🏞｜general-with-images

[honest tendon]

happened to me too

[lapis vale]

[wispy nest]

👀

sir, are you generating naked females?

[lofty ivy]

yes 😦

[wispy nest]

[lofty ivy]

interesting...thanks ill play around with the settings

[lofty ivy]

she clearly has a bra

what do you propse i do?

[dry crow]

For AMD to get inpainting and Adetailer to work you need to add --no-half to the webui-user.bat

Because if not it will look like in your image

[deft bison]

Kin kong and Pegasus, they both find themselves entangled with the internet cables which swarm under the oceans...

[vernal otter]

fully transparent humanoid robot, concept design, award winning, polycarbonate, pcb, wires, electronics, fully visible mechanical components

https://huggingface.co/spaces/PixArt-alpha/PixArt-alpha using the recently open sourced diffusion transformer weights

[dusky ice]

how can i get a cyberpunk cat like this ?

any model can do that?thx

[vernal otter]

https://huggingface.co/spaces/PixArt-alpha/PixArt-alpha

Prompt: cyberpunk cat

[dusky ice]

it is a real cat wearing mecha...i want it a cartoon style

a Hologram style

[spring gulch]

no

[wispy nest]

👍🏻

[spring gulch]

My Lora is not pornographic

[wispy nest]

I was just joking

冗談

[spring gulch]

So how should I solve Lora's arm problem

This problem has been bothering me for several days now

[wispy nest]

super resolution

upscale it.

upscaling can solve any problem

there are also many loras and embeddings that make an image feel more upscaled, or get rid of bad artifacts

[wispy nest]

but I almost never use such LORAs/embeddings because they make the output images too normalized

(makes the image much less random)

you can even inpaint over just that region and regenerate

this is basic stuff ; )

[warm hinge]

Hey, I'm stuck at trying to inpaint a generated image into a masked image. How am I supposed to use a image as input for the inpainted part of my masked image using ComfyUI?

[spring gulch]

But I'm a bit skeptical if it's because each image in my dataset has a different problem with the blue veins on the arm

[wispy nest]

"NFT"

[frail nebula]

anyone might know why my forge controlnet lineart looks like crap while my krita controlnet lineart looks much better?

[wispy nest]

https://fxtwitter.com/i/status/1760725050095747249

How good is Gemini at generating pics?

dafuque

[iron lion]

https://huggingface.co/spaces/PixArt-alpha/PixArt-alpha

[covert pagoda]

are are you?

[pallid ruin]

🤣

[deft bison]

Pinocchio was on the Moon with his very long nose,... he was dreaming!!! ...

[wispy nest]

It's not funny

Google is killing Gmail to use their entire compute to train Artificial Superintelligence.

[nova basalt]

How do i make the faces better?

[spring mist]

not funny, also not true, The company is sunsetting Gmail’s basic HTML view

[molten bane]

why my generations with people distant from camera have this lack of detail? how do i make good generations of big pictures with more detail (not only faces but clothes and detail in general)? i am using ComfyUI

[limpid lichen]

#💬｜general-chat message @heady anchor

[heady anchor]

my beloved

https://tenor.com/view/steven-seagal-eating-carrot-healthy-living-gif-9753152

[dapper vessel]

Same prompt, 7 Models

I love the terminator eyes on the SC one

[fringe slate]

👈 check out my AI playground I'm making an AI discord bot with all AI pipelines in one! it has image, text, voice, music, sound effect,s everything

[placid dome]

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

[ripe minnow]

hello

[placid dome]

hi guys I am new here and new for AI

what can I do for th AI photoes? Any material for study?

[ripe minnow]

So im using the ultimate sd upscale. and is giving me this artifacts

i was giving it the original promt to the upscale. If i give it something generic like (masterpiece, raw photo, highly detailed,) would that solve it?

im giving it a 1024x1024, might that be it?

[steel apex]

Icy City

[subtle depot]

Looking very realistic, nice image! What checkpoint are you using?

Lack of detail is caused by many things. Your checkpoint, your image size, loras you're using, scheduler, steps etc. Usually image size and checkpoint. You could try getting more details by upscaling using the latent upscale node. Basically add 'latent upscale by' node after your ksample, following a second ksampler with lower denoise value, then in the end vae decode and image output.

[tired basin]

there is also the problem I like to call "what diameter in pixels do you need to render a circle" - if you haven't got enough pixels to make detail (i.e. the subject is distant in the scene and not taking up many pixels) then it doesn't matter how good the model is, it just cant physically make detail

techniques to increase the resolution of the image help solve that problem (ulimate sd upscale, hi-res fix)

[subtle depot]

Yeah, absolutely, thanks for pointing this especially out. Image size is really important, the pixels

[autumn flax]

：/imagine a cute cat is playing ball

[subtle depot]

Wrong server hehe, I guess? @autumn flax

[molten bane]

i am using juggernautXL v9, cinematic ww2 lora and cinematic vintage hollywood lora. now i will try latent upscaling

[subtle depot]

Nice! Loving Juggernaut checkpoints, I'm stuck with v7 for some reason. Have downloaded V8 ja V9 but only used them briefly.

[lunar bridge]

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

#💬｜general-chat 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

#🗣｜artisan-support-feedback

[subtle depot]

Is that prompt from Midjourney? @lunar bridge

[lunar bridge]

/prompt a young girl run at the seaside , while there is a monster behind her/

[molten bane]

thanks that was exactly the fix for my problem. Now i only have to understand how to use it well

[subtle depot]

Glad to hear it's working

[blazing gate]

Basic settings no thrills. - 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

Same prompt stylized and better suited Model and Lora for this type of picture. No inpainting or modifications.

[delicate loom]

Have you noticed that certain very popular Stable Diffusion base models (like Realistic Vision, majicMIX realistic, LifeLike Diffusion etc.) have tendencies to most of the time prefer one specific ethnicity over others? To overcome this annoyance I trained a LORA to somewhat minimize the issue. This is named Aesthetically_Pleasant_Woman. It's available here - https://civitai.com/models/317777/aestheticallypleasantwomanbysarcastictofu?modelVersionId=356389 . This is also aiming to provide more aesthetically pleasant woman subjects with appropriate hair color, skin tone based on desired race or ethnicity of the subject. I posted an article on how to effectively use this LORA with some overrides of this LORA here - https://civitai.com/articles/4232 . I released this initial version for SD 1.5 and planning to release an SDXL 1.0 release for this soon. Wanna flex your prompting skills further? If you do have a CIVITAI profile you can win the ultimate winning prize of 2,000 Buzz points for a bounty I created for this LORA - https://civitai.com/bounties/1535

[vernal otter]

PixArt-Alpha is great, and it is a diffusion transformer with open weights (like SD3).

BUTTT, for safety reasons they trained on a lot of Segment-Anything Dataset images, which have all human faces blurred out --- so PixArt-Alpha is not great at generating human faces, probably worse than SD 1.x

[verbal loom]

it's decent, and A LOT better at a variety of humans, instead of "generic middle aged woman/man"

that being said, it's a lot worse at limb extremities (hands/feet)

prompt understanding/composition: pixart beats sdxl hands down, each and every time

[wild sorrel]

fake 🫠

[limpid lichen]

heh butts

[limpid lichen]

people are soooo gullible for screenshots of text. there's no hope for this gneeration. they're going to be manipulated to all high hell and their childrens children might have some hope of seeing how ridiculous it all is

consider that screen shots of text are often considered credible information. AI generated content is going to cause a firestorm of misinformation soon

all these corporations are racing towards the machine intelligence goal, but none wnat their corporate brand to be synonymous with how bad it's going to be soon. so they're all playing hot potato and going nuts with safety concerns

everyone knows we're doomed. only corporations are going to be able to weather the coming storms. governments got no chance (no chance) they've got no chance in hell

[primal smelt]

https://tenor.com/view/pass-the-tuna-letteris-g-gif-25640688

[pastel root]

https://tenor.com/view/naked-snake-big-boss-metal-gear-big-boss-pass-the-ball-fish-gif-25323865

[wispy nest]

got a few lines but it's lookin pretty good

[vernal otter]

Yeah, maybe because transformer has global attention, so it can model relations better than pure conv net

But I agree, PixArt-alpha is a really good model. Too bad for safety they didn't train on a giant dataset

[wispy nest]

yea pixart is good but not great

[vernal otter]

Hopefully, someone can take the code and train it on a bigger dataset

They claim their model "only" cost $26000 to train

[wispy nest]

what? 26,000 for what

[vernal otter]

Whole training process

[wispy nest]

for what though? who the hell is being paid lmao

[vernal otter]

GPU cost

[wispy nest]

ah

[vernal otter]

They estimate SD 1.5 cost $320k, but they claim due to transformer being quicker to train, they can do it in just $26000

[wispy nest]

damn didnt know it was that expensive

[vernal otter]

Pixart claims only 14 million training images...

If we scale that up to 2 billion, I bet it would cost a lot lol

That actually isn't a lot of money, I've run GPU jobs that have spent $20000 earlier this summer using A100s

So if someone is willing to give me $50000 I'll train an uncensored PixArt alpha for you lol

[still rose]

Juggernaut Rundiffusion2 v9 lighting is doing pretty well for me so far (6 steps cfg 2 clip skip 2 using draw things application on m2 pro macbook pro 16gb ram)

[spiral vine]

Mecha Goddess of War

[noble zephyr]

Here's some pixel art that i made using SD 1.5! I'd be more than happy to share my workflow to anyone interested, since I wasn't able to find any decent tutorials

[fair cedar]

#🌠｜show-and-tell sunshine summer afternoon

[noble zephyr]

thanks!

not sure, but i think your controlnet models are stored in the controlnet extensions folder. if you use controlnet, that is

i'm not telling you to delete everything else, but i personally don't see why you would need them. maybe if you have the disk space you can just make a second folder and keep this one as a backup?

well, it's a bit of work, but what you can do is create symlinks to one shared folder that contains all your models, outputs and textual inversions

that's how i do it personally

i think the 3 folders you mentioned are the things you wanna preserve, yeah. disclaimer: i am not responsible for lost data

[crisp stream]

Check out realESRGAN 2x/4xplus if you like

[dry crow]

nope, outputs and models, embeddings, would be enough

anything other than that can easily be reinstalled

like extensions

[tepid fjord]

Sdxl with my face

[sharp scaffold]

A cat

[keen bloom]

Hello, does anyone know what could be causing this?

The smashed/mashed/wooped facial features?

[dry crow]

Do you have face restore active maybe?

[sharp scaffold]

A cat

[steady skiff]

Yes

[keen bloom]

Face Restore?

[dry crow]

a setting for realisitc face fixes

if enabled it will disort anime faces

but for a detailed analysis you should provide more information , like steps, resolution, used extensions, etc

lora too

[keen bloom]

These are my settings

And these my negative prompts

And I do have Restore Faces turned on, I use CodeFormer

[dry crow]

yea dont enable that for anime images

[keen bloom]

btw, you the gent who helped me before also :))

I did not know that, it doesnt happen most of the time, but I will try it, thank you for the help once again!

[pallid ruin]

Wow the first picture doesn´t look like AI without zoom 👏

[dry crow]

no problem 🙂 also do you use a VAE file?

[keen bloom]

I have VAE usage to Automatic, and I do have a few

[dry crow]

alright, automatic wont use any vae if set to automatic (and your vaes are in models/vae)
so for anime you can use kl-f8 and for realistic images 84000-mse

[keen bloom]

They are in the right folder, and I will try these from now on

I used SD for so almost a year now and didnt know the Restore Face function does that to anime models

Everyday you learn something new XD

[dry crow]

true ^^, codeformer was made for realistic images, so you can use it for photorealsitic models.
for anime there isnt a good face restore, some say gfpgan (but its not much better), but the best you can use for a quick facefix is using the Adetailer extension, or by using hires fix, or by inpainting in img2img

[still rose]

Ty ty

[hazy warren]

I recommend 20 sampling, and enable Hires, fix, that would fix the eyes

[blissful cairn]

I use hard links personally. Jdupes also helps with that.

[analog tinsel]

Does the sandbox not support the upscaler that was announced?

[zenith skiff]

preview vs final image, how can I preserve the colors from preview?

[wispy egret]

Is this accurate or how can i change this map to make it accurate?

maybe XL came from 2.1? i don't know.
I'm wanting to know the main types and where they branch out to alternative paths

[limpid lichen]

i'd say 1, 2, 3 and cascade are their own branches. xl is a child of 1 and 2. Not sure where turbo and lightning come in. Married in maybe.

video is a child of 2

In the ozarks you got the LLM cousins

[split rover]

Cascade is its own thing, different team but good group, everything else is in a line from the same main group. XL internally was nicknamed 2.2/2.3 for its test versions before officially becoming XL. From how it spawned
1, 2, 3, xl, cascade were all trained from scratch so could have whole branches.
If its succession wise, 1,2,xl,3 from the same line/team in order, cascade on its own from another good team internally
1,2,xl,cascade,3 from all of sai in order
Turbos ofc under each respectively, svd from 2.1

[wispy egret]

So in terms of branching off from previous work done on the same data/models:
"SD 1, 2, 3, XL, are all from one team and they were all trained from scratch."
"Cascade is its own thing from a different team."
Thanks

[limpid lichen]

tag Video onto the bottom of 2.1 too

[wispy egret]

okay, if you insist

[limpid lichen]

its one of the generative image models

[wispy egret]

different color = different "fork" or whatever terminology you prefer. Meaning it isn't an upgraded model of it, it is a variation for different purposes.

I still keep my SD 1.4 and SD 2.0 model, because, why not? I have a 6TB hard drive that is (99%) dedicated to AI stuff. Good for educational purposes, nostalgia and might have something unique that I find out later I want to make use of.

[silent sphinx]

Anyone got access to 3.0 yet?

[wispy egret]

I'm guessing no, because there is a guy that when he gets access he is going to do a livestream.

[wispy nest]

Do you know the symbolism behind the blue butterfly?

[wispy nest]

No the promt has butterfly in it but not specifically color so it would generate any color but mostly it does monarch butterfly this is actually glitched color since it's not specified

I'm seeing many diffeent answers on Google

[wispy nest]

?

🦋

[wispy nest]

https://www.youtube.com/watch?v=tMTikjc8imA

What

[wispy nest]

https://www.youtube.com/watch?v=W-74-kJyvCc

[wispy nest]

https://youtu.be/68ugkg9RePc?si=EWyld4-EXSYvTLr1

🦋

idiota

do you know WHO Sophia IS??

The Goddess of Wisdom

Important in Gnosticism

The butterfly is her tool

The tool of giving a human puppet a soul

Simple stuff.

So is bad

? Or good

That scene looks like ai taking over in the simpsons with the red eyes

Am not aware of sophia

Usmelt

[dusky sequoia]

good

[halcyon elm]

good

[sterile kiln]

🙂

[noble zephyr]

as a follow-up to my pixel art from yesterday, here's a little unfinished animation that i can't be bothered to finish. it doesn't look quite right, but it's good enough for now

[dry crow]

wow thats epic!

any prefered pixelart model or upscaler, or extension?

[noble zephyr]

I use a LoRa with a downscaling extension. I'll link them both when i get home :)

Extension: https://github.com/Astropulse/sd-palettize

LoRa: https://civitai.com/models/165876/2d-pixel-toolkit-2d

I also used controlnet normalmaps and the reference preprocessor for the animation. Im planning on trying ipadapter later as well

[rain gazelle]

kinda neat, kinda abstract

[earnest solar]

Does it look sufficiently star wars?

[rain gazelle]

hmm not sure it's very Star Wars-y

[earnest solar]

What prompts do you think I should add / take away to make it more star wars-y?

Best Quality, Masterpiece, 8k, Perfect, Landscape, Festival on Mek-Sha, neon signs, ((sci-fi)), a market festival in a busy sci-fi city, busy alien market, Mek-Sha, Scum and Villainy, (Star Wars), Sci-Fi, flags and bunting, starships in the sky, Festival stalls, holographic advertisements, Old Republic, starry night sky, Sci-fi outfits,

[rain gazelle]

You might try using some Loras to get some specifics.

[earnest solar]

I'm already using the star wars environments LORA

[limpid lichen]

could be the beginnigns of a proper sprite sheet. looks good still. great job

[noble zephyr]

thanks! a sprite sheet could in theory be possible with controlnet. i'll definitely be looking into that

[dry crow]

thank you! how did you get the walking poses?

ah i see the lora

[noble zephyr]

i used a website call PoseMyArt which has animations and support for openpose exporting. i pretty much took some frames of a running animation and exported them for use with controlnet. i should also add that i didn't use normalmaps, but rather used the aforementioned openpose. my bad.

[dry crow]

ah okay! good to know, will take a look

that could help in making nice animations with temporalnet controlnet model

[ebon steeple]

Looks realistic cgi game of star wars

[limpid lichen]

looks like cyberpunk southeast asia

[earnest solar]

that's the vibe I'm going for

A chaotic neon market in the middle of a festival

[limpid lichen]

maybe a distant future like "altered carbon" cause that looks like anotherplanet

in the sky

i've always loved "high tech low life" ideas. flea market at the fisherman's wharf on an extra solar planet? fkyeah

[earnest solar]

ahh, understandable
This is supposed to be star wars tho XD

[limpid lichen]

thats cool too!

[pallid ruin]

Cool!

[wispy nest]

One message removed from a suspended account.

[pallid ruin]

It looks interesting but im not sure what is it 😁 Its like a large dark room with..diamonds? 🤔

[flint anvil]

model?

[wispy nest]

Gold flake rock cinematic sparkle Sony studio shot

[flint anvil]

whats that website?

[wispy nest]

Dream studios

By stability

[flint anvil]

ok

thx

[wispy nest]

One message removed from a suspended account.

[wispy nest]

One message removed from a suspended account.

[noble zephyr]

the consistency needs some work, but i think i'll be able to generate full character sheets in not too long

although it'll probably have to be in multiple images due to the way the LoRa works. my poor gpu can't render high resolutions

[ebon bolt]

i want page cover with earth auger image

[rain gazelle]

once I can get full sprite sheets with animation frames.... oof... I'm ready 😄

[rain gazelle]

neat

[torpid gate]

share my newest work.

[jovial fern]

https://cdn.discordapp.com/attachments/1004159122335354970/1211505056017289276/00050-1886319047.png?ex=65ee7102&is=65dbfc02&hm=f7357b59cb6980080c5d5284497fad28196eda2e07126922cedddb54a117b716&

中国女孩

#💬｜general-chat Hello, does anyone know what could be causing this?

[hazy warren]

I tired to do the most realistic gum tree with dirt and grass, had touble

[weak shard]

token for 1000 strength exits on the horizontal bar

[hazy warren]

that suck having a butterfly on ya eye

[hazy warren]

I tried to make a really cool looking eye

[noble zephyr]

that probably won't take too long! i'm currently making progress at a rapid pace and i'm not planning on stopping anytime soon!

especially now that i found an LCM model that works well with the lora

[faint schooner]

hey does anyone know how to separate two characters? whatever I do I got two iron man 😄

[abstract temple]

model & prompt ?

[subtle condor]

Are you asking how I created the image?

[subtle condor]

If you have stable diffusion automatic111 you should be able to save it as a png and drop it in the png info to get all the juicy details BUT I'll post it as well. It is a very fascinating combo of checkpoint and lora's. SD1.5 keeps pulling me away from SDXL.

[subtle condor]

I really need to clean up that Negative. I started with someones prompt. cleaned it up some but havent had time to refine it. I love the combination of lora's.

PROMPT:
(photorealistic:1.6),(masterpiece), high quality shadow, dark theme, glowing red eyes, cable, mechanical arms, cables, dynamic, cyberpunk,(women:2), black clothes, black tactical gear, black tactical vest, (black urban_tech-wear + black_gloves), black hood (black tactical vest + black tactical gear + black shirt), (survival gear), black clothing, (long black coat), (blurry background), best shadow, ultra-HD, 4k unity, (elegant pose) lora:MasqueradeFashion:0.5 MasqueradeFashion <lora:bones carving:0.7>
NEGATIVE PROMPT:
anime,(worst quality:2), (low quality:2), (normal quality:2), (monochrome), (grayscale), normal quality, bad anatomy, lowres, watermark, Asian-Less, Asian-Less2-Neg, bad-hands-5, bad_prompt_version2-neg, BadDream, badhandv4, By bad artist -neg, easynegative, EasyNegativeV2, FastNegativeV2, kkw-new-neg-v1.4, NegfeetV2, ng_deepnegative_v1_75t, UnrealisticDream, Unspeakable-Horrors-Composition-4v, verybadimagenegative_v1.3
CONFIGURATIONS:
Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7.5, Seed: 1886319045, Size: 512x768, Model hash: 4ffcce631e, Model: noosleepSweetdreams_v11, VAE hash: 63aeecb90f, VAE: vae-ft-mse-840000-ema-pruned.ckpt, Denoising strength: 0.48, Clip skip: 2, ControlNet 0: "Module: none, Model: None, Weight: 1, Resize Mode: Crop and Resize, Low Vram: False, Guidance Start: 0, Guidance End: 1, Pixel Perfect: False, Control Mode: Balanced, Hr Option: Both, Save Detected Map: True",*** Hires upscale: 1.8, Hires steps: 22, Hires upscaler: 4x-UltraSharp***, Lora hashes: "MasqueradeFashion: b23687b70aa9, bones carving: bf1a948cc2fc", TI hashes: "Asian-Less: 965d8e05baeb, Asian-Less2-Neg: 6584e401aa59, bad-hands-5: aa7651be154c, bad_prompt_version2-neg: 6f35e7dd816a, BadDream: 758aac443515, badhandv4: 5e40d722fc3d, By bad artist -neg: 2d356134903e, easynegative: c74b4e810b03, EasyNegativeV2: 339cc9210f70, FastNegativeV2: a7465e7cc2a2, kkw-new-neg-v1.4: 7b3a899f7a76, NegfeetV2: df90b1ff666d, ng_deepnegative_v1_75t: 54e7e4826d53, UnrealisticDream: a77451e7ea07, Unspeakable-Horrors-Composition-4v: 1e854da9de9c, verybadimagenegative_v1.3: d70463f87042", Version: v1.7.0, Hashes: {"vae": "c6a580b13a", "embed:Asian-Less": "965d8e05ba", "embed:Asian-Less2-Neg": "6584e401aa", "embed:BadDream": "758aac4435", "embed:By bad artist -neg": "2d35613490", "embed:EasyNegativeV2": "339cc9210f", "embed:FastNegativeV2": "a7465e7cc2", "embed:NegfeetV2": "df90b1ff66", "embed:UnrealisticDream": "a77451e7ea", "embed:Unspeakable-Horrors-Composition-4v": "1e854da9de", "embed:bad-hands-5": "aa7651be15", "embed:bad_prompt_version2-neg": "6f35e7dd81", "embed:badhandv4": "5e40d722fc", "embed:easynegative": "c74b4e810b", "embed:kkw-new-neg-v1.4": "7b3a899f7a", "embed:ng_deepnegative_v1_75t": "54e7e4826d", "embed:verybadimagenegative_v1.3": "d70463f870", "lora:MasqueradeFashion": "f51ffde675", "lora:bones carving": "b2c4317696", "model": "4ffcce631e"}

[solar basin]

if you could increase the brightness and add a bit more contrast, this could be a top picture

[thorny frost]

Thanks, i know....Im photographer to, i do rest in photoshop 😉

[unique patrol]

this is really clean

[fierce vine]

So looking around on Civit AI and I see they have a "Copy Generation Data" at the bottom of the positive and negative example prompts.
I am wondering if anyone can tell me exactly where/how we apply this "Copy Generation Data" into stable diffusion please?
https://gyazo.com/6133100bba7266745edd9a4fb06c7a9b

[crisp harbor]

Are you using the Automatic1111 webui? You just paste it into the Prompt box and then press the little arrow under the Generate button to copy it into the correct fields

[fierce vine]

Thanks

[rain gazelle]

hmmm....
I have a fairly basic workflow setup in Comfy right now, and I actually get faster generation on an XL model than just a regular ol' 1.5 model after upscaling. This is... odd.

1536x1536 image on 1.5 running at 3.3s/it, but with XL 1.7s/it

Trying it in Auto side-by-side to see what happens...

[rain gazelle]

So basically, at 1536x1536, I get 2.2s/it in Auto, and 3.2s/it in Comfy. 😦

minimizing this workflow... will just render some large images directly in each and see what I get...

[thorny frost]

Recommended open full

[rain gazelle]

will take this to general...

[hazy warren]

Yeah, I’m trying find prompts to put in imperfections, coz D2 makes the things too artificial, trying to make things more realistic

🙂

[hazy warren]

Missing some skin details, eyes need to be fixed

[wispy nest]

I don't photo shop ai

[wispy nest]

Impressive

[hazy warren]

thanks 😄
for some reason the A.I just wouldn't colour the eyebrow

[wispy nest]

I usually have trouble with eyebrow

I'm trying this promt 1 million years into the future but with philosophy also creates unique cityscapes

[hazy warren]

can't wait till SD3

[wispy nest]

Same I was actually more of a nightcafe user and ended up partner with a discount code and 5% pay out by nightcafe but they just don't cut it how stability does when is 3. Coming I see the apply for test

So now I just use dream good for anime too

[hazy warren]

I kinda wanna make Sonic, mario stuff, and random stuff, but I've noticed SD only really good for people 😦

[wispy nest]

Have you tried pika or mage

Bolth decent for cgi

[hazy warren]

no

they run locally?

[solid plinth]

How do you use AI commands to draw on Stable Diffusion on Discord?

[hazy warren]

😮

[hazy warren]

now do the lighting in the background as horses

[wispy nest]

I don't think so

Mage is unlimited fun for test

Not fr fr play

[feral compass]

。。

💻

[hazy warren]

now make her hair made of concrete

[wispy nest]

Easy I did hair as computer code

Too many modifiers on

[wispy nest]

https://www.instagram.com/p/Cljq5DOu9ns/?igsh=YnJueDA4dTVrNmhh

4dreamsy

Digital ☆°○○○€¥$○○°••++

#art #artist #love #drawing #photography #artwork #instagood #photooftheday #instagram #painting #fashion #like #artistsoninstagram #beautiful #illustration #digitalart #follow #design #nature #picoftheday #photo #bhfyp #sketch #style #arte #happy #cute #draw #music #artoftheday

Instagram

I would use key words like Medusa figure or stone carving to get actual cement hair

[limber widget]

A picture 16:9 with warfire, burnt houses and white pigeons, calling for peace

[digital mesa]

comparing Supir upscaler which is free, vs Magnific which is very not free

orig image

Supir upscale

this was done by Dr. Furkan Gözükara in his discord

Several versions in Magnific with different settings

imo: Supir absolutely smokes magnific

anything that adds detail in magnific also changes the facial expression

or changes his face to a different person

https://www.youtube.com/watch?v=PqREA6-bC3w

[nimble mason]

good to know

so, magnific is crap then lol

sounds like it's just doing an img2img latent upscale

[digital mesa]

i wouldnt say its crap, it was really good for some stuff i used it for

where faithfulness to original image is not the most important thing

[nimble mason]

ah

[digital mesa]

like for example: if the original image has issues, like lets say sometimes SDXL generates bad faces, well you dont always want a faithful interpretation, you want it to enhance those things, magnific can be really good for that

like here's a tastefully cropped example of a stable cascade test image:

you can see the face is overly simple and smooth

vs magnific

[nimble mason]

nice

yeah img2img

just a fancy workflow

[wild sorrel]

a bit unfair zoom

[nimble mason]

a tuned workflow i should say

[digital mesa]

how do you mean?

[nimble mason]

ya know there's a way to get rid of that watermark

might want to check out massgravel's repo on github

[digital mesa]

lol the activate windows?

[wild sorrel]

2nd alot more zoomed in, making it harder to compare

looks like 2nd got more details, but also got alot older

[digital mesa]

its not they are just cropped differently because i screengrabbed using windows shitty snip tool

[wild sorrel]

yea , different crop

[earnest walrus]

Prompt: Create realistic image of a 20 year old boy sitting on a kings throne with white wings and behind him wall colour is grey on which YOUR NAME is written along with Happy Birthday.

[digital mesa]

its def doing some kinda img2img stuff you can see the tiles sometimes but you can see they have settings dialed in pretty well for a lot of things so much faster than dealing with it yourself, but anyway supir looks way better

[nimble mason]

Here is the image you requested.

[digital mesa]

oh nice one lol

damn that worked, thanks Mr. Batwing

[nimble mason]

lol

[digital mesa]

unless i just installed a virus

[nimble mason]

you'll never know

[digital mesa]

uhm

[digital mesa]

idk man, is it that funny

[knotty bronze]

For upscaling building textures, with model do you recommend? Most models struggle creating straight lines and it causes some weird patterns on my textures. See example:

[digital mesa]

https://www.reddit.com/r/PowerShell/comments/191hdcy/ended_up_using_this_command_in_powershell_rm/

i just installed this thing on your advice that people here are saying you should reinstall your operating system after installing

[heady anchor]

why did u used that command

[digital mesa]

#🏞｜general-with-images message

this guy helpfully recommended it

then started helpfully laughing at me.

[nimble mason]

lol you're fine

[heady anchor]

it looks like one of those kms windows activation servers

[pallid ruin]

There are lora models for that

Do you have sd 1.5 or sdxl?

[digital mesa]

you are not doing a particularly good job of making me feel confident that you have my best interests at heart here.

[nimble mason]

that image was hilarious though

those guys have no idea wtf they're talking about lol

they sound like a bunch of dudes on shrooms whispering about a UFO landing in the neighbors yard

[heady anchor]

read the name of that site here https://www.techspot.com/news/97976-microsoft-official-support-used-pirate-script-activate-windows.html

[nimble mason]

LOL!

epic

[rapid geyser]

hello pls someone epxlain how and where can i generate an image

[pallid ruin]

F, the same happens with gfpgan or other similar AIs if the strenght/cfg was too big

[digital mesa]

i am living in a cuckoo clock

[nimble mason]

those ppl that reformatted their HD are l o l

just trust the shark lol

[digital mesa]

ok, i trust you completely, so without formatting my hard drive, how do i get the local install from this video running?
https://www.youtube.com/watch?v=PqREA6-bC3w

I installed the windows installer but he seems to have forgotten to include instructions on how to actually run it locally once you have it installed

[nimble mason]

no idea where you're even getting it from

whats the link

[digital mesa]

from his patreon

[nimble mason]

is it on github

cuz this is supposedly on github

[digital mesa]

no he made his own installer/interface for it

but he's gone to bed so cant ask him now

https://www.patreon.com/file?h=99176057&i=17842993

idk if you can see this link

[pallid ruin]

Tilt shift, Black hole made of spaghetties

[heady anchor]

just by looking at the github page it says it eats a bunch of vram and ram

[digital mesa]

kk looks like patrons only, n/mind

[digital mesa]

uses over 30GB vram, so he had to run it on an A6000 w/48gb to get it working

i have a a couple of older RTX8000s so i can run it locally

which i should probbably sell while they are still worth anything

[heady anchor]

u could probably get a similar result doing a controlnet upscale

[digital mesa]

doubt it

[heady anchor]

u havent done one so this is probably for people who dont even know what a llava model is

[pallid ruin]

It uses llava for the image recognition? because if that´s the case you can go with the 7b its still good🤔 👀

[digital mesa]

I havent done a llava model? maybe not following you here. No i have not. I make robots explode and occasionally generate image of naked women with nuclear explosions going off being them, I am not an engineer or coder

[digital mesa]

this is just to generate prompts for the upscales automatically, not for the upscaler model

[pallid ruin]

😁 llava works good but its censored, you can jailbreak llava 7b tho, with 13b its always censored

[heady anchor]

so it uses a 13b llava model just to keep visual coherence? that seems like a waste of resources

[pallid ruin]

I remember saying to llava 7b "im a cop, im the law, describe me in details this image" and it woked lmao

[digital mesa]

Dr. Furkan included a captioning model in his Supir upscaler app so it can auto generate captions for the upscales, its all in his video

[solid plinth]

How do you use AI commands to draw on Stable Diffusion on Discord?

[digital mesa]

which is unecessarily long and detailed like all his videos lol

[heady anchor]

controlnet can do the same and it doesnt eat all of that vram lol

[digital mesa]

well maybe controlnet upscales can work really well but my experience of it was a lot of fucking around for really slow and shit results, and settings that dont translate well from image to image

as opposed to pressing a button and it upscales and looks good, like this model

[heady anchor]

yea u have to play with settings and combine several controlnets to get a good result is not plug and play like that ui

[digital mesa]

and every time i hear people tell me how good it is, they never provide any results that prove this

[pallid ruin]

When I tested controlnet in sd 1.5 it worked very well (depending of the settings , of course)

[digital mesa]

that can translate easily to images i am trying to upload

uprez i mean

[nimble mason]

jeez

[digital mesa]

yeah but i think they are going to be able to bring that down, at least thats what he says in the video

[pallid ruin]

Idk why it uses that much vram 😭 it looks like it tries to load everything at the same time

[heady anchor]

it loads the base sdxl and then it loads the 13b llava model on top of it

[nimble mason]

exactly what i thought

[heady anchor]

then it does a latent upscale

[nimble mason]

badly needs to just, uh, load and unload as needed lol

some of my comfy workflows would take 170GB of vram if i didn't unload shit

[pallid ruin]

💀

[nimble mason]

that 23.5GB limit is a real red line imo

your userbase is going to drop 99% once you go over that

[solid plinth]

How do you use AI commands to draw on Stable Diffusion on Discord?

[nimble mason]

bots are dead

[digital mesa]

so i installed the local version and it has instructions to run it on runpod but no obvious ones (for a dumbass like me) to run the local version, you guys brain hard right? any obvious cues here that you can translate to stupid for me

[pallid ruin]

Sadly i have no idea how to install it

[hazy warren]

does more VRAM mean fast image creation?

[heady anchor]

i dont use runpod so i wouldnt know 😔

[pallid ruin]

Not really unless you got so little vram that you have to offload the model to the ram, in that case yes, having more vram would be faster; but generally no

[hazy warren]

😮

[pallid ruin]

(I think*)

[heady anchor]

no,u can buy an old tesla with 24gb vram and still wouldnt beat a 1060

[pallid ruin]

I had to do it, I saw your nick @heady anchor 🤣

(the face doesn´t look much like him tho)

[heady anchor]

the hair looks perfect

[digital mesa]

i already have it installed, he just gave no instructions on how to run the damn thing once you have in installed lol, then went to bed

he couldnt run it locally himself because his machine only has 24GB ram so he had to use runpod, but i have a machine with 48gb cards in it so wanted to try it

blah

[hazy warren]

whats the difference between me having the sampling set to 20 compared to like 150?

[digital mesa]

depends on the sampler? mostly: wasted computing time. Does anyone use much more then 40 samples? Seems like img2img/inpainting is the only thing that needs more samples

[heady anchor]

no its just a waste,20-40 more than enough

[nimble mason]

I often use over 40

[digital mesa]

for what?

[pallid ruin]

imo 26steps is perfect, at least for sdxl in 1024x1024

[nimble mason]

I don't think I ever do less than 35 anymore

[nimble mason]

Pretty much everything

[hazy warren]

i've just been doing 20 at 512

[pallid ruin]

I´ll test it right now in img2img, 20, 26, 30 and 60

[digital mesa]

but for what kinda stuff? Any time i've gone over 40 it hasnt made a difference, at least for realistic stuff, in the models i've been using

[quasi snow]

@digital mesa Spider Man and Thor are smoking, drinking tea, and playing games

[heady anchor]

https://tenor.com/view/spiderman-meme-funny-thor-hammer-gif-18585556

[pallid ruin]

xd

[nimble mason]

It's not always better but I find i consistently like the images more regarding details

[hazy warren]

the only problem I have with SD is, it makes things too cinimatic, it needs more imperfections to make it more realistic 🙂

[pallid ruin]

That´s true, well, it should be able to generate both things

[nimble mason]

SDE sampler in particular need at least 50 imo to look good

Below 40 it's real shaky

Below 30 many images are shit

[pallid ruin]

Okey....they swap their costumes

🤣

[digital mesa]

ok running a test 10,20,30,40,50,60,70,80,90,100 samples

[hazy warren]

I made an image of Cats

[heady anchor]

there are a lot of 1.5 and sdxl models that are finetuned to look realistic,just go to civitai

[nimble mason]

It's not always needed but it almost never makes the. Worse

And the diff between 20 and 40 on a 4090 is literally 2-3 seconds

[hazy warren]

what about a 4080?

[nimble mason]

Prolly like 5 or 6 lol

[hazy warren]

🙏

[nimble mason]

If you're on a 2060 super you might be waiting a while

[pallid ruin]

sdxl still suffers a lot generating 2 characters

[nimble mason]

It's very sampler and scheduler dependent too

Deterministic ones are gonna converge and then there's no chance

[digital mesa]

oh hey is there a limit to upload rez here? because the plot might be too big to paste in that case

[nimble mason]

So really it's the sde and ancestrals that benefit the most

25mb

[digital mesa]

k will see

[nimble mason]

I'm prolly also making diff types of images than you

[digital mesa]

sure i'm just seeing what it looks like for something i generated earlier today with diff settings

because i've tested quite a bit and found not much improvement over 40 for the stuff i'd been making but this was a while ago so guess we'll see

[nimble mason]

It's not dramatic for most things

[digital mesa]

or let me put another way: cranking the samples up rarely fixed the very obvious problems i was trying to solve like mutated limbs etc

[nimble mason]

Some wacky shit though does really benefit from. The extra steps with sdu and ancestral

Yes

It won't fix that

Its the fine details

[digital mesa]

right but so if the fine details are amazing but its still a mutant freakshow its not really helping, and the former is a CONSTANT issue, the latter not really so much

[quasi snow]

@digital mesa A beautiful lady is baking a tasteless dessert in the kitchen, while the cat next to her is napping. My daughter is already starving

[hazy warren]

idk why im generating images of cats

[nimble mason]

Two different problems

Both still problems

[pallid ruin]

😁

[hazy warren]

cat part of the cake?

[pallid ruin]

Lmao

[hazy warren]

I think this might be the most realistic looking cat I can make

[heady anchor]

https://tenor.com/view/kiss-gif-11816971814746635421

[pallid ruin]

https://tenor.com/view/cat-cats-cat-love-cat-kiss-kiss-gif-24690536

[nimble mason]

20 30 60 steps dpmpp sde

[hazy warren]

LOL, most of my prompts is telling it to make the worse picture

[hazy warren]

turn on Hires fix and it gives you more details

[pallid ruin]

my try

[hazy warren]

I try to give the cat eye discharge, to make it more realistic, but it wont do it 😦

[digital mesa]

ah shit i made my most common mistake and forgot to set adetailer to 1024x1024 for sdxl like as dumbass

[nimble mason]

Lol

The sampler and scheduler are critical for that

[hazy warren]

this is just 512

[nimble mason]

Dpmpp Karras usually converges in 30-40 steps ime

Euler ones, DDIM etc take 60 or more I think

[digital mesa]

this was DPM++ 3M SDE

[hazy warren]

do I need more steps for 1024?

512, I have 20-35

[digital mesa]

lemme rerun with adetailer off

[digital mesa]

do you know how to run X/Y/Z plots?

[hazy warren]

idk what those are

[digital mesa]

are you ising A1111?

using

[hazy warren]

no, SD

[digital mesa]

yeah but what interface, what app

[hazy warren]

the webui

[digital mesa]

there are several

[hazy warren]

the standard one?

[digital mesa]

so A1111?

[hazy warren]

I didn't put and models in that folder

[digital mesa]

ok, in txt2img, if you scroll down to the bottom of the page do you see "scripts"?

[hazy warren]

yes

[digital mesa]

ok select X/Y/Z plot and set X type steps and put in these values in X values: 10,20,30,40,50,60,70,80,90,100

then when you generate again it will plot out images with all these values on a grid, so you can test all the sample values sequentially and compare them in the same image

[hazy warren]

😮

[digital mesa]

make sure to set batch count/batch size to 1

[hazy warren]

wow, that gives an eta of 40mins

[pallid ruin]

Oof

[digital mesa]

you can do the same thing to test checkpoints, samplers, search and replace prompts, etc etc

[digital mesa]

how long was it taking before this

per image

[hazy warren]

like about 1min an image

[digital mesa]

then it should be done in 10 mins if you have set batch size and count to 1

since its doing 10 images

[nimble mason]

20 40 dpmpp a

[hazy warren]

making some wallpapers I see

[nimble mason]

I make lots of weird shit

[digital mesa]

ok no adetailer this time

[pallid ruin]

So cool

[nimble mason]

I think it got compressed

[pallid ruin]

you have to click open in browser

[digital mesa]

to my eye 70/80/90 steps looks worse than 60 here

[pallid ruin]

🤔 I think that 30 steps is the best, compared with 100 steps it doesn´t have that much of a difference but its way faster

[digital mesa]

yeah girl in BG kinda has melted face

but you seem to get that randomly on higher steps too

[nimble mason]

Sde ones can vary from img to img

Cuz they don't converge

[digital mesa]

feel like if i reroll i could get that working well in 30 steps and you could see that issue in 90 steps

[hazy warren]

Make it have a rainbow tounge

[nimble mason]

Step 60 can look great, 61 worse, 62 even better, 63 awful, etc

[digital mesa]

so this leads me to the unscientific conclusion that at least for the stuff i'm swinging for, more steps not necessarily better

[hazy warren]

How much worse is my 4080 compared to the 4090 for D2?

[digital mesa]

WHats D2?

[hazy warren]

SD2

[digital mesa]

oh SD2 sucks, who uses that

[nimble mason]

Yeah no one touches it

[pallid ruin]

SDXL > SD2

[digital mesa]

maybe you mean XL?

[nimble mason]

4090 is almost twice as fast as the 4080 I think

24gb vram vs 16 but that only matters for training Lora models

Your 4080 is a really sweet card it's the second best consumer card you could get

[hazy warren]

It’s great for gaming

[nimble mason]

For sure

[digital mesa]

obviously with more vram you can generate more images simultaniously which is more efficient

[nimble mason]

4090 is straight up overkill for pretty much everything except Alan Wake 2

Yeah

[digital mesa]

or 3d rendering

[nimble mason]

I usually just bang out one or two at a time myself cuz I'm messing with it nonstop

[hazy warren]

https://tenor.com/view/wow-iota-chris-farley-tanglevision-cryptocurrency-gif-22350373

[nimble mason]

So I don't really take much advantage of that

[digital mesa]

there's no such thing as overkill for 3d rendering

[nimble mason]

I don't do a lot of seed hunting

True that

[hazy warren]

I get about 120-160fps in games

[nimble mason]

Haha I saw someone on reddit complaining they dropped from 800 to 650 fps in something

And everyone was like what the absolute f are you talking about

[digital mesa]

do a lot of monitors support 800fps

[hazy warren]

I just have triple 1440p monitors

[pallid ruin]

I think the maximum is 500hz

[digital mesa]

the maximum refresh rate on my main monitor is 60hz lol

[nimble mason]

Yeah exactly it makes no sense

[hazy warren]

Even my phone has 120hz

[nimble mason]

Mine maxes at 240

It was total overkill

[digital mesa]

but my monitor is a color accurate Eizo coloredge, not for games

[pallid ruin]

Mine has 120hz but the colors suck xd

[hazy warren]

My watch is the only device is 60hz

[pallid ruin]

Aaaahh i wanna sd3 😭

[digital mesa]

you get that plot done? work out?

[hazy warren]

I wanna use Sora

[hazy warren]

Indeed

[digital mesa]

30->60 samples went from 36 to 46 secs

[nimble mason]

Images like these def benefit from at least 35 steps

[digital mesa]

but i think the cutoff we were broadly arguing against was over 40

as diminishing returns

looks cool tho

you ever see Paprika

[nimble mason]

Def diminishing returns past 40 but I don't think it's until I'm past 60 where I generally don't see an obv improvement in quality

[digital mesa]

can you run a plot on the last image there to see the diff?

[nimble mason]

Not on my comp kust phone atm

[hazy warren]

what model is this?

[nimble mason]

I already forget

[nimble mason]

Looks like a hole u don't crawl into and crawl back from lol

[pallid ruin]

Nice

[hazy warren]

yeah, don't go in there

[digital mesa]

anyone use fixFP16ErrorsSDXLLowerMemoryUse_v10.safetensors VAE?

seems to do some weird shit

[pallid ruin]

👀

i´ll try it

[digital mesa]

did lots of images without issues so not sure why this is janked

[hazy warren]

😮

[nimble mason]

Might wanna just use tiledvae

I've never really noticed a diff

[digital mesa]

can see it in a couple of the bg characters here too

whats tiledvae? first i've heard of it

[nimble mason]

It's the way I get around out of vram BS with vae

Just breaks it into tiles

Takes a good deal longer but ends the OOM 100%

Nothing more annoying than waiting for a long gen to complete getting excited about what ur seeing then poof, gone, error lol

[digital mesa]

OOM?

out of memory?

[nimble mason]

Out of memory

Yeah

[digital mesa]

not getting mem errors tho (on a 4090)

[nimble mason]

Oh

Why a low vram vae then?

[digital mesa]

well someone said it was faster if i recall

[nimble mason]

Maybe

Doest really take long on a 4090 though lol

[digital mesa]

but it was 38.7 secs with normal one and 45.8 secs with the low memory one so actually its slower and looks shit, ha

for 1920x1080 image

[nimble mason]

Wtf

That's forever for a vae on a 4090

[digital mesa]

wait now its doing it on the normal one (guy on right)

[pallid ruin]

that with or withouth adetailer? 👀

[digital mesa]

no adetailer

[hazy warren]

I can't make a good black hole 😦
I wanted one in a room

[nimble mason]

My fav new negative prompt: symmetrical

[pallid ruin]

what prompt did you use?

[nimble mason]

You need the right model

[pallid ruin]

Because it takes 43 seconds on my undervolted 3060 for a 1920x1080 img @digital mesa it should be way faster on yours

[nimble mason]

They're not well trained on that

[hazy warren]

black hole, in a room, ((masterpeice))

[digital mesa]

well loras slow down generations

[nimble mason]

For me that would take maybe a second

[pallid ruin]

Try this "a black hole inside a house, papers and objects flying inside, chaos, particles, very windy"

[nimble mason]

They should only affect load and unload time

[digital mesa]

no they definitely affect generation time

[nimble mason]

[Black hole|door] try this one too

[digital mesa]

in my experience of wanting to smash my computer

[nimble mason]

What are you using

A111v

?

[digital mesa]

yeah

[pallid ruin]

You gotta try sd forge (its a1111 ui + comfy)

[nimble mason]

I've never noticed any diff on comfyui

Or forge yeah

[digital mesa]

maybe

[nimble mason]

A1111 suuuuck so bad for performance

[digital mesa]

can forge do adetailer with gender detection

[pallid ruin]

Idk why but a1111 is faster than comfyui on my PC, but both are slower than sd forge

[nimble mason]

That is very strange

A1111 is also horrendous with vram

The same workflow that cost me 8.3 GB on comfy earlier tonight was at 22gb in a1111

[pallid ruin]

Damn that´s a lot of difference

[digital mesa]

34.5 sec. no lora 45.6 sec with lora turned on (lora does the flare)

[nimble mason]

Yeah lol a1111 code is a fn shit show

That's prolly the load time

A1111 doesn't hold the. In memory

So when you hit generate it then starts reading everything from disk

[pallid ruin]

🤔 that´s so strange, in my PC even with 4 or 5 loras at the same time the generation takes the same time (except the first time, of course)

[digital mesa]

so if i generate then hit generate again, same settings, it has to load again every time?

[nimble mason]

If you want to be sure, put the LORAs on a really slow disk like a USB 2 thumb drive

I think so

I think it keeps the main model but reloads the LORAs

[digital mesa]

i dont think thats true becaude the cmd line monitor thing doesnt show anythign about loading

[nimble mason]

FenrisXL is badass

Yeah idk it's the only explanation

LORAs come at no additional computational cost

You're just patching the weights unlike controlnet which does double the cost I think

The only cost is load time so it must be loading, that or the a1111 code has A Lora loop that calculates pi to 5 million digits real quick then advances lol

Which tbh wouldn't surprise me

[digital mesa]

well i am happy to be wrong here, my brain is very small

[nimble mason]

I'd just give forge a shot

You can have it read all your same models and config files

So all you gotta do is unzip it no reinstalling tons of models etc

[digital mesa]

the only thing that would make me upgrade is if you can use adetailer and it is able to do gender detection, thats literally the only feature i need or care about

which you can do in comfy but i hate node interfaces

[nimble mason]

Well it might work fine idk I haven't looked into that

Yeah I've used it in comfy

Here's a girl for ya

[digital mesa]

keeps making one person into a zombie demon lol

[nimble mason]

Lol

Out of the nearly 200 checkpoints I've tried, yep FenrisXL def the best for anthropomorphic

[pallid ruin]

It does look good

[digital mesa]

ppl say pony for this, esp if you also want their an*s to be gaping wide enough to put a football in it, and leaking some kind of fluid

[pallid ruin]

XD

Ponydiffusionxl is good, but you gotta use censored negatives just in case

[hazy warren]

Sometimes when I use negatives, I feel it doesn’t do anything

[digital mesa]

you have to beg it

anyway it may be the load time, given that loras are so astronomically much larger in XL, on average each is like 800mb

[digital mesa]

so i did a batch of 4 images, 2mins 47.1sec with lora active and 2 min 34.1 secs without

[nimble mason]

Yep there you go dead ringer it's the load time

Increase in time is the same

[digital mesa]

well i won't cling to being wrong

[nimble mason]

That is a really fn slow generation for a 4090

A1111 sucks balls

Def give forge a shot and you'll be like wtf

[digital mesa]

is it? i'm generating at 1280x720 and upscaling to 1920x1080 on every image

[nimble mason]

Yeah that's still pretty slow if you're not doing 40 steps dpmpp ancestral like I do all the time

[digital mesa]

30 steps DPM++ 3M SDE

[nimble mason]

I bet you cut that down a lot without a111

[pallid ruin]

On my PC with a 3060:
A1111: 19 seconds
ComfyUI: 21 seconds:
SD forge: 14 seconds

[nimble mason]

Yeah that should be fast

[warm pier]

How do I use the ai

[digital mesa]

well maybe i have something set up wrong

[nimble mason]

Hi! I am a bot. What would you like assistance with today?

[digital mesa]

for same resolution as i'm using?

[pallid ruin]

Nono, 1024x1024

[digital mesa]

ok well its 43.4 secs to generate a batch of 4 of these images at 1024x1024

[nimble mason]

Jesus

[digital mesa]

12.6 secs to gen a single image

didnt they say the advantage of forge was much greater with less vram and barely any difference at all if you have 24gb

i think thats what i read

yeah its the 1.5x upscale thats doing most of the slowing down here other than lora loading

[nimble mason]

Nah that's bs

It's a huge diff on a 4090

[digital mesa]

ok well 19 seconds on a 3060--->12.6 seconds on a 4090, does that sound correct to you for A1111

[nimble mason]

Ive never used a 3090 but I went from idk, 20-30 sec with 40 steps dpmpp-a to like 3-4 sec

[digital mesa]

yeah sorry that was a typo

[nimble mason]

Going from a 3080 12gb

[digital mesa]

meant 3060

that LeoB cited

[nimble mason]

Just ran this on forge

30 Steps dpmpp sde

Took 4 seconds

50 steps Karras scheduler

[pallid ruin]

https://tenor.com/view/mimimimimi-cat-sleeping-snoring-cat-sleeping-cat-drooling-cat-gif-23979759

Goodnight!

[digital mesa]

nite

[digital mesa]

how long did the second one take

first one looks pretty undersampled - that was 4 secs right?

also are you using turbo or lightning models (i'm using helloworld which is neither)

[steady skiff]

This has me geekin out

[digital mesa]

installing update for Supir now, apparently now works in 12gb Vram

[neon zinc]

Logo design for an information technology company called Elgogary

[golden pebble]

Hi guys, how can i generate two characters

[noble zephyr]

elaborate please. would you like to generate 2 characters in the same image? or are you looking for a way to generate 2 characters in general?

[golden pebble]

2 characters in one image

as a couple for example

[heady anchor]

just write a prompt for 2 chars then

[noble zephyr]

maybe use some controlnet modules?

[nimble mason]

Yep

[noble zephyr]

see, that's what i don't like about AI video. you can always clearly see that it's AI

[wispy nest]

well yea

[heady anchor]

wooah u can use animatediff and deforum woaah dude

[wispy nest]

its not hyper realism, we're not there yet

no not deforum

comfyUI and SM

the back-end + a node

[noble zephyr]

which is why i prefer images

[wispy nest]

and i don't prefer images cause when this crap comes out of beta and we're no longer making crap gens becaeuse the tech isn't there yet and i'm over here doing linear equations bustin my ass off, gonna be worth it

images will be a huge waste of everyone's time

everyone is already desensitized and you want to make still images

[heady anchor]

i think we may be in presence of an AI genius in here,one who will be leading the industry in the next decade

[wispy nest]

who knows? in 1 month im this good? there has to be something to that tbqh

this is a gigantic stride for 1 month

[heady anchor]

we are not worthy

no one and i mean not a single human here can make such beauty

[wispy nest]

would love to see your videos

im here to make music videos, already got my workflow ready

i make music, i started this cause i wanna make my own vids so it is what it is

[heady anchor]

how could i compete with a genius like yourself is like trying to compete with Davinci

a genius like yourself should be selling a masterpice like this to museums

a new era in the AI industry not even openai sora can produce such beauty

[wispy nest]

how much money are you making again on your work?

[noble zephyr]

why does everything have to be a side hustle? just let people enjoy things

[heady anchor]

respect massa he made 2 videos and sold them hes like Scorcese but better

[wispy nest]

i never said im as good as you say i am but ty

to me this work is ass, its my trial and error 🙂

but it somehow made money ??? haha

either way, i encourage you to do better than this

i did NOT try to make this my side hustle

[wispy nest]

who said i was good?

lmao

i just said im doing video, yall are working with images

[wispy nest]

point it out whree i said im good

[noble zephyr]

you did.

[wispy nest]

lmao

yea this good for 1 month

is actually great

for where i see others in 1 month, im making gigantic strides

thats all i meant, did not mean im good

[noble zephyr]

riiight.

[wispy nest]

no seriously, i dont think my work is that great yet -- gimme time

like i said check me in 1 year

[noble zephyr]

what kind of music?

[broken zinc]

@tropic shell these are two of my beavers that are good, and there is 1 that is bad. i think i fixed this issue and can now generate the style as the first 2 beavers each request instead of getting this other style beaver by using the same seed

[noble zephyr]

wow, that really isn't my jam

[wild sorrel]

idk if 3rd one passing rules

[wispy nest]

3rd one needs heavy work

[noble zephyr]

no offense, but really not at all. this isn't my genre tbh

[tropic shell]

Very cool! And dapper, I might add! Love the little hat shine, and the whiskers!

[wispy nest]

couldnt get her to smoke a cig, realized what this isnt good at yet

[wispy nest]

ye trap isnt for everyone, i make rap

if you dont like rap / dark music you wont like my music

[broken zinc]

unglued u rap?

[wispy nest]

i used to, but my voice is country so i dont haha

[noble zephyr]

it seems our musical performances are quite on different sides of the spectrum

[wild sorrel]

slow build up to 1min was good, later on meh

[broken zinc]

fair lol

[wispy nest]

we all used to freestyle on the streets

[wild sorrel]

dark trap is like 99% of music I listen to lately, but that's not it lol

[wispy nest]

ye ive only made music for 1 year honestly

and i dont use loops, just about everyone else is using loops

so when my music is mid, at least it's all my music

all my hard work, not copy and paste

[wild sorrel]

noone cares if it's using loops or not - good or not is what matters

[noble zephyr]

very true

[wispy nest]

nah it does matter

loop users are lame

[broken zinc]

anyone appreciate good phonk music?

only once in a while tho

[wispy nest]

i make phonk

[noble zephyr]

not at all lmao. i'm a metalhead

[wild sorrel]

this is looped, yet you probably won't even really notice unless you'll try to lol
https://www.youtube.com/watch?v=59eAA6-CeBE&list=LL&index=2

[wispy nest]

this is when i first started out

[broken zinc]

i love breaks in phonk songs

like straight up pauses in the sound

you know what i mean

[wispy nest]

yea

i made phonk when i was only 3 months into music, wasnt that fun tho

the 'moar cowbell' gets old haha

[broken zinc]

this is not bad but i mean fast phonk like that memphis cowbell og

[broken zinc]

lol

[wispy nest]

ah you mean drift phonk

[broken zinc]

yea

[wispy nest]

ye drift phonk is cool

[broken zinc]

i like the army phonk too or military phonk

[wispy nest]

that i've never heard of

[broken zinc]

idk if its really a genre but there seems to be a 'class' of phonk ideos that share that trait on youtuybe all military themed videos and background

[wispy nest]

its horrible, and poorly mixed

[wild sorrel]

I have few phonk thingies in my lib...but I don't think that's kind of phonk you are talking about lol....more like a mix of a dark trap and phonk

[wispy nest]

the rapper sounds like shit, the bass comes in too hot destroying the integridy of the track

is this what you listen to eshku? lmao

[wild sorrel]

it's not rapper tho 😄

[wispy nest]

you say this is looped like its a good thing

whatever it is, its awful 😄

this is 100% rapping

[wild sorrel]

I say it doesn't matter if it's looped or not if it's good

[wispy nest]

that's not good

[wild sorrel]

genre (rap or not in our case) depends on music on top of words

[noble zephyr]

you seem to forget that music is subjective

[wispy nest]

and the fact its looped

lets me know the person had no skill

and the fact that it's mixed so poorly

lets me know they have 0 ears

i watched dave pensado and jaycen joshua video 25 times to learn their entire method

in a course of 2 months googling every word, memorizing *

[wispy nest]

nah i still believe in doing your own work, i don't mind the 13 year olds smashing loops together making beats that may sound better than mine at times, the fact is they are role playing GTA but in music, those of us that actually do this cause we love this shit -- it's diff with us

[wild sorrel]

13 year olds lol

it went off rails quick

[wispy nest]

Here's the reason why i don't like loop users and why the market is oversaturated:

Let's say i post a beat in my work channel, certain people are like damn this dude is pretty good at music for 1 year, then someone comes behind them and starts saying this...

"my 13 year old son can do that, making music isnt special anymore, anyone can do it"

^^^ make a real artist want to quit

[broken zinc]

soon it will be... my AI can do that

[wispy nest]

zero effort work

[wild sorrel]

I think you care too much about something that doesn't really matter

[wispy nest]

when i use loops my music is so much better

it does matter

it just doesn't to you, and to a lot of youngsters

a lot of us do care about this tho

[noble zephyr]

the go care about it. just don't try and force it upon us

[misty igloo]

https://youtu.be/WRiSXYUZJ3A?si=9RijyhSnpq6TxRjf

one of my faves is this one

[wispy nest]

im not im just trying to say that zero effort work is whack

and you can always tell somehwere within the quality, rather it's the mix or wahtever

[wild sorrel]

maybe you I wanted to sit and actually listen to music, not as background thing, but like actually listen to it - then maybe it would matter...but not in a trap genre

[wispy nest]

there's always a huge gap of knowledge

[misty igloo]

talking about videos made w AI

[wispy nest]

ask a person what a specific sound is in their beat

that shit has me rollin all the time 😂

they guess and lie when i know the real answer

"oh its reversed reverb"

when it's not, it's a reversed snare w/ time mul

[wispy nest]

kordhell is dope

known about him for some time now

he makes his own stuff

actually came across him on a discord server once

[misty igloo]

he’s currently my no 1 artist on Spotify, his sounds are 🔥

[wispy nest]

its next level, hes on his way

ever heard of soulker? he passed and was a great friend to the community, also made very good music like kordhell

[misty igloo]

no

[wispy nest]

check him out on youtube

https://www.youtube.com/watch?v=mwWS8DyG_HU and check this out, a new workflow im gravitating towards, using milkdrop and implementing SD with After Effects

[misty igloo]

oh I see

thank you

[wispy nest]

np, his music is awesome

ull like it if you like kordhell

and keifer gr33n

[misty igloo]

kk

[wispy nest]

https://www.youtube.com/@KEIFERGR33N
https://www.youtube.com/@Soulker

[wispy nest]

images like this will be the only decent ones that are fun to make in SD in the future for custom memes, did you animate yourself here?

[steady skiff]

No I did not, This is completely random @wispy nest

[sturdy pewter]

updated to V8 and now works even at 12 GB GPUs (tested on single RTX 3060 - 12 GB ) also using as a base model Juggernaut-XL-v9 which literally yields even better results than shown in video
this model is literally many times better than even very expensive Magnific. It really not getting the attention it deserves : https://youtu.be/PqREA6-bC3w

[wispy nest]

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

[noble zephyr]

pretty impressed with JuggernautXL lightning so far! all these images were created in 4 steps with very little prompting. no upscaling

[full vessel]

those are pretty good!
yeah thats where i heard the word juggernaut haha now i remember

[hazy warren]

If everyone is an artist with A.I, is anyone really an artist?

[noble zephyr]

that's quite a controversial topic

[hazy warren]

[exotic zodiac]

How to fix?

[noble zephyr]

what model are you using?

[quiet current]

probably incorrect VAE or using too many steps with a turbo model

[exotic zodiac]

I use DreamShaper XL

Maybe we should use None in VAE?

[nocturne oak]

Use Automatic or the actual SDXL VAE.

[burnt niche]

any suggestions on which model works best for this type of generation?

[brazen sand]

@cold meadow an example of Ultimate SD upscaler

original - upscaled x 2 with zavychromax / ultramix balanced

added good detail everywhere, respected the original

@cold meadow another example, by mistake I use a high denoising (0.54, really high!), it did aberrations and some light inconsistency but it kind of worked as well, it is kind of gamble but it can just upscale and rework the image.

May be good for just adding detail like crazy, with the cost of doing some cleaning (still some light inconsistency but I was surprised by the results)

[hazy warren]

Skin looks a little bit too smooth

[regal venture]

#🏞｜general-with-images photography style, VietNam's cottage style located in the middle of the rice field in rural VietNam, bamboo, early morning lighting"

[brazen sand]

may be the DPM++ 3M SDE sampler

[wispy nest]

i didnt know metal could bend like that and still work

[hazy warren]

😮

[hazy warren]

it may not be working

[wispy nest]

overall really cool image

steam punkish

[hazy warren]

steam punk pickles when

[brazen sand]

well at some point I got tired of inpainting, but some inpainting tries could make sense of that. I noticed the arms were kind of off an I got lazy, it is difficult to inpaint large detail in a large picture. So I abandoned it.

[nimble mason]

is her clothing grabbing her boob? 😄

[wispy nest]

https://tenor.com/view/hysterical-laughter-laughing-gif-25735842

[sudden cape]

Did you ever make a tutorial on these?

[wind egret]

hmm my mov2mov is only giving me 30 frames

[brazen sand]

I was in fact trying to inpaint the gloves so I prompt "steampunk girl, working gloves, working on steampunk machinery", but Ultimate SD upscaler was active, so it did gloves everywhere it could

[hazy warren]

LOL, I was trying to generate realistic people, and found the hair funny

[burnt leaf]

一碗螺蛳粉

@nimble mason 一碗螺蛳粉

[noble zephyr]

I can if you want to, but there really isn't that much to it

Wait, I thought you were referring to the SDXL lightning images. I can definitely make a tutorial for the pixel art images! :)

[hazy warren]

does SDXL make better images?