🆕｜sd3

[dry wave]

you can always merge a lora with a checkpoint

[vast condor]

and usually the reverse also

extract a lora

[dry wave]

that's also possible, yes

[sage burrow]

runs off to merge my own lora with Flux dev

[sage burrow]

oh?Q

[sage burrow]

How do you specify which aspects you are extracting?

[vast condor]

flux is so new, I'm not just sure on where that's at...but maybe

[sage burrow]

Experimenting is fun 😄

that crazy guy on yt who merged 1000 or something loras; I would have done that out of pure curiosity for fun LOL

(the resulting images weren't that great from what I recall) lol

[vast condor]

all I've done with flux so far is play with workflows and create a few loras, so I'm confident there, the rest surely will come if it's not available yet

[sage burrow]

😄

[sacred jewel]

Damn that's hot

[sacred jewel]

Looks like Tilda Swinton

[bitter hearth]

thanks , playing with schnell nf4

experimenting with some quality prompts in the positive ofc

[rain current]

A figure with striking features. The individual has pale skin, pronounced facial bones, and closed eyes. Their hair stands upright in a wild, wavy manner, resembling flames or tendrils. Art by Jeremy Geddes

[bitter hearth]

Lmao

[sacred jewel]

#AwesomeSauce

[sage burrow]

Hmm, this isn't correct is it? The vae and clip probably need to interact before the very end?

[bitter hearth]

Flux

[sage burrow]

This reminds me of 1.5

[bitter hearth]

[bitter hearth]

Sd3 is too good

I only get 20 per day

That's it

[sage burrow]

Why only 20 per day?

[bitter hearth]

Cause daily limit

[cerulean hill]

1.5 is a fossil at this point

[sage burrow]

If you mean glif, just go to my profile, see all my glifs, hit the remix button on the ones you like. Then ask for a creator pass...
(they might go for it lol)

[bitter hearth]

[errant dust]

The VAE is last and used after the image has been processed through the main model

The clip takes place before

simplified it is clip -> model -> LoRA -> model processing -> VAE -> Image output

[mortal mesa]

This works in ComfyUI now, Long-CLIP - Long Input length Increase the maximum input length of CLIP from 77 to 248

https://github.com/beichenzbc/Long-CLIP https://huggingface.co/BeichenZhang/LongCLIP-L/tree/main

[old falcon]

Is SD3 good yet

[sacred jewel]

It was never bad... it's just that Flux is better.

to most

but in the end, it is always subjective to the end user

[noble coyote]

Which folder does it go? Embeds?

[mortal mesa]

clip

[sacred jewel]

Pirate Female Xenomorph Doctor ? O.K. 🤭

[sage burrow]

Looks like lapis

[bitter hearth]

generated with flux

FaceDetailer with SDXL

finally img2img upscaled

[errant dust]

Here's a THIRD example of ChatGPT failing on image description compared to Claude. 3/3.

ChatGPT: Black and white ink sketch in a comic book style. Panoramic desert landscape with mountains and cloudy sky. Foreground shows three men in profile: a bald muscular figure, a long-haired individual, and a robed person with covered head. Background depicts three distant mounted figures.

Claude: Black and white comic book style illustration. Expansive desert landscape with distant mountains and cloudy sky. Four muscular men in foreground wearing varied primitive clothing. One has long hair another wears a striped headdress. Three riders on horseback visible in far distance.

[rain current]

alien or Godzilla

[errant dust]

Definitely curious to see it. Wonder how the name of the founder of Rome fits in the narrative, but at least it will be entertaining

[rain current]

Florence-2: The image is a black and white illustration of a group of people standing in a desert-like landscape. The people appear to be of different ages and ethnicities, with some wearing long robes and others wearing headscarves. They are standing close together, with one person in the center of the group.\n\nOn the left side of the image, there is a man with long hair and a beard, wearing a long robe and holding a staff. He appears to be looking off into the distance with a serious expression on his face. On the right side, there are several other people, some of whom are walking towards the man with their backs to the camera. In the background, we can see a vast expanse of sand dunes and mountains, with a cloudy sky above. The overall mood of the illustration is somber and contemplative

[errant dust]

Ask it how many people are on the left side and how many on the right

[toxic bone]

spoilers would be nice not to do

[sacred jewel]

Aldzilla
Gollien

[sacred jewel]

Groq: None... there is no one name many on either side

😛

[bitter hearth]

a woman with a basket of flowers. she is wearing a white tank top, and red pants. her hair is long, blonde, flowy. she seems to be peacefully absorbed in thoughts. scenic outdoor background.

[sacred jewel]

There's a basket of flowers in the image? 👀

[bitter hearth]

yes ofc

[rain current]

the happy maiden

[bitter hearth]

wasn't getting precise results with schnell, then tried the dev for that image

[errant dust]

I asked because if Florence can tell the number, I will install it and use it instead. I am not in love with Claude, nor a subscriber, but it simply proved to be competent where ChatGPT is not

[toxic bone]

https://www.youtube.com/watch?v=p3PfKf0ndik

[rain current]

I am using it with the Gradio demo with captioning, so the chat mode is not available

[errant dust]

That's actually good enough since it clearly shows F2 distinguished the four and then the three

Ok, will install locally and see if I can get it up and running here

I know there's a Comfy setup for it

[rain current]

There is a node for Florence2, I tried it in Comfy. Now I'm using it in Spaces by Forge locally

[toxic bone]

https://github.com/kijai/ComfyUI-Florence2

[bitter hearth]

can you run florence2 with ollama?

[toxic bone]

a github discussion that came up on google says no, but i could be wrong. stuff changes all the time.

[bitter hearth]

havent found a single guide on installing florence on ollama, i took the alternative for img to text... using llava

works fairly with ollama vision comfyui node

[toxic bone]

https://tenor.com/view/bolin-lavabending-legend-of-korra-lok-korra-gif-2003126247303149610

[errant dust]

Didn't realize there were so many models of Florence 2 already

ex:

https://huggingface.co/gokaygokay/Florence-2-Flux-Large

[mortal mesa]

i liked moondream for that stuff

[sacred jewel]

I got placed in TIMEOUT... not sure what I did but hope to learn why so I can stay away from whatever I did to trigger it 😦

anyone else get a TIMEOUT from the server?

It may have been the NIP SLIP

[bitter hearth]

no time out ... odd that you got hit by it

[mortal mesa]

Moondream2: The image is a black and white illustration depicting a group of people walking across a vast, snowy landscape. The people are wearing long, flowing robes and carrying various objects, suggesting they are travelers. The landscape is dotted with mountains and trees, creating a sense of depth and vastness. The sky is filled with clouds, adding to the dramatic atmosphere of the scene. The illustration is rendered in a realistic style, with attention to detail in the people's clothing and the landscape.

[bitter hearth]

so i got this with flux

notice the facial defect

but im glad i could correct that with a very simple seperate workflow with SDXL model

reason i seperated face detailer from rendering cause i dont wanna waste time with every images, only those i think look good in posture

and the new comfyui makes it a breeze to switch between all your saved workflow templates, without losing any of the work in progress

[toxic bone]

what did comfyui change to make it easier?

[bitter hearth]

in settings

[sacred jewel]

It may have bee nthe image I posted that showed nipples even though I had SPOILER on it. My bad. If that was it

[toxic bone]

if it was just one nip slip, that's a really weird enforcement approach. time you out for a minor one time 'offense' and then not tell you about it. Am i wrong to say that seems like lazy modding?

[hexed dirge]

working on some new lors. A classic manga one is on the way

[toxic bone]

i guess i'll go look at what this whole new menu thing is

i rarely update comfy becuase it breaks so many nodes

wanna be sure the new changes are worth it before i drop bombs on my dependency environment

[bitter hearth]

the new UI update is what i enjoy comyui even more

[sacred jewel]

It's all I can think of because the image is missing... my bad... I didn;t see it as too offensive,... it was sor of a tasteful image but not my turf. I follow rules.

[bitter hearth]

there are functional improvements too i'd assume, but the UI is very sleek on itself

[toxic bone]

I'm the guy who should be getting timed out if anything. That whole situation with the "slap funk bass" guy. Shoulda caught me a suspension if they're giving you a timeout for a nip. They even allowed that guy to remain on the server and have given him warnings an shit.

very inconsistent mod decisions. It highlights laziness.

I wouldn't worry about your own behaviors. There's nothing you can do to adjust who you are if they're being this inconsistent and lazy with their decisions.

[toxic bone]

looks like a good time to update

[bitter hearth]

btw it took me a bit of self persuasion to make the shift to new UI but i figured all the menus are logical and seems easy to use

[toxic bone]

Every new workflow is a new UI. No comfy user should fear a new UI. In the past i've only ever tolerated CUI's lack of UX. this looks like a significant UX upgrade

[bitter hearth]

i just want to add... it took me about 1 year to swtich from Auto1111 to comfyui lol

but never looking back now

[oak gazelle]

Hello everyone, I'm encountering a small problem in sd3 that I can't find diffusion_pytorch_model.safetensors file anywhere

[toxic bone]

I go back and forth all the time. Comfyui is good for workflows that you can't get out of a webui.

memory issues have always been solved by adjusting a setting somewhere

[oak gazelle]

Does anyone know where I can find it or how to solve it, sd3 keeps giving me error about the file

[bitter hearth]

diffusion pytorch model doesnt say much about what you are looking for, you could ask in #🤝｜tech-support with your error log

[oak gazelle]

It only asks for the file, but I can't find it anywhere

Not even in the repos

[sage burrow]

They clearly missed all our posts when sd3 came out and everyone was trying to debate nipples lol

[errant dust]

Anyone know what node I am supposed to use to collect the Caption from Florence 2?

[sage burrow]

I'm convinced that @noble coyote is the expert at llm comfy workflows..

[sacred jewel]

Definitely the nip slip... my bad ... oopsies...

[sage burrow]

I fortunately started with comfy. Someone tried to talk me into running 111 or Forge also. It didn't end well rofl

[mortal mesa]

text node to see, or directly into clip encode

[errant dust]

yet

nyet

[bitter hearth]

if I remember rightly it just comes out the caption node as a string

[sacred jewel]

Back to LoRA testing 😄

NoveauAnime LoRA

[bitter hearth]

its a rly nice node

when I used florence 2 to copy an image it got rly close

[errant dust]

I get it, but there is no caption node

[bitter hearth]

on the node you have, the caption dot on the right

[errant dust]

yes

[bitter hearth]

string comes out there

[errant dust]

and goes....?

[bitter hearth]

what did you want to do with the string

I missed the context

[errant dust]

output the caption text

[bitter hearth]

install this node pack
https://github.com/rgthree/rgthree-comfy
and then take node called Display Any

[errant dust]

it connected and is now downloading

[errant dust]

Maybe I am doing this wrong, but the output is utter nonsense

I mean it is text and English, but makes my complaints of ChatGPT seem overdone to this Mr Magoo performance

[errant dust]

"The image is a black and white illustration of a group of people standing in a desert-like landscape. The people appear to be of different ages and ethnicities, with some wearing long robes and others wearing headscarves. They are standing in front of a large pyramid-like structure, which appears to be a temple or a temple complex.\n\nIn the center of the image, there is a woman with long hair and a crown on her head, standing in the foreground. She is looking off into the distance with a serious expression on her face. To her left, there are two men, one of whom is holding a staff and the other is looking at the woman with a concerned expression. To the right of the woman, there appears to have a man with a beard and a hat, who is standing with his back to the viewer. In the background, we can see a line of people walking on a sandy beach, with mountains in the distance. The sky is cloudy and the overall mood of the illustration is somber and contemplative."

[bitter hearth]

Florence2 is much, much weaker than ChatGPT

[errant dust]

Not weaker. BLIND

Bah

[bitter hearth]

but I also have a theory (this is probably wrong) that the line style you are using might not be good for vision models

they aren't good enough yet to not struggle with certain stuff

[errant dust]

unless their name is Claude

ok, well, was an experience in any case

[bitter hearth]

I will try Claude, thanks for the recommendation

oh hey, and i get this from llava ....

The image is a black and white illustration featuring a scene from the television show "Lost." At the forefront, there are three main characters: Jack Shear (portrayed by actor Matthew Fox), John Locke (portrayed by actor Terry O'Quinn), and Sawyer (portrayed by actor Josh Holloway). They are standing on a rocky beach, with waves crashing in the background. The characters appear to be observing something in the distance.

In the top right corner of the illustration, there is an additional character depicted, although their face is not visible due to the angle and perspective of the image. This character is wearing a hat and seems to be interacting with or directing the attention of the other three characters. The overall tone of the image is dramatic, capturing a moment of intrigue and mystery that is characteristic of the show "Lost."

[errant dust]

What it meant to write, and I am editing for accuracy "The image is a black and white illustration featuring a scene from the television show "Lost." in which I am the protagonist."

[bitter hearth]

yeah its messing up pretty bad

The image presents a tranquil scene of nature's beauty. Dominating the foreground is a black snake, coiled elegantly on the right side of the frame. Its body is adorned with intricate patterns in hues of blue and silver, adding a touch of mystery to its appearance.

The snake is nestled amidst a carpet of snow, which is dotted with small red flowers that contrast beautifully with the white snow. The background is a serene winter landscape, characterized by a blanket of snow that extends to the horizon.

In this picturesque setting, the snake's head and tail are pointed towards a bright orange flower in the center. This vibrant poppy stands out against the monochromatic backdrop, drawing the viewer's eye to its striking color. The overall composition of the image creates a harmonious blend of wildlife and nature's elements.

[errant dust]

It had me going until the orange flower, aka sun

it and I have differing concepts of "monochrome backdrop" too

[sage burrow]

Thanks for the warning

[bitter hearth]

Florence2 was ok for me but my images are extremely basic

fixed their teeth defect

[errant dust]

The prompt was for anorexic chicks?

This was just announced elsewhere:

Hey, we are announcing a 4-5x increase in speed for LoRA training in this new endpoint while having a superior quality for identities, try it and let us know how you feel: https://x.com/jfischoff/status/1829209970317169030

Jonathan Fischoff (@jfischoff) on X

BIG @fal announcement!

Proud to release our new FLUX LoRA trainer https://t.co/81FbnL5f0E

Now you can train at higher quality under 5 minutes! Great job @cloneofsimo.

Twitter · Aug 29, 2024, 5:30 PM

[errant dust]

https://www.tomsguide.com/ai/ai-image-video/move-over-midjourney-freepik-mystic-is-the-most-realistic-ai-image-generator-ive-tried

Sounds to me more like an elaborate and sophisticated workflow in Comfy than a genuine new model

[bitter hearth]

any word on lora support for any of the nf4 models?

[hexed dirge]

https://civitai.com/models/699491

[cunning lintel]

hard to care when the price is $5/run, you'd expect more efficiency == lower price, but this is just increased price. But maybe it works a lot better than current trainers, i'm not going to test it

[errant dust]

Ahhhh... I was not aware and merely passed it along

[cunning lintel]

i saw it teased on twitter earlier, was so disapointed when i saw the cost 😉

[desert garnet]

5$ per lora training? holy moly

[hexed dirge]

it's 2$ and some cents on Civ

[desert garnet]

even there still too expensive specially if u go with higher steps it can go up to 5 or 6k buzz

[bitter hearth]

https://www.reddit.com/r/StableDiffusion/comments/1etqaea/trying_the_realism_lora_with_flux_1_dev_nf4_cfg/ ?? this reddit post claims lora works with dev nf4 model

[desert garnet]

they dont work? i tried two and they work on forge,u only need to enable the never OOM extension and increase weight of the lora to like 2.5 or 3

[errant dust]

It is 2000 Buzz?

[sage burrow]

@errant dust now I wonder if my loras should be higher steps?
Only ever $2 each

Also, you must have more than an 8gb gpu!

[hexed dirge]

also more. I spent 2200 per lora

[errant dust]

So it was with not a little wonder that i discovered that the horse in this image is brown. My eyes not being what they used to be, thank goodness Florence 2 was able to explain this.

[bitter hearth]

i get a page list of error when trying to run lora with dev nf4 checkpoint

but its working with unet dev model

[errant dust]

"The horse is dark brown"

[desert garnet]

i tried a style with default settings but didnt get good results so had to do 15+ epochs and price went up

[desert garnet]

try to update forge

[bitter hearth]

i only use comfyui

[desert garnet]

then idk

[sage burrow]

Did you get better results with 15? I've used 10 a few times but still the first few end up best

[desert garnet]

yea lower epochs were deformed and didnt apply style properly

[bitter hearth]

btw this is with unet dev using 8 steps

[desert garnet]

i only had 50 imgs so maybe it was that

[noble coyote]

Florence2/Flux i2i

[errant dust]

Truly a good likeness of me reacting to Florence's descriptions

[noble coyote]

They are a tad weak

[sage burrow]

I made 3 flux loras; 64 images, 500 and 200. The one with 64 turned out best!

[noble coyote]

They need some style-selector to go with the rather banal text

I eventually found that in ComfyUI Manager - the config.ini file - my security settings were too high
Which prevented the Florence2 and tensorops installation. Lowering the security level from normal

to weak allowed me to progress

[sage burrow]

So fal us $5 per lora?! No free trial I see darnit.

How much is Shakker per lora btw?

[noble coyote]

"Tell LoRA I Love Her ...!" 😄

[sage burrow]

I think she's been cheating on you, just sayin 😉

[noble coyote]

Florence2/Flux i2i

[bitter hearth]

cheating, jealousy are emergent destructive concept from a predatory economic model

[noble coyote]

My new favourite prompting word (from MJ Theme of the Day) - s u b m e c h a n o p h o b i a

Or even - c e l l u l a r a u t o m a t a

[bitter hearth]

i seriously hope that allowing flux to generate life like celeb photos doesn't backfire

and all those celeb loras at civitai are pretty impressive

[sacred jewel]

#AwesomeSauce

[sacred jewel]

Frank Miller LoRA

Correction: I never enabled the LoRA so this is without it.

[sacred jewel]

WITH the Frank Miller LoRA actually enabled

[bitter hearth]

sdxl couldn't do bike images properly

im tired lol missing words

[fleet meteor]

Lmao

[bitter hearth]

face detailer

[sacred jewel]

I had not thought of that before... let's see how FLux does Zombies in bikes

[bitter hearth]

flux dev does it fairly well

make the zombies do a hula hoop dance

[sacred jewel]

Agreed... Hoolas coming up LOOL

[bitter hearth]

the front guy came out pretty cool with the bike

[sacred jewel]

John Wick theme Frank Miller LoRA

LOLOLOL

Maybe the LoRA got in the way. I am doing it again without the LoRA

[bitter hearth]

awesome backdrop and character looks epic

[bitter hearth]

if thats about hula hoop lol im having diffculty with it

trying to simplify prompt but not getting right on my end

[sacred jewel]

couple more in the queue with Frank Miller and then the ones without the lora should generate...

[bitter hearth]

maybe i should try w/o loras

kinda better w/o lora

[bitter hearth]

ahhh looks nice 🙂

[sacred jewel]

This one is very creative

[bitter hearth]

lol

[sacred jewel]

This one looks photoshopped afterwards (I can't believe I am saying that)

[bitter hearth]

cool, zombie with blood red hula hoop

[runic tusk]

All these images with whoa misspelled. 🫢

[sacred jewel]

If you do a search for my name on the channel, you can go through some of the tests I have been making with various LoRAs. It does get it right a lot 😄

[bitter hearth]

lol....

The correct spelling is "whoa", while "woah" is an alternate spelling that is considered nonstandard or informal. "Whoa" is the older spelling and is more commonly used.
"Whoa" is an interjection that can be used to express surprise, awe, joy, or to get someone's attention. It can also be used to calm someone down or to tell a horse to slow down. "Whoa" is traditionally used to command a horse to slow down or stop, but it can also be used to command a person.
"Woah" is more popular in UK English than US English, but many dictionaries do not consider it an acceptable variant of "whoa".

[sacred jewel]

Flux Eva - aiai 3v4 LoRA

[sacred jewel]

And also worth noting this is a prompt that I was asked to test for someone else in a different Discord so I use it VERBATIM. Never really bothered to use one version vs another 😉

zombies running screaming with giant billboard in the background that reads " W O A H "

[runic tusk]

Anyone can make whatever they want; no hating here. It's just funny to have so many of these posted and they're all the wrong spelling

[bitter hearth]

yeah no big deal 🙂 but also woah is like informal version of whoa

[runic tusk]

That doesn't make what I said less true.

Even the explanation you posted said what I said. 😄

[sacred jewel]

I don;t necessarily think it's misspelled or wrong spelling. As pointed out above, it is a popular way of spelling it in the UK. The person who's prompt it is originally, is actually from the UK so maybe that's it.

Never took it as hate 😄 I personally LOVE being corrected if I misspell anything as I walways want to improve my language 😉

[runic tusk]

But again, I'm just finding it humorous.

[bitter hearth]

hand got a little messed up there

[errant dust]

Not sure what prompt is, but these super skinny chicks with their rib cages sticking out.... is not a healthy look

[sacred jewel]

Here it is with the formal spelling

m100-style v.02 LoRA

[runic tusk]

https://tenor.com/O1rn.gif

[sacred jewel]

Oorah is a battle cry common in the United States Marine Corps since the mid-20th century. It is comparable to hooah in the US Army and hooyah in the US Navy and US Coast Guard. It is most commonly used to respond to a verbal greeting or as an expression of enthusiasm.

(from Wikipedia)

[bitter hearth]

Now we know the meaning! @sage burrow

.>

[toxic bone]

woah is how keanu says it

[runic tusk]

truth

[toxic bone]

https://tenor.com/view/gif-gif-19494627

[sacred jewel]

https://tenor.com/view/double-truth-do-the-right-thing-gif-9268013136674592813

[runic tusk]

https://tenor.com/2K6n.gif

[toxic bone]

https://tenor.com/view/hoes-lord-of-the-rings-lot-r-samwise-gamgee-sean-astin-gif-5727957

[sacred jewel]

https://tenor.com/view/yes-nodding-golden-girls-betty-white-rose-nylund-gif-1065076748312954815

https://tenor.com/view/true-dat-october-snoop-dog-dancing-gif-6121493

[toxic bone]

https://tenor.com/view/right-on-joe-dirt-nodding-gif-6009141

[sacred jewel]

Damn you all got me crying laughing here... ROTFL

[runic tusk]

https://tenor.com/cxwUYrxgUjG.gif

[sacred jewel]

https://tenor.com/view/money-gif-9430651

[runic tusk]

https://tenor.com/rvZ7.gif

[sacred jewel]

Syd Mead LoRA

[toxic bone]

stop trying to make fetch happen

[errant dust]

Yes, think of it as a modernized "Tally-ho!"

[toxic bone]

https://tenor.com/view/in-living-color-homey-the-clown-sit-down-mad-clown-hit-gif-15467259

[runic tusk]

https://tenor.com/9KSZ.gif

[sacred jewel]

OH YES!!!!

In Living Color needs to be resurrected... I know the originals are much older now but I bet they're even funnier now

[toxic bone]

damon's son is basically him

[runic tusk]

That whole family is basically the same person.

[sacred jewel]

https://tenor.com/view/hated-it-hate-gif-13364410

[runic tusk]

https://tenor.com/sOyn.gif

[toxic bone]

Jim Carey would go harder on sketch comedy if he did it again

https://tenor.com/view/fire-marshals-gif-19744892

[sacred jewel]

https://tenor.com/view/fire-marshals-gif-19744892

you beat me to it

[runic tusk]

https://tenor.com/bewMz.gif

[toxic bone]

https://tenor.com/view/wands-in-living-color-gif-21010144

[toxic bone]

yeah theres that, but damon jr really is his father lol

https://tenor.com/view/moist-cake-moist-cake-so-moist-new-girl-gif-5610947

[coral sable]

is there date for sd3 2b refresh, or it's pushed to later?

[sacred jewel]

"in two weeks" has been the running joke since they made an announcement a month or more ago...

More steps with Flux definitely add "flavor" to the resulting image ...

40
30
20
15
10

top to bottom number of steps... same seed and settings

40 steps clearly looks more saturated and defined. At least with this prompt and LoRA and latent dimensions. Probably better test to be done with the base model alone and at a sweet spot resolution and A:R

[sage burrow]

Merging my lora with Flux dev did NOT work 😭 😭 😭 the image results definitely prove it. DARNIT!

[vast condor]

I found the command, it was in the readme on the github, is this what you did? python networks/flux_merge_lora.py --flux_model flux1-dev.sft --save_to output.safetensors --models lora1.safetensors --ratios 2.0 --save_precision fp16 --loading_device cuda --working_device cpu

[trim talon]

a

[sage burrow]

I left the settings at 1. Trying 4 now just to see :D. Will try 2 next.

Oh, I just tried flux base, apparently my merges did do something, but not enough for undeniable proof.

More experimentation needed 😄

Just the comfy flux merge (which definitely works for sole things.

For fun and curiosity mostly. The fun if having models I can change 🙂

[obsidian junco]

Chris Christie?

[foggy cloak]

What's everyone's favorite Flux checkpoint so far?

[silver bridge]

Pixelwave imo both the dev and schnell variants

[foggy cloak]

Checked it out, the preview images look good, downloads are super slow on civitai for some reason though

[proven cipher]

Does controlnet work with SD3 ? Any tutorial for comfyui ?

[noble coyote]

https://www.youtube.com/watch?v=PppS2WzZwmI

Florence2/Flux i2i

Using A1111 it is possible perhaps - see this YT Video - it is VERY technical!
https://www.youtube.com/watch?v=sBFGitIvD2A

[sage burrow]

But, but, but, comfy! 😦 lol

111 doesn't support flux yet tho right?

[noble coyote]

If you find that 8Gb is too little VRAM - try this nVidia workaround
https://nvidia.custhelp.com/app/answers/detail/a_id/5490/~/system-memory-fallback-for-stable-diffusion

[sage burrow]

That's awesome, thank you! 🙂

[noble coyote]

I'm just trying it on my RTX 2070 8Gb

[sage burrow]

Let us know how it goes 🙂 Though you prob have a lot of regular ram to help?

[noble coyote]

I have 64Gb RAM - but this nVidia workaround can help with smaller VRAM sizes

Florence2/Flux i2i

They too are on limited RAM/VRAM?

[sage burrow]

@errant dust and perhaps even @bitter hearth 😄

Now of course I'm wondering, when google colab can't find an A100 to use, we can't just use that Nvidia differnet driver method on collab can we? lolol

[noble coyote]

Nope 😄

[sage burrow]

Darnit! Didn't think so LOL

[noble coyote]

I tried it - on X-Flux-IPAdapter - but got OOM - so I reverted to Driver Default setting

[sage burrow]

I've tried so many things on mine that I've lost track, including my 200gb swap file. It seems I can run some stuff that those with the same machine can't. But boy does it take forever when pushing it!!!

[sage burrow]

Some things are difficult to train with Flux, so I thought I'd let others figure it out LOL

[bitter hearth]

comfy is pretty good for minimising hardware use such as VRAM but its not the optimal way to go for minimising hardware use

its really not bad though

for such a big program

normally devs don't optimise node-based software but comfy did

[sage burrow]

Anyone who randomly wanders in here will see all our flux images and think SD3 is the best thing ever!!!
(not that it isn't, but it doens't look like flux)

[noble coyote]

We Flux bcuz SD3 disappointz 😄

[sage burrow]

There aren't any loras for SD3

[alpine summit]

Sd1.5

[sage burrow]

can auto111 or Forge users open our comfy workflows (which are embedded into images) into their UI?

[bitter hearth]

You get a bunch of words

And nothing makes sense

At least on the version i used

[errant dust]

Such a big program?

[sacred jewel]

#AwesomeSauce

[visual mist]

The virgin SD3 vs the chad Flux

[bitter hearth]

Says the furry

[visual mist]

That's the cat from the Ghost Stories english dub 😂

[bitter hearth]

I'm starting to believe that you funny is a bot

[sacred jewel]

https://tenor.com/view/twiki-buck-rogers-25th-century-theopolis-gif-14882036468294386025

[sacred jewel]

...and I am starting to believe you have 48GB of VRAM

that 4gb thing is a typo isn't it 😛 😛 😛

[sacred jewel]

If anyone deserves a pair of 48GB A100s, it is you 😄

[alpine summit]

Sd 1.5

[sacred jewel]

Looks like R2D2 ...

was that on purpose? very cool

[sacred jewel]

Love this... I am not jealous 👀 😛

Looks more like a Macy*s Day Parade Balloon than a realistic scene

[alpine summit]

Sdxl

[alpine summit]

Sdxl

[alpine summit]

Flux

[sacred jewel]

short dude

[noble coyote]

Florence2/Flux i2i

[sage burrow]

Best buy loans agrees with you lol. However my budget does not lol

[alpine summit]

Sd3

[sage burrow]

Well this is neat. How to do 15 min versions painlessly would be nice! https://x.com/CharaspowerAI/status/1824822023647846827

Pierrick Chevallier | IA (@CharaspowerAI) on X

🚨Dreaming of making 100% AI-generated videos with Sound like this?🚨

I’ve prepared a complete workflow that reveals all the secrets to creating incredible videos like this.
PUT THE SOUND
(Bookmark and get ready to be blown away) 🧵🔽

Twitter · Aug 17, 2024, 2:54 PM

[sage burrow]

This sort of reminded me of the Houichol beaded artwork!
(in person is even brighter, and not dragons lol)

[bitter hearth]

well, sort of medium-sized I guess

anyway what I was thinking was the optimal minimal VRAM usage for a given diffusion performance requirement would probably involve some sort of exotic compiler

with a lot of careful optimisations

if you were doing that, you would want to have a fairly narrow scope

on some level NVIDIA® TensorRT™ is already pushing in that direction a bit

[errant dust]

More like tiny

Zipped it's what? Less than 5MB?

[bitter hearth]

wow that's so small

I guess mentally I always (completely incorrectly) include the storage of the models

even then
I do not go over 100GB

some people do apparently

TBH I don't actually make images that much, if I did then maybe I too would have downloaded half of Civit by this point

[sage burrow]

That looks like a chocolate chip cookie gun 😄

[low stone]

yeah well, have to make it safe for the kids. 🙂

[hexed dirge]

https://civitai.com/models/699467

[sacred jewel]

You keep killing it... NICE!!

LOL

[hexed dirge]

thanks

also publishing 2 or 3 loras in next hours

one very dark

[bitter hearth]

I'm always amazed by loras

Even tho they are the same for me since 1.5 lmao

Just cool stuff

[bitter hearth]

Make some tea

[bitter hearth]

I'm sick

[sacred jewel]

Sometimes, Flux gets scale really really well...

[bitter hearth]

Big meow

Actually looks like a sheep

Lmao

[sage burrow]

Try out making some of your own, it's hella fun 🙂

[noble coyote]

ELLA

[noble coyote]

Can you use SDXL LoRAs on SD1.5?

[low stone]

needs more vintage anime amirite?

[sage burrow]

Unfortunately not. But you can use a SDXL refiner in your workflow.

[sacred jewel]

Mattias Adolsson LoRA

[low stone]

like with the vintage anime lora, I've found jacking up the fluxguidance on illustrative sharp line stuff to help a lot. I'm using vintage anime lora at fluxguidance of 6.0

reinforces the style but also cleans up some of muddyness of things.

[sacred jewel]

Awesome to know. I am terrified of trying anything above 3.5 lol as I did try way at the beginning of Flux release and got nothing buto horrible results. I need to expand my horizons 😄

Vintage anime lora is great...

Anatomical Sketch LoRA

[bitter hearth]

me as a kid loved making boats walk in robocraft

it looks so stupid lmao

[bitter hearth]

bad pc

[sacred jewel]

I don;t think so... I think it's cool... 🤭

[sage burrow]

There's better options, but this one is by far the easiest. Does cost $2 per lora though.

[sage burrow]

What, you can't train loras on 4gb gpu?! Kidding. I definitely; use external methods, and I have 8gb

[sacred jewel]

Anatomical Ink Sketch LoRA

Toy Camera LoRA by @hexed dirge

[young blade]

anyone have a better controlnet img2img workflow for Flux GGUF schnell? OUCH

makes my UPS beep using 900 watts 😦

[hexed dirge]

wonderful try with 0.8

[sage burrow]

Flux ip adapter! I haven't tried it yet, but cool! https://www.shakker.ai/modelinfo/7eebb942a4504d6abef007ddee70cc8d?from=search

[bitter hearth]

[sacred jewel]

Dark landscape LoRA

[runic tusk]

W"ᴼ ᴬᴴ"

[sacred jewel]

w00t w00t

6.0 guidance

12 Guidance

2.0 Guidance

mechwhite LoRA

[bitter hearth]

Woah

[sacred jewel]

Larry Sultan LoRA

[bitter hearth]

No woah

:<

Put woah on people's foreheads

[bitter hearth]

Thas me

[sacred jewel]

First try: FAIL

[bitter hearth]

Lul

@sacred jewel for the woah!

[sacred jewel]

Sketch V1 LoRA

[sullen moss]

I see no one is interested in two weeks anymore. LOL

[sacred jewel]

I WILL be interested in two weeks if two weeks materializes

[bitter hearth]

People need drama to spread negativity on the internet, not much drama happening lately here

[sacred jewel]

When peeps have a good model, everyone Shuts The F Up 😛

I have not seen ONE SINGLE image of woman laying on grass ROTFL

[sullen moss]

I've almost completely switched my pipeline to FLUX. The only thing I can't figure out is how to get inpainting to generate more details, especially on the skin.

[sacred jewel]

I have been having so much fun with Flux.1 Dev and LoRAs that have not even considred img2img or Inpainting ...

[bitter hearth]

[sacred jewel]

You did that right now on paper with crayon didn't you...

[bitter hearth]

[bitter hearth]

runs away

[sullen moss]

base gen in Dalle

rework with FLUX

[bitter hearth]

Handy gen

[sacred jewel]

Shaun Tan LoRA

Clone Storm Trooper LoRA

[desert garnet]

help me get up

[bitter hearth]

How would that work

I'm under her

[desert garnet]

[bitter hearth]

Death by butt

[sage burrow]

Describe it in detail. It does work. Though if you are talking as detailed as skin pores I don't know.

[desert garnet]

would u rather die by the sword instead?

[sage burrow]

Brazil banned x hmmm

[desert garnet]

elon in the mud

[sage burrow]

@errant dust no X for you!
(If I'm remembering people's lications correctly)

[errant dust]

I wouldn't pay it any attention

There is always some idiot judge pulling a stunt like this

not the first time

and to show how efficient the justice system is, tomorrow or day after anothr judge will revoke it

No one pays them any attention except the media

for this sort of thing

As to VPNs..... Frankly, I use one all the time. 24/7

How else could I use Imagen 3 for example, which is exclusive to the US?

The no-VPN stuff is nonsense, because the whole point is that you are unnoticeable

[sage burrow]

Aw, I thought Brazil was going to have a life quality improvement 😅

[desert garnet]

they could but a lot of ppl are addicted to it like crack

[sage burrow]

What?! So weak! hides my discord and glif windows

[low stone]

Do check out controlnet though.

[sacred jewel]

I am reluctant to dive in to this version... waiting for IPAdapter from Matteo... let's see what happens.

[sage burrow]

anyone know which folder clip_vision_l.safetensors goes into?

[desert garnet]

goes in the text_encoder folder

[low stone]

comfyui\models\clip_vision @sage burrow

[sage burrow]

Thank you very much 🙂

I even installed my 50th instance of python (or it seems like it lol) darn thing still doesn't work 😛

[desert garnet]

use forge its easier there

[sage burrow]

Then I have to learn Forge 😦 (tried that last week didn't end well lol)

THough I guess it was the whole trying to get Forge to use my comfy folders that was the problem

[desert garnet]

nothing to learn just install it,then put checkpoints in the folders and open it ,works better than comfy right now because u can use loras with nf4

[sage burrow]

Dont have enough HD space to double of those 25GB models!

[desert garnet]

yea the t5 and clip goes into text_encoder folder in forge

[desert garnet]

u can just add this to your webui-user.bat --ckpt-dir "H:\AI\models\Stable-diffusion"

[sage burrow]

Anyoe remember that Dice guy? lol He talked me into installing Forge, then Tried to show me how to write the code for Forge looking in my comfy folders for my files........

Where did that file go... I remember it being far more convoluted than that!

I know I got it wrong, the example file had shorter directories:
"set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=

@REM Uncomment following code to reference an existing Comfyui checkout.
@REM set C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI/=Your comfyui checkout dir
@REM
@REM set VENV_DIR=C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI/venv
@REM set VAE_DIR=C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\vae
@REM set COMMANDLINE_ARGS=%COMMANDLINE_ARGS% ^
@REM --ckpt-dir C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\checkpoints ^
@REM --hypernetwork-dir %A1111_HOME%/models/hypernetworks ^
@REM --embeddings-dir C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\embeddings ^
@REM --lora-dir C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\loras
call webui.bat
"

[desert garnet]

its better to delete all of that,look at mine

but in case you wanna use it like that just remove the @rem set

[toxic bone]

@REM is remove i think. commented out

[sage burrow]

er, I'm embarassed to admit that your example doesn't make any more sense to me

[desert garnet]

i couldnt make it work with the comfy venv so i just cleaned it up and let it gen its own venv and it works

[sage burrow]

my directory structure is too long in that right?

[desert garnet]

no that doesnt matter

[sage burrow]

and the REM aspect?

Anyways ,I put all that in the corresponding file, but it did not work 😦

Why no one like comfy? 😭 lol

[desert garnet]

here its fixed ```set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=
--ckpt-dir "C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\checkpoints"
--embeddings-dir "C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\embeddings"
--lora-dir "C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\loras"

call webui.bat```

[sage burrow]

Thank you so very much!!! 🙂 🙂 🙂

All I've ever coded before is html lolol

[desert garnet]

[bitter hearth]

im gonna touch u bro

[desert garnet]

https://tenor.com/view/broccoli-in-hot-tub-gif-7208850437699132279

[bitter hearth]

nvm hell nahh

[sage burrow]

Darnit, my checkpoints still dont show up in Forge

[desert garnet]

try to add this instead to your webui-user.bat ```
@echo off
set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=
--ckpt-dir "C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\checkpoints"
--embeddings-dir "C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\embeddings"
--lora-dir "C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\loras"

call webui.bat```

[sage burrow]

oh FFS

the consorship on here has never heard of copy pasting!!!

[sage burrow]

Did you literally mean copy / paste the above? Or maybe take out the "?
(it didn't work)

I was just being cranky, not important

[desert garnet]

literal copy use the screenshot i sent to see how it should look like

[sage burrow]

c alash p is banned

am I missing something?

[desert garnet]

thats all,try to open it now

[sage burrow]

did

a few ties

times

[desert garnet]

ok it could be the extra spaces lets try again with this ```
@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --ckpt-dir "C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\checkpoints" --embeddings-dir "C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\embeddings" --lora-dir "C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\loras"

call webui.bat```

[sage burrow]

You are the second person who has recommended Forge to me within a week, it must be useful!

[desert garnet]

well loras work on bnb nf4 idk if they fixed them on comfy

[sage burrow]

I finally read the command prompt window, perhaps this is the problem? "=====
INCOMPATIBLE PYTHON VERSION

This program is tested with 3.10.6 Python, but you have 3.11.9."

[desert garnet]

oh yea thats the problem try to open forge from the run.bat file the one thats next to the environment.bat not the webui-user.bat

[left relic]

#🌠｜show-and-tell monky in banana

[sage burrow]

Gemini advanced and I discussed this the other day, there is no run.bat 😦

I'd just double up my models if I didn't have so many

[desert garnet]

this is how my forge folder structure looks like

u prob did a git clone and didnt download the zip from the releases page thats prob why u dont have the system folder and the environment.bat file

that file tells forge to use the python version inside the system folder

and not the one u have installed

[sage burrow]

Er maybe blush

ssshhh don't tell CW lololol

[desert garnet]

https://tenor.com/view/oocw-outofcontextwrasslin-oh-sad-sad-emoji-gif-15462661027727635708

[sage burrow]

I need that version of python for Kohya anyways, sooo

51st python instance coming up 😄

[desert garnet]

[sage burrow]

Could have hand drawn them all by now rofl

[desert garnet]

only use forge cuz its faster than auto1111 but i do have 2 forge installations cuz new one broke old abandoned extensions so i have one for old XL stuff and one for Flux

[sage burrow]

so this is prob not the way to go? ;0

[desert garnet]

yea its better to use the one from github release pages cuz it has its own python so dont have to install anything

[sage burrow]

How come not comfy? Comfy is so awesome! (exacpt when I'm trying to figure out which folder to put things)

[desert garnet]

i have like 5k loras so trying to remember how all of them looks its a pain cuz no previews in comfy

[bitter hearth]

I'm fever

[sage burrow]

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

[desert garnet]

yea go to releases theres this package webui_forge_cu121_torch231.7z

[sage burrow]

I should make a YT video "what it's ACTUALLY like to install Forge" 🤣

Just refering to the installation process (but Flux didn't include the word installation)

[desert garnet]

Flux Flex

[sage burrow]

why did I not keep using mac, why?

[desert garnet]

😔

u suppose to open webui-user.bat not webui.bat

[sage burrow]

oops!

[hallow lion]

happy friday.

[desert garnet]

balls

[noble coyote]

(Installing Forge to see what all the 'noise' is about!) 😄

[noble coyote]

First image using Forge

[desert garnet]

https://tenor.com/view/so-good-tasty-chef-gif-13228799

[noble coyote]

Sheesh. Forge is not worth it. No Refiner, No ADetailer ... hum ho!

No Textual Inversion ...

[bitter hearth]

huh

not for flux, right ? Forge has all of that

what we tend to call "refiner" is just sampling a second time TBH

I always get confused by the terminology that came from A1111

like I am not 100% sure what "hi-res fix" is

same with Adetailer, I know the idea is it is drawing an inpainting mask for you, but I am not sure exactly which models it has

some services instead of Adetailer say "hand-fix" or "face-fix" so I guess they chose a specific model in that case

[desert garnet]

after detailer is called in comfy i think

[noble coyote]

Yolo8v.pt

Forge is riffing on A1111 - and A1111 seems to work!

[bitter hearth]

ah ok its yolov8

[noble coyote]

I shall stick with Automatic

[bitter hearth]

Impact pack has a lot of detailer nodes yeah

as far as I know "hi-res fix" is a second pass, after an upscale
but whether the upscale is pixel or latent I am not sure
and for SDXL I am not sure if the second pass is with normal SDXL or with the refiner

[desert garnet]

you can chose latent but it sucks,best latent upscaler is on comfy with nn upscaler,also forge doesnt support HAT upscalers

[bitter hearth]

ah yeah its a shame

personally I love HAT

[desert garnet]

yea hat is the best upscaler out there,other ones are outdated

[bitter hearth]

yeah I agree

as well as the similar ones to HAT like ADT, DAT and RGT etc

best results I have seen

[desert garnet]

oh yea forge does support DAT

[bitter hearth]

that's a bit funny, to support DAT but not HAT 🤔

[desert garnet]

yea looks like they dont care to add that to forge,same with the nn-upscaler

[bitter hearth]

I think if they add too much to Forge their target users might stop liking it

so its fine

[noble coyote]

Forge img2img

I got ADetailer to work!

[bitter hearth]

nice

[noble coyote]

Forge img2img + Generative Fill (PhotoShop) of the Wolfhound

[bitter hearth]

I really like it

its nice to see unique stuff

[sullen moss]

upscaling experiments

[bitter hearth]

this is amazing

I downloaded and zoomed on 4k monitor

really nice

[alpine summit]

Sd1.5

[sage burrow]

Which can fix a model's limitations 😉

[real terrace]

that is crazy, how do you do that with SD 1.5 only? some crazy upcalling?

also I wonder if it only works for rooms full of let's say incoherent stuff

[bitter hearth]

how do you get such good detail like this out of SD 1.5?

its amazing

[sage burrow]

Ooooh, this looks interesting 😄

[bitter hearth]

haha yeah there are a few nodes that do lora in comfy

[sage burrow]

Even more interesting 😄

[bitter hearth]

I'm working on a pytorch workflow for lora training

decided I want to avoid Diffusers or the ones like Kohya or One Trainer

TBH if you are used to comfy code
its not that far removed from a lot of pytorch diffusion repos

cos comfy is based partly on this repo called K diffusion

and K diffusion was pytorch workflow

that's why they call it K sampler

[rain current]

upscale

[noble coyote]

Good, but horribly oversmooth; and there is fringing on her thumb

Going in the right direction though ...

[rain current]

That's because of the second upscale pass; the first one does it well

[sage burrow]

Damit! lol "This error message suggests that there's an issue with allocating memory on your GPU device when trying to initialize FluxLoRA training in ComfyUI. It's likely related to insufficient GPU memory for the operation you're trying to perform. Here are a few potential reasons and solutions:?

[bitter hearth]

I really like this one

is your method Epic Photonism and then Ultimate Upscale?

the detail level is rly high in your images

[sullen moss]

Sometimes I downscale after upscaling and then upscale again

[bitter hearth]

this works well yeah

or sometimes if I don't want to have a big image, i just upscale then downscale and leave it at that

[noble coyote]

Forge img2img

[real terrace]

that looks like Buenos Aires's Caminito

[real terrace]

At this point I don't know if I'm too trained at looking at AI img's, but usually none of them look realistic at all, the ones that tries to be realistic. There's like this SD texture or something, and the faces, poses, the subject vs background...

for example this one is kind of cartoony but (to mee) it could be a real photo

[sage burrow]

What does that do?

[real terrace]

it seems this flux lora is trained with a photo of this guy or something

I haven't played with the weight much

[young blade]

Has anyone here tried this yet?

[sacred jewel]

Some tests of a new LoRA I am working on - Rene Magritte style

Various strengths

[sage burrow]

I keep getting errors that suggest an 8gb gpu may not be quite enough... or perhaps I need a smaller dataset.

[young blade]

Seem pretty straight forward to use?

[real terrace]

I see some loose connections to Magritte only, like some poses exactly like the paintings or the clothes

but I guess it must be really hard

like, creating a lora to make something like this would need a special training

[sage burrow]

More straightforward than a Forge install and run! (For me at least)

[noble coyote]

Magritte-y

[sacred jewel]

If I had ANY IDEA what I was doing, it might be possible 😄

[noble coyote]

Brendan Fraser!

[sage burrow]

The workflow is in the examples folder (I think it was). They have a github page also for info.

[sacred jewel]

Strength 1.0

[sage burrow]

My dataset was 500 images, will try again with 100!

[real terrace]

that looks Magritte-like yes

[sacred jewel]

It's EXTREMELY strong... works with or without trigger and anywhere from 0.10 to over 1.0 strength. It does get out of whack a lot but then again, so was Magritte. I did not restrict the dataset to a particular period of style of Magritte. I threw in all of my favorites 🤷‍♂️

[rain current]

And I think we only consider 'realistic' the images that are of low quality, boring, VHS, etc. When an image is sharp, it's already assumed to be AI. If you now see real stock photos in this chat, most people would probably say they are AI.

[real terrace]

yes that's true, high quality over produced photo might look AI

also real photos sometimes has some horrible nightmarish hands and finger that doesn't make sense lol

[sacred jewel]

Same seed and settings. Strengths 1.0, .8, .6, .4, .3, .2, .1 and no LoRA

[young blade]

Will set it up in a bit, wonder if you can train from the gguf models or if it's only able to use the 17g dev original release. I've got a pair of 4090 24g idle right now and I refuse to pay 3rd party for loras. I need to research more, I've only managed to successfully make 1 sdxl Lora with koyah a while back.

[sacred jewel]

0.60 strength seems to be a good spot