleet]

Pentadoggy!!! 😄

Successfully creepy?

[languid pebble]

A.I. knows us pretty well. I asked for a "No shit sign"

[candid surge]

lol

[languid pebble]

"No human, no cry!" sing

[steep sleet]

Flux@nf4

[languid pebble]

There has been a Lady at my last exhibition who would love your lighthouse pictures ...

[steep sleet]

Send me her email? If you can ...

[languid pebble]

I'm sorry ... I don't have it. But there are lovers for pictures like that

[steep sleet]

Flux@nf4

[steep sleet]

What sites do you exhibit/sell on?

I am at Society6, fineartamerica, redbubble and Printler

[languid pebble]

Like this 🙂

[steep sleet]

Physical, and not online?

[languid pebble]

Trying to sell NFTs only scammers stole my time ...

[steep sleet]

NFT - too many scammers and time-wasters!

[languid pebble]

That's why I wanted to create the "No shit sign" to stick a post on X that I don't have ETH and don't need promotion ....

[candid surge]

fooling around with masking

[errant zenith]

Hey guys, I'm using schnell Q4_0, and honestly this is worse than normal 1.5 models

Something about the contrast and exposure puts me off

[steep sleet]

Use Contrast Correction Node in ComfyUI

Or specify cloudy/overcast look?

[errant zenith]

I'm using stable diffusion forge

I don't think correction will help, the image is almost blended

[dull flame]

lol, this is too fun

[dull flame]

i wonder why flux is so good with interfaces

it almost never fails

[errant zenith]

I'd play that game

[steep sleet]

Flux@nf4

[steep sleet]

Ouch! Gloriously good! I like the light across his face

[candid surge]

semi animated scanning stuff with mask & inpaint workflow in comfy

using increments and some math nodes to shift the masked area

[errant zenith]

😭 Willy Wonka tea factory

[celest sigil]

No, the Mad Hatter's tea party

[carmine warren]

Can someone make a prompt that would generate this girl

Pls

[errant zenith]

white

[candid surge]

that is a good start, small addition:
white lilac color theme anime image

try something like:

white lilac color theme anime image of a small girl with very long hair holding a big rifle, (big rifle in front of her:1.05), (walking with her cat:1.08), sport shoes, black garters, (monochrome:0.6), cat ears, (flat colors, water painting:0.75), empty apocalyptic white background with broken buildings, simple anime style, hand drawn:0.8), bright

change the weight depending on your model, and maybe add some negative prompt

[errant zenith]

lol I wasn't being serious

[candid surge]

though it was a good start. 😛

[sterile kiln]

what did you use for the caption?
Personally, I use the tool in Kohya_SS, Caption W14, with the prefix $ebl4_Flx_v1, .
This gives :
$ebl4_Flx_v1, solo, looking at viewer, short hair, shirt, black hair, 1boy, closed mouth, male focus, black shirt, facial hair, thick eyebrows, portrait, beard, mature male, realistic

You then need to retrieve all the caption prompts and fill in your json like this:

after that, I personally put it on pastebin, then copy the RAW link, and fill it in on the site, in the advanced settings.
The photos should be good, but so should the prompts.

[magic zephyr]

hey guys , what should i do if i get constantly deformed images

[wispy nest]

deep shrink res adapter hidiffusion

[lusty otter]

Tell the alien overlords you need a new cat 🐱 but actually you are using too big of dimensions or the wrong dimensions for the model.

If you got the model off civitai check the page for the suggested dimensions

Or google the typical dimensions for the model type you are using

[candid surge]

this is fine, it's typical cat behavior

try dogs

[magic zephyr]

thanks all

[magic zephyr]

what does this do ?

[wispy nest]

they allow you to generate at different dimensions

[magic zephyr]

it actually doing good job but how

[candid surge]

with sd15 model its hard to get both the rifle and cat right. cat's gone here.

[magic zephyr]

with and without it , same key

[errant zenith]

she got neural link, elon already have flux in his claws by the grok 2 partnership

I can't get it to make realistic art.. What am I doing wrong?

[candid surge]

nothing, flux is not specialized in realistic stuff. it's always a bit over-realistic in a way

[magic zephyr]

guys do you use comfy ui , i only use stable diffusion , i get really confused in compy

i want to use LivePortrait does it work in stable diffusion ?

[wispy nest]

are you using A1111?

[magic zephyr]

yes

[steep sleet]

LivePortrait freely available on Pinokio2

[magic zephyr]

is it worth the effort to migrate to comfy ui ? i am very comfortable in A1111 now

[wispy nest]

yeah comfy ui is worth it

[errant zenith]

Uhm- I'm using a lora that's supposed to enhance realism and it generates half anime half realistic images.. not.. over realistic?

I just find it really unstable

[candid surge]

idk much about flux loras

[errant zenith]

maybe I'm using it wrong..

[languid pebble]

Haven't tried a LoRa with Flux ... have you played around with the weight?

[errant zenith]

Well I'm starting to think schnell is not what I want to use

quality is just not good...

[wispy nest]

I can't stand schnell

meme model

dev is incredible though

[languid pebble]

I remember of having read LoRo only works with Dev ... but not 100% sure

[wispy nest]

they might be able to make some limited loras for schnell

[errant zenith]

I'm not sure why they would bother 😅

Dev is not directly bigger than schnell

[languid pebble]

Never tried schnell 🙂

[rain gazelle]

interesting style here, almost an american style like from Animated Batman shows...

[errant zenith]

Hey guys.. what is wrong with my generation?!?
First one is the official dev version,
second one is a quantization of the dev version,
and the third is schnell?? they are identical!

I swear I'm going crazy

Surely they are not all the same..

[rain gazelle]

hmm

have you checked the metadata?

[errant zenith]

hmm..

[rain gazelle]

first one is schnell

[errant zenith]

hmm..

[rain gazelle]

they're all schnell

[errant zenith]

...

[rain gazelle]

you downloaded schnell and named it dev 😄

[errant zenith]

I think forge isn't loading the right model..

[errant zenith]

😭 who is pulling a prank on me

[rain gazelle]

First two are exactly the same metadata

[errant zenith]

how do I check metadata?

[rain gazelle]

First image is metadata for the first two images. Second is slightly different with the different module.

[errant zenith]

hmm

[rain gazelle]

You can check metadata on the "PNG Info" tab in your Forge page.

(Or use a program like the above... shameless plug. 🙂 )

[errant zenith]

thanks

[rain gazelle]

no prob

[errant zenith]

png info doesn't display metadata for some reason

[rain gazelle]

yeah it does

[errant zenith]

huh..

strange..

is it just forge?

[rain gazelle]

Oh I thought you said you were using Forge.

[errant zenith]

I am using

I think?

[rain gazelle]

maybe you need to reinstall it? I dunno

[errant zenith]

I reinstalled it yesterday but I don't know if I missed something

"webui_forge_cu121_torch231"

[potent vapor]

/stable

#👥｜roles Hey guys.. what is wrong with my generation?!?
First one is the official dev version,
second one is a quantization of the dev version,
and the third is schnell?? they are identical!

[errant zenith]

what-

[rain gazelle]

weird, I dunno. 😐

[potent vapor]

hlw how to use this stable diffusion

[errant zenith]

I'm starting to think you are a bot...

[rain gazelle]

you ca'nt use it in this channel.. I think you can still use the artisan service in those channels.

otherwise you can do a local install

[errant zenith]

it's fine I guess... I could just open it in a text file, tho it is not as pretty

Okay.. what the hell.. forge isn't changing model 😭

I think I got to check my console or something..

[rain gazelle]

yeah something is broken

[errant zenith]

or is it just a bug in the latest version..?

[rain gazelle]

I'm on latest version and seems fine.

i'm still on pytorch 2.1 though

[errant zenith]

I think my client was bugged, probably should reload the window more often

[rain gazelle]

I haven't really had much luck with schnell.. never seem to like the results much.

Doing a couple of tests between nf4, nf4 v2, and dev, now.

[errant zenith]

Not sure- why forge says "[Unload] Trying to free 953674316406250018963456.00 MB for cuda:0 with 0 models keep loaded ..."

[rain gazelle]

that's a lotta megabytes LOL

[errant zenith]

yea-

[rain gazelle]

restart the thing

[errant zenith]

idk if that's the issue

but it does load the right model now

[rain gazelle]

make sure you're using the right version of the model

[potent vapor]

it need to subcribe ha

give a link of stable diffusion

[errant zenith]

yea this looks right, I'm pretty sure it's dev now but

[errant zenith]

you should probably not generate ai if you can't find the stable diffusion repository

[rain gazelle]

that's the generation? lol

[errant zenith]

Not sure what happened

[rain gazelle]

was your prompt "Blurry photo of a carrot" ?

[errant zenith]

Yea exactly

maybe just the seed

[rain gazelle]

20 steps?

[errant zenith]

hmm, yea not 20, I wasn't sure how many

probably why

[rain gazelle]

nf4 vs dev

I think I might actually prefer the nf4 one in this case, but just goes to show they can both yield pretty good results.

[errant zenith]

yea nf4 seem to be a bit more logical

what does bnb mean?

[rain gazelle]

not sure, what context?

oh, in the model name

BitsnBytes (more info: https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981 )

[errant zenith]

😅 wait it says GTX 10XX/20XX might not support NF4 but my gpu is rtx 2060?

it's not a typo right?

[rain gazelle]

not sure, I use a 3060

looks like you're right... might not work with nf4

use original dev version 🙂

[errant zenith]

uoh

[rain gazelle]

render at 512x512 first to make sure it's working... then you can bump the res

[errant zenith]

I do have the original dev version but I also have a quantized version, does quantization make things worse?

[languid pebble]

Nobody using Cubes finetuned Model?

[errant zenith]

It's rather I don't know of it

[rain gazelle]

quantized .. you mean nf4?

[errant zenith]

dev Q5_0

is flux good with words or with sentences?

[rain gazelle]

I've heard natural sentences are best.

[errant zenith]

yea seems a bit better, but lacking in realism but maybe that's what I get for doing an anime character

[sterile kiln]

the propagation of light is so incredible...

[rain gazelle]

yeah, it is probably just drawing the character

[errant zenith]

hmm.. I didn'ty even name the character though

strange..

well it does put patterns together I guess

[rain gazelle]

yeah just try something simpler, should spit out something more realistic typically..

[errant zenith]

hmm

[steep sleet]

NF4 on my 8Gb RTX 2070 is OK, along with 64Gb RAM - 90 seconds/1024x1024 image

[rain gazelle]

90 seconds, nice, not too bad

[errant zenith]

currently takes me around 3-4 minutes on 6gb with 32gb ram

[rain gazelle]

practice at low-res first. 512x512, save some time

I run about 3.6-3.7 s/it. about 1 minute 10 seconds to do 1024x1024. (3060, 12GB VRAM)

[steep sleet]

I think the extra RAM makes a lot of difference

[errant zenith]

it does probably

[errant zenith]

something.. is wrong with the rem cosplay

[errant pulsar]

@jaunty zephyr simple wedding-ring with a broken metallic heart and a red gem as blood

[fleet goblet]

Same boat no matter what model you choose tbh

I'm on an Arc a770 16gb. All of the models from 4-bit GGUF to fp8 safetensors are in the range of 2.7-3.7IT/S.

[fleet goblet]

What was this produced with?

If this uses comfyui-3d-pack

I am now sad

[languid pebble]

Picture with Flux, Animation with Stable Video Diffusion, and some work afterwards

[fleet goblet]

stable video diffusion

I can't use that I believe.

At least I don't know a method that doesn't go through the 3d pack.

[languid pebble]

SVD you never know what you will get ... but if I like the picture it's worth a try

Later I add a SloMo cause SVD only gives you 4 Seconds, and pump the Video up to 25 Frames/Second instead of 6/Second

[clever oar]

👋

[languid pebble]

Hope you are fine!

[clever oar]

i start watch yor film

[smoky vigil]

https://www.youtube.com/watch?v=gO3Mk3le0qs

Please developers, please listen to that.

[clever oar]

batter not include

the beginning looks strange😃

[wispy nest]

to a certain extent yes but
a ton of very popular movies have lots of massive continuity errors

[smoky vigil]

Of course but nowhere near as much as AI does now.

[clever oar]

I wrote promt that he throws shurikens😃

maybe it cant do weapon

[smoky vigil]

How it would work is- do NOT generate your scene at once with prompts, that will never give you the degree of control a pro filmmaker will need )or graphic artist for that matter.) Generate all characters doing their thing separately against a "blue screen". Or neutral grey background... Also generate major objects against a neutral background in comfy, then render the background. The remove all the elements and recomposite them against the new background. Then relight everything using C-Light or whatever the future equivalent of it will be- much like a real director, arrange your scene and light it... then the generations can begin... basically they are the equivalents of "takes" in movies, you'll need several and pick the best. And then paint in motion and lip-synch and music.

More or less this could be the process, youd still need a script and storyboard of course.

[wispy nest]

compositing yeah

[smoky vigil]

All this 3D stuff gives me hope because it shows that AI is starting to understand things 3 dimensionally. That means it will eventually udnerstand that character how it would look from any angle.

[wispy nest]

yeah its a good sing

sign

[fleet goblet]

Downloaded q8_0

since that's literally as close to fp16 as you get with slightly more vram compared to fp8

[languid pebble]

Flux is going to text2video next ....

[fleet goblet]

with gguf tech thats gonna be nuts

flux video at 4_0 quant

[languid pebble]

Flux beated SD3 .... so let's hope the best ...

[clever oar]

what are you talking about?

[clever oar]

its better but not in all things

[smoky vigil]

movies

[clever oar]

i know but you talk about svd?

[smoky vigil]

well whatever the tools will be, svd is definitely not there yet, its just how the workflow and pipeline will look like conceptually

[clever oar]

Mojo you start watch gremlins?

[languid pebble]

Watched first part .... second still waiting ...

[clever oar]

i watch 50 % your move

about ufo

oh not true

40 min

want them in my home

repair broken things

[languid pebble]

Next trial ...

[fleet goblet]

Just a question for anyone else using this node.

Do you have to place the base prompt in prepend_tags and then the LLM prompt in text?

I ask because I put an identical prompt in both slots and it gave me a full prompt instead of leaving out the main prompt in the output.

Damn, wrong figures but the shadow is accurate.

Wario's up to somethin' folks.

[raven mirage]

生成一张风景图

[hallow vault]

own

[limpid lichen]

be wary of people trashing on forge. subversive efforts are afoot by people that want to monetize workflows that do everything forge does

[royal monolith]

you do know who posted that, correct?

[limpid lichen]

Creator of control net framework

[edgy root]

Flowers, cars, water, space, these elements design a poster with the theme of "No waste"

#🌠｜show-and-tell Flowers, cars, water, space, these elements design a poster with the theme of "No waste"

[languid pebble]

Good Morning Coffee

[steep sleet]

Using Dice_AI's Platinum SD3 DD Checkpoint

(It's based on SDXL 0.9)

[languid pebble]

Happy with it?

[steep sleet]

It's definitely geared to a more artistic look - but it made claims for more coherent text - which didn't produce!

[steep sleet]

AuraFlow0.3

[languid pebble]

Moonface 🙂

[steep sleet]

SD3 GOLD checkpoint by Dice_AI

[viral frost]

Flux: Planet of the Grapes

Flux:

[steep sleet]

I went to the Organ Donation Centre. Didn't have the heart to go in!!!

[tropic shell]

Love those grapes--what a fun image!

[celest sigil]

I'm sure nobody will notice him...

[steep sleet]

"Mr Mothbee?!" 😄

AuraFlow0.3

SD3 "Leg Thru The Bathtub!!!"

[steep sleet]

Boys gone fishing! SD3 via SwarmUI

[robust sparrow]

flux.1-dev

[misty forge]

Bro really loves bathtubs and crabs and bagpipes on the beach

@celest sigil the model cars look great

[pallid ruin]

lol!

[languid pebble]

Good wine year here.... Prost!

[sterile kiln]

🙂
(its me on the picture (LoRA))

[fleet goblet]

Trying to get stuff like SVD to work on arc is more painful it seems

I can't do full res.

Doing 768x432 instead.

[languid pebble]

But nice!

[fleet goblet]

If this video could be continued constantly until it loops and maintain consistency

That'd be fascinating for 3d model usage

[languid pebble]

That would be nice!

[fleet goblet]

I wish there were lower vram versions of svd xt

lmao

[languid pebble]

At least I found out how to make them longer ...

[pallid ruin]

@limpid lichen

The text was not perfect, but still, 256x256p 😅

[fleet goblet]

256

doesnt like 256

[pallid ruin]

photo of a cat driving a car on the moon, cinematic, holding a cardboard with the text "FLUX ON 512P"

there are 3 moons lmao

[fleet goblet]

this isnt just flux at 256

this is flux 4_0 at 256

lmao

[pallid ruin]

Lol

[limpid lichen]

https://tenor.com/view/smol-gif-18164930

[fleet goblet]

Ope, I got one.

[limpid lichen]

256 nice though

[fleet goblet]

the modularity of resolutions with flux is a nice change

Let's see what happens if I do 256 with the 8_0 variant

Clean? Yes.

Is it the same as 4_0 at numbers at 256? Yes.

8_0

4_0

[buoyant stream]

2001 Chrysler Aspen

[unique condor]

That's part of why there's a base and max shift, vs just having a single shift value, and why width and height are also factored into the equation. In the comfyui section of this channel, I was explaining it a day or two ago and how it all factors into the output shift value.

Probably let's it navigate the network better, and it obviously shows in the results

[misty forge]

That's clean af

Some of the clearest ai art I've seen

[clever oar]

🖐️

[fleet goblet]

And yet it was done with the 4_0 quant of flux.1 dev

used ollama prompt encode for better prompts

[misty forge]

I'm just hearing about flux now, I'm watching a video as we speak, it is very impressive

Dare I say almost perfect

[fleet goblet]

Sadly not. It is missing a large percentage of NSFW data and certain things that would make it all-round better

However it is fantastic for a start.

[misty forge]

If that's your mission

[fleet goblet]

NSFW not just for NSFW.

It helps with anatomy.

And fingers.

[misty forge]

True

[misty forge]

I like the results I've been getting with SDXL, it blows early stuff like dalle out of the water but still has those trademark glitches pretty often

But sometimes I like that chaotic style, it feels like the signature of ai that we will soon throw in the garbage

[fleet goblet]

There will be little reason to use it once flux catches up on the tools front

But it is still a good model nonetheless

[misty forge]

I like the ethereal vibe you get from glitched out sidewalks and dramatic lighting and color

Let me share something

Made in fooocus with negative prompts but no adjusting weights or inpainting

It's really not bad

But maybe I can go back and fix the back of the jacket and the bottom of her hand

[unique condor]

You can bake in anatomy that's NSFW and just make it inaccessible (likely through captioning and censoring images in the dataset), which is what they did. I've had flux periodically spit out people missing tops or pants and it's babrie/ken dolled. You'll see it mostly when you prompt for things involving many people in the scene, like crowds

Sd3 screwed the pooch by attempting to redirect concepts instead and they likely just removed a lot of NSFW stuff from the dataset rather than censor it(costly)

[steel yarrow]

i2i with controlnet is pretty spiffy

[fleet goblet]

\

[languid pebble]

Good Morning Coffee ready ...

[round flame]

我是一个微博的运营编辑，我要写一篇关于厨房安全教育的博文，详细内容如下。请帮我生成内容封面，抽象风格，不要有文字出现。 博文：标题： 😱厨房大“雷区”！这五件事千万不能做！#安全教育 #厨房禁忌 内容： 😱妈妈们注意啦！厨房里面有几件事千万不能做👇 😲第一，拧燃气一次没打着，别习惯性多拧，不然可能会爆炸，燃气会喷出，等散开再试😰 😨第二，有明火时别在旁边倒面粉，面粉粉尘遇火会爆炸，烧伤可不是闹着玩的😫 😖第三，做饭时手机别放灶台边，电池温度过高会爆炸的😵 😣第四，没解冻的冷冻食品别直接下油锅，会让油噼里啪啦沸腾，引发火灾甚至爆炸😫 😩第五，别把鸡蛋放微波炉加热，蛋壳阻止气体膨胀，压力大就爆炸啦😖

[nimble mason]

Here is the image you requested.

[fleet goblet]

I let the AI prompt itself.

[haughty roost]

Its so buggy for me right now but I wanted to test the new imagen 3 with some random prompts I found online and check this out, only about 7 seconds to generate

[steep sleet]

Long trains of text have yet to find an able-enough Gen AI generator

[celest sigil]

More likely to be the prompt. The clear text is the only bit I prompted for 😄

[dense compass]

INK GRAPHIC STYLE COMIC, psychedelic hippie style 1960s advertising poster, magazine cover,, Norman Rockwell style, Arthur Rackham style, woodcut, bright rich colors COLORS,, She poses for the flirts, straight hair, , beautiful detailed face, a little smile happyness girl SUNGLASS , on big sup board
american young woman girl half-turned SHORT BIKINI, lively natural pose, camping on the beach, twilight , SUNSET, big waves,
ART-DECO, Hokusai style, Egon Schiele style,

[steep sleet]

SD3

[steep sleet]

AuraFlow0.3

[dry saffron]

Hi

[dull flame]

Hello

[dry saffron]

Monkey drive car

Hey brave

How are you doing

How they work?

[royal monolith]

read the information in this channel -> #artisan-faq

[languid pebble]

Want cake? 🙂

[merry iron]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[royal monolith]

https://x.com/AISafetyMemes/status/1825798663983214898

AI Notkilleveryoneism Memes ⏸️ (@AISafetyMemes) on X

So it begins

Twitter · Aug 20, 2024, 7:34 AM

[scenic eagle]

prompt?

[wispy nest]

dogshit take right here

but only grandpas and lazy people fall for AI generated shit

if anyone outside of that demographic cares about debunking something as AI, they can do so with a bit of critical thinking and observation

[unique condor]

You see tons of it online and on social media and just don't recognize it as you scroll by because a lot of it has already jumped the uncanny valley gap.

Obviously, if you spent time analyzing every single thing you see, you might start to notice it more, but the point is that a lot of it doesn't stand out like a deformed thumb anymore at a 0.5sec glance

[royal monolith]

and fail badly

[nimble mason]

\

[misty forge]

Hahaha what is this

Bro tried to make an Apple commerical

[haughty roost]

This might one of my best images ever created with ai

[royal monolith]

something silly he made and posted

[misty forge]

Even if ai had real feelings I would still oppress them

Technology was built to serve us and to suggest the opposite is madness. That's why I don't like social media

[royal monolith]

then why are you using discord?

[misty forge]

Discord isn't optimized for user retention like Instagram

I use this like sms

It's not physiologically addictive like other platforms

[royal monolith]

neither are a lot of other social media sites. but discord is still social media

[misty forge]

Semantics.

[royal monolith]

what you mean is that you dont' like instagram

[misty forge]

Or TikTok, or Facebook, or YouTube these days, or Snapchat

But those apps all have something in common

[royal monolith]

the fact that they aren't apps? just websites that you can use an app to access?

[misty forge]

I'm finished discussing this

You're trying to win an argument with semantics instead of seeing or refuting my point

[royal monolith]

the only point i see that you have is that you don't like instagram. not sure why you're feeling the need to discuss that on this discord, in this channel

[misty forge]

We were talking about technology as tools

[royal monolith]

you were a little off the deep end then, i think - if you think that it would be acceptable to oppress an intelligence that had real feelings and was aware of them just because it's your slave

[haughty roost]

Guys I just found the best solution to cutting down image generation time. I was using flux schnell fp8 for a little bit now and while it is my favorite by no doubt, it takes forever compared to my previous main generator. After following a tip from Reddit saying to reduce your desktop resolution, I went from 70-100s to 22 on my first try

[ivory badger]

Two snakes, with extended tails, not so symmetrical. One tail extends to the left and wraps around a pine branch. The other tail stretches to the right, coiling around a flower

[nimble mason]

how do you know it's not the other way around?

[unique condor]

While that does reduce GPU overhead by a tiny amount, I highly doubt it would make that level of difference for most people. Realistically, you were running out of vram probably by a hundred megabytes and hitting the shared memory pool. By lowering the desktop resolution, you lowered the frame buffer size (think windows defaults to triple buffering, so desktop height * width * 3 color channels * 8 or 10 or 12 bits per channel * 3 frames for the buffer).

Comfyui has an added launch flag to reserve more or less vram. It defaults to 500mb, so raise it to like 0.6 or 0.7 (the launch flag is in gb)

[haughty roost]

yeah so I wrote that too early, I realized after a bit it was because I ran the same prompt.

for some reason with flux, on average it takes me 40 seconds for the prompt processing then 20 seconds or so for the actual image

[unique condor]

Oh that long portion was it loading everything then. They don't load the models until the hit the prompt encoder and ksampler, at least with the gguf versions

Open your task manager and look at the drive graphs

When it's going slow

[haughty roost]

Well I stopped it a while ago cause I was done making images, but do you know how I could decrease that time to process the prompt?

I understand that the reason flux is so good is probably because of how it comprehends prompts, but it's just annoying how long the wait is when I could cut it in half and generate twice the images

[unique condor]

move your models to a faster drive like an ssd or ideally, an nvme drive. your average 50USD nvme drive can load like 3-5 gigabytes per second. regular sata3 SSDs cap out at around 400-500 megabytes per second

[haughty roost]

I have only one drive, it's a 512GB nvme ssd. I have an external 4TB drive however it's not actually that fast

[unique condor]

oh and mechanical HDDs cap out around 100MB/s

the external might be limited by your USB level

like if it's only USB3 and not 3.1 or something

it might not be doing superspeed mode

and they cap out around 100MB/s in that slower mode if i recall correctly

so a big 12GB model will take a a while. Ten seconds per GB, so 120 seconds for 12GB

(if only loading at 100MB/s)

[zealous void]

Create a logo for "Grills Inn" that embodies the essence of a modern, inviting grill restaurant. The logo should feature elements that convey warmth, flavor, and a passion for grilling. Incorporate imagery such as a stylized grill, flames, or a sizzling steak. Use a bold, approachable color palette that reflects the rich, hearty nature of grilled food. The font should be clean and contemporary, ensuring that the restaurant's name is prominent and easy to read. The overall design should feel both trendy and welcoming, appealing to a broad range of diners.

[steep sleet]

Using the rbchar0al Flux.Dev LoRA - it adds an inimitable style!!! Hats off to its creator ⭐

For some reason, Flux.Dev has started to work at 2 minutes/image instead of 20!!! Kudos

[steep sleet]

Luckily, I keep all models on an SSD

[unique condor]

Same. I actually want to move the bigger ones over to one of my nvme drives though like the t5xxl and flux models since they're big. They'll load in like 2 seconds on my 5GB/s nvme

[steep sleet]

Flux.Dev and rbcharc0al LoRA

[steep sleet]

Flux.Dev with rbcharc0al LoRA

[wispy nest]

what's the theme of the lora?

[steep sleet]

#🆕｜sd3 message

[wispy nest]

ty

[steep sleet]

Flux.Dev rbcharc0al LoRA

[misty forge]

last one in this group I think is top 5 AI images I've ever seen

[obtuse vine]

XD

Thanks

[misty forge]

[steep sleet]

Flux.Dev

[steep sleet]

Flux.Dev and rbcharc0al LoRA (red and black charcoal)

[wooden trench]

a Gundam head

[languid pebble]

"Don't mess with me!" ^^

[nimble mason]

lovely

[steep sleet]

Just started up Flux+IPAdapter ...

[fleet goblet]

q4 is definitely not weaker lmao

[serene ridge]

@fleet goblet Ty for cool images

[serene ridge]

Cool image!

I'm new here so any tips would be appriciated

[frank drift]

So cool!! What'd you do for them?

[frank drift]

How'd you make the video?

[fleet goblet]

Used LLM-Party with a set systemprompt and base prompt to endlessly make it make images

[frank drift]

Very interesting! I've not heard of those systems. May I ask your prompt?

[fleet goblet]

...The prompt is AI generated. That's what I am getting at.

Lol

[frank drift]

Oooh gotcha gotcha. Sorry, I misunderstood

[fleet goblet]

Also, if I remember right it should be embedded straight into the file.

[frank drift]

It's been a long day, my heads totally not in the game rn 😅

[haughty roost]

This is my first result after trying the new nf4 model of flux. How come it's so bad?

[fleet goblet]

What's the prompt?

[fleet goblet]

I'd recommend the GGUF models over the nf4 one

4_K_S will do you a far better job

[haughty roost]

I think I installed the gguf thing or something

[fleet goblet]

That's the nf4 safetensors

https://huggingface.co/city96/FLUX.1-schnell-gguf

https://huggingface.co/city96/FLUX.1-dev-gguf

you put these in the unet folder in models

and run them with the gguf nodes

https://github.com/city96/ComfyUI-GGUF

[haughty roost]

I also want to say that I was told nf4 was faster as well, and I literally got like a 20s difference. It still takes forever for me to generate one image

[fleet goblet]

i took your image and florenced it for a prompt

then ran that through my llm that enhanced prompts

[fleet goblet]

It's gonna take forever if you don't have the VRAM required for the model, quantized or not

[haughty roost]

Im on the schnell files page, which one do I get? I'm super inexperienced with flux

[fleet goblet]

Less to do with flux and more to do with quants

[haughty roost]

Ive been doing this in lowvram mode

Someone said it's better for 8GB

Although in settings I think it says 16GB are dedicated so I'm confused

[fleet goblet]

That's paged mem

if I remember right

[haughty roost]

Ok then

So which model do I download?

[fleet goblet]

I'd first try 4_K_S and go from there

[wispy nest]

Hey guys

[haughty roost]

Whats the difference?

[fleet goblet]

Perplexity, mainly. Lower quant = less bits per weight.

[wispy nest]

i made a live screen share in comfyui hope u like it

[haughty roost]

Dan do you have the workflow for this new gguf thing?

[fleet goblet]

Are you saying you don't know how to place nodes?

Lol

[haughty roost]

No I mean is there a workflow, also where do I put the gguf file cause its not a checkpoint apparently

[fleet goblet]

I told you to put it in unet

I also said to download Comfyui-GGUF node

[wispy nest]

yeah

[haughty roost]

Is that downloadable via manager? Thats how I got the nf4 loader

[fleet goblet]

You would know this if you had simply just looked at the github page i sent

[wispy nest]

@fleet goblethave u made gguf work with new control net ?

[fleet goblet]

I don't use the current controlnet because it's still in alpha

Also why wouldnt the controlnet work

[wispy nest]

yeah i get error

[fleet goblet]

the weights haven't changed

[haughty roost]

Oh mb I didnt look at that yet, I was too busy on trying to download the file from hf

[fleet goblet]

oh the flux controlnet union is beta now

[haughty roost]

Ok so I ran the portable comfyui command in the terminal as it said on the github but I dont think Im seeing it

Do I have to restart the comfyui

[fleet goblet]

ye

[haughty roost]

Ok so I got the node up, but theres no clip to drag, I don't think I downloaded any from the hf or github

[fleet goblet]

Yeah, because it isn't embedded into the model.

You need to download clip_l and t5xxl seperate and load it with dualcliploader for these gguf quants

https://huggingface.co/stabilityai/stable-diffusion-3-medium/tree/main/text_encoders

Specifically Clip_l and t5xxl safetensors

[haughty roost]

Omggg so much workkk

[fleet goblet]

No shit, sherlock. This is AI.

[haughty roost]

😭

[fleet goblet]

https://tenor.com/view/wazowski-mike-mike-sulivan-meme-monster-inc-gif-19634164

Also @haughty roost

Do you want to use

the t5xxl gguf

It's probably faster

[haughty roost]

I can download the fp8 clip right? It should be faster for me since I'm on a mid-end pc

[fleet goblet]

lol

[haughty roost]

YEAHH

I cant deal with these 40+ second wait times

It's so much better than realvision for everything besides realistic stuff, but its legit 12x longer

[fleet goblet]

https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf

[fleet goblet]

Flux will take 40 seconds no matter what

[haughty roost]

WAIT

[fleet goblet]

At 1024x1024, schnell will at minimum by itself

[haughty roost]

BEFORE YOU SAY WHY

It's cause of the prompt right? I did the tests and I realized that almost 70% of the time taken to generate an image is spent before the actual sampling process

[fleet goblet]

use t5 8_0

OR

use t5 at 5_k_s

Anything below gets you quality loss

[haughty roost]

Ok hnold on

Is there a specific order

[fleet goblet]

nope

[haughty roost]

Can you send me that enhanced prompt thing?

I tried scrolling up and its not there

[fleet goblet]

Huh? That's a different install entirely alongside LLM-Party

I have ollama set up

https://github.com/heshengtao/comfyui_LLM_party

[haughty roost]

No I mean the prompt you made

I want to try out this new setup

By the way are these settings right? I didnt know if that was part of the reason that t-rex earlier looked absolutely cooked

[fleet goblet]

Oh. You want the enhanced T-rex prompt.

I don't have it anymore since I enhanced

a different prompt

[haughty roost]

Ok thats fine

But are those settings right?

Thats from the workflow from forever ago

[fleet goblet]

schnell can do simple, normal and beta scheduler fine I think

Yep

that's basically it

[haughty roost]

Ok

Still taking forever on the prompt node

[fleet goblet]

Which it will no matter what

[haughty roost]

Is flux only long because they have some super advanced prompt understanding thing cause I swear it takes 2/3 of the time just stuck there

[fleet goblet]

you are on --lowvram mode

clip encoders are run on cpu in --lowvram

[haughty roost]

Yeah cause 8GB

Wait my memory is not even 60%

Does that mean I shouldnt be in lowvram

[fleet goblet]

It's not loaded into your vram

[fleet goblet]

Lowvram loads clip text encoders into CPU, under sysram

[haughty roost]

Oh hold on then

Should I restart comfyui in normal or highvram then?

[fleet goblet]

Normal, if you want it to load into vram
However it might try to keep models loaded. That's the problem

[haughty roost]

What does that mean?

But I'll set it normal, high is prob way too much for me until I finally upgrade my gpu

[fleet goblet]

As in, it might try to keep models loaded into VRAM despite it going to the next node.

[haughty roost]

Ohhhh

Is there like a slightly lower normal vram mode?

[fleet goblet]

nope just lowvram

lol

[haughty roost]

Ok

Also just wanted to check the previous generations before I got all this new stuff, turns out that this is actually my slowest in a while at 108 seconds

But prob due to the cpu thing

Wait what happens if I dont set a vram mode?

[fleet goblet]

It just goes to normal

or whatever was previously set

[haughty roost]

ill just set it to normal

I started it again after setting it to normalvram

YO

I dont know if it kept my prompt remembered so it didnt have to take long on it but if it didn't, it took 55 SECONDS

Dan thank you bro

You literally cut it in half

Give me a new prompt so I can see if it was a fluke or noty

@fleet goblet

You will not believe this. I gave chatgpt the prompt to make a new version in a random style, and it did it this time in 33 seconds not counting that little wait time at the vae decode thing when it turns the data into an image

Also sorry for spamming, I'm just excited this works now

[wispy nest]

my rtx 2060 is fluxing

@haughty roosti am as happy as u

[haughty roost]

Thank you kind sir

Anyways after putting this version through firefly to remove the 6th and 7th finger, as well as some random floating sword, check this paper mache version out:

[wispy nest]

flux on 2019 gpu in less then 26 sec

[haughty roost]

WOW

I do think you should lower your denoise thing though. I did that and ever since, my photorealistic pictures have come out way better since the noise looks like texture to the skin

Or you can manually add noise but I like keeping mine between 0.85 and 0.95

Look how cute this little guy is

This is way better actually, adds more depth I think

[fleet goblet]

I'm happy that it does.

I basically do the chatgpt thing you did but locally

🤷‍♂️

[haughty roost]

Dude you're literally a life saver

[haughty roost]

Did you make that just now in case I said that to you

[fleet goblet]

Nope. That's an older image.

LOL

I do have it running right now, though.

I run the 8_0 variants since I have the VRAM for it. I don't have the VRAM for the full precision ones

But unlike you I'm on Intel Arc.

[haughty roost]

Btw, I'm using firefly to edit these images since some things look weird, but I cant figure out how to get this license plate to say something like "GOFAST"

[fleet goblet]

What was the prompt for this image?

[haughty roost]

See intel arc might be one of my favorite gpu's however I'm sticking with nvidia due to its compatibility with the games that I play, especially this space sim I bought a while ago that is NOTORIOUSLY buggy with AMD, so I dont even wanna think about it trying to run on arc

A car with rocket boosters emitting purple gas speeds through a space tunnel with swirling red, yellow, and black colors, all captured with intense motion blur to emphasize the high-speed movement

[fleet goblet]

This is gonna seem funny to you

Would you like to know what I added to the prompt to add those words?

[haughty roost]

🤔

[fleet goblet]

The license plate of the car has the words "GOFAST" on it.

Same prompt. I only added that.

[haughty roost]

Weird

Is it possible that whenever I get those like annoying errors when everything else works fine that it's just a seed thing?

Are some seeds destined for me to fail?

[fleet goblet]

Some seeds can straight up be not good

however, keep in mind you are running lower-precision quantizations of the models

[haughty roost]

Oh right

[fleet goblet]

Though I will say the differences are not nearly as big as you'd expect...

https://imgsli.com/Mjg3Nzg3

As you can clearly see they really aren't that different at all lmao

Some people have gotten better images than expected with 4_0, directly on par with FP16.

[royal monolith]

not fail, but since the AI does math and uses the seed for the starting point of that math, the math could land the final result in a strange area of latent space. and then you wind up with odd, strange, and bazaar data

[haughty roost]

Are there seeds that are just godly and give you most of what you want

[fleet goblet]

There are certain seeds that do have the right values for the latent space diffusion to appear a certain way.

[royal monolith]

no. the AI is doing matrix multiplication to find locations on an XYZ grid and retrieve the data that it stored at those locations (called vectors) when it trained. if you use the term apple, it's going to find all the information it learned about apples. but if the seed just happens to be the wrong random number it could land at a vector where some data about apple, some about apple tree, and some about a dog named apple is stored right next to each other. and it might give you a furry apple

[haughty roost]

Can you sum that down in like 1 sentence

I started reading and my brain overheated

[fleet goblet]

So it's a no, but you then explain yourself in such a way that makes it seem like a Yes

[royal monolith]

and because every AI stores the data itself, and decides the vectors itself, when it trains, there are not 'special seeds' that work for all ais

[wispy nest]

woh, i made an image to image workflow of flux by mixing two workflows

[fleet goblet]

because I'm pretty sure a furry apple is a failure when you try to generate a normal apple.

[royal monolith]

it's only a yes IF you explore one specific AI enough to know that seed xxxxxxxxxxxx gives you good results when used with teh tokens for yyyyy

that seed will only work with that specific AI and those specific tokens along with the all teh other specific settings such as cfg, sampler, scheduler, etc

so save that if you run across it so you can use it again

[fleet goblet]

But that is an impossibility, as one seed that gives you good results regarding one specific set of tokens might be good with a different prompt.

Too many factors.

[royal monolith]

it's a random number. when you hit geneate, the first thing that happens is that all your prompt words are turned into numbers we call tokens. then the AI starts doing math. and it finds ALL the data for those tokens. then it rolls another random number and picks a chunk of that data to use

[haughty roost]

Okkk so a no then

[royal monolith]

it's a no

[fleet goblet]

Which is just an elaboration of what I just said I assume?

[royal monolith]

it's a clarification of what happens for those that aren't aware of how this works

[fleet goblet]

so it's both.

Since context is multiple and many.

👍

[royal monolith]

there are no magic numbers. period

you can lock your settings and math in place by using the same seed each time, which will let you change the prompt and see how the tokens work

[fleet goblet]

There are magic numbers in very niche, very specific cases where you know a seed outputs what.

That's basically what you said.

Lol

[royal monolith]

welcome to the world of actual mathmatics. i said there are no magic numbers. and I'm very very very tired of you trying to twist what I said to make it mean what you want. i'm sick and tired of trolls.

so i'm done with you. i wasnt' talking to you, anyway.

[fleet goblet]

Supposedly

I am a troll

and not just stoned and having a good time

🤷‍♂️

[royal monolith]

go jump in a lake and don't come back out

[fleet goblet]

I did not even think this was an argument.

[fleet goblet]

Alright.

Telling me to end my life over something like that, is pitiful.

Ironically, it seems it's my turn to block you.

[haughty roost]

Yall I just wanted to know how seeds worked regarding prompt accuracy why you fighting over calculus 😆

[fleet goblet]

We aren't fighting over anything. He's mad I explain things in a different way.

I never even insulted him. Seems just questions were enough for an individual who dare not have the patience.

[haughty roost]

I didnt really read any of it cause I got too confused on the first message so

Yeah he took that WAY too farf

[fleet goblet]

Also the last part is actually just me memeing on him since I noticed his pointless anger

Aka this

[haughty roost]

By the way, I generated the image again and it gave me this

I think this is the best of the three

[fleet goblet]

Nice.

[haughty roost]

But how can I get it so it makes a space time tunnel and not an actual tunnel:
A car with rocket boosters emitting purple gas speeds through a space time tunnel with swirling red, yellow, and black colors, all captured with intense motion blur to emphasize the high-speed movement. the license plate reads "GOFAST"

[royal monolith]

when you put the words in " " you tell the AI to write them.

[royal monolith]

don't use the word tunnel

[haughty roost]

I wanted the thing from like the peabody and sherman movie if you know what I mean

[royal monolith]

try wormhole

[haughty roost]

But like then what do I replace it with?

OH SHOOT thats so much smarter

[royal monolith]

think about sci-fi and the terms used in movies

[haughty roost]

But again thank you Dan for helping me generate like quality images

[royal monolith]

and add the term "motion blur' to your prompt

[haughty roost]

Im now averaging 30s an image, way better than 100

I did, I put intense motion blur

[haughty roost]

Still giving me a road

[wispy nest]

i screen shared ur render using flux

[royal monolith]

cars run on roads, the AI knows that. what's your prompt?

[haughty roost]

Sandland how did you end up choosing 0.74 as your denoise, its so specifid

A car with rocket boosters emitting purple gas speeds through a space time wormhole with swirling red, yellow, and black colors, all captured with intense motion blur to emphasize the high-speed movement. the license plate reads "GOFAST"

Maybe I should write flying

[royal monolith]

try: rocket car flying through a wormhole, motion blur

see what just that gets you

[haughty roost]

A flying car with purple gas rocket boosters speeds through a space wormhole with swirling red, yellow, and black colors, all captured with intense motion blur to emphasize the high-speed flight. the license plate reads "GOFAST"

Let me try that first

I gotta redo it cause it made some weird flying contraption

Let me try 1 more time

[royal monolith]

try the short prompt i suggested first

[haughty roost]

Yeha it didnt work, Ill try yours

See I like that tunnel but with a flying car

Do you think you can inpaint it for me and add the car I showed earlier?

[royal monolith]

so now you ahve a good idea of how the AI is thyinking

so i would take that into photoshop, remove the rocket, take the other into photoshop, remove the background, and then composite

[haughty roost]

Got it I can try

[viral frost]

Flux:

[haughty roost]

Ok, this is the best I could get it

Still pretty sweet

[royal monolith]

not bad

[haughty roost]

If the tires worked and maybe I colored the car slightly cooler, it would be perfect

[royal monolith]

you have photoshop - generate a few more cars

[haughty roost]

Too lazy

I think it looks good like that

[fleet goblet]

Ideogram v2 released (The below image is from ideogram v2, the rest are flux)

Seems quite good as well.

^

[viral frost]

Flux: Taking the dog for a walk

[haughty roost]

What resolution is this

[wispy nest]

thanks for WF

[fleet goblet]

Seems luigi found a sparkler

Other than the L being replaced with the utorrent logo

[wispy nest]

screen share live is the best way to test ai options like sampler

[languid pebble]

Stable Video Diffusion

[wispy nest]

new ideogram 2.0 looking good

[steep sleet]

AuraFlow0.3

[viral frost]

4032x2304 It's a 3x upscale.

[lime star]

Flux: Taking the dog for a walk

[wispy nest]

how is the speed on ur gpu ?

[steep sleet]

about 110 seconds/image

RTX2070 8Gb

[steep sleet]

Flux.Dev with rbcharc0al LoRA

[wispy nest]

wow nice

I like this lora more now

[misty forge]

Wtf is this flux

Some of those look real

[celest sigil]

Yes, flux

[flat sand]

you can get much more realism with flux. It's hard to prompt, though

[steep sleet]

Flux.Dev with rbcharc0al LoRA

[misty forge]

Should I even try flux with my chipset?

i7 9750h 32gb 2060m 6gb

[steep sleet]

Flux works on my RTX2070 8Gb VRAM PC

[wispy nest]

yeah the smallest quants may be ok

[steep sleet]

6 biggy bytes might be a tight squeeze ... try it!

[wispy nest]

unless you are doing extremely "secret" images
would recommend cloud

its like 1-2 dollars per hour for 40-80GB VRAM gpus

[misty forge]

Secret images 😂

[wispy nest]

lol

[misty forge]

Ngl I'm not gonna pay, I don't have a commercial use for this at the moment

[wispy nest]

yeah I understand. cloud fees having been costing a lot lately

[misty forge]

SDXL is already leaps and bounds past what I remember ai generation looking like

[wispy nest]

oh I am still on SDXL lol

[misty forge]

It is frustrating when it can't get something but sometimes I like the glitches

I think when our outputs become flawless it will become less interesting to me and I will be more likely to side against it

Companies like Adobe are already trying to leverage ai into workflows but idk if I like that

[wispy nest]

yeah I spent most of my time trying to break the model in funny ways

[misty forge]

That honestly looks real

I just find that every model makes the pictures a little too high contrast/almost beveled looking

Some grading/filtering might fix that though

[wispy nest]

there is a genre of lora called boring reality that you might like

otherwise stock photo or amateur photo models

[celest sigil]

I used https://civitai.com/models/652699/amateur-photography-flux-dev

[misty forge]

it looks pretty good

yeah I don't have anything aftermarket for fooocus

it comes with a regular model, realistic and anime

22GB

[celest sigil]

You can download others

[misty forge]

so I can just look for any SDXL model right

because Fooocus is my front end/GUI

[celest sigil]

Yes

[misty forge]

just needs to be a .bat file right

[frank drift]

I wasn't aware SD had video capabilities! I must investigate! Thank you! Any YT videos that might help me on my way to get started?

[steep sleet]

ipiv motion

[lapis perch]

Hey guys! I have been trying to generate Iron Man Prime for over a year now but, every model available cannot make anything other than MCU/Movie iron man. Any tips?

[wispy nest]

a lora would be ideal

[steep sleet]

Flux.Dev

[celest sigil]

No, they're .safetensor files

[misty forge]

ok

I launch my program through bat files but I see the safetensors in advanced settings

[celest sigil]

Flies through the air, but never has any bugs splattered on the armour!

[still rose]

Loving Flux 1 so far

[misty forge]

damn that looks real

[wispy nest]

soogsx lora

[nocturne oak]

Isn't nanotechnology great? (Plus, he's got a good ceramic coating applied. 🤣 )

[languid pebble]

Sorry in Cologne at the moment so can't answer everything.

[still rose]

fr

runs on macbook quite decently

[misty forge]

wow that's impressive

[still rose]

wait ima need 2nd image prompt thats good

[languid pebble]

Real photos from Cologne 😂

[still rose]

im fucking stupid i didnt read the top