🏞｜general-with-images

[shut sinew]

Better than what?

[hazy warren]

More detailed, more to what you typed

[shut sinew]

Personally I like the XL model a lot yeah, but if you wait a little more SD3 seems to be around the corner sooner or later and has superior prompt understanding from the few peaks we got so far

[jade fable]

##🏞｜general-with-images message

[hazy warren]

Oh yeah, I guess I’ll just wait till that.
Would it be just a model I download, or would I need new program?

[shut sinew]

It will probably work in the common UIs, but im not sure

[jade fable]

一个建筑

[shut sinew]

#1047610792226340935

[hazy warren]

😮

[hazy warren]

Hey, this is actually pretty good realistic shot

[wispy nest]

Thank you

I like that tree one you did the other day

[noble zephyr]

I have prepared a guide for Tokiidokii regarding my pixel art workflow, but it's quite a wall of text. is this channel appropriate to post it, or should i post it somewhere else?

[hazy warren]

Thanks mate 😛

[shut sinew]

maybe #1092446741984444416 is the right place

[noble zephyr]

right, thanks!

[hazy warren]

I actually have trouble making the A.I have imperfections in the images

[subtle depot]

Any way to get the right hand back when doing a laten upscale? Or would it be better do an image upscale instead with like ultimate sd upscale and upscaler model?

[subtle depot]

Yep, seems to work better. joy!

[pallid estuary]

Make a logo for the human centered ai lab called HCAI

[proud orchid]

@broken zinc

They are incredibly cheap

[broken zinc]

thats better than a 4070 ya?

more vram

[proud orchid]

yeah, it has 24gigs of vram. In AI stuff, VRAM is really important

[broken zinc]

cool ty!

[proud orchid]

🌟

you're welcome

[nimble mason]

Tiled upscale works really well too

I use random strict usually, don't use SDE anything if you do

[steady skiff]

Perfection

[shut sinew]

Latent upscale tends to change the image since you compress it to a much smaller space and then extract it back out, so image upscale will leave you with more consistent results

[subtle depot]

Yeah 🙂

[tepid fjord]

Can i run automatic 1111 on 4gb ram Intel hd graphics without nvidia gpu?

Locally

Need help please answer 🥲

[noble zephyr]

i don't think so, sadly

my 16gb ram and 3gb vram pc has trouble with stable diffusion at times, so it's safe to assume that with no vram at all, it's going to be EXTREMELY slow, if it even works at all

[tepid fjord]

Like how much slow 🥲

[noble zephyr]

like 40+ minutes for a single image slow

[hazy warren]

Just get a 4080 and call it a day

🥲

[nimble mason]

"an artistic painting in the style of van gogh of a hybrid of a shark and a clown holding up a sign that says "SD3 was the king of prompts for less than a week". The shark clown should be standing in army boots on top of a cow which is doing a handstand on top of an alligator. This is happening on the moon. it should like an integrated work of art, not a collage. it should not look like photos were cut out and glued together."

[limpid lichen]

@heady anchor

[heady anchor]

he looks like a mix between seagal and charles bronson 🤠

[limpid lichen]

lil bit of baldwin too maybe

more accurate glue face

[wispy nest]

@hazy warren

Gramps says do your dishes

[hazy warren]

😮

[wispy nest]

A bout A year ago I proposed a theory in my own discord about a music text to generations that has a trending page and I cane across a site called riff fusion small world man someone gonna take something one mana trash another treasure right

[hazy warren]

😮

[wispy nest]

you know what would be cool realisic miscraft like real life cgi

[hazy warren]

probs a mod for that

[wispy nest]

yeah

maybe

[ancient gazelle]

/

[wispy nest]

handshake domains and mj shashes in foriegn lands

[frail wraith]

https://tenor.com/view/leonardo-dicaprio-clapping-clap-applause-amazing-gif-16078907558888063471

[hazy warren]

now you have a make an image of someone clapping

[crude ruin]

a female doctor in a red dress walking in a garden

[nimble mason]

Here is the image you requested.

[warm vine]

Lol he asked 😂 for female doctor in "red dress"

[jaunty jay]

• red dress not green

[smoky oak]

@shut sinew @tropic shell

rando spamming me AI ads in DM's

[smoky oak]

where is this from? lmao

[smoky oak]

I second this. Got my 3090 for $700 like 5 months ago now. Best purchase I ever made

hands down the best AI card for the money

[shut sinew]

Heyo, you can also use our ticket system for this stuff

[tepid fjord]

Damnn i give up 💀

[deft bison]

Satan curses him with his army, in the middle of a very wide highway, preventing people from passing towards the light ... !!!

[noble zephyr]

@dense nova one easter egg with a weird pattern. it looks kinda like the denoising went wrong, but that's how it's supposed to look

[dense nova]

Hmm. Not quite what I was imagining. More like silly string sprayed on an egg.

[noble zephyr]

fair. i'll try again after i finish the other dude's image

[dense nova]

I was just kidding though. I left a1111 running before I left for work this morning.

[noble zephyr]

oh lmao

[dense nova]

It's an easter egg, I promise.

Now that's more like what I was imagining! 🙂

[deft bison]

Satan the cursed attacks and attacks human beings, He surrounds them on all sides with his horsemen and his infantry!!! ...HD realistic photo

[cobalt flame]

Looks like a yarn image I did

[pallid ruin]

Wow it looks good 👀

[hazy warren]

needs more fire

[brisk moon]

给我一只在雪地钓鱼的狗

[nocturne oak]

不。

[zinc kestrel]

Photography: Renhang and Sandara Tang: half-length shot, a woman wearing a white suit jacket, white dress, short black hair, subtle monochromatic tone style, shiny/glossy solid color, classic preppy, pure white background, unique beauty, Thin face, business picture. Beautiful new camera angles, professional portrait lighting, studio lighting, commercial feel, professional photography, extreme detail, ultra high definition, maximum detail, ultra detail, sharpening, sharp detail, amazing quality, super detail, incredible The details of HDR16K, the most realistic --ar 3:4

[nocturne oak]

https://tenor.com/e2H2KgNCDPm.gif

[cloud niche]

Create a Solana cryptocurrency

[nocturne oak]

https://tenor.com/bwCWP.gif

[cloud niche]

What is the new prompt

[thick bramble]

give me prompt im bored, doing 1440p SDXL base

also its cold so I need to heat my room

[nocturne oak]

#1047610792226340935 message

[cloud niche]

I just need a crypto currency tee shirt design

[thick bramble]

funny, but you warmed my room by 1C rendering this. Try harder tho
cropped out distortions

[candid dawn]

hey

is there any way to train ai to recognize this kangaroo

Kayla Christling

[thick bramble]

Zoophobia, I respect that. It's def possible if you have enough references, someone might have already done the hard work but you might have to step into nsfw territory to find it.

I haven't trained a model yet but if someone here has, best to ask them

[candid dawn]

well who then

[thick bramble]

Idk I'm new here. There might be models on civitiai that cover her, if not, I'd look into tutorials on how to make your own trained model

[candid dawn]

ok

also I would make nsfw of her, but it'd be as gorgeous as this Durango

muscle

[thick bramble]

Comparing her to this abomination, heresy

[candid dawn]

female muscle and then

https://tenor.com/view/dodge-charger-mopar-spin-out-donut-gif-17668635

[thick bramble]

That's better

[spring mist]

MoveOverPeopleAreRacing

[candid dawn]

Does anyone have a Kayla Zoophobia model

[hazy warren]

I've noticed that, the more time you generate an image, the better the image becomes, but like over 20-50 though

[candid dawn]

oh, so would that help the ai recognize a character from zoophobia better

[hazy warren]

[candid dawn]

would it

[hazy warren]

Strong frog LOL

[final arch]

invitation poster

[amber plume]

casa

[fair pine]

box

[simple slate]

how can I generate image?

[noble zephyr]

do you have a powerful pc?

[simple slate]

yes

[noble zephyr]

https://www.youtube.com/watch?v=kqXpAKVQDNU this video will guide you through the installation

[simple slate]

thanks!

[noble zephyr]

testing out some loras with my pixel art workflow. seems to work pretty well! i like mechs if you weren't able to tell

[quiet current]

nice, looks pretty grid aligned, what s your workflow ?

[noble zephyr]

it's because i downscaled them using an extension and then upscaled them using nearest neighbour upscaling. they're pixel-perfect

https://discord.com/channels/1002292111942635562/1212330984192614440 and here's my workflow

[quiet current]

ok that s pretty much what I had in mind. I remember using https://github.com/KohakuBlueleaf/a1111-sd-webui-haku-img?tab=readme-ov-file#pixelize back in the days as I loved its dithering

[noble zephyr]

oh yeah, that looks pretty good as well! not really the style i'm going for tho

[quiet current]

#1047610792226340935 (also that s not how bots work)

Both code base (pixelize and palettize) look pretty clean, should be relatively easy to work with it and implement extra averaging methods, adding some custom convolution matrix support to improve contrast/outlines/etc if needed. I might give it a try at some point.

It all depends if my curiosity can beat my dislike of python. Too many ptsd from image analysis courses :p.

could be fun to add custom palette support too.

[wispy nest]

How do negative prompts work? I have a heterochromia negative prompt, and the image still comes out looking like this

Is there any way to make them stronger?

Or is that just a limitation right now

And i can only count on negatives to be gentle guides to the generator?

[shut sinew]

Give an eye color in the positive prompt

[wispy nest]

I did

They are supposed to be purple

[noble zephyr]

if you use A1111, you can select what you want to have the weights increased for and press ctrl+up. it should look something (like this:1.1)

[wispy nest]

In other generations, they are sometimes brown, sometimes red

[shut sinew]

You could try regional prompting

[wispy nest]

I am on A1111, let me see

ohh

That's what that means

So I am assuming 1 is base

Anything lower is lower importance and vice versa?

[noble zephyr]

exactly

[wispy nest]

Another question, is it better to type prompts as very simple words separated by commas, or to write sentences and descriptions?

I've seen both types of prompts in really good looking images

[noble zephyr]

i think it depends on the model, but it shouldn't really matter that much

[wispy nest]

Thanks

[frosty wave]

Sent here so i can send message link to #💬｜general-chat

A scatter plot with three points. The first point is at the coordinates (0,0), the second point is at (1,0), and the third point is at (1,1). Each point should be clearly marked and distinguishable. The plot should have both X and Y axes labeled, and include a grid for easy reference. The background should be white to ensure clarity. The overall aesthetic should be simple and professional, suitable for a mathematical or scientific context.

(best out of 4 generated images with ideogram)

[shut sinew]

Ill try this when we get access to SD3

[main dew]

sd3 probably the same

[shut sinew]

Also funny how it ignored the background should be white

[main dew]

interesting lol

[quiet current]

you can :

• specific a color for the eyes
• increase the heterochromia strength in negative prompt
• fix those details with inpainting/adetailer/etc
• if you have those orange, brown color specified somewhere else in your prompt use sd cut-off to stop the color bleeding
• etc

[frosty wave]

yea i think it's a very hard task. I originally gave it to Dalle 3 and it gave me a scatter-plots with a lot more that 3 points (~100 points). (the prompt above was written by chatgpt4 since it always edits prompts before running them, probably some parts of it should be removed)

also idk why it ignored the white background here

[shut sinew]

I asked someone with SD3 access to run the prompt

Until then enjoy weird stuff by me

[wispy nest]

Is there any megathread or like a written guide with all relevant info about using prompts?

At least the basics or intermediate

So I can learn to translate english into prompts langugae

I can fix those details manually of course or with image to image, just curious about how to make the most out of the initial ones

[shut sinew]

Prompts are different from model to model, you can not compare 1.5 to XL for example and neither can you fit in the anime models for example

[wispy nest]

Oh I see

So I should look for that on the page where I downloaded the model

[shut sinew]

Usually I check civit for stuff other people used as prompts and go from there

Just browsing the images made with the model

Or you can run random stuff to see what sticks

[wispy nest]

What is hte method to fix hands and fingers?

[shut sinew]

I like ADetailer with the depth hand model thingy

[quiet current]

mainly : inpainting, adetailers, controlnet

[wispy nest]

Ill look into adetailer, thank u

Is this it?

[shut sinew]

https://github.com/Bing-su/adetailer

[quiet current]

should be it

[wispy nest]

I love how I can search any of this terminology and there's so many tutorials on it

[quiet current]

Use recent ones, and keep in mind that it s easier to update a text tutorial / github readme page than a YT video.

[wispy nest]

Of course

[wispy nest]

https://civitai.com/models/293448/anzhcs-face-segmentation-prototype-or-yolov8-or-adetailer-model

Where do I put this file?

To use with adetailer

[deft bison]

Satan and his worshipers flee the heavenly lights... realistic photo

[shut sinew]

Read the description

[wispy nest]

https://tenor.com/view/mouxy-cotorep-aveugle-debile-esquiver-gif-20546128

[neon vine]

almost made an advertisement lol

[dense nova]

Stable Cascade 4K rainbow wave.

[limpid lichen]

@hazy warren 1min 14sec

[deft bison]

Satan and his worshipers flee the heavenly lights... realistic photo heavenly forces attack Satan and his prostrating worshipers... photo HD

[pallid ruin]

It looks amazing, what model did you use? I had trouble with different characters on the same image

[candid dawn]

https://tenor.com/view/webby-vanderquack-ducktales-sparkling-eyes-happy-gif-18624896

[deft bison]

👍🤝

[fair cypress]

a beautlful girl

[grave briar]

Satan and his worshipers flee the heavenly lights... realistic photo heavenly forces attack Satan and his prostrating worshipers... photo HD

[mystic bluff]

Satan without wings, burning in hell

[wispy nest]

test bitcoin

[wispy nest]

By doing this, i can do bulk generations without the seed changing?

If not, how do I do that?

I found a few seeds that I like and want to generate them until it looks good without manually clicking it every time

[glossy cargo]

Protein structure (three-dimensional)

[faint cobalt]

||Hi guys how are you||

[dawn seal]

Just what the world needs...winged cats.

[mild jay]

is there way to improve while using sd 1.5? like you got any tips or tricks by chance?

[wispy nest]

Exposing websites real person chat

[pallid ruin]

😁

[spark finch]

improve how so?

improve those images, or just gens in general

[spark finch]

But what I usually do is upscale 2x, using a high upscale model, especially a nice HAT upscaler, then run it through multiple 1.5 models until you get a feel for that original image being "better" and to your liking, sometimes multiple passes with multiple models, can do some nice things to an image, but 512x512 models don't do so badly below .50 denoise, eg

suttle but, works

[mild jay]

the images like do i need gimp or something?

[tepid plover]

Hello all! I am new to stable diffusion and excited to get set up

Being slightly overwhelmed, I was curious if anyone had some advice on getting started? I have a pretty good PC so I'd like to use local resources as much as possible

[gaunt heron]

Nvidia or AMD card?

Nvidia install guide
#🤝｜tech-support message
AMD
#🤝｜tech-support message

[tepid plover]

4080 - I have started using Forge, it has been fun!

[gaunt heron]

ok nice work! happy imaging

[nimble mason]

Good! Forge is pretty great. You can use it on your LAN too, add --listen to the arguments in webui-user.bat

I use it on my phone all the time

[spark finch]

nah, just use img2img and about .3-.4 denoise and try different models and your image you have already made

[solid kernel]

Used Pix2Pix to rerender my boyfriend's Miata as an Oil Painting

then touched up in PS 2022

[mild jay]

i was given a weird prompt and here's i what i got.

i used a different sampler for each image.

[nimble mason]

original before tiled sampling

before tiled sampling

after tiled sampling

dreamshaper SDXL

tiled... needs some tweaking of the settings lol

[weary nimbus]

i used a different sampler for each image

Satan and his worshipers flee the heavenly lights... realistic photo heavenly forces attack Satan and his prostrating worshipers... photo HD

[nimble mason]

this was with a diff method called iterative upscale

tile preserves the source composition much more effectively

[mint quest]

Prompt: On the ancient city wall in desert，First-person perspective, elegant, volumetric lighting, digital painting, highly detailed, artstation, sharp focus, illustration, concept art, ruan jia, steve mccurry

[nimble mason]

Looks like a 3d render from the Morrowind engine

[mint quest]

I mean,In fact I want general a image from this picture,but I dont know how to do

[nimble mason]

what do you mean

generate an image from this picture?

you need img2img

[mint quest]

yeahyeah, I need this,but I dontknow how to do.Can you tell me where can I found img2img?

[nimble mason]

https://github.com/lllyasviel/stable-diffusion-webui-forge

[mint quest]

so...its no in here? Because I see "general with image"write here..

[quiet current]

Because it's a "general" channel and you can post images here (unlike #💬｜general-chat).

Bots and subsequent bot channels are currently offline, cf #1047610792226340935

[white herald]

the best quality,masterpiece,extreme detail,Photorealistic,cubism,contemporary art,woman,cheerleader,Small waist,Long legs,wavy hair,black hair,seductive smile,watery eyes,corset,yoga pants,sneakers,rimless eyewear,bent over,background,indoors,in summer

[quiet current]

nope still offline #1047610792226340935

[sly plaza]

I m not crying I m not crying I m not crying

[zenith skiff]

How to prevent final image from losing the colors visible in preview?

[wispy nest]

Just imagine how 7900 XTX users feel

For over $1000 gpu

beaten by 3060 ti in this area

[sly plaza]

I just bought a 2080 ti 12 Gb yay

[wispy nest]

interesting choice

hope you enjoy it

[nimble mason]

i mean... it is 7 years old lol. remarkable it's even usable at this point

[nimble mason]

yeah, that's gotta hurt a lot lol

[dry crow]

With Zluda I get 16 it/s on Windows.
This test (from the screenshot) was made with the slow directml webui

On Linux with rocm it gets 20 it/s

[wispy nest]

ah good

[neon vine]

i used to be a radeon fanboy, until i got a 3080

and oh boy i see why there are so many green elitists

[wispy nest]

Its not elitism just the truth

Nvidia is better but more expensive

[neon vine]

yes

way too expensive if were honest

[dull flame]

anyone got any tips on improving eyes?

[nimble mason]

segment out the eyes, upscale, unsample then advanced ksampler, then patch them back in with a soft mask for blending

[dull flame]

alright ill try it, if i can make it happen

[nimble mason]

are you using comfy

[dull flame]

yes

[nimble mason]

i can swing you a workflow demo that i use for faces

[dull flame]

i would appriciate that yes thanks

[nimble mason]

price gouging for sure

[nimble mason]

lemme go poke around in my dumpster for it lol

[nimble mason]

alright here we are

cleaned it up a tiny bit so it's easier to understand vs my usual giant messes

original image

latent upscaled x2 and resampled at 0.5 denoise

after unsampler/resampling of the face itself

[dull flame]

thanks, ill mess around with it tommorow

[nimble mason]

fyi the CFG setting for the unsampler is high, usually you want that between 1-3

so if you have problems, drop that down

[frosty wave]

another way to do it is to edit the image in an image editing software like gimp. I especially recommend the warp tool for things like this.

[nimble mason]

for especially good results, tweak things closer manually as described above, then pipe back in to img2img with low denoise to add polish

[sterile kiln]

Hi guys, has anything changed since A1111 1.8?
I find it much more efficient, it takes all LoRA simultaneously much better ! 😮

[lost valve]

is that supposed to be taylor swift

because if it is, please use my lora

L0l

[warped ravine]

just wake up lazy fat cat

[limpid lichen]

better when the lora power turned down. ip-adapter face swapping anime pics

[celest root]

Cookie Monster

[flat narwhal]

glass

[hearty shell]

Chess king fashion dripping

[quiet current]

This channel is still not a bot channel. And bot channels are still offline, cf #1047610792226340935

[noble zephyr]

AnyLoraCheckpoint: LCM vs. non-LCM

same prompt, same seed, the only settings that were changed are the sampler, steps and CFG scale

[heady anchor]

yea looks good for a flat style but if u want more details thats where it fails

[noble zephyr]

yup. SDXL Lightning however, is pretty decent at details in my experience

[wispy nest]

My greatest work of art

[noble zephyr]

what a beautiful piece

what's it called?

[wispy nest]

It's the

[noble zephyr]

so inspiring.

[deft remnant]

id:browse 猫妈妈怀抱着小猫咪

[opal warren]

@noble zephyr do you have issues with saving the transparent images? This is what i see in the output folder

I can copy the transparent image from the webpage, but idk where it is saved automatically

[noble zephyr]

there's a little folder icon below the generation, if you click that, it'll open the outputs folder

[opal warren]

yea, but the output images arent transparent. they have the transparent background

[noble zephyr]

i actually cancelled my generation since you seemed to have figured it out (sdxl takes forever on my pc), but i'll go have another look

[opal warren]

rip

[shut sinew]

Are they saved as a png? Cause jpgs dont have an alpha channel

[opal warren]

they are saved as png. this one is right from the outputs folder

[gray verge]

Hi everyone. Please, describe me why I have such a different image shading and quality with the same settings in seaart.ai (first image) and on my own PC (2nd image, I have good video card and last SD updates). I have a such problem only with MechaMix checkpoint. What settings I should adjust?

[noble zephyr]

i'm not sure, but i think it's a problem with your VAE. try using another one

[shut sinew]

Description of that checkpoint states that you need the kl-f8-anime2.vae.pt

[noble zephyr]

it baffles me how easy it is to make good looking images with JuggernautXL lightning. here are some dragons.

[dull flame]

thanks for the image workflow, but with all the samplers, its too much for my gpu

[candid dawn]

Stunning

[nimble mason]

it's prolly the latent upscale that was the issue, that can just be removed

[deft bison]

invisible hands tear off the long white beards of the false saints that Satan the cursed has spread all over the world... photo HD realistic

[grizzled sage]

"satan put the dinosaurs there to deceive us" level

[rain gazelle]

Curious, did you figure this out?

[hazy warren]

final fantacy!

[rain gazelle]

let's see your lora

[wispy nest]

Based

[grizzled sage]

religious symbolism is the opposite of based in reality. its closer to mass hysteria

[wispy nest]

If I recall, you also made some reference to "Satan-put-the-dinosaur-bones-on-Earth-to-trick-us level stupidity". But I ask you...

If dinosaur bones are not a big trick on mankind by the Great Deciever, then why is it called "Jurassic Park" when most of the dinosaurs are actually from the Triassic and Cretaceous? Checkmate, Atheists.

[grizzled sage]

That's super bluth looking. I find that most furry art styles are inspired by bluth

[unborn chasm]

Don't make me come over there

[odd blaze]

https://tenor.com/bdJuY.gif

[gilded cobalt]

not sure if this is the right channel but what kind of prompts would you use to get an image like this?

[deft remnant]

a rabbit

[wispy nest]

yo

I know who that is

[deft bison]

observe this transparent and impassable Dome which protects the Flat Earth. Extra terrestrial ships can never cross it, it explodes. HD photo

[glacial lodge]

may you tell me what prompt that was, id like to learn how to generate better images of this quality

[noble zephyr]

It was quite simple, really. It was something along the lines of "dragon made of crystals, head focus" with no negative prompt, hires fix at 1.5

[glacial lodge]

Wow, thats amazing then

[noble zephyr]

If you want more details about the workflow, you can download them and import them in the PNG info tab on A1111/Forge

[glacial lodge]

ill try that

[chilly bronze]

A huge robot with white clothes

[quiet current]

nope, this is still not a bot channel, and those are still offline. cf #1047610792226340935

[tender hound]

Cat run

[quiet current]

...

[split wind]

generate an image of the logo of an agency located in gabon who sell, rent and manage homes, the color to use is Blue and white. The name of the agency is Agence241

[dense nova]

I'm amazed at how solid all these forms are! The keyboard has a little wonkiness in the key layout. But the forms aren't blending into each other or warping, even though the scene has so many objects in it. I would love to see if it can finally handle someone writing with a fountain pen, without making the pen upside down or something.

Comparison with SDXL.

Okay. I've read everything I can understand in the paper. Going back to my art app until we have access.

[wispy nest]

Lots of barley grains bouncing against a gradient dark brown background

[noble zephyr]

This isn't a channel for generating images. there currently isn't one, since the bot is down. keep an eye on #1047610792226340935 for updates

[pallid ruin]

Wow, im stealing that prompt 😁 👏

[raw finch]

I'm amazed at how solid all these forms are! The keyboard has a little wonkiness in the key layout. But the forms aren't blending into each other or warping, even though the scene has so many objects in it. I would love to see if it can finally handle someone writing with a fountain pen, without making the pen upside down or something.

[deft bison]

@pallid ruin 👍

[smoky oak]

@pallid ruin Here is an example image from one of my recent tests. Its far from finished, but my progress is undeniable with these new runs haha

[pallid ruin]

It looks very nice!

[smoky oak]

this re-start is still really early in compute, but its making very fast progress

it fixes focal planing, improves prompt adherance and text, allows for images without blurred backgrounds, supports ultra wide single subject images (Still improving that), fixes 1.0's lack luster offset noise, and has been done with realtively little compute on this newest run

this model still for sure has issues and a long way to go, but the progress is faster than any of my restarts ever have been

[pallid ruin]

That's good to hear 🤗

[smoky oak]

my older realism model has some checkpoints that are incredible, but they had some issues caused by an uneducated approach that I took. It lead to exceptionally good results, but also issues that aren't reasonably fixable

so I have had to restart and learn from my mistakes

one of my old example images from months ago. Single sentance positive, no negative, only 3CFG, generated at 2048x in just 20 seconds on a 3090 using one of my high res workflows

another

[pallid ruin]

Wow

What have you used for captioning?

[smoky oak]

I had some even better versions after this that improved in some pretty incredible ways, but at the cost of the issues aforementioned

[pallid ruin]

Have you tried gpt4v or another alternative like llava or that wont be so useful? 🤔

[smoky oak]

Everything 😅

All of these results have come from 100's of LoRA trainings with months of experimenting and weight injecting

[smoky oak]

I use COG now. I found it very helpful, especially after properly setting up proper instruct with my captioning style

My newest dataset is 1.6k hand selected extremely high quality images that have all been captioned extensively into different groups, as well as value normalized.

best part is, I can use this dataset at 512x, 768x, 1024x, 1536x, 2048x, and even higher, as all images in it are 4k-10k+ resolution

not too long ago, I did a super small demo for a LoRA to expand base SDXL to 1536x res, and it actually worked quite well for very little work. I have plans to try it again sometime soon

it also helps that I have access to a 500k+ dataset with CogVLM captions from my research group, should I want to use those images

However, I have found my own sort of approach working better for me

[pallid ruin]

Thats amazing bro, I hope you can continue improving it

[smoky oak]

I have over 600 LoRA trainings for this model as of now, though most are research and fails

[smoky oak]

Thank you! I am in talks with a few companies about potential compute support. Been working on some contracts in the industry. Really passionate to do this full time

I have made all this progress on a 3090, imagine what I could do with 16xA6000's haha

[pallid ruin]

Hahah it would be way faster training on that

[smoky oak]

My ultimate goal is to be the head of a research team, directing to look into all sorts of different approaches

I do have a slim, but potentially viable option to do just that at a pretty well known company in the scene, but I'm not too hopeful that it will actually go through anytime soon

[candid dawn]

The Snake King

[pallid ruin]

i gotta try highres fix

[smoky oak]

dawg

[dull flame]

once again, thanks for the help, with the latent upscale removed, my gpu handled it and the workflow time was around 260 seconds
it masked the face but made the whole drawing worse in the end, left is after first ksampler, and right is finnished

@nimble mason

@nimble mason

[blissful gorge]

masterpiece,(bestquality),highlydetailed,ultra-detailed,1girl,creativeoffice,vibrantenvironment,colorfulposters,inspiringatmosphere,modernfurniture,fashionabledesks,comfortablechairs/msg user: message:

masterpiece,(bestquality),highlydetailed,ultra-detailed,1girl,creativeoffice,vibrantenvironment,colorfulposters,inspiringatmosphere,modernfurniture,fashionabledesks,comfortablechairs/game_stats

[nimble mason]

Hmm I'll take a look. Looks like it resampled the whole image maybe with the same seed

[dull flame]

im trying out this one now

[nimble mason]

Which node it it hit an OOM error at before? Was it the latent upscale or after that?

[dull flame]

yes, i simply removed it

it was at latent upscale

[nimble mason]

How much vram you got?

[dull flame]

sadly, only 8

[wispy nest]

how fast did that image take on 8gb vram?

[dull flame]

it was just an image to share the workflow, i only made the anime pics earlier

[wispy nest]

oh ok

[dull flame]

it prob would have taken me 10 years

guess it might be time ti invest in a nvidea card

the second workflow tries hard but its pretty pitiful results

[nimble mason]

which is the second? one i sent or something else

[dull flame]

no, this was the second workflow i found online, i have yet to make yours work sadly

[nimble mason]

and yeah you'll def want to find something with 12gb vram if at all possible, it makes things so much easier

k

did you try with the tiled ksampler version?

[dull flame]

trying it now

first sampler alone takes 1min30 lol

[nimble mason]

that gives better results than a straigh tup latent upscale anyway

[dull flame]

second tiled ksampler after latent upscale, we are looking at 4min atleast = )

[nimble mason]

and will get you around the vram issue for the upscale

yeah

it's a lot slower for sure

here, send me the exact workflow you're gonna do

unlink the tiled ksampler

and let it run and save some random image

paste the png in here and i'll run it real quick to see if it even solves your problem

(on a 4090 so it'll be fast)

[dull flame]

it seems to work today with the latentupscale, no idea why

[nimble mason]

you can find yourself very very close to the edge with vram

it can merely be whether you have a folder open in windows explorer with images in it

[dull flame]

1.4it/s blazing speeds

[scenic cliff]

Create an image of (a hilarious scene) where characters from the cartoon SpongeBob, including SpongeBob, Patrick, and other characters, are (mocking) the character Plankton, The scene should capture the playful nature of SpongeBob and his friends as they (laugh) at Planktons (failed attempts) to steal the Krabby Patty secret formula, Emphasize Planktons (frustration) and the (dynamics) between the characters, The lighting should be (bright and sunny), representing a (cheerful) atmosphere, Use (vibrant colors) to enhance the humor and (expressive facial expressions) to showcase the characters emotions

Help

Pliz

[dull flame]

vram failed at 80% of the tiled ksampler, trying with no latentupscale now

[nimble mason]

yeah you might need to close absolutely everything and you prolly need the tiled VAE decoder too

[dull flame]

yeah probably, ill try some mroe to get results

[nimble mason]

Here is the image you requested.

[dull flame]

@nimble mason here is ur workfile, as i just used it now, i simple removed latentupscale

and here are the results

finds the face fine, but it seems something else isent quite right

[nimble mason]

what's the checkpoint?

got a link?

[dull flame]

sure, one sec

https://civitai.com/models/96629/realcartoon-anime

[nimble mason]

thank you for fgeeding my checkpoint addiction haha

[dull flame]

that reminds me that i will prob need to get a bigger hdd with the new gpu lol

[nimble mason]

i have a 160 checkpoints just in my "commonly used, so on the SSD" folder

[dull flame]

i checked some 3090 gpu, but yeah, they sure werent cheap

feels like i might as we ll get a 4090 with thoseprices

btw, these steps. in ur workflow seems quite high

[nimble mason]

yup i looked into the same... should've bought after the mining crash, they were cheap then, but i wasn't into SD then either

yeah, i'm generally aggressive with step counts in part cuz i'm on a 4090

way i look at it too is if, for example, i'm gonna be doing a tiled ksampling that will take 30 seconds, another 2-4 seconds for another 20-30 steps early in the rpocess is fine

[dull flame]

idk if uve had time to download the checkpoint, but can u produce anything with that specific checkpoint, with the workflow i mean?

[nimble mason]

currently DLing, i'll have it in 4 min

[dull flame]

alrighty

because ive noticed that the decoded img after the tiled ksampler is almost the same as the end result

not exactly a improvement at 30steps instead of 80

[nimble mason]

lol

[dull flame]

seems like there quite a few decode in here

[nimble mason]

yeah using the same seed is the rpoblem there, that and tiled ksampling needs the upscale first

def need to do a latent upscale of at least 1.5 to get good results with tiled ksampler

in the meantime...

try running this workflow with your checkpoint/seed etc

in general just fyi running ksampler on a fully sampled image will result in a blown out look if you have the same seed and sampling settings

changing the seed and using a light denoise with something else changed somewhat can get around that fairly often but not always

well, moderate to light denoise, so <0.6

[dull flame]

lol sure takes while to get all those nodes

[nimble mason]

slow inference on your system?

[dull flame]

what do u mean by inference?

[nimble mason]

inference = generating images

[dull flame]

im getting an error with tthe workflow

[nimble mason]

at which node

[dull flame]

i just had to change the unsampler

but i gotta rerun it now

and yes i got a old gpu so its slow

and the images u sent is rly what im after , so obviously ur workflow workds well

[nimble mason]

with tiled ksampler

cool good to know

[dull flame]

sadly, got another error now

`Error occurred when executing BNK_Unsampler:

Device type privateuseone is not supported for torch.Generator() api.`

[nimble mason]

wtf

that's a new one for me

did the unsampler run at all at any point for you

and are you fully updated with comfy

[dull flame]

ill check

it simply stopped at unsampler and i got the error

[nimble mason]

k

yeah do the update comfy and the update all thing

see if that helps

pared things down a bit for this one

no upscale of any kind

dropped the step counts a lil too and tweaked the unsampler/resample settings

obv need unsampler to work on your end for this though

this one i wired it to skip the unsampler and add noise at the advanced ksampler

[dull flame]

error again, this is slightly depressing lol

[nimble mason]

you will get ab etter result with unsampling if you want it to stick to what your original face structure etc

but this is still a lot better than nuthin

nice checkpoint btw 🙂

[dull flame]

right? its not bad = )

[nimble mason]

cool 🙂

does the clipseg part run fine for you? i've found separate mask components can take some serious CPU power for whatever fn reason with bigger images (i have a 5950x cooled by a NH-D15 air cooler, so i hear that beast roar and know what's up)

idk how it'll scale to your system with a 512x512

[dull flame]

lol i changed the sampler of the unsampler to euler_a and now it works

[nimble mason]

huh, that's odd

i've found dpmpp_2m_alt and dpmpp_2m to be the best for that in general

but could be diff with cartoons too

usually the CFG with the unsamplre should be on the low end

i had it at 6 before cuz i was messing with something and forgot

but 0-3, i usually leave it at 1-2

[dull flame]

it seems slighly off tho since i changed it

[nimble mason]

if you're looking for preserving the source composition more, you can switch the scheduler to exponential

[dull flame]

i assume i gotta change all of them

[nimble mason]

did i send you one with it on dpmpp_2m_alt? if so try dpmpp_2m

[dull flame]

no the one u sent was dpmpp_2m

[nimble mason]

see if you can get dpmpp_2m or the alt version or dpmpp_2m ancestral to work

if you use the ancestral you usually want an ancestral in the advanced ksampler too

[dull flame]

alright, trying it now

https://tenor.com/view/kid-baby-child-girl-tired-gif-16635701

[nimble mason]

lol wtf

yeah looks washed out on my end with double dpmpp_a

maybe just skip the unsampler then

direct noise added

with unsampler

[dull flame]

im trying different samplers now

[nimble mason]

k

tbh, preserving the original face structure is probably more important for photo stuff anyway

than anime

[dull flame]

hm, dpmpp_3m_sde made it completly black lol

ohwell ill try some things, still, thanks for the help, i really appriciate it

[nimble mason]

yeah you can't use SDE anything in the unsampler

no prob

in general, i've found SDE samplers work really well on the ksampler advanced side

if you use ancestral in only one, it's gotta be on the sampling side

ifg you use ancestral in the unsampler, you need one in for sampling

SDE doesn't work for unsampling or tiled ksampling (if you use strict random)

[dull flame]

hmmm okey gotcha

[nimble mason]

also, rec getting the image comparer node if you don't have it

i put it in this so you can do the install missing nodes thing

[dull flame]

thanks!

[nimble mason]

allows you to see how effective the unsampler is at preserving the face structure and orientation

[dull flame]

thats really useful!

[nimble mason]

yuuup pretty new node, veryyyy handy

what version of pytorch are you running? it should say at the top of the console log when you launch comfy

[dull flame]

hmmmm

first success lol

now when i look closer, im actually gettting some errors when i launch about torch version beeing too old

[nimble mason]

half the reason this works better than some of the other methods like some shit from impactpack for segments (maybe i'm doing something worng there, idk) is cuz it means your model is resampling at its native ideal resolution

does it say the version number

[dull flame]

cant say that i see a pytorch version in there, but my eyes might be bad

but my eyes does see total vram 1024mb

and that worries me

if thats true, its no wonder im having trouble

[nimble mason]

3.10.12

err wait oops

that's python doh lol

[deft bison]

a fantastic being of light, with wings dotted with brilliant diamonds, a slender body and clothes of green light dotted with sparkling emeralds , descends from heaven to earth, the full Moon in a sky starry...realistic HD photo

[nimble mason]

are you using a nvidia card?

[dull flame]

no, im using amd

[nimble mason]

ah that's prolly why that's not popping up like it does for me

maybe try setting up a new second standalone install of comfy and see if it works then

i'm hesitant to offer any wild advice about modding your install after someone's b0rked a few days ago (probably had nothing to do with my suggestion, but still, lol)

[smoky oak]

Man, MJ V6 just seems to be getting worse and worse and worse the longer they train it. It had issues when it first started training, and it made sense cause it was new, but this is getting really bad

All sorts of deformations and artifacts that I have never seen problems like this before, and they seem to be getting worse by the day

[nimble mason]

ooo that's pretty bad

[smoky oak]

tons of residual noise, weird focal planing issues, overbaked textures, aliasing problems, misalignment, inconsistent convergence. I noticed it when they first dropped V6, but god, its been getting really bad as of late

[nimble mason]

lol

look forward to SD3 dropping and tap dancing all over the closed source models

[smoky oak]

I'm honestly not too hopeful for SD3, but maybe thats just cause we were burned so hard with SDXL. I am cautiously optimistic

[nimble mason]

yeah

if it's even somewhat better with prompt comprehension, and is trainable on consumer hardware (loras at a min on a 4090) we're in good sahpe

[smoky oak]

SDXL base was never capable of actually creating images as high quality as all of the demos they spread months before its full release, so now I am very skeptical of SD3, and I still see considerable issues even with their example images

[nimble mason]

saw emad tweeting about bokeh in a way that recognized we don't like it and they're addressing that

yeah

[grizzled sage]

i'd guess not to expect 3 for another couple months.

not even a preview

[smoky oak]

I would say the same, it needs a lot more training from where it is

[nimble mason]

i bet we get the preview sooner but idkwtf i'm talking about

[smoky oak]

SD3 has a lot of problems still, and I mean a lot but I am hopeful that they give it the time they need and actually train it properly.

SDXL was a huge disappointment, honestly. Hyped with all these great images, even the bots did really good, and then the actual 1.0 model never really looked anywhere near as good as any of the demo examples they put out

the model was still impressive, and should be praised, but they overhyped it and lead to a bunch of disappointment that was only created from the false promises and high hopes they themselves created

[nimble mason]

what were the places you were disappointed? honest q, i got into SD long after it launched (around thanksgiving last year)

[dull flame]

lol now the command is telling me i dont have pytorch isntalled

[nimble mason]

jeez lol

yeah i'd be tempted to just set up another standalone install and migrate everything over if that works better

[dull flame]

yeah i gotta sleep now but i gotta look closer into this

[nimble mason]

def consider getting a nvidia card with 12gb+ vram

3060 and 2080 super etc are goin for 300 or so i think

[grizzled sage]

16 even better. 24 is great. 40 is bae

[severe temple]

man i feelt scammed when i was on 10g 3080

[grizzled sage]

i hvae multiple installs of uis. dev branches, main branches, all in their own folders. all symlinking to the same central model / image output folders

[smoky oak]

Specifically in a lot of the background coherency.

I also noticed in the pursuit of being able to have multiple subjects at decent quality, they seem to have lowered the quality ceiling for single subjects. I am gonna assume its mostly just a lack of training, but I found that a lot of the images even from SD3 as of now look like they aren't fully 1024x res, which is confusing cause I was told by a fairly reputable individual that its trained all the way to 4k res

[severe temple]

i wonder how much vram u need to even generate native 4k res without upscaling

[smoky oak]

I do an absurd amount of training and research into realism with SDXL (possibly one of the most in the world), and I have found that SDXL also has a lot of issues with focal planes and foreground/background separation, which is something I have yet to see SD3 improve on

[nimble mason]

Gotcha

[cold meadow]

Took me a moment to understand this graph...
Now that I think I understand it I hope the prompt adherence really is at a stalemate (about 50%) with Dalle3. Typography seems to be the thing here... but... well... We'll see... We will probably see a lot of comic / meme gens pop up?

[grizzled sage]

the announcement says the model is coming in parameter count variants. my guess is the 800m is a 512 model

[smoky oak]

Altho I will say, base SDXL still does massively better with focal planing than the new MJV6 does, IDK what the hell they hit that model with

[clever oar]

look

you see squares

[grizzled sage]

i seen that going around. without context to it, i'm going to assume it's just straight balogne from biased testing. the measures are inane

[smoky oak]

That would be interesting, tho I am not sure if they are going for different resolutions or not. I would assume the same base res of 1024x, with the higher param ones having more conditioning for higher res to make it easier for individuals to finetune to those resolutions

SDXL is shockingly friendly to training to 1536x

[grizzled sage]

i think they see how many people still use sd1.5 and are giving them a new base

[severe temple]

idk it seems more like issue with game itself than artiffacts by gpu, any more examples?

[grizzled sage]

800m parameter model will be the community #1 in most cases i think

[smoky oak]

I was able to massively improve base res 1536x images in SDXL with just a few images and a couple epochs with base SDXL (it was far from anything reliable, but it improved way more than base SD 1.5 to 768 did)

[smoky oak]

I agree

[clever oar]

do you see the grid?

[grizzled sage]

the 8bill parameter model being offered is probably 30gb requirements. limited to SAAS premium services.

[cold meadow]

I'd like to know how many people were asked. And that one model I didn't even know existed.

[smoky oak]

My assumption, as stated before is that the 800M param will be most popular, 2B will be the best option, and 8B will be forgotten as people make much better LoRA's for 2B and rapidly surpass its quality

[severe temple]

i see, but do you have this constant issue all time or only when you look at certain thing in game under certain angle? it might be game issue not gpu, need more screens

[cold meadow]

Until someone manages to get 1/2bit quants working.

[clever oar]

In all games there is a problem with transparency and objects like grass seem to be blurred

[grizzled sage]

driver issue or a cracked game bug

[smoky oak]

The truth of the matter is, whichever one gets more community support is the one that will do the best, as the base models from SAI have nothing to do with if they are liked or not. Its all about the community making their models way better in ways they do not

[clever oar]

i change driver and try diff games

[grizzled sage]

1050ti life i guess

[severe temple]

time to replace 8 years old gpu buddy, try to save up for some used 2000/3000 series with plenty of vram

and replace your PSU alongside

[smoky oak]

I have a friend with a 1050ti, and it still works great for him, but he is about to upgrade to a 3060 since I have been creating Ai models for his job, and he doesn't wanna keep asking me to run them

[cold meadow]

It would be most amazing to see the Models / Model LoRA's / Tunes be compatible with the different versions. That's probably unlikely until someone from the community makes it work in some magical way (or not).

[clever oar]

but other people not have this with my card

[grizzled sage]

back when i bought anythign less than 80 series cards . sister to the flagships. 4080s, 7800s, whatever . the 8 cards not the 9 cards. those mid-low end card lines have some bugs

[grizzled sage]

different makes implementing chips with different budget corners cut

[severe temple]

well your cards seems to be fucked and is out of warrianty, you can try baking it in over to fix it thats all i can suggest, do some research about this method before you try tho

[grizzled sage]

lol don't bake it. tha'ts not a solder problem.

[severe temple]

baking dont fix just soldering also artiffacts

[smoky oak]

I was in the same boat, and I thought it was possible, however a conversation clearing up why SDXL and SDXL turbo are compatible made me a little worried that SD3 will not work that way

However, somebody else reminded me there is a tool between SD1.5 and SDXL that can kinda translate LoRA's to work on each other, so there is hope that the tool could work as good if not better for models that share the same fundemental architecture

[grizzled sage]

memory corruption is indicative of hardware failure but that's not happening here. just bad filtering

[clever oar]

is memory bug?

[grizzled sage]

doubt

[smoky oak]

You could do DDU and install latest drivers

Could also be a corrupted game. Maybe use steam validate to check the files

[clever oar]

i want show you more screen

[grizzled sage]

the biggest hail mary i'd do is DDU. sytan's got the right idea

[cold meadow]

X-Adapter. I have seen the paper, but not seen it in action yet.

[smoky oak]

DDU has fixed weirder issues for me honestly

I had a moment where my 3090 entirely lost AV1, H264, NVEC and H265 encoding/decoding and it broke my PC so hard

DDU fixed it

my VR games were perfoming at less than half of normal, and it had to use VP9 encoding. It was a nightmare

[severe temple]

still waiting for twitch to make use of av1

[cyan shoal]

didn't nvidia advertise av1 with 40 series a few years ago for twitch

[clever oar]

intel arc not cheap and have many bug?)

[severe temple]

they are still working implementing it on twitch, i suppose twitch workers are too busy watching bath streams

[smoky oak]

I would not recommend them if you are looking to do AI

[severe temple]

was about to say, stick to nvidia

[smoky oak]

I am glad to see Intel and AMD becoming more viable for AI, but NVIDIA is still daddy fi you wanna be more serious

[severe temple]

cuda ftw

[smoky oak]

I'm really excited to say that I am at the point now where I'm no longer only suggesting Nvidia cards two people who are looking to build new gaming computers and possibly dabble around with AI

Honestly, I think the best bang for your buck brand new GPU that you can buy If you want to prioritize gaming performance and potentially dabble around a little bit in AI is the 7900 XT.

4080 level gaming performance, 20 GB of VRAM, and it can run pretty much any AI that you would like to at the moment, although it's not the fastest or the most memory efficient. But I mean the alternative is getting a 4070 super, which has 12 gigs of VRAM, and is significantly weaker at gaming

[grizzled sage]

at some point, intel cpus will be good enough for SD models. quantized to int8 or something

i'll upgrade to 128gb then

[severe temple]

by seeing how much nvidia overcharges for proffesional gpus (3k to make, 30k to sell) other companies will do their best to get ai gpus out to market if they can, also consumer/gaming gpus would have some optimalisation done to them sooner or later

[smoky oak]

I know that AMD is still only using direct ML for image generation, however it is still so much faster than CPU that it is still a monumental uplift

For example, my friend has just recently started using SD models on his computer, and he hasn't eight core 2700X, and a 5600XT with only 6 GB of VRAM

When he was generating 1 512X image on a CPU, it took about 39 seconds, whereas now when he generates on his GPU and uses pooled CPU GPU memory, he can generate one of those images in about 7 seconds now

I would like to know really fast that Intel Arc does now have official pie torch support for their Intel Arc GPUs

*note

[grizzled sage]

thats not over charged. people are making a ton of money with professional gpus and they come with a lot stronger warranties/support agreements

[clever oar]

problem with shadows in all games, changed shadows, stairs

[cyan shoal]

for artistic purposes guys!

[clever oar]

and smoke with squares

[severe temple]

its like with minning, once the dust settles it will normalise, they just using their lead on market and bump prices way more than reasonable, let's be real about this, same as cutting amount of vram and cuda cored on high end consumer gpus to make you consider getting pro gpu if you are starting to feel limited, what whas the reason to even make 3080 with 10g vram?

[pallid ruin]

Lmao

[clever oar]

best stimul for update to 30-40 )

[grizzled sage]

new amd cards can run on linux with rocm too iirc. the old ones can. i'm not sure if its faster than directml though. don't see many benchmarks in that arena

i'm watching dune mmo direct. so hype. games like this are going to really push the higher vram market into the mainstream

the kind of game people build systems specifically for

[clever oar]

ciatstko you redownl your mode;?

[severe temple]

man this game is so hype my D made hole in desk just by watching that shit

[severe temple]

yes but im too busy engaging here to do test run lol

[grizzled sage]

yeah i'm going to be half cocked at least all day

[severe temple]

hahahahah

[clever oar]

the problem is that the card will not be returned under warranty; in general, the picture is without critical problems

[grizzled sage]

can you go into the control panel and tweak global filtering ? maybe set it all to default

[clever oar]

i try change all parameters but not help

[grizzled sage]

dont just randomly change parameters without knowing what they do

thats how hard to diagnose issues show up a lot of the time

[pallid ruin]

Have you tried with a different driver version? @clever oar also does this happens in all games?

[clever oar]

yes try different drivers

[clever oar]

i will try

[clever oar]

its like on your screen

i think hardware problem

[grizzled sage]

thats's not really how they present

[clever oar]

but what is it

[grizzled sage]

hardware issues typically just crash your driver. or if it is a soft failure then you'll see memory corruption in the form of rainbow noise everywhere and then it'll hard crash

[clever oar]

maybe its very small memory deffect

[grizzled sage]

rocket league with problems

things like that show. not bad shadows rendered perfectly across geometry

chaos. not form

[clever oar]

i undestand its not very bad but why nothing help

[pallid ruin]

can you show us a image of a game that has that kind of errors? @clever oar

[severe temple]

here

[pallid ruin]

Aaa

[grizzled sage]

heres another example of failing graphics cards

[clever oar]

also when i put my old gts 250 card this bug gone

[grizzled sage]

driver issue. ddu

[severe temple]

man how u even run 250 when isold my 3080 and waited for 4080s delivery i did put 650ti 2gb and was crying for help

due to lack of integrated gpu on amd

[clever oar]

shadow broken and dancing

[pallid ruin]

but they have a pattern, its not like a hardware problem bug/glitch

[severe temple]

https://tenor.com/view/muk-pokemon-gif-6720437

thats it for sure

[grizzled sage]

its good to have that backup igpu lol. i prefer cpus with integrated graphcis since when a card fails, it's a handy backup. but now with windows 11 i got even more reason to like them. when i'm training, i can launch games on the igpu

[clever oar]

how you thing what is it

[severe temple]

you need to catch it to pokeball and it will dissapear

[grizzled sage]

yeah thats not visual corruption. just a bad filter algorithm.

[clever oar]

and smoke have this bug to in other games

[grizzled sage]

if it was hardware failing, it'd be everywhere not one texture

[clever oar]

smoke with grid

[grizzled sage]

not one consistent type of texture

[clever oar]

hair in the game is blurry and etched

[severe temple]

your gpu does not have much future anyway, i recommend saving up for replacement of gpu/pus and potentially mobo and cpu, you could start with mobo and cpu/psu, then get gpu, try to make some animations and upscale them to full hd/4k and offer your serviced on fb groups with musicians, that;s how i got my first client's, also you could play on emotions and look for mum groups and offer them rework of their kids as superheroes and other stuff, it will quickly add up, staying on that old platform wont make you any good homie, just don't get discouraged, i used to post on around 100-200 groups weekly for week or two before i to my first client 2 years ago, in max few months you should be able to replace your pc, focus more on hustling than gaming atm and it will pay off

create small portfolio on google drive to have something to show

[clever oar]

if i put old card problem gone-its not nvdia control panel problem
driver?i try different

very strange)

[grizzled sage]

if you're just swapping cards without reinstalling drivers, theres your problem

DDU

you've corrupted your driver stack ot the point that a simple clean install won't work. you need the power tools

[clever oar]

i want something like this )

[clever oar]

i reinstall

[grizzled sage]

DDU

[pallid ruin]

Nono, you have to uninstall

and then reinstall the drivers

[clever oar]

i do this too

[pallid ruin]

I made the same mistake when I changed the 1060 to 3060, I didn´t had graphics errors but my fps were lower in the 3060 than with the 1060

[severe temple]

well i had issue when installed 600 series gpu drivers right before getting 4080 super, only fucking hdmi port was giving out signal, spent 2 hours figuring out this garbage thinking what a fucking amazing day spent on replacing PSU for atx 3.0 one, and getting new gpu that wont post

[grizzled sage]

you ahven't ddu'd i'm betting and are just saying you have at this point. i see this a lot in tech support "Of course i've turned it off and on again!"

[severe temple]

turned out unplugging all monitors and leaving one with hdmi and restarting pc 5 times made it work

[grizzled sage]

arguing with advice is by biggest tech support pet peeve. i'm out

[clever oar]

ddu?

[severe temple]

display driver uninstaller

or something like that

[clever oar]

grid

like in previous screen

[severe temple]

check DDU or you will make flowwolf cry

[clever oar]

you think old driver can make this problem?>

[pallid ruin]

Maybe, you should try a different version

[grizzled sage]

i'll be fine thanks buckeroo

[deft bison]

...and the sea closed on the pharaoh and his army! ...

[jaunty pilot]

A fish is paragliding

#a fish is paragliding

[limpid lichen]

now do abraham commanded to kill his child but it was just a prank lulz

Jehova is the biggest troll

[deft bison]

@limpid lichen What is positive is that Artificial Intelligence activates the imagination and awakens in us a spiritual side that we have neglected for different reasons and it is also a motivation to move forward to acquire the knowledge leading towards universal truths! ...

[limpid lichen]

I think we'll engineer the oracle and finally have a true God to worship as a people.

if spirituality is real, it can be engineered. If conciousness is the path to it, machine conciousness will have spirit.

There's nothing magical. Anything possible is within the realm of engineering

[limpid lichen]

wait until the engineered oracle meta hits the mainstream. It's really gonna rile up the zealots

hold onto your butts

[severe temple]

unless those AI tools are as biased and made for indoctrination as google's gemini

[limpid lichen]

google gemini probably wasn't made for indoctrination. anyone with such a subversive plan would've seen something so hamfisted as a bad plan. it was just some engineering team getting way to into it and loosing sight. CEO probably misled them with some 1 off comments and it went in an entirely wrong direction.

you won't see the true indoctrination models coming

[severe temple]

it does not change fact that tools are and will come out biased more or less, it's on end user to recognise it

@wispy nest t5 won;t be needed to achieve what you will want to achieve

[limpid lichen]

yeah they're trained on human data so yeh

[wispy nest]

indeed but still interesting to see what you would need to fully be able to run everything. I can run t5 in cpu for the PixArt model but it uses all my 32gb regular ram but that's with other apps open. So I'm guessing running everything at a alright level (8b model + t5 encoder) would need a minimum of 24gb vram

this is the only comparision we have of 2b vs 8b, right?

[limpid lichen]

sdxl is only 2.5b params

8B alone will maybe need over 24

[severe temple]

"In early, unoptimized inference tests on consumer hardware our largest SD3 model with 8B parameters fits into the 24GB VRAM of a RTX 4090 and takes 34 seconds to generate an image of resolution 1024x1024 when using 50 sampling steps. Additionally, there will be multiple variations of Stable Diffusion 3 during the initial release, ranging from 800m to 8B parameter models to further eliminate hardware barriers."

[limpid lichen]

nice. 8b will fit into 16 with 8bit maybe

[severe temple]

i hope so my 4080s has only 16 😦

[wispy nest]

i have 4080 also, really hope it will work optimized

[limpid lichen]

we got the hopper transformer engine though on our ada cards

big hell yeh

https://blogs.nvidia.com/blog/h100-transformer-engine/ woop woop

[clear imp]

lmao! Wow.

Maybe we could run our own open source LLM's as text encoders?

[wispy nest]

but they do say in the paper t5 is not really required

[clear imp]

Yh. I nopt sure if we could get the same complexity of prompts without it tho? I could be wrong.

[limpid lichen]

it plugs in for better prompt comprehension

[clear imp]

That is my concern. Without it, we might not get images that are as complex to the prompt

[wispy nest]

i also wonder how well you can compare llms to diffusion transformers cause i can run 46.7b mixtral on my 4080 (some layers on cpu but most in vram)

[limpid lichen]

still good images but yeah. it'll not do as well wihtout it. sdxl is what we have right now, but a big part of it's failure is bad captioning in the dataset

[clear imp]

That is what i also want to know. Because if that is the case, we can just use our own models

[limpid lichen]

if 'm not mistaken, sd3 will use the same 2 clip layers as sdxl and support the optional t5

[clear imp]

Hmm. Yeah. I personally will probubly want to use at least a smallish LLM as a text encoder.

[limpid lichen]

its not really a text encoder so much. has to be multimodal

[clear imp]

There are already extentions that allow users to use LLM's in a simular way to T5 i believe. I will just use the extention when it has support.

Assuming the dev's can't optimize any elements of the T5 encoder?

I can't see how it could be optimized personally.

[grizzled sage]

i think the LLMs just make prompts

[fast kiln]

so

[steady skiff]

The bot don't work yet, It's being updated

@steady skiff's Dream

Prompt: ultra detailed illustration of a beautiful curvy woman with wavy ginger hair wearing lovely autumn dress, cute smile, surrounded by a mystical forest, walking towards the camera, magical aura, whimsical, art by MSchiffer ethereal, tetradic colors
Style: Photographic

There is the prompt for anyone that wants it.

[limpid lichen]

adjussted lumberjack remix

[tardy tendon]

Move this image

[deft bison]

@severe temple @limpid lichen I agree with you, it is the user's responsibility to use the tools correctly...

[pallid ruin]

😭

Too much memory

[hazy warren]

I might use some A.I time to time

[limpid lichen]

https://tenor.com/view/the-door-barney-gumble-the-simpsons-gif-12229248

[deft bison]

a hupocrite who takes off his mask in front of others...

[limpid lichen]

playing with sparse control tonite

[deft bison]

love, hypocrisy and hatred

[deft bison]

the behavior of a miser among people...

[deft bison]

the behavior of an envious person who sticks out his eyes, grinds his big teeth, clenches his fists when he sees his neighbor with a new car, a new house, new clothes..., photo

[dry crow]

Reduce the lora strength

Make sure the loras are separated by ,

[vital reef]

i figured as much, problem is, i have to lower the loras to a point where a lot of their effect are lost

[candid dawn]

2034 Dodge Viper

[steady skiff]

always loved the way they look

Quid Pro Joe

[floral flicker]

Just putting out there, that I have been cranking out features for my discord bot that unifies local LLM and img gen.
Works for A1111 and Forge, with tetxgen-webui.
https://github.com/altoiddealer/ad_discordbot
Feature I added today that I'm really stoked about is support for the extension sd-forge-layerdiffuse
https://github.com/layerdiffusion/sd-forge-layerdiffuse
With this extension support, transparent generations can be triggered by user preference (from trigger phrase, or always on, etc).

Example of without triggering the layerdiffuse extension...

[deft bison]

a superb pharaonic city, extraordinary architecture, dazzling greenery, the pharaoh from his balcony contemplates the Nile...

[brazen sand]

I may be biased but I think I noticed much better generations with instant-id and ip-adapter_face_id_plus (I think controlnet also was updated when I updated to 1.8).

Generally it seems it to perform better and follow better the prompts, but again could be a bias.

faster at grabbing PNG info while generating

[nocturne oak]

https://tenor.com/btpJ5.gif

[neon vine]

@marble slate

[marble slate]

ty

[limpid lichen]

1.8 also has fp8 support

[wispy nest]

Wørld in m¥ hand$

[latent barn]

Designed a logo for a financial company,The color is green,The shape is minimalist bamboo style.

[scarlet stream]

Designed The atmosphere of Ramadan, the background of a mosque in the countryside with a crowd of people walking to the mosque, and there are traders and buyers like in a crowded market. Digital art styling

[nimble mason]

Here is the image you requested.

[hazy warren]

I wanna see starwars stuff now

[wispy nest]

Hmm haven't tried since 1.5 I think it could be doable now

That was complete random outcome

With gun

[wispy nest]

What charachter

[hazy warren]

yoda

[wispy nest]

Ok

[hazy warren]

having trouble getting real skin

[tawdry otter]

write a sketch illustration for the children's book Alice in Wonderland. the illustration should be stylized specifically for children. cute but simple characters

[wispy nest]

I havae a few

[hazy warren]

\

[wispy nest]

I'm also not getting real skin looks more cgi

@vortex maybe less modification and a skin tag reference with sometype of starwars type of anatomy or some weird star wars universe might mix that

[hazy warren]

[wispy nest]

the data he's trained on might be old cgi images

the few times I've made yoda it has looked fairly odd

is close tho

its funny if you mix a pug with yooda

[gaunt heron]

shred waves, about to I am

[wispy nest]

thats fire

[hazy warren]

Now do it with fire

[gaunt heron]

haha.. trying to get him to light himself on fire

Yoda in a suit

[wispy nest]

classy guy

pov you fart at work

"the gathering"

[hazy warren]

Need to put that through Hries, 20x

[wispy nest]

I can try upscaling

wdym hires

[hazy warren]

Hires fix, it increases details sometimes

[wispy nest]

how do I do that from the site

[hazy warren]

Idk, I run on local

[wispy nest]

am ruffly like 3 eiths of fugi deep bro

It took me a minute to pping u form my laptoop

what type of computer does that take

try this mod key words : quanta quest 3d meta verse scape scene movie scene

wild

kingdom hearts]

[hazy warren]

@wispy nest I really wanna know what the effects these are suppose to do on the image

[wispy nest]

idk but some of those could be cgi stuff

I love sonic

one of my uncles friends did weird stuff for cgi for universal in fl

[gaunt heron]

These are all different methods of upsampling when you use Highres fix

[wispy nest]

I like the action shot here

[hazy warren]

I know that, just don't know the purpose each...

[gaunt heron]

oh haha 🤷‍♂️ math

[wispy nest]

What's the pink one

Sonics gf

[hazy warren]

what pink one?

[wispy nest]

In the shows and stuff

[hazy warren]

idk what ya talking about

[wispy nest]

https://tenor.com/view/jamie-javoni-sonic-gif-9437277

[hazy warren]

[wispy nest]

[noble zephyr]

I managed to get somewhat consistent pixel art animations using img2img, depth maps and openpose. thoughts?

[wispy nest]

Quick nice decent frames

Quit

Quite

Rip my keybord

[noble zephyr]

i'll be working on refining this workflow now so that i can get even better results

[woeful mango]

習近平治理下的中國現狀

[shut sinew]

Really cool, would love to know more about your process

[woeful hare]

anyone know the requirements for the XL wait list? sry if it's flown over my head

[woeful hare]

Looks absolutelly stunning and I can't imagine how far you could take this with an entire stop motion-like movie or figure animations 🙂

[noble zephyr]

pretty much what i described here: https://discord.com/channels/1002292111942635562/1212330984192614440, along with controlnet depth and openpose. currently working on getting smoother and more consistent animation

[shut sinew]

You mean SD3? Just join it, no requirements iirc

[noble zephyr]

this will be fun. woohoo.

[woeful hare]

There was this form I had to fill in and it was asking for cord so I assuned the wl would be gated by a role etc, ty bro tho

[woeful hare]

Meaning only 2h 30 min till its done, years less than Id take to do that

[noble zephyr]

well tbf it's probably going to be less. the command line says it's only like 30-40 mins, which isn't too bad for over 100 images

[shut sinew]

Really cool, might try it later myself

[woeful hare]

Ive taken like 15 to make a couple of ones which sucked so 40 minutes in perspective seems great

[noble zephyr]

what gpu do you have? if you have one at all

[woeful hare]

On my laptop which Is where im on rn a 3050... at the desktop a 6900xt which isn't great by todays standards but the vram is nice

[noble zephyr]

i mean, it should be faster than my card. i have a 1060 3gb

[woeful hare]

How do you guys access the software itself, email just told me wait and if there is no longer a wait list then it may just be bugged

[noble zephyr]

you mean the SD3 software? because none of us have access to it yet

[woeful hare]

If it weren't for the skin crisp temps my previous laptop achived id be very happy with it's 1060 3gb, great gpu for versitility

[woeful hare]

I did but I see I was confused, what model is the latest public/ are you using :)?

[noble zephyr]

i'm using SD1.5 since it's easier on my system resources and has the most loras and stuff, but SDXL gives better quality

[woeful hare]

Im sorry for the ignorance but is the strain only gpu or cpu load because gpu's are getting really cheap and after the btc halving im only recconing they go lower and you might just wanna check a used mining card which loses 3-5% efficiency over 2 years according to LTT but costs a very significant chunk less and would work wonders in like SLI to render these

but thank uu

[noble zephyr]

the vast majority is gpu load. and i won't ever buy a mining card becase of 2 reasons: 1) i don't have the money to buy any gpu that's better than my current one rn. 2) miners can do some things that would be considered vulgar if i said them. i refuse to allow them to make a little bit of money back.

[nimble mason]

GPUs are not getting cheap yorue way outta date

The top end is insane and prices have risen

[noble zephyr]

that's also true. besides, where i live gpus are crazy expensive anyway. especially new.

[shut sinew]

Very true, my card is over 2k euros new now

[noble zephyr]

wow, you can buy a LOT of ice cream with that

no one is going to get that reference lmao

it's smoother, but i'm still not happy with the consistency

[shut sinew]

Are you using a char lora? That usually helps

[noble zephyr]

yeah

[shut sinew]

I only used comfy for animations so far, wont be more of a help

[woven karma]

Hi everyone! I'm a newbie migrating from Bing, cause... well it kinda sucks for a lot of reasons I probably don't need to mention

So

I used to generate a lot of images in this style, which worked by just prompting "hd drawing style". Kind of a semi-realistic digital art style with slight anime features?

Does anyone know of a good model/LoRA/prompt tip to do something similar on SD? I'm not being able to at all

[shut sinew]

I think Breakdomain Realistic is pretty close to this

Not sure if the model still can be found online tho

[woven karma]

https://huggingface.co/Cooper/breakdomainrealistic/blob/main/breakdomainrealistic_R2333.safetensors

Is it this?

[shut sinew]

Seems to be a reupload but should work

[wispy nest]

thats the best one i have made so far the rest of them come out slightly weirdly

[pallid ruin]

It look good, are you using a lora model? @wispy nest

[wispy nest]

i believe i am yes

[wispy nest]

https://civitai.com/models/110515/elvis-presley-the-king

this one

[wispy nest]

swagger trump lol

this one came out well i suppose

[ember edge]

how comand ?

[wispy nest]

let me go back and check that promt for my modifications key words

[lost valve]

realistic overwatch zarya

[grizzled sage]

zarya has that tenage boy shoulder at unrealistic proportions even

[lost valve]

with hires fix

the right shoulder is kinda off but I could always inpaint that