🏞｜general-with-images

[nimble mason]

pick style

[crisp stream]

Ah, I see, then where do I find the weight (what weight to begin with :-D) and "pick style"?

[nimble mason]

click on the ipa nodes and click the weight drop down

cat asleep in my arms so typing like shit 😄

with clownschedule B

[crisp stream]

can´t see anything alike (when right-clicking the IPAdapter Embed node)

[nimble mason]

screenshot?

or worklfow

[crisp stream]

ah, you are referring to choosing the weight type? Because they renamed it?

[nimble mason]

yep

[crisp stream]

like before it was style transfer SDXL and now it is simply style transfer (and strong style transfer)

[nimble mason]

yup

[crisp stream]

yes, thank you 🙂

[nimble mason]

cuz it works with sdxkl and sd15 now

np!

[ruby marten]

can someone pls do mexican mike tyson

[pastel root]

You'd have to give me AT LEAST 50 dollars to sleep with Mexico Mike

[ruby marten]

damn not in that way

prompt

[clever oar]

dreadnought warhammer 40 k

[thick tiger]

Hi. Tried using your workflow but one node is missing. Which one is it?

And where could I find sdxl ai creator checkpoint ?

[nimble mason]

go to comfyui manager, "install missing custom nodes"

[nocturne oak]

Or don't. We don't judge here.

😛

Actually, we already did judge. You failed.

[thick tiger]

Yes it usualy works but no nodes show up there

[civic osprey]

can you help with next pront:a photograph of a snake, with large ice wings, an iridescent skin, a hypnotic look, which makes it a mythical creature, primordial force, elegant, captivating, aura of mystery, tail with sting, is on top of a snowy mountain, clouds around him, with little light on his right side.

[nimble mason]

i think it's the extra samplers node

https://github.com/Clybius/ComfyUI-Extra-Samplers

err, pack i mean

samplercustomnoise maybe

or maybe it's just flat out ksampler and your shi tis fucked up

[thick tiger]

I have that installed

[floral maple]

Hi, can you help me with this promt: A photograph of a fantastic animal that is a western dragon, with angel wings, its body, translucent and dark, reveals its skeleton in a play of light and shadow. This being, its food is souls, it has an indomitable and mysterious power, it is a duality between light and darkness, between the celestial and the earthly, it stands guard on a rock on top of a mountain at sunset. -- ar 3:2

[thick tiger]

nah that's something else, haven't had problems before

[nimble mason]

have you updated comfyui in the last 24 hours?

[thick tiger]

Yes

[nocturne oak]

last 24 hours last 24 minutes
ftfy

[thick tiger]

Just tried again 😄

[nimble mason]

try just adding a ksampler and manually hooking it up in place of the dead node

[thick tiger]

Sure will try

[nimble mason]

Here is the image you requested.

[nimble mason]

Here is the image you requested.

[floral maple]

Thanks you

[dense bison]

Hello, can you help me with this promt: It generates a realistic photo about a creature, it has a large humanoid body with imposing musculature, crowned with curved buffalo horns that radiates a sense of power in a majestic enchanted forest. His large, majestic wings gracefully reflect the celestial light as they unfold. Masterfully, in her hands she controls fire and makes flames dance at will in the place where ancient trees and shadows move to the beat of ancestral magic. Her intense and penetrating gaze reveals the wisdom of ancestral beings, embodying strength, magic and majesty

[nimble mason]

Here is the image you requested.

[dense bison]

Thank You 🙂

[arctic spear]

Hello, can you help me with this promt: An art piece in photographic style, full-body view, featuring a mouth resembling Orbea variegata with sharp teeth, the tongue like that of the Xenomorph's Facehugger, the head of the Xenomorph, and the body of the Demogorgon from Stranger Things, with a tail like that of the Xenomorph.

[pastel root]

Your image is now loading.
Please Wait Patiently.

[nimble mason]

Here is the image you requested.

[pastel root]

[nimble mason]

@jovial tiger In an old-world kitchen, a coal-burning coffee maker of polished ebony metal, with intricate silver filigree, commands attention. It's perched on a stone countertop, its firebox aglow with glowing coals. A delicate porcelain cup sits beneath the spout, catching the rich, dark coffee that drips slowly into it. Around it, the air is filled with the robust scent of coffee mingling with the earthy undertone of burning coal, as morning light streams through a nearby window, highlighting the steam that curls gently from the spout.

...almost 100% adherence

only thing missing is coffee dripping dirctly into the porcelain cup

✅ old-world kitchen
✅ coal-burning coffee maker
✅ polished ebony metal
✅ intricate silver filigree
✅ stone countertop
✅ firebox aglow with glowing coals
✅ delicate porcelain cup sits beneath the spout 🤨 catching the rich dark coffee that drips slowly into it 🤨
✅ morning light streams through a nearby window
✅ highlighting the steam that curls gently from the spout

insane

no seed hunting. dpmpp_sde is really good with sigma

[keen gate]

@nimble mason A professional headshot of a Chinese woman in a business suit. She has a confident and approachable expression. The suit is a dark color, perhaps navy or black, and she is wearing a crisp white shirt underneath. Her hair is styled neatly. The background is grey, not distracting from the subject. The lighting is soft and flattering, highlighting her features without creating harsh shadows. The photo is taken with a high-resolution camera, ensuring every detail is captured clearly. The composition is centered, with the woman's face in the middle of the frame. The photo should be hyper-realistic, highly detailed, and high-resolution 16k. --ar 3:4 --v 5.2 --style raw --q 2 --s 250

[nimble mason]

Here is the image you requested.

[arctic spear]

Hello, can you help me with this prompt? An art piece in photographic style of a creature that is depicted in full-body view, with a mouth resembling the flower of Orbea variegata with sharp teeth, the tongue like that of the Xenomorph's Facehugger, the head of the Xenomorph, and the body of the Demogorgon from Stranger Things, with a tail like that of the Xenomorph.

[nimble mason]

Here is the image you requested.

[arctic spear]

Photo where a hybridization is made between the demogorgon from the Netflix series "stanger things" and the xenomorph from the horror movie "Alien"

[nimble mason]

Here is the image you requested.

[arctic spear]

Photo where a hybridization is made between the demogorgon from the Netflix series "stanger things" and the xenomorph from the horror movie "Alien"

[floral maple]

Hi, can you help me with this promt: A photograph of a fairy, with brown skin, a bright green dress clinging to her body, she has translucent turquoise wings. She who is sitting in the middle of a magical forest, is greeting a spirit that looks like an orange fox. -- ar 3:2

[nimble mason]

Here is the image you requested.

[arctic spear]

Photo where a hybridization is made between the demogorgon from the Netflix series "stanger things" and the xenomorph from the horror movie "Alien"

[nimble mason]

Here is the image you requested.

[floral maple]

lol

ty so much

[nimble mason]

Here is the image you requested.

[arctic spear]

Photo where a hybridization is made between the demogorgon from the Netflix series "stanger things" and the xenomorph from the horror movie "Alien"

[pastel root]

https://tenor.com/view/budgie-parakeet-yay-bike-ride-cookie-gif-17064900

[nimble mason]

Here is the image you requested.

[wispy nest]

can someone help me prompt makeup

[wispy nest]

im trying to mainly get eyeshadow/eyeliner but no matter what weight or what keywords i use it all comes out badly or without any makeup

[keen gate]

@arctic spear A person wears a Chinese female professional outfit. This dress is deep-colored or black, and she is wearing a loose white shirt inside. Her hair is neatly arranged. Pay attention. The lighting is soft and attractive, it highlights her features, but it won't produce a shadow on the eye. The photo is taken by a high-resolution camera, so it can clearly capture every detail. The photo should be realistic, highly detailed, high resolution 16k. --ar 9:16 --v 5.2 -- original style --q 2 --s 250

[nimble mason]

Here is the image you requested.

[crisp stream]

let´s say both are mechs ;-D

This looks more like one 🙂

[frigid tree]

@arctic spearA girl in a kimono stands on the beach and releases a goldfish

[nimble mason]

Here is the image you requested.

[nimble mason]

is that from the big eggs thing?

great stuff as always

[crisp stream]

Thank you, then I don´t know about the big eggs thing 😄

[nimble mason]

here they are

wasn't sure if you were doing any img2img stuff

i toss in crap from here all the time and see what i can do with it

[crisp stream]

😄

It was simply prompting

didn´t prompt for big eggs though

[nimble mason]

awesome

[nimble mason]

are you still editing stuff with photoshop liquify? i still need to try that

[crisp stream]

yes, I still do, yet not with the recent ones, nostly for special images only 🙂

[nimble mason]

^used this again

[crisp stream]

looks more like this though 😄 #🏞｜general-with-images message

[nimble mason]

exactly, i threw that img into my workflow here

more obv here

[crisp stream]

I was referring to this comment 🙂

[nimble mason]

oh haha

[shut sinew]

Thing is I cant share images

[nimble mason]

frozen nuke

[pastel root]

Jelly fish drill

[languid pebble]

Excat the same problem here ...

[languid pebble]

Thank you! Found it!

[frail walrus]

Gold

[nimble mason]

Here is the image you requested.

[wind jungle]

How to generate these images gusy?

[silent junco]

Can you help me with the next promt: Photograph of a half spider half woman creature, arachne-like, tarantula bottom, scorpion tail abdomen, huge mouth on abdomen with fangs, beautiful woman, blood on her hands, skeletons on the ground, dark forest background, dim light, cobwebs around--ar 2:3

[nimble mason]

Here is the image you requested.

[iron remnant]

good morning

[languid pebble]

Good morning 🙂

[arctic laurel]

Made with SD3

[lament spruce]

Hello, could you help me generate an image with the prompt: Skin color---Olive-toned complexion; Hairstyle---Dark brown, chin-length bob with bangs; Jewelry---Large, sculptural gold earring; Photo style--- Close-up, high-resolution portrait; Lens--- Shallow depth of field, focus on the face; Background--- Simple, monochromatic pale blue; Light--- Soft, diffuse lighting with a subtle highlight on the cheek and earring.

[languid pebble]

lol My Youtube video has negative views ^^

[cyan shoal]

pixart-sigma

[shy eagle]

pixart-sigma

[jovial tiger]

Really happy with it so far. Even single subject images are way more prompt following.

[cyan shoal]

im using llava 1.6 mistral with an input image then literally plugging the resulting prompt in the T5 conditioning

[jovial tiger]

Haven't thought to do that yet. Sigma goes up to a 300 token limit instead of 75 from clip. Clip couldn't handle the longer output from llava but now this can. Neat stuff

[cyan shoal]

exactly

I set it to a low token limit of ~200-250

and even those are extremely long

[jolly flower]

Hi, Can you help me with the next prompt:Photograph of a creature, inspired by the myth of the leshen in its version of the video game the witcher 3, has a body composed of elements of light poles, the size of a tall lamppost, emits its own light of a warm and dim color, wears black clothes and broken, his head is a cow skull with light pole supports as horns, werewolf claws with metallic texture, lives in alleys of cities, at night, only illuminated by its own light--ar 2: 3

[clever oar]

portal 2

[languid pebble]

YaY!

[nimble mason]

Here is the image you requested.

[clever oar]

crow sharing bathing

[languid pebble]

Need a picture? 😄

[clever oar]

yes 😆

[clever oar]

spirit of forest

[clever oar]

ahahaha 😆

[languid pebble]

Wait! Is that Crowman in the background?

[covert pagoda]

Muahahahahaha

[neat lagoon]

What is everybody's favorite model for prompt adherence, to get a result as close as possible to the input prompt? For example this prompt: "closeup of a grey cat wearing a blue suit, a red hat and a green tie is sitting on a white table in a room with big windows"

[clever oar]

crows bathing lol

[neat lagoon]

i'm doing some testing and I'm curious what checkpoints are usually considered best for prompt adherence

[clever oar]

I didn’t expect that my words would be taken as a request 😃

[languid pebble]

We asked you 😛

[clever oar]

bad water

[shy eagle]

Might be sd3

[clever oar]

crow cant drink it

[neat lagoon]

yeah but that isn't available yet

[shy eagle]

I have another one

[languid pebble]

Is it peeing? o.O

[exotic pelican]

cat

[clever oar]

bad water again very fast comeback)

toxic water

[languid pebble]

Here's your picture

[jovial tiger]

At this point, pixart sigma.

[ashen nest]

//

[nimble mason]

Here is the image you requested.

[clever oar]

24 gb is 4090?

[languid pebble]

If not Notebook version

[clever oar]

only 4090 have 24 gb?

in invidia cards

[languid pebble]

I don't know ... but it's the benefit of AMD to have more VRAM ... just not so supported by A.I.

[clever oar]

i try chap gpt for tech support

chat)

it doesn't give a direct answer, but it does suggest interesting thoughts that you hadn't thought of.

[languid pebble]

Not sure whether it knows the latest models ...

Looks like only 4090 with 24 GB but there are AMD with 24GB for half price

Ohhh... see: Gigabyte GeForce RTX 3090 Gaming OC 24GB GDDR6X Grafikkarte GV-N3090GAMING OC-24GD Schwarz

But 3090 not really cheaper

[nimble mason]

4090 and 3090

yeah it doesn't make much sense to buy a new 3090 right now

[languid pebble]

Pray for Taiwan or prices will be higher 😄

[clever oar]

one time i thought asus +nvidia give me quality...

[nimble mason]

prompt: "high quality, detailed photograph, woman"

[languid pebble]

I still hope so ... Asus Rog Scar Strix here ^^

[clever oar]

and i have 2 dead motherboards and video card with problematic shadows

now i buy gigabyte

[languid pebble]

I won this from ASUS ^^

[clever oar]

south bridge die

[clever oar]

not electro?

[languid pebble]

Electro worth 7k-9k ...

[clever oar]

i have electro bike

[languid pebble]

45KM/H ... so no E-Bike ....

[clever oar]

50

[languid pebble]

You can tweak but not allowed here

[clever oar]

is dangerous you every time test speed)

[languid pebble]

I asked ASUS for a 5k Coupon for their online shop ... I can't drive the bike ...

[clever oar]

but is very fun

[languid pebble]

Big city here ... to much stress

[clever oar]

only big problem with cells...

[languid pebble]

I prefer a new notebook in 1-2 years 🙂

[clever oar]

cells is problem for e-transport

expensive and often die

[languid pebble]

Bike is beeing produced in my city 🙂

[languid pebble]

Another woman?

[clever oar]

my battery fell into the water and some of the cells died

[cyan shoal]

pixart sigma

[languid pebble]

Bad luck! There are companies that fix it ...

much cheaper than a new cell

[clever oar]

i try fix

in service

not work afrer

[languid pebble]

Sorry for you

[clever oar]

work but 30 % capacity)

i want new battery technology

[languid pebble]

Water motor 🙂

There's no excuse for a bad outfit ^^

[jovial tiger]

it really is pretty great. I'm loving the new composition over normal sdxl

[nimble mason]

yeah, it understands fewer concepts for sure, but the ones it understands it understands really well

i'm hoping a finetune can add a lot of those in just like with sdxl base

[languid pebble]

DJ Neelix at an OpenAir 🙂

[jovial tiger]

@nimble mason @cyan shoal the one thing it's horrifically bad at, is hands. I'm trying denoising, I'm trying hand detailer... they're so bad that neither of these things can fix what pixart is doing.

[nimble mason]

yeah it's REALLY bad

worse than SD15

worse than SD15 base

i actually haven't seen it generate good hands even once

arms merge into each other too

[jovial tiger]

if schedule == 'DPM-Solver':
if not isinstance(pipe.scheduler, DPMSolverMultistepScheduler):
pipe.scheduler = DPMSolverMultistepScheduler()
num_inference_steps = dpms_inference_steps
guidance_scale = dpms_guidance_scale

so I don't know if I'm going in the wrong direction, but I was looking at teh code.

it's mentioinging a dpmsolvermultistepscheduler. you know where I also saw that scheduler?

ella. so i wonder if we need that here somehow

[nimble mason]

huh

[jovial tiger]

lol

i'm at the point where I have to throw out the depth controlnet because it's picking up the 57 fingers so it doesn't force the 2nd sampler to do bad fingers, with a 0.7 denoise to fix it all.

[nimble mason]

i think that might require some aggressive unsampling if it's fixable at all

probably the clownschedule

[clever oar]

where is generate

[cyan shoal]

@minor fractal

[jovial tiger]

good to know whatever we get will be the lower quality. 🙂

but i guess that's the nature of it. when we have it downloaded then we can go nuts on it with comfy

[cyan shoal]

exactly

I bet the magic workflow behind this is literally just highresfix (upscale -> img2img)

[jovial tiger]

clownscheduler™ hah

probably. I thought I heard him say at one point that with sdxl there's lots of samplers. with sd3 there's mainly just 1.

[cyan shoal]

hmm

[jovial tiger]

but i don't know how true that is.

[cyan shoal]

hmmm

but whatever the workflow is, it must have highresfix in it

[jovial tiger]

pixart sigma

[minor fractal]

hmm, that'd explain it

[crisp stream]

@nimble mason

[crisp stream]

Looks more like DJ Harry 😄

[nimble mason]

holy shit that's awesome

[crisp stream]

Thank you, thought you might like it 🙂

[nimble mason]

wonderful

[crisp stream]

😄

[grave scarab]

Guys, if I drag one of your pictures to stable diffusion, can I see the prompts that were used for that image ?

I remember someone mentioning this kind of thing in one of the guides

[signal wasp]

Cómo generó imágenes ???

En qué canal ???

[grave scarab]

But I can't remember

[nimble mason]

with comfyui yes

click open in browser and drag the png in

[grave scarab]

This is not like mid journey, you have to install on your machine

[grave scarab]

Thank you

[grave scarab]

is this an extension ?

[nimble mason]

no lol

it's way more powerful than a1111

[covert pagoda]

Henry Swinton

I hope someone gets that.

[grave scarab]

im not learning anything after all

jesus christ this is hard

[nocturne oak]

Why are you trying to generate such a huge image with a SD1.5 model?

[grave scarab]

no idea

i'm just trying to generate 1 good image, can't get it right

[nocturne oak]

SD1.5 models are trained on 512x512 images; some can go up to 512x768 or even 768x768 on occasion.

You're getting shit because of that.

[grave scarab]

so i should reduce that to 512

[nocturne oak]

Yeah.

Try it.

[grave scarab]

i give up

fuck that shit

[nocturne oak]

Do you always give up so easily on things?

[grave scarab]

trying for a week now

and i have till tomorrow only

[nocturne oak]

Post your settings, show another example.

[grave scarab]

i need to replicate this, let me show u

into this

need to do that exact same thing for 3 other images

[nocturne oak]

Are you saying you need to add the stuff that is on the left & right sides of the first image?

[grave scarab]

nono, just enhance the character, add all the details on the face mask, details on the clothing

basically just upscale and make it super detailed like that 2nd image

the background doesn't really matter

[nocturne oak]

So, you are looking for something called "generative upscaling". I don't use Automatic1111 much at all anymore, but I'm sure there are tons of videos on YouTube for A1111 generative upscaling. There's probably also a bunch of people here that use it to do that. The thing is, if you haven't gotten anywhere in a week and need it by tomorrow, you're probably not in the best position right now.

[grave scarab]

generative upscaling, okay thank you ! at least that's a direction i can follow to try and get what i need

[nocturne oak]

Best of luck.

[nimble mason]

is this kinda what you have in mind

[grave scarab]

no ?

[nimble mason]

what's wrong with that one

[nimble mason]

here lemme upscale this one for ya

[grave scarab]

you don't have to waste your time buddy

i'm trying to learn

[nimble mason]

here ya go

[grave scarab]

[deft bison]

a place that is not easy for these poor old people! ... 😀👍

[nimble mason]

is this better? i tried to keep the composition more like the original this time

[grave scarab]

sure, thanks

[nimble mason]

got any others for me to upscale

[hoary hearth]

Can you help me with they next prom: "A creature from ancient mythology, its body is a fusion of the majesty of a dragon and a serpent, adorned with delicate, large butterfly wings, wearing medieval attire. Its body is adorned with resplendent scales, and its eyes are large and hypnotic. The iridescent and ethereal butterfly wings gracefully unfold in the wind, carrying with them an aura of mystery and charm. Its gaze is penetrating, a mixture of ancient wisdom and indomitable ferocity. Its hair is of a fantastical, long color."

[arctic laurel]

a w h o w s w e e t

[south temple]

lol

[nimble mason]

1000x1000? weird resolution

[nimble mason]

you wanna set it at 1024x1024 just fyi

better trained on that resolution

[south temple]

ok ok ok

better?

[nimble mason]

workflow embedded

[south temple]

well huh

lol

[arctic laurel]

certain brands of frozen pretzeldogs are capable of asexual reproduction

[south temple]

ok, ill play

[arctic laurel]

Meow Wolf would be all over this stuff for an art installation

[south temple]

this one... might be too much for even me

well not so bad

[arctic laurel]

looks like…

someone vomited their pandan milkshake

[south temple]

I like that

[south temple]

nifty there

[unique condor]

The brother of the rabbit from Monty Python's: The Holy Grail. Dude's jacked... Jokes aside, I like that hybrid of a rabbit and a panther

[south temple]

its about to get wierd

sorry

[south temple]

this was worth staying on this lol

getting there

[nimble mason]

that one is awesome

[south temple]

give me 4 pics

the first... sets the tone

[south temple]

next

[languid pebble]

Neelix Sister always says he looks like Harry. That's why I prompt him. Good eyes!

[languid pebble]

Looks like my A.I. doesn't know GLaDOS ...

[south temple]

here we go

ok

sick

lol

this thing is gonna be stupid with nature

I dont think I want to fiddle with anything

this might be perfect

whoooops

lol

I mean.... common.

lol

[nimble mason]

very cool 🙂

[hazy warren]

Not enough aliens!

[south temple]

agreed

and forrest mushroom people

ON IT!

[unkempt dock]

993 911 Speedster with wide body driving fast on the street, Car is blue and has HRE wheels. Location Times Square New York in the background, dramatic lighting , rain is falling and lightning , --style raw --ar 9:7 --v 6.0

[whole elm]

生存一张星空的照片

[languid pebble]

It looks like infinity with selfcreating holes ...

[whole elm]

993 911 Speedster with wide body driving fast on the street, Car is blue and has HRE wheels. Location Times Square New York in the background, dramatic lighting , rain is falling and lightning , --style raw --ar 9:7 --v 6.0

[short plaza]

Design an GPU with 'ionet' text

[wicked charm]

lovely ,3d,ai robot.

[humble python]

hello, does anyone know how to generate different emotional face angles from a base image?

in batches

around 100

example:

[split ginkgo]

Can someone help me with a prompt template for garment generation with specific fixed attributes required in Fashion Garment industry.

[weary light]

[wispy nest]

what kind of specified attributes? youre best bet is probably going to be some regional prompting and inpaining

[wispy nest]

wildcards/lists (few different ways to go about this) with different expressions (or run images like that through lava or clip interrogate or something for the expression prompt), do the same thing for pose/angle. probably is a way to tie in control net open pose or depth or similar with this, but im not sure how. add xyz plot, columns for angles and rows for expressions.

or vice verse

8kx6k. absolute overkill upscale 🙂

[wispy nest]

so many weird artifacts when zoomed in

[shut sinew]

Get a lora of said person and generate a list of expressions with chat GPT, then feed the list into your prompt as dynamic prompt and let it generate a grid

[wispy nest]

yeah thats probably way easier than the way i said!

[shut sinew]

Look at the eyes lmao

[shut sinew]

You could even fix the seed, that way only the expression should change and the image should almost stay the same

[wispy nest]

i saw that, i got one even better, his bowtie

[shut sinew]

Hairy

[languid pebble]

Hairy eyes ... and I thought the song was about hungry eyes ...

[wispy nest]

im dying 🤣 its so detailed but so wrong

[shadow cove]

/help

[wispy nest]

sighs i guess its time to ditch tilingfor the 8k pictures. its always hair , faces, or eyes all over the place

[humble python]

do you know any cloud platform with these available tools where it can make the process faster?

[shadow cove]

A painting depicting two stylized figures in an outdoor setting, suggestive of a street or café. The figure on the left has an elongated body, dressed in a dark suit and wearing a fedora. The figure on the right, also with an elongated body, wears a flowing dress and a wide-brimmed hat with a red ribbon. Render the scene in a style that mimics a detailed mosaic or stained glass, with vibrant colors and intricate tile work, offering a stunning visual complexity and transforming the artwork into a visually rich composition.

[shut sinew]

#1047610792226340935

[shut sinew]

Im not into cloud services sorry, I was throwing ideas for local installs

[shut sinew]

You could try lower denoise, not sure what you used to upscale

[wispy nest]

theres some websites that let you use loras, but your control is going to be limited. as for the prompt, just use chatgpt i guess. im pretty sure there is a free version

[wispy nest]

yeah i might lower it a bit, but it was pretty low already. im about positive that tiling VAE is responsible for the artifacts. it was supir

[humble python]

which model would be good with lora?

[wispy nest]

its a delicate balance with noise on there when going that high up in res. need a LITTLE bit to get details, but too much and you get luscious eye lashes and hairy bowties

[shut sinew]

Depends on what your character is, anime needs anime model and realistic stuff well...

[shut sinew]

Which Supir workflow are you using

[humble python]

for now anime and realistic

[wispy nest]

standalone local gradio app. i havent run it on comfy backend yet

[shut sinew]

Ah was about to say

[shut sinew]

Then Id check civitai for the top models respectively

[humble python]

yeah i am currently checking it there

[wispy nest]

have you run on comfy? the thing im skeptical about is that it will increase gpu usage, and with 20gb vram and fp16 settings, 115 steps, 8kx6k, and llava, that took nearly an hour. im not sure what benefit would be had by running in comfy

[shut sinew]

Beyond 5k you seem to need FP8

Im using a 4090

[wispy nest]

and that was with tiling vae 😮

[humble python]

thanks guys for the help

[shut sinew]

115 steps is too much imo

Like way too much

[wispy nest]

id agree normally, but i was aiming for ridiculously overdone detail. just, not necessarily with the oddities i got

[shut sinew]

I doubt that anything beyond 30-40 changes a lot besides render time

[wispy nest]

ill try lower steps, lower denoise, and fp8. is odd though, i normally aim for around 100 when inpainting

[shut sinew]

Thats the comfy workflow I found yesterday, not mine

[wispy nest]

although thats usually around 1024

nice, new toy 😄

[shut sinew]

[wispy nest]

the weird thing here, and one of the reasons i used slightly more than normal denoise, is around the mouth where i did some editing in gimp. had assumed it would get fixed. so, somehow hair gets added to eyeballs but a little blurriness goes untouched

amazing

[split ginkgo]

Attributes like - Create a female model of height 5'8 is wearing a size S Top of - Fabric/Material: White Viscose Rayon, Pattern/Print: Monstera leaf print, Leaf Print Motif size : 2 inch, Leaf Print color: Teal #008794 and #90E4C1 on white color base fabric, Cut: Shirt Style, Fit: Regular, Occasion: Casual, Weather: Summer, Length: Regular, Neckline: Mandarin Collar, Sleeves: Three-Quarter Sleeves, Hemline: Curved, Closures: Opaque Buttons, Background Color Palette: White Studio Background, Lighting: soft diffused light, Composition: symmetrical balance, Perspective: Two point, Focus: shallow depth of field, Lighting: Studio white, Color Palette: Bright Catalogue Studio, Activity: Model Shoot, Camera Techniques: ISO: 100 Aperture: f/1.8 or f/2.8 Shutter Speed: 1/125th

[wispy nest]

not sure you will be able to get that specific with prompting alone.

[split ginkgo]

Can you suggest some alternatives?

[wispy nest]

studio photoshoot clothing model advertisement, aim for getting it close with prompt & controlnet for initial generation (pos prompt:female mode,slim,(try diff words for specifying size like 2 inch, not sure if it will work) teal leaf prints on white shirt, casual summerwear, white background,soft diffused white light( you can specify lighting direction around here), shallow depth of field, (you can try adding your camera specs in too but its not going to be 1:1 with a real photo

controlnet for the pose on that first gen

inpaint the finer details in like collar, button opaqueness, hemline etc

[haughty thicket]

can I turn an image like this into (for instance) a painting by van Gogh, using img2img and a LORA?

[wispy nest]

you can try specifying that stuff in prompt but i doubt youre going to get every single detail how you want just doing that, hence the recommendation for inpaint. if youre struggling with that part, try editing in somerthing like gimp. just basic stuff, like if youre changing the neckline you could use the heal or clone tool and make a very rough neck line fix, and then run it through img2img with low denoise or inpaint with same

[wispy nest]

1.5 model, controlnet inpaint , independent image upload. clip interrogate your pic and use it as prompt. delete anything from interrogate related to detail, photo, realistic, etc. add in oil, painting, artwork, (in the style of van gogh:1.(whatever))

inpaint and controlnet are always the answer

[haughty thicket]

can I use SDXL?

[junior sky]

You can use the COSXL edit model to prompt a style

[haughty thicket]

thanks

[junior sky]

But that requires a special workflow.

[nimble mason]

generally the case, but that's different with res_momentumized

highest quality images i've made yet have been about 400 steps (200 for initial generation, 200 for upscale) using this noise schedule via img2img

[shut sinew]

Were talking about SUPIR tho

[nimble mason]

oh oops

still working on morning coffee here 😛

[chilly reef]

Can anyone create an image with this prompts

Gamer playing Water game,hands over the white table, Realistic Water Wave with surfer coming out of gaming mobile phone, Animation,Ultra realistic, Horizontal mobile, both hands

I want a more good image can anyone 🙂?

[vapid sundial]

"I would like to create a poster for Wahaha AD Calcium Milk. The poster should have a vibrant and attractive design that catches the eye.
At the center of the poster, I would like to see a large, prominent image of a refreshing bottle of Wahaha AD Calcium Milk. The bottle should be positioned in a way that highlights its shape and design, making it the focal point of the poster.
Surrounding the bottle, I would like to see a variety of bright and colorful elements that complement the overall design. These could include splashes of vibrant colors, abstract shapes, or even illustrations of happy children enjoying Wahaha AD Calcium Milk.
In terms of text, I would like to see the brand name "Wahaha" prominently displayed at the top of the poster, with the product name "AD Calcium Milk" below it. The font should be clear and easy to read, and the colors should contrast well with the background to ensure maximum visibility.
Finally, I would like the overall tone of the poster to be upbeat and cheerful, reflecting the refreshing and enjoyable nature of Wahaha AD Calcium Milk."

[clever oar]

here your image

your muhahaha valhala milk

[clever oar]

.

[chilly reef]

Yeah but where is mobile 🙂?

[clever oar]

ai ignore promt)

[hoary hearth]

Can anyone create an image with this prompt female creature from ancient mythology, her body is a fusion between a dragon and a snake, adorned with large and delicate butterfly wings, dressed in medieval attire. Her body is adorned with resplendent scales and her eyes are large and hypnotic. The Iridescent and ethereal butterfly wings unfold gracefully in the wind, carrying with them an aura of mystery and charm. Its gaze is penetrating, a mix of ancient wisdom and indomitable ferocity. It is found in a fantastic forest with exotic flowers

[chilly reef]

Okay thank for your try 🙂

[chilly reef]

Leave that I have generated with Leonardo

[atomic plinth]

hello, can you help me with the next promt: full body photograph, heroic pose, mermaid woman, blue scales, with crimson red dragon wings, fully extended, slender body, white skin, short black hair, round face, small thick lips, large black eyes, slender hands. in 2:3.

[covert pagoda]

No.

[nimble mason]

I've noticed lots of ppl are coming in with that same exact quote, but with different prompts

Starting to wonder if it's bots generating shit for some paid service, idk... Would be pretty funny though considering what they tend to get from us

[languid pebble]

It's like they come with Midjourney prompts asking us to change them to SD prompts ...

[clever oar]

give them crow sharing bathing 😆

[bleak nimbus]

Hello, can you help me with the next promt: Mythological creature that combines characteristics of a dragon and a crab. I want it to be giant, with shiny scales and sharp wings. The creature should be in a pose that conveys aggressiveness. Use mostly purple and yellow. The environment in which this creature is found is one of war--ar 3:2.

[rapid ibex]

Hello! Could you help me with this promopt: Photograph of a Cyber Tiger, white tiger, white armor, metallic claws, white fur, black stripes, neon lights, unicorn horn, mystical aura, intense and penetrating eyes, reindeer antlers.

[pastel root]

i need some bokoys boxers

[languid pebble]

First time SUPIR worked here

TBH it was a test how to create a guy with shorts hanging at his knees. Tried that cause I wanted him to sit on a toilet ^^

[patent trail]

hello, nice to meet you. I read your post that are looking for gradio dev. I have rich experience that have built gradio app. I want to work with you.

[clever oar]

here your image what you requested

[nimble mason]

Here is the image you requested.

[junior sky]

Here is the image you requested

[clever oar]

japan animation

[crisp stream]

had to look it up, not seen then movie and weeeeeell... 😄

[covert pagoda]

they couldve done a better job casting

[nimble mason]

Here is the image you requested.

[meager kiln]

hello, can you help me with the next promt: female beast, she is a fusion between a dragon and a snake, she has large and delicate butterfly wings, she has a floral outfit. It has glowing scales and its eyes are large and hypnotic, along with long glowing hair. Butterfly wings are iridescent and ethereal. It is located in a fantastic forest, the color palette is fantasy, full of exotic flowers.

[clever oar]

why all say help me with promt

is bots

[languid pebble]

Maybe we should stop kidding around and not answer ... I don't really feel comfortable with it ... but we can't help them all ...

[shy eagle]

Imagine it is question bot and anwser bot

[clever oar]

clownashark can help all)

[clever oar]

you bot?

[nimble mason]

Here is the image you requested.

[nimble mason]

Here is the image you requested.

[clever oar]

its bots for train midjourney 😄

[pastel root]

No NSFW

[clever oar]

iron man

[clever oar]

extra leg

[pastel root]

he is spiderman

[languid pebble]

Long Dingdong ...

[cyan shoal]

@jovial tiger stable diffusion 3

[clever oar]

fell asleep in discord on the keyboard when I woke up it froze due to the fact that 8000 characters were entered there

[nimble mason]

you should see what happens when you use it as a prompt

[jovial tiger]

you got it yet? 🙂

@cyan shoal @nimble mason so I've been playing with trying to get better quality and faster inference with pixart sigma. I may have stumbled upon something that's fast AND good.

that image is raw, with no upsampling or sdxl refiner.

[nimble mason]

oh nice

yeah that looks a lot better

[cyan shoal]

naaahh

I would tell you guys

if I were allowed to tell you guys

but if I would be then absolutely

if stability allowed me I would also do one of these "I got access, give me prompts"

[cyan shoal]

interesting, and it's ddim too

like generic ddim so its also faster than res momentumized

any cfg rescaling?

[jovial tiger]

yeah, but it's got its own scheduler. batwing lives for schedulers.

[cyan shoal]

ooh

I only mainly use exponential and karras

interesting

[jovial tiger]

I'm doing lots of upscaling etc afterwards, but I was just trying to get stage 1 down.

[jovial tiger]

i still need to do prompt following tests first.

[cyan shoal]

nice

[nimble mason]

a few interesting noise schedules

[jovial tiger]

yeah ok fine, closeups look amazing with that ddim. arms look all messed up. sigh. can't win with this mdoel.

model.

[south temple]

A priest, he is sticking his fingers into a light socket, electricity is blowing out his face, he likes tree frogs and the Bible, pancakes

[nimble mason]

maybe the first chunk of denoising with another sampler, then switch?

[south temple]

That man is electric

[jovial tiger]

you didn't specify if he was a man

[south temple]

A duck, walking up to a lemonaid stand, he says to the man, HEY, got any grapes? 4k, cinema, bum bum bum

[cyan shoal]

epic

yeah no text for pixart sigma and SD1.5 ELLA

somewhat surprising from pixart sigma, but its not like they promised text at all

[south temple]

I understood just fine

[cyan shoal]

but the rest of the prompt was nicely followed

[south temple]

Then he waddled away.

[jovial tiger]

it's a 3 gig model.

can't expect much.

maybe with the bigger versions

[cyan shoal]

0.6B parameters

smaller than the smallest SD3 model

[jovial tiger]

there we go. finally a lemonade vendor

[cyan shoal]

I was about to compared this to muse but holy hell, Muse is 3 Billion paremeters, close to SDXL

[south temple]

World war 3, ducks are fighting cats, the cats are taking over the litter box, the ducks have balloons, the fat kid needs to poop, it’s a sunny day, plenty of mushrooms, who’s hat is this?

[cyan shoal]

and bigger than SD3-2B

[south temple]

This is how it starts

[jovial tiger]

ww3 was always gonna start with balloons.

it seems to think ww3 equals a kid in a old timey fisherman's outfit.

[south temple]

Ducks fighting baby kittens, the ducks are holding balloons and the kittens all have catnip, giant smiling faces with moody peanut shaped clouds, 8k cinema 35mm film, Hot and scratchy

[south temple]

Well hmmm

[south temple]

lol

On the mark

[jovial tiger]

"the aristocrats!"

[south temple]

It is the world record dart throwing competition, the crowd is massive and they are all silent waiting for the award winning dart to be thrown, a kid farts very loud as the dart is being released, the man pees out cotton candy for the kid, everyone is please with the gas, unicorn

Please indeed

[jovial tiger]

I think i broke it

[shell sleet]

that is horrifying

that is oddly horrifying

[nimble mason]

that is awesome

[south temple]

A Kitten is addicted to catnip, the dark army descends on the streets of New York lighting up all they see, cinematic, close up, f/1.8 aperture, mirrorless, iso 400, slight color grade, skeletons of the dead unicorns

[south temple]

This is how you win awards

[jovial tiger]

I think that's too random. doesn't give me much more than a cat

[south temple]

Its the close up stuff I bet

[jovial tiger]

@nimble mason this has been supir'ed

pixart sigma actually has tiny subjects, like REALLY tiny ones that still have clear details unlike regular sdxl. so supir actually makes sense for this.

[nimble mason]

interesting

this is a pixart sigma one and a 4x upscale

it's almost great but the structure needs some clean up

[jovial tiger]

wow that's really neat. what aspect ratio is that?

i'll do this one in that.

[nimble mason]

comfyui could really use an image browser...

[nimble mason]

that's 3.61 i think

close to 3.56 which is my monitor (5120x1440)

a really aggressive denoise with res

[jovial tiger]

that's really good

[nimble mason]

looks very good, just... some messed up shit and lost some control over the composition

i was using ipa tiled with composition, high strength

yeah, looks really nice

i'll try the clownsched next

200 steps, ready or not here we come lol

[jovial tiger]

wow.. these really wide aspect ratios are a way to cheat more resolution out of this thing

hah

[nimble mason]

yeah

if we can come up with a way to stabilize the composition and upscale these things... it'd be pretty great

pixart's the only model i've seen that generates >3 AR images without repeating elements

[jovial tiger]

supir

[nimble mason]

awesome

another thing i'm wondering about here... maybe a single pass of a very light clownsched on these would be the cherry on top

at the same time though it might fight the original comp

i've found when it wants to go a slightly different direction, you end up with a fuzzy image

not badly blurred, but a subtle effect, looks almost like low cfg

4.5-5 or so

heavy clownsched

oops, i added noise

[jovial tiger]

so is that pixart or just regular sdxl?

[nimble mason]

this is sdxl

pixart seems to not handle upscaling well at all

tiled ksampler... only did one test, but eww

got this

[jovial tiger]

right

[nimble mason]

well, stabilized the composition, but lost shit in the background

[jovial tiger]

maybe see if pixart can do it. depth controlnet etc so you don't lose it.

[nimble mason]

Yeah, I forgot about that, doh... Should be using de.th here for sure

[jovial tiger]

I found it much better than canny

left more open for sdxl to do its thing

[nimble mason]

Yeah everything I've tried to do with pixart for upscaling has been awful

[jovial tiger]

without exceeding the bounds.

[nimble mason]

Yeah for sure same experience

The exact depth model is huge too

Haven't tried many

But I know I've noticed that

[jovial tiger]

want my flow?

[nimble mason]

Sure

[jovial tiger]

should be in there.

of note, there's 2 positive text boxes.

if you know of a way to have the text from one auto copy to the other, let me know.

[nimble mason]

There is I'll show ya in a min here

[jovial tiger]

i have 2 because they each have their own source of clip or t5 in this case, so can't usually share

[nimble mason]

oooo this turned out pretty nicely

[nimble mason]

you want the node pack comfyroll studio to have that same node

but just right click on the text encode nodes, convert text to input

[jovial tiger]

ok cool, thanks

[nimble mason]

np

[jovial tiger]

ok, so i have a "positive" node as part of the fooooooooocus nodes. so that coupled with the convert text to input should do it

[nimble mason]

yeah anything that spits out text/prompt types will work

also, btw qrcodemonster can also be helpful for stabilizing a composition

as can the optical illusion one

[nimble mason]

that's awesome

[jovial tiger]

a menacing squid, positioned as a battlefield commander, with its tentacles gripping a large cannon that shoots rockets. Its eyes gleam with a mischievous laugh, set against a dark, smoky background.

expanded of course.

as expected, i can't quite get teh cannon to fire anything

[nimble mason]

diffusion_pytorch_model.safetensors: 9%|███▊ | 31.5M/335M [00:26<04:01, 1.25MB/s]
diffusion_pytorch_model.safetensors: 0%| | 0.00/3.46G [00:00<?, ?B/s]
model.safetensors: 5%|███▏ | 73.4M/1.36G [00:27<07:20, 2.93MB/s]
diffusion_pytorch_model.safetensors: 2%|▉ | 83.9M/3.46G [00:27<16:10, 3.48MB/s]

i swear, i want to throttle any dev that does this shit

matteo is great cuz he refuses to do it

now i'm closing comfy and restarting

[jovial tiger]

doesn't rename their stuff from the default? yeeeeeaaahhh

[nimble mason]

that, and auto downloads

my connection isn't great

so right there that means comfy is outta action for 30 minutes minimum

i'd like to actually be able to download the shit in the background via my browser, instead of having comfy locked up that whole time

PixArt: Not using xformers!
Expect images to be non-deterministic!
Batch sizes > 1 are most likely broken

just noticed that

and i just remembered reading that with pixart alpha, image quality with xformers was supposedly significantly better

looking pretty good now

[nimble mason]

unsampling version

interesting... it's less faithful than simple gaussian noising... prolly cuz res does so much crazy shit with each step

[jovial tiger]

yeah but definitely filled in

[nimble mason]

which looks better to your less biased eye?

out of the last two

also, do you have xformers running with your comfy env?

[jovial tiger]

are frogs supposed to have 3 or 4 toes on their back feet.

because other than a little bit of leopard spot difference, that's pretty much the only change.

[nimble mason]

no controlnet, noise only, composition strength = 2

[jovial tiger]

hah at the graffiti'ed island house

[terse nacelle]

@bright harbor i dont see the thing for embedding

[clever oar]

you know what is it?)

[nimble mason]

idk

[clever oar]

lord of necrons

warhammer 40 k

[nimble mason]

ahh cool

[clever oar]

😃

[keen gate]

@nimble mason A bottle of white wine is placed on the table. It may be a clear and translucent white wine, such as wine or other fine wines. This drink is often associated with relaxation, celebration, or socializing. The wine bottles are placed on the table, which may be made of wood or other materials, giving a sense of stability and neatness. Through the window, you can see the outdoor scenery, which is full of spring and vitality. The windows may be glass, clearly reflecting the view outside. The scenery outside your window might include green leaves, blooming flowers, a sunny sky, and perhaps fluttering butterflies or busy birds. These elements convey the warmth and vitality of spring. The quietness of the room and the white wine on the table contrast with the vibrant scenery outside the window, showing the beauty of the changing seasons. Create a 9:16 poster

[clever oar]

lol

promt like book for clownshark

imagine result

he buy happy meal

[nimble mason]

Here is the image you requested.

[clever oar]

as always there is a small nuance... 😄

[wispy nest]

whats does res mean here? i have an idea from context but not sure 😄

[hazy warren]

@tired basin

[wispy nest]

i hate when my arms turn into ps1 graphics. always at the worst times

[tired basin]

what in god's name happened there?!

[hazy warren]

dunno, i put in your settings XD

[tired basin]

comfy or forge?

or a1111?

[hazy warren]

forge

[tired basin]

freeU 1.1, 1.1, 0.9, 1.1 then

this is the sort of thing I was getting yesterday with it

[tired basin]

cfg 3

6 way too high for realism, and doubly so with PAG on

[tired basin]

a bit grainy and theres consistency issues with the pole and the hand, but otherwise looking pretty good

[hazy warren]

this is without prompts

[tired basin]

dang.. not shabby

the meerkat was Meerkat, fish-eye lens, macro photo with a wide angle lens, animal close up with wide field of view lens

swap meerkat for any baby animal and it works awesome, no neg prompt req

[hazy warren]

hey a girl popped up this time

[tired basin]

yeah, I try to keep the dataset balanced so it doesn't skew either way, and definitely doesn't skew nsfw like some do

[tired basin]

I'll let you be the judge, but you can see why I threw it into the mix!

[hazy warren]

lol, I typed in Pepe

[tired basin]

haha.. that's not a frog, nor a frenchman

[hazy warren]

im having touble getting a unicorn, its just doing fish

[tired basin]

da fuq

[marble brook]

https://tenor.com/bBmA5.gif

[hazy warren]

LOL

[south temple]

🙂

[tired basin]

I mean that is cool, but it is not a unicorn

[hazy warren]

ye, im having trouble getting the unicorn

[hazy warren]

oh ok, so I had to add the word rainbow to get it

[south temple]

we need 4 more of those

[south temple]

last one

[quaint quarry]

How would I get the ai to make pixel art simliar to this style? I've already done the pixel art stuff I just need help getting it to follow that pictures theme

[south temple]

running

godspeed

[hazy warren]

oooo, thats a good image to make

[south temple]

well this is atreu acid

[hazy warren]

OMG LOL

[south temple]

and its upscaling this?

oof

[hazy warren]

wow, godspeed is interesting

[hazy warren]

idk, I never ticked upscale

[south temple]

I love upscaling

lol

great detail

this is a test

[echo skiff]

draw a picture with cute penguin in a egg

[south temple]

heeeere we go

dont poop...

[hazy warren]

heres an upscale one

[echo skiff]

which channel we should try

[south temple]

HOOOOOO

[hazy warren]

not a unicorn

[south temple]

nope.

correct

lemmmmmmmme try somthing

[south temple]

still a magestic my lil pony

ding

you get what you get sir.

[stiff yew]

Nature

[bronze oar]

/drea

[cyan shoal]

@minor fractal @jovial tiger

[languid pebble]

At least the arse can fly 😄

[cyan shoal]

So we'll get increasingly better base models maybe

But at least SD 3.0 won't be delayed

But it won't be a permanent problem if it ends up being rather shoddy quality

[wispy nest]

what model are these from? looks nice

[languid pebble]

😛

[clever oar]

😃

[languid pebble]

Happy bee day! 🙂

[languid pebble]

I think you missunderstood the concept! Just kidding! 😂

[deft bison]

how can you not understand in the middle of this universe of varieties! . we tend more or less towards the extreme but we always fall towards points of balance in order to regain momentum...

[languid pebble]

Let's hope for a balance ...

[languid pebble]

He will never run out of honey!

[clever oar]

he dont want they friends)

[languid pebble]

We've had bees at school ... I know how to handle them 🙂

No need to harm them for honey ...

[clever oar]

I once caught a bee with my hands, released it and it stung a boy in the face 😄

but this summer karma overtook me, I crashed into a flying wasp with my lips....

[languid pebble]

Awww... I'd prefer a bee ^^

[clever oar]

for you

[languid pebble]

Old creation for a competition. Theme was: Spring

[clever oar]

not ideal leg proportion

right cat

[languid pebble]

Yes ... didn't win with this one ... but it's one of these I still remember 🙂

[languid pebble]

Dancing bear 🙂

[clever oar]

is not bees 😲

[languid pebble]

Alien bees ... sounds like prompt meaterial 🙂

[languid pebble]

Doesn't look healthy 🙂

[sterile kiln]

Hmm summer, cockroaches ...

[languid pebble]

Some eat them ...

[sterile kiln]

shrimps and cockroaches are in the same family, so we've all eaten them.

[languid pebble]

Why not ... it's more a brain thingy ....

[languid pebble]

with filled pockets? 🙂

[clever oar]

yep 😃

[languid pebble]

Honey!

[clever oar]

i understand now what wrong

in promt was beear )

[languid pebble]

🙂

It's not homeless ...

... whatever it is ^^

[jovial tiger]

After lykon's tweet yesterday mentioning the changes, my first reaction was that it's going to be months before we can download this thing.

[languid pebble]

I want summer!!! 😛

[forest olive]

Amy is a beautiful girl with braided light brown hair and wearing a light blue dress and A wise, old kindly woman handing a red silk ribbon to Amy beneath the Tree of Dreams

[languid pebble]

We all would want to meet Amy ^^

[forest olive]

#🏞｜general-with-images Amy is a beautiful girl with braided light brown hair and wearing a light blue dress and A wise, old kindly woman handing a red silk ribbon to Amy beneath the Tree of Dreams

[junior sky]

No she isn't.

[clever oar]

need remember winter...

[languid pebble]

I'll tell my bees to attack winter!

[forest olive]

Amy is a beautiful girl with braided light brown hair and wearing a light blue dress and A wise, old kindly woman handing a red silk ribbon to Amy beneath the Tree of Dreams#🏞｜general-with-images

[languid pebble]

Bots offline here

[valid grotto]

Xi Jinpingprison

[wanton linden]

#🏞｜general-with-images message

[cyan shoal]

nah its like 2-3 weeks probably, if API stuff won't come this week then 3-4 weeks lmao

I just don't know how long these RLHF stuff last

but no longer than 1 month tbh

after 3.0 releases: "I can't wait for 3.1"

[jovial tiger]

well, from the demo video, it does text really well, about 50% to 25% of the time, so a lot of generations are needed. same for hands, same for objects that people interact with. so maybe some of that is in the architecture change side of things.

[cyan shoal]

I think 3.0 coming out in 2-3 weeks should happen as we are dying for a new model

and if we really are getting further trained or modified models then there's no reason to say "but this is the last model and they rushed it, what will we do now"

unlike games, a rushed model is not forever bad, cause we can finetune easily ourselves

and from lykon's images I think as a base model it looks good already

[jovial tiger]

so sd3 is going to segment the userbase. that's always bad. whatever architectural changes that come later, they can't make people have to retrain their checkpoints because that'll just segment the community even more.

[cyan shoal]

^

well yeah different model sizes

[jovial tiger]

and then there's that

[cyan shoal]

800M for SD1.5 users
2B for SDXL users
8B for enthusiasts

2B already looks good and iirc it's almost finished training too

with DPO and stuff

[jovial tiger]

which personally, I think will be the worst part. lykon mentioned you need more than a 4090 to train the biggest version of the model....which means very few will have checkpoints for the biggest one. sad face.

[cyan shoal]

well if you mean dreambooth stuff then we'll still be getting massive finetunes just like before, just maybe slower(?)

but loras will sadly fade away for 8B

[jovial tiger]

yeah

[cyan shoal]

2B and 800M will keep going

Loras, dreambooth, enough vram to train those of course

and 6B is like, the black sheep tbh

like idk what to say about that one

[jovial tiger]

probably targeted at all the major consumer cards

to match each one

[cyan shoal]

but at that point just use 8B since that's their flagship model, its guaranteed to be better finetuned and trained and everything

it is guaranteed to run on 16GB and 24GB and with T5 quantization around 12GB

6B is also 512px iirc

not like its a problem cause it has the amazing 16 channel VAE (thats why 2B also looks good at 512)

[jovial tiger]

ewww 512

[cyan shoal]

no its good at 512 no joke

[jovial tiger]

I've been fooling around with LLMs a lot, and cutting a language model in half has a big impact. many will say 'minimal quality loss' but that's garbage. the reasoning ability and understand what you wanted is severely impacted through all of my testing.

[cyan shoal]

have you seen models at 4-bit quant using stuff like gguf and exl2?

we're not talking about HQQ 2-bit or whatever

[jovial tiger]

I'm just talking about going from fp16's to fp8's.

[cyan shoal]

ew fp8

[jovial tiger]

something can understand my instructions at fp16 that just flat out can't at fp8.

[cyan shoal]

im literally using int8 with pixart sigma T5 and it works perfectly

[jovial tiger]

and most people say fp8's for llms have hardly any loss.

[cyan shoal]

who the hell uses fp8 for llms

I literally never see people do that

[jovial tiger]

I do all the time. 🙂

[cyan shoal]

only for GGUF they use K quants

4-bit, 5-bit, 6-bit

[jovial tiger]

if you use ollama, 8's and often fp4's are the default

[cyan shoal]

3-bit if you are desperate

bruh

thats weird

[jovial tiger]

The normal mistral 7b is 14-15 gigs. the default ollama run mistral7b is 4 gigs.

I've been running the dolphin-mixtral q8 for a long time now. 46 gigs. fp16 is 96, so I can't fit that. the difference in reasoning is huge between the 2. I've been running against the full size via api and it can easily handle stuff the 46 gig can't.

[cyan shoal]

I just use koboldcpp which has overpowered features such as min_p and dynamic temperature

and I mostly use Q4_K_M and Q5_K_M

[crisp stream]

Nerds 😉 😄

[shut sinew]

Text is really good from what ive tried yeah

But i dont know if everyone will be able to run the big version with T5

[cyan shoal]

@jovial tiger

https://stability.ai/news/stable-diffusion-3-api

its getting CLOSE

API IS RELEASED

https://tenor.com/view/sleep-time-gif-2822468507363798988

[minor fractal]

id believe that

oh neat api

[cyan shoal]

especially now that the API is out

[minor fractal]

(t5 can be quantized to 4bit)

[cyan shoal]

maybe 2-3 weeks isnt so unrealistic

especially with those sample images

I understand cherry picking but you cannot cherry pick quality, only how much it really adhered to the prompt

you cannot suddenly have something look like 320px big and then one that is 4K

I just dont get how the other stable assistant images look like if they were from 800M or something

[crisp stream]

😄