🆕｜sd3

[bitter hearth]

do you guys notice a big difference with th ehuge master model?

this is incredible!

[bitter hearth]

thanks bro. Thats using my Flux Hyper Checkpoint

[bitter hearth]

can i dl it somehweres?

yes here https://tensor.art/models/759856135286068673/FLUX-HYPER-TRAINED-DREAM-DIFFUSION-BY-DICE-V-1

noice!

[bitter hearth]

https://tensor.art/articles/760080809324596639

[bitter hearth]

so i just drag and drop into the model folder? no other files needed?

read the discription as flux models need , Vae, T5xxxl, Clip-L and a workflow if you use it in comfy. All the download links are in the model description

oh ok

[bitter hearth]

or run it on tensor art website generator.

i just got this rig so i ant to test it out haha

thats what i like to here

hear

[bitter hearth]

im new to comfy and this stable stuff really... for the files like the vae do i just need to dl this: flux1-schnell.safetensors? other stuffs i can ignore? 🤔

the vae is this one

thats the t5xxxl to download toand the clip_l

the 16 is for the grand big model right?

yes if you have more than say 20GB vram use the 16 if not use the 8

[bitter hearth]

can you remind me where to place the files?> which folders? sorry to bother you im overwhelmed and probabloy just not seeing the directions somewheres haha

[bitter hearth]

nm i found the infos

[sage burrow]

Or if you are patient and don't mind waiting 10 mins cause you only have 8gb gpu

compared with 2 mins for the 8 clip lol

[bitter hearth]

mine render in like 20 to 30 seconds lol

[mortal mesa]

i do T5 fp16 on CPU, feels worth it

[sage burrow]

Showoff! lol

[sage burrow]

Hpw long does that take?

[mortal mesa]

well i dont have like numbers for you but its faster than it half loading on GPU

[sage burrow]

Any reason you aren't using 16?

[bitter hearth]

I use both but a advertise the checkpoint with 8 as most people have a meduim size Vram . Im running a rig with 2 RTX 4090s witha out put of 48GB vram

But as I do this for a living I need the big rig for work so it pays for me to have a powerful set up . I built my own rig from scratch . If you want the full info and how to build the same rig i made a youtube video , https://youtu.be/jEcQqhcAtOI?feature=shared

anyone help what to do here or am missing? Prompt outputs failed validation
VAELoader:

• Value not in list: vae_name: 'vaeae.sft' not in ['ae.safetensors', 'taesd', 'taesdxl', 'taesd3']
  UNETLoader:
• Value not in list: unet_name: 'flux1-dev.sft' not in ['flux1-dev.safetensors']

[bitter hearth]

put the vae and the t5xxxl in clip vision folder

and create a new unet folder and put my Flux hyper model in there

[bitter hearth]

should flux1 also be in its own folder there not just dropped into unet folder?

[bitter hearth]

but i mean in unet you dont just drop the file in? you have to make a new folder within the unet folder?

just drop the check point in there

ok

[bitter hearth]

does just clip_l.safetensors go in the clip folder?

sorry for all the questions

[bitter hearth]

yep both clip_l and T4XXL in clip_vision folder

t5xxxl sorry typo

but the txxl goes in both clip and the other clip folder right?

only in that folder

unet folder is for the dream diffusion flux model

in clip_vision i thought i needed the tfxxls too. maybe not im all confused as you can tell haha

the vae just goes in the Vae folder

see my screen shot above

clip_l and t5xxxl both go in the clip_vision folder

think i got it. a lot of files you know and two clip folders haha. sorry for everyoen who had to read through this

[bitter hearth]

takes a bit of getting your head around at first with flux . once there in there tho the fun begins

[bitter hearth]

If I wnat to do image2image, what do I click to start that?

you will need the img2 img workflow for flux

and last Q i hope... for dual clip loader can they be a combo of any two or does the clip_l one have to be selected

download that with the download option

oh awesome thanks

ip adaptor workflow when you get more expericed with the comfy nodes

whats that one do?

and the beast

all you will need to start with is the simple workflow to get started with

thanks

download that last one and create a folder somewhere called workflows and save those all in there but use the last one to get up and running which will look like this

[bitter hearth]

your welcome buddy

gonna go get some dinner. will test it out when i get back. thanks again for your patience and making a model. must have been a lot of work

[bitter hearth]

all my models can be downloaded from here

https://www.shakker.ai/userpage/8b0d2aadaa2a4f2592cbb367c329ea51/publish

[sacred jewel]

You're running the 4090s in SLI? Or similar? How do you get ""48GB VRAM"" if you meant pooled. ??

Thanks

[bitter hearth]

1 rtx 4090 is 24GB vram my system runs 2 rtx 4090s giving it a combined vram of 48GB

[sacred jewel]

I know I saw that. My question is, you say 48GB VRAM as if it were one single pool. I thought you could not do that in either comfy or other UIs. Or maybe you can with special drivers for other software.

I also thought that Comfy can only inference on one card per generation. You can send stuff to multiple cards at once but not one single workflow over both/multiple GPUs ?

(P.S. I have a 4090 and a 3090 as well but they run separately)

[bitter hearth]

Yes i run multiple webui split across 3 monitors , forge, swarm and auto1111. all running renders at the same time

[sacred jewel]

That's what I thought. You don't have more than 24GB at a time 👍👌

[bitter hearth]

ill show you a video that i made for my sd3 gold check point . thats running both simultan

[sacred jewel]

So you can run a 30GB model then. Nice.

[bitter hearth]

https://youtu.be/FrVITtD0q_Y?feature=shared

[sacred jewel]

You must show me how you get the pooled VRAM. I always thought it wasn't possible.

[bitter hearth]

everything is possible with the right coding , takes a lot of messing around learning it as self taught as i couldnt find any help online on how to do it

[sacred jewel]

So you can run a training of say batch of 20 and as long as it fits into the combined 48GB of VRAM across the cards is that what you do?

Since 4090 has no SLI (nvlink not supported) how did you get the system to treat both cards as one?

[bitter hearth]

sorry i posted the wrong video this is the side by side video i ment https://youtu.be/zu3x2DUyng0?feature=shared

[sacred jewel]

I'll check it out. Thanks.

Nice build video too btw 👌

My build is 3 years old now 😭😭😭

[bitter hearth]

id never ever make another watercooled sysyem. them pipes were a frekkin headache

[sacred jewel]

I hate hate water-cooling unless someone else does it.

My team manages a high performance computing cluster where I work and it is water cooled racks and nothing but fucking problems. No thanks.

If I could afford someone to manage / build it for me, fine.

And I mean custom. I don't mind AIO.

[mortal mesa]

3 dead AIO's later back to air cooling

i mean its been years but ya that happens

[sacred jewel]

I also hate monster air cooled CPUs

I've gotten used to air cooled GPUs

[bitter hearth]

my main problem is that my rig is situated by my south facing window in direct sunlight. if the rig temperture hit over 40c then the pipes get warm and leak from the fittings

[sacred jewel]

I've been lucky so far.

[bitter hearth]

right now im running 3 screens all rendering and the temp is at 33c

normally sits there unless its a boiling hot summers day

[sacred jewel]

Have you all seen next gen cooling solutions?

https://tenor.com/view/wb640vgv0x-inverter-frenchbottomfreezer-hitachi-tủlạnh-gif-21552375

[bitter hearth]

lol

[sacred jewel]

Props on the AORUS ... that is my favorite MoBo mfg...

And of course, RGB will always add a few more FPS, amarite? LOLOL

I am assuming you upgraded the PSU for the dual 4090s? And I also assume you custom water cooled the 4090s too?

STUD if you did. I would be TERRIFIED to take a part a 4090. lol

[sage burrow]

Do you put ice in it? 😄

Is there any freon involved?

[bitter hearth]

yeah when i first built it I started out with 1 RTX 3090

its gone through more changes than king rolo

[bitter hearth]

somedays I think ice might help lol

Ive been working on somthing a bit funky check this image out and tell me what you see

walk away from the screen then re look at the image

[sacred jewel]

a car front end

easily see a car... looks like a Supra

or a ferrari

[errant dust]

ControlNet?

[bitter hearth]

Yeah a toyota supra , pretty neat hey

looks epic if you squint your eyes

[lapis fractal]

feel terrible after breaking up with boyfriend

[sage burrow]

Render his face on a dartboard, you'll feel better! 😄

[errant dust]

or print him out on toilet paper

[sacred jewel]

Brutal ...

[errant dust]

lol

[sage burrow]

Therapeutic!

[sacred jewel]

well, it finished in a couple of hours... here is the results page

Haven't been billed yet so not sure how much it was.

With and Without LoRA (same prompt, settings, seed etc.)
So far I do not see much change... let me keep testing seeds/prompts

without/with

[errant dust]

I just got a super interesting reply precisely about the lack of effectiveness of LoRAs on the GGUF models. Apparently there is a known issue if you use the LowVram flag in Comfy, which prevents full effectiveness of the LoRA.

[sage burrow]

Oh!

[errant dust]

https://github.com/city96/ComfyUI-GGUF/issues/33

[sage burrow]

Glad I use regular

[errant dust]

This is news to me and as it so happens applies to me as well.

[sacred jewel]

without|with

[errant dust]

Oh phew. Twin images... lol

Thanks for confirming though.

[sacred jewel]

W-O | W

[sage burrow]

I can 100% say loras work with flux dev regular! I'll spare everyone the proof lolol
The loras add specific aspects which are missing when no lora is used!
🥒🍒

[errant dust]

Actually, I'm going to test a little the lowvrsm on the regular also but I've been using the fp8 model because I was shocked at how fast it runs on my machine and how well

[sage burrow]

2 mins for me per image

[errant dust]

I take your word for it man. I was convinced the very first couple of images

[errant dust]

But I'm hating you a little bit because of the speed at which you are producing and posting these

😆

[sage burrow]

Wany my wirkflow?

[errant dust]

I meant him because he's producing the images at of the rate of maybe 10 seconds

[sage burrow]

Or you prob mean you funny guys, I haven't posted t9day really

[sage burrow]

Try glif 😄

[errant dust]

Yes, I mentioned the problem and then a minute later he has two side by side images. One with low vram and the other without.

And that's only possible to do locally

[bitter hearth]

https://tenor.com/view/keep-it-local-keepitlocal-lutzu-keep-it-local-keep-it-local-tgrnation-gif-7347509927330789500

[errant dust]

I'm greener than the Hulk with envy. Sans the muscles. 😆

But this is good news. Because it's an easy workaround.

[bitter hearth]

Subtle One lol

[errant dust]

lol

It's a regular nick I use in homage to a friend's description of my style in chess. (he was being ultra ironic)

[sacred jewel]

I must admit it is nice to have a 4090 to play with here at home.

uploaded V3 for those who want to play

https://civitai.com/models/670153?modelVersionId=768786

Above without any LoRA

Now with LoRAs

Blue Future

[sacred jewel]

inkpunk LoRA

Simon Stalenhag

Giger

Frazetta Pulp

Niji

OUE Style

RB Charcoal

[bitter hearth]

so i generated these two images with flux schnell and used schnell model for img2img to combine those two images into the 3rd pic

[sage burrow]

MJ (it's still free trial for a few more hours). I edited it for discord lol

[short thicket]

EldritchPunk 🙂

[hallow lion]

I love all the AI related jargon seeping into our collective nomenclature.

[bitter hearth]

https://tenor.com/view/meow-cat-gif-27161697

[hallow lion]

Awww. The 4GB cat. :< He needs help.

[sacred jewel]

I am willing to be Cat w 4GB has a pair of 96GB VRAM A100s and is just trolling us 😛

[brittle nexus]

woman in bikini warning

[bitter hearth]

a little cherry picked but love flux and it's lora training

[hallow lion]

The hero cats journey to embark on the long journey, to seek more VRAM.

[sacred jewel]

https://tenor.com/view/cat-moar-more-i-want-more-mooaar-gif-23786940

[hallow lion]

[obsidian junco]

I think I know what the prompt was for this one

[hallow lion]

These are so nice.

I only tried realism. I have no clue how to get style stuff.

[hallow lion]

I am also glad finally mechanical things are looking ok.

Before AI just did organic stuff well.

[obsidian junco]

the Medusa pics are fire

[sterile pendant]

DEIS is a really good sampler for flux.dev. shit gives good quality images vs euler at even really low step counts like 15

DEIS vs Euler 15 steps, simple schedule, flux.dev q8 gguf. Inference times are within +/- 2 seconds of each other

i know i was talking about it a few days back, but i hadn't really tried to see how low in step counts it can go before crumbling

[void fog]

Tribes of people lived, fought and made alliances. Everything was fine until the great winter of Fimbulwinter came, it was supposed to last three years and so that Midgard and all living things in it would not perish, and the Aetuns would not populate new lands, the Asses and Baths agreed to give people a source of life, which, like the sun, warms the settlement of people. But the great winter did not end in three years, but continued for decades to this day.

[torn wharf]

So , avatar?

[dusky thistle]

Here is the image you requested.

[craggy crest]

@bitter hearth

[bitter hearth]

DEIS has always been super under-rated yeah it was one of the big sampling papers

beta scheduler also under-rated although I do see more people using that

at least to my eye beta-scheduler is just a better Karras schedule

not for ret flow though

[bitter hearth]

when this gun fires does it shoot over his left shoulder 🤔

[dusky thistle]

just in case someone sneaks up

[dusky thistle]

don't trust that mouth, it's not small

[bitter hearth]

most enthusiastic burger salesman

[dusky thistle]

you WILL have a BIG MACCC

[bitter hearth]

why is one hog normal

[dusky thistle]

8B

[bitter hearth]

I like how the Mcdonalds logo sneaked in there

its an official Mcdonalds advert now

[sterile pendant]

i use beta almost exclusively these days and it works really well with flux usually

well flux dev obviously

[bitter hearth]

ye I use beta almost exclusively too

I thought ret flow would not like it

cos SD3 didn't seem to like karras

[sterile pendant]

well comfyui forces the flux.dev sigma curve and the beta modifies it

[craggy crest]

flux is heavily overfit for: women, dogs, fantasy, and furries (specifically anime cat girls)

[bitter hearth]

this was flux's attempt at my R2D2 prompt

to be honest this is not better than ones I've been making with older models

[sterile pendant]

see what i mean

[craggy crest]

SD3 also doesn't like ancestral anything

[sterile pendant]

comfy forces the curve for dev and the straight line for schnell, but beta can still modify them

you can also get schnell to have the dev curve if you use the modelsamplingflux node as well, at least i think, lemme double check

[sterile pendant]

yeah see, gives it the curve. this one reason why people have issues with schnell

they think they should have the modelsamplingflux node, but it actually causes harm

[craggy crest]

you get the same curves if you're not using gguf?

[sterile pendant]

yeah

with the regular models, it does the same. it's in the comfy code

[craggy crest]

grab my workflow for the donut and see what you think

[bitter hearth]

I wish comfy didn't work this way
and just took sigmas explicitly by default

[craggy crest]

go open an issue on the github for that suggestion

[sterile pendant]

damn i have to unpeel this lol

[craggy crest]

heh - i hide what i'm not gonna change 😉

[bitter hearth]

I'm moving off of comfy so I'm not rly invested in changing it

[craggy crest]

oh? don't tell me you're moving to forge

[bitter hearth]

haha

no I'm moving to raw pytorch / JAX workflow

[craggy crest]

you have fun with that

[bitter hearth]

ty

[sterile pendant]

there's also the one diffuser.cpp i think, cant remember the exact name

its fast, but still a big WIP, but it supports flux

[bitter hearth]

yeah someone on L2 discord uses it

I love C++ so that might be good

[craggy crest]

I'm boggled that they're writing it in c++

[sterile pendant]

yeah it's a pretty standard workflow. you don't have to do the complicated custom sampler anymore though. you can just use a ksampler with cfg1.0 and just connect the prompt to both pos and neg since it ignores neg

[sterile pendant]

c++ is usually around 1,000 to 10,000 faster than python for certain types of operations

[craggy crest]

not till i get done with this one word prompt walk through of the data training set

[craggy crest]

i'm not complaining, tha'ts my prefered language, i'm just surprised they're using it

[sterile pendant]

a lot of the ML world uses it. well the bulk actually. llama.cpp

but then thats why they also make things like llama.cpp python that acts as a wrapper for those that want to interface with python

[craggy crest]

i've just heard so many people looking down their noses as the very idea over the last 2 years that I didn't think any one would dare mention it, much less code in it

[sterile pendant]

python is used a lot because it's easy and is honestly more of a scripting language. i mean don't get me wrong, you can definitely still make classes and other complex things, it's just one of the highest level coding languages. but that's also why it's one of the most common as well. it's super friendly to newcomers, due to the very lax syntax of it

you almost dont even have to be a coder to use python

[craggy crest]

what flux thinks Dorsal means

[sterile pendant]

try it with schnell

dev is overcooked for people and shit like that, schnell is a lot more creative

[craggy crest]

i'm walking through with dev. you can try it with schnell if you want

[sterile pendant]

lemme give it a run

[craggy crest]

not overcooked, overfit

[bitter hearth]

I first started doing machine learning in the 90's when it was all C or C++ lol

the python stuff is recent historiy

[craggy crest]

i was coding a mud back then

[bitter hearth]

wow nice

[sterile pendant]

yeah i meant fit, i'm making food and it was on my mind

[craggy crest]

heh. your dinner is overfit for charcoal

flux dev is overfit for: fantasy, women, furries (specifically anime cat girls) and dogs

[bitter hearth]

I'm kinda getting the sneaking feeling that flux has been over-rated

due to its very stunning initial aesthetic

[craggy crest]

it's the community that over-rated it if it is. black forest labs didn't make any spectaular claims, they just released it

[sterile pendant]

got it on the first try

[bitter hearth]

BFL don't say much of anything either way yeah

[craggy crest]

that's a mean looking fish

[bitter hearth]

I actually quite like the "drop the model and run" style of release

[sterile pendant]

yeah, but dorsal is definitely included

"of, on, or relating to the upper side or back of an animal, plant, or organ."

[craggy crest]

the BFL lab guys are all research scientists - not one of them has a clue how to run a company, and they have no one that does. so they're doing what they normally do - drop the release and move on

[bitter hearth]

ye

on some level I prefer that

[sterile pendant]

ran it eight more times and it gave me decent fish every time

[craggy crest]

as long as you can pay the bills, why not

[craggy crest]

cool. what do you get for the word yawn

[sterile pendant]

oh wait, there's an odd one

but that is technically the top of an almost organic surface

[craggy crest]

prompt: Encaustic <-- spot on

[bitter hearth]

these look nice

[sterile pendant]

yawn

[craggy crest]

run it several times

[sterile pendant]

i am

[craggy crest]

yeah. and that's exactly what the word means. when it understands something, it's on target. but it has no clue what elastic means. however it knows what a very specific elastic/rubber coating is

[sterile pendant]

elastic isn't the right term for that, say rubber

[craggy crest]

i'm doing one word prompts to see what the top of the bellcurve for them are

it knows exactly what Elastomeric is. it does not know what Elastic is

[sterile pendant]

elastomer != elastic though

[craggy crest]

Elastomeric coatings are fluid-applied roofing membranes with elastic properties that allow it to expand and contract with the substrate. Elongation is important because roofs expand and contract and this allows the roof coating to move with the substrate.

it's a very niche word

[sterile pendant]

that's an actual noun though

elastic is an adjective and isn't always a visible one

[craggy crest]

but elastic is a very common word. used for all sorts of things

[sterile pendant]

i know it is, but people and AI aren't likely going to use it to describe something for a dataset

[craggy crest]

this is what it thinks Elastomeric means - and is what it means

this is what it thinks elastic means

go search google for elastic and look at the images tab

you won't find knitted stuff

[sterile pendant]

the human body is warm, an AI or person isn't going to use it to describe them in a thermal sense. but you can associate the term warm with a scene composition or a smile though

it's all about association

and yarn is elastic

[craggy crest]

one word prompts, remember?

and yarn is not elastic

[sterile pendant]

i know

[craggy crest]

it might be a little stretchy but it's not elastic

[sterile pendant]

you're missing the engineering world

[craggy crest]

nor is it ever described as such

[sterile pendant]

elasticity is used to describe even metal

same with plasticity

<---engineer

[craggy crest]

i didn't give it elasticity, i gave it elastic

[bitter hearth]

I like how Elastomeric is in the training data

[craggy crest]

i didn't give it plasticity, either

[sterile pendant]

im using the broad version of the term

metals can undergo plastic deformation or elastic deformation

and it is worded as such

[craggy crest]

yes, well - i'm using single word prompts. so what I get back is going to be the absolute best answer, what has the most weight. not the broader defintion of a word i didn't even use

[sterile pendant]

but if you understand how these models work, you then understand that there are tons of concepts that can be associated with a single word. playing the seed game just picks with it leans towards

[craggy crest]

what the dictionary defines a term as, has little to do with what it has learned

[craggy crest]

i teach this stuff

[sterile pendant]

youre forgetting where they pull all this data from

[craggy crest]

it looks like they pulled it from commercial videos, movies, advertisements, etc

[bitter hearth]

its complex cos its also going through T5 and so the training of T5 will colour this also

[sterile pendant]

from millions and millions and millions of items from datascraping the web

from facebook posts, to dictionaries, to wikipedia, to youtube, to game forums, you name it

[craggy crest]

however, to quote matteo, it was likely trained on coglvm - meaning that the lables are stuff lke "this appears to be a woman in her 30s or 40s" rather than lables that actually fit the image and state what's in it

[bitter hearth]

its kind of an issue that CogVLM is not very smart

[craggy crest]

elasticity is even worsse

[bitter hearth]

does not take a long time of looking at CogVLM input-output pairs to see errors

[craggy crest]

it's no where close to your defintion or the dictionaries

[sterile pendant]

cogvlm is another network that also had to be trained on mostly scraped data from the web, with some handcaptioning as well

[bitter hearth]

cogvlm seems to "get confused" easily
I use it for some agent stuff and it can be a bit squiffy

[craggy crest]

and this is where all the probelms lie - the images used needed lables that were hard definitions "shoes, red," not "this appears to be footware in the color crimson which might work well when dressing for dinner"

to quote matteo again - if you give claude or chatGPT a copy of starry starry night, tell it to describe it, then hand that to flux, you will get the exact painting.

[craggy crest]

so i guess that means it's best to get an LLM to desccribe the scene you want and hand that to flux

[craggy crest]

you really need a vacation

[bitter hearth]

I always thought that OpenAI did the right thing by fine tuning an LLM to prompt the model

they should have made it optional though

but I think its the right way to go

if the model creators make the llm

[craggy crest]

it is optional. just go to microsoft's interface and talk to dall-e3 directly

[bitter hearth]

ah ok

[sterile pendant]

all of these are for elasticity and all of them technically fit various definitions of it. from https://www.thesaurus.com/browse/elasticity
strongest matches:
adaptability
flexibility
resilience

strong matches:
fluidity
give
malleability
plasticity
pliancy
springiness
suppleness

[craggy crest]

which has absolutely nothing to do with what I was trying to point out.

[sterile pendant]

what i'm trying to prove is that you're thinking too rigidly about the single word

you think the word elastic only means rubber band or something

[craggy crest]

what i'm trying to show you is what flux dev thinks the word means - not what the word means, what the dictionary says it means, or how to prompt correctly to get it to give you a picture of something that's elastic

[sterile pendant]

and what i'm showing is that schnell is definitely completing the task

correctly

just because it's not giving me a rubber band, doesn't mean it's wrong

[craggy crest]

and? i didn't say 'oh look at this, dev does a much better job than schnell" i said "this is what flux dev thinks this word means"

[sterile pendant]

and you're thinking that they are wrong or something, when they aren't

[craggy crest]

those images are what the word elasticity get flux dev to create

i don't think those are wrong, they ARE wrong

not one of them fit any of your words

those are sculptures at the very best

[sterile pendant]

they all look rubbery and the top right dawg has stretchy looking arms, not to mention the top left with all the tendrils that are likely inspired by:
fluidity
give
malleability
plasticity
pliancy
springiness
suppleness

[craggy crest]

they're ceramic sculptures

[sterile pendant]

i mean you can argue it all you want, but they definitely imply some of the associated concepts with the word plasticity

[craggy crest]

i

did

not

tell

it

plasticity

anyway, this is turning into an argument so i'm leaving

we don't need that

[sterile pendant]

elasticity*

[craggy crest]

elasticity IS NOT THE SAME AS ELASTIC. ELASTIC IS A NOUN

and i didn't give it elasticity

[sterile pendant]

they are heavily associated as strong matches

[craggy crest]

I DID NOT GIVE IT THE WORD ELASTICITY! for the last time, i gave it the word elastic. that's a noun. an object. something you sew with that makes it possible to streatch your pants wen you put them on

it's not a concept

it's this stuff

[sterile pendant]

and i'm saying that you're not understanding how these models work under the hood. concepts form a tensor map essentially of all possible concepts in their dictionary. 10s of thousands, if not hundreds of thousands. each one of those concepts has a correlation with other concepts in the n-dimensional tensor

each concept has a weight with every single other concept. like with LLMs and their next token, it calculates EVERY SINGLE possible next token probability and then picks from the top 40 or so or however you set the temperature

[craggy crest]

yeah. i do understand how they work. and IF i was using a longer prompt where it had the leeway necessary to go off to areas of latent space that wasn't specifically what the work's absoltue highest weight was that would be fine. but i am giving it ONE WORD - which returns the absolute highest weight for the meaning. and what it should return is that stuff. not something estoeric

[sterile pendant]

that single word still goes into the concept map and still forms associates with every other concept

it doesn't work that way due to seeds

and how the models work in general, with noise

[craggy crest]

sigh. you're wrong. and i'm done

[sterile pendant]

you cannot guarantee it will give you some top highest weight, that's just not how it works

that's the whole point of the noise in the process

[bitter hearth]

its going via T5 which was trained on disambiguation, which implies that T5 may unpack words into trees of synonyms in the latent space (we don't actually know how T5 works within)

[sterile pendant]

Yeah, I was just lumping it all together as a single package to save explanation, but this mostly falls on the LLM and how it creates the embeds. But even after that, those token concepts then get mapped to data in the dataset, which is in itself, also a very noisy process as well(training)

[bitter hearth]

I also think one word prompts is an ablation

its too far out of distribution

[sterile pendant]

I think what they were trying to produce were essentially unconditional generations

Which is a whole different topic

[bitter hearth]

ye at least on SDXL if you do a very short prompt you get that unconditional distribution look (very soft and feathery, pastel colours)

[sterile pendant]

Well it doesn't use cfg or anything and the images always look like a soft blurry mess

Want to say you essentially use cfg 0 or something, but I could be wrong

[bitter hearth]

depends on code base anyway, where the cfg scale starts

[sterile pendant]

right, might be zero or one

[bitter hearth]

https://about.meta.com/realitylabs/codecavatars/sapiens

new meta

[noble coyote]

Try these words: tenegrity, voronoi, fibonacci, zentangle, rococopunk, retrofuturism, cubism ...

[sterile pendant]

If I get some time in a bit and remember, I'll try them. I use retrofuturism a lot and know it works a lot of the time

I'd also assume Fibonacci and voronoi would also work well, since they are extremely common themes. I use voronoi all the time for world creation in game related work

It's a common base noise used for terrain making, before things like erosion, has a good natural feel to it

But it's also good for cellular and organic things

[noble coyote]

In strictly artistic terms - voronoi produces triangles and polygons

Zentangle is a fine art effect, almost like lace

[sterile pendant]

yeah, but the random pattern boundaries are inspired by how cells in things like plants divide, which is why they call it a more "organic" pattern

[noble coyote]

Rococopunk adds a baroque look

How fatigue in metal propagates; how lighting 'decides' which path to ground ...

[sterile pendant]

i've used rococopunk a few times in the past, it's a neat style

another fun one is reaction-diffusion or turing pattern. it's got a few names

it's that wild pattern that plecos and pufferfish have

this one

[noble coyote]

Tesselated

[sterile pendant]

well in the case of this pattern, it's directly due to the physics of reaction-diffusion. things like bacteria will form patterns like this as they consume food and divide

same with cells, and it's why and how a creature like a pleco gets the pattern on their skin

fibonocci, voronoi was the prompt

[noble coyote]

Pleco = catfish (another gap in my education closed) 🙂

[sterile pendant]

yeah they are a type, but a lot of people keep in in aquariums as pets. they also get called suckerfish as well by some. here's the type i was directly refering to

just off wikipedia

but do you see the turing pattern?

[noble coyote]

Yes

Mebbe mackerel the same

[sterile pendant]

another fun pattern is the lichtenberg pattern

it's the lightning strike looking one with all the small feeler paths in it

like this (from google)

[noble coyote]

I once worked in a scientific institute which spent a lot of time researching how cracks from metal fatigue propagate ... I guess it too is Lichtenbergian

[sterile pendant]

yeah, it's actually closely related to pathfinding algorithms like A*

(a-star)

at each instantaneous moment, it's searching for the next path of least resistance. sometimes, they hit dead ends like in the cube above

you'll also see it in super slowmo clips of lightning strikes before the actual main discharge happens as it finds the completed path to ground

[noble coyote]

You mean it "sends out a scout" before the main event?

[sterile pendant]

essentially

in the case of electric discharge, those feelers create ionized paths of lower resistance, so when they do find a surface of lower potential, the main discharge goes through that ionized path to ground. otherwise, lightning could never strike from miles in the air since it would require voltages in the 1x10^really big number range

[noble coyote]

Like how does the scout path tell the bolt "we're ready for you?"

[sterile pendant]

it doesnt, its forming a path of least resistance. at each instantaneous moment, they wander always to the lowest path of resistance

but the split second they make contact to "ground" the discharge happens because the resistance is low enough

[noble coyote]

OK, but if the main bolt is holding back until the PoLR is established - there seems to be some 'communication'

at which point it lets rip

[sterile pendant]

nope, it's no different than a ball constantly wanting to be pulled to the ground. the tug is always there

[noble coyote]

Acceleration is to the centre

[sterile pendant]

entropy always wants to go from high to low, so with electricity, if there is a potential difference between surfaces, they want to equalize. so they generate a field between them (and well everything else in the universe technically)

same goes with gravity. you could have two protons at opposite ends of the visible universe and they are still tugging on each other, along with every single other mass in the universe and vice versa

[noble coyote]

Entropy is increasing ... bringing with it the natural decay of our world

[sterile pendant]

max entropy means absolute zero, it's a confusing term

[noble coyote]

Quantum Entanglement - I'd love to be able to understand that! My math is long division, feet and inches 😄

Entropy = chaos

[sterile pendant]

at any rate, as long as there's a potential difference in voltages, there's a field between them

[noble coyote]

I used Lichtenberg Pattern in this one

[sterile pendant]

oh damn, that actually does kinda show there

[noble coyote]

She has an extra arm!

[sterile pendant]

well nobody is perfect

[bitter hearth]

flux does seem to do extra arms fairly often

in the last 2 days I did a few hundred flux gens

and third arm was not rare

its still a base model after all

[sterile pendant]

yeah i wouldnt know, dont make people that often or if i do, they're in boring poses like arms by their sides standing

[bitter hearth]

ye I don't make people that often either

[sterile pendant]

except when i do old men posing like tiktokers

dressed in skirts and uniforms lol

thankfully, people chilled out with the waifu spam, but i do miss making my anti-waifu memes

[bitter hearth]

this server has gone quiet in general

its better though

[sterile pendant]

well there's always a surge when some new model is coming out or when it comes out. typical hype falloff that applies to pretty much everything

[noble coyote]

Grieving over 'stillborn' SD3

[sterile pendant]

i still use the shit out of sd3, just not for people lol

[noble coyote]

Clownshark uses this space for some amazing Cascade Creations

[bitter hearth]

they deleted the cascade channel even though half the regulars here use cascade lol

[noble coyote]

I have a full-year sub to SD3@ClipDrop - but I hardly use it 😦

[bitter hearth]

wow

money well spent

[noble coyote]

I much prefer Flux and AuraFlow0.3

[bitter hearth]

not sure if auraflow aesthetics are quite there yet

[noble coyote]

The LoRAs for AuraFlow are at a minimum

[bitter hearth]

nice

its got a great license

so if it gets good aesthetics then it would be a good model

[sage burrow]

It'll be back 🙂

[sage burrow]

Has it improved? Last I used auraflow it was kinda terrible

[bitter hearth]

Its getting better

[sterile pendant]

It's going to take a bit of time to get better. I think now with the gguf quantizations, people should gguf it and that way, people with 8gb vram gpus can actually use it

Since 8gb and under makes up like 70% of all pcs

Well I'd have to double-check steam hardware survey to be sure

But if more people can use it, it can gain more traction

[noble coyote]

AuraFlow0.3

Its very capable

[bitter hearth]

thats seems a high %

these are all SD3, with SD1.5 face fix
the model was never bad even with this really
it needed face and hand fix though

[noble coyote]

SD3 is good - but the eyes need work - agreed!

[bitter hearth]

Ive made 2 SD3 checkpoints with both having the hand, face and poses fixed if you want them bud

[noble coyote]

Give me the link, svp?

[bitter hearth]

https://tensor.art/models/753956516668680191/SD3-PLATINUM-Dream-Diffusion-By-Dice-V1 https://tensor.art/models/751454663859189150/SD3-GOLD-Dream-Diffusion-By-DICE-v1

[noble coyote]

Random Message appearing in Flux.Dev

[sterile pendant]

it's not, https://store.steampowered.com/hwsurvey/Steam-Hardware-Software-Survey-Welcome-to-Steam

[noble coyote]

I have these two - thank you - one is based on SDXL0.9?

[bitter hearth]

average hardware is hilariously low yeah

[sterile pendant]

71.51% have 8gb and under

just did the math

and steam hardware has hundreds of millions of samples

[bitter hearth]

platinum is the full SD3 10GB hyper trained, Gold was a SD3 Meduim reframed to run in auto 1111

how did you train the larger SD3?

[sterile pendant]

in the game industry, we use steam hardware survey like it's the obelisk out of 2001 a space odyssey

[bitter hearth]

I have hyper trained Flux to . I have created my own Training platform that will be dropped for public use in the next 2 weeks as my open source project

https://tensor.art/models/759856135286068673/FLUX-HYPER-TRAINED-DREAM-DIFFUSION-BY-DICE-V-1

but when you say the full SD3

do you mean the 8B model?

[bitter hearth]

10

is there a 10B SD3 model?

[sterile pendant]

anyways, the point is that the bulk majority of local pc setups are not flexin much vram. so if you want more people to use a model to help it gain social traction, you gotta get it to run on 8gb.

and sd3 8b could run on weaker hardware if they gguf it. flux is proof of that

[noble coyote]

I suspect the future of PiXart is difficult?

[bitter hearth]

this model bro https://www.shakker.ai/modelinfo/244fdd507f1a438cac25eaa6ecb7497e?from=personal_page

[sage burrow]

They didn't have options for more last i checked! OK so bestbuy isn't the best source (pun intended), but still.

[noble coyote]

[What does GGUF stand for?]

[sterile pendant]

i honestly suspect you'll start to see a lot of these other models doing collabs with each other. but pixart might go somewhere since nvidia is backing them

[bitter hearth]

okay this is SD3 2B

[sterile pendant]

gguf is a form of quantization that came from LLMs

[sterile pendant]

a q8 quant of an LLM is usually within a single percent as accurate as the bfp16 version of the model

but at a fraction of the size

[noble coyote]

Ah, OK - smaller and really just as good

[sage burrow]

Gpu gouging underrated feature?

[sterile pendant]

but they can also have more or less levels of quantization

like q4 q5 q6, and then with some K variations like s, m, l

where they use different precisions for important blocks

[noble coyote]

So GGUF was developed so that the model size was manageable, without compromising quality?

[sterile pendant]

yes

pretty much anyone running local models in the LLM world use them or some similar technique

[noble coyote]

The dichotomy is - produce massive file-size models which can do everything - but demand expensive hardware!
Or make models smaller models - with no compromise in quality - so no need for the community to shell-out on a new PC?

[sage burrow]

So, every 2 days lately? 😄

[sterile pendant]

here's an example list of different quants and some sizes with them:
https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF

[noble coyote]

I'm tempted to go for the 32 bit

[sterile pendant]

and here's an example of the perplexity curve of them (a fancy way of saying how lobotomized a model becomes)

smaller PPL = better

q5 km is probably the best tradeoff in size vs quality

but anything q4 km and up is usually deemed the standard for most models

[sage burrow]

If it was always 4 arms, that would be awesome and useful. 3 though, too much post processing

[noble coyote]

I have Q4 and Q8

[bitter hearth]

could inpaint the third arm with an sdxl model and turn it into a snake or a ribbon or something

[sterile pendant]

yeah those are what i keep usually as well

[bitter hearth]

For users who still only use SDXL models, I've Created a SDXL checkpoint that runs on the Turbo base. Its hyper trained to a Maximum 100%. The adherence isnt that of flux but its very very good and the output image is the best youll see from any SDXL checkpoint. See what you think of the images posted in the gallery. Cheers https://tensor.art/models/751494452436224082/TWIN-TURBO-Dream-Diffusion-By-DICE-v1 https://www.shakker.ai/modelinfo/3ed729873f034271bef568cfc92176aa?from=personal_page

[sterile pendant]

but with LLMs, a lot of people want to pick the maximum quant size they can fit into their GPU to prevent it getting like 10x slower due to offloading. but you pick a quant usually with a gb or two room left for context size

[bitter hearth]

matching quant to GPU is important yeah

[sterile pendant]

but if it's a big enough model that you know you can't fit, even all the way down at q4, then just say f it and get the q8 lol

it's one of those all or nothing type deals

like any time i have to use 13b models and stuff with 8gb vram, q8 it is lol

[sage burrow]

What's 10?

[noble coyote]

Dare I say it, but as an artist, 'poor quality' actually becomes part of the art!

[bitter hearth]

10GB download

[bitter hearth]

its the 2B, I checked

[sterile pendant]

well he might be thinking of the version with t5 and clips

[noble coyote]

If I was more a photographer, then it would be quality quality quality

[sterile pendant]

which is around 10gb

[bitter hearth]

yeah its the difference between download size in GB and the number of model parameters (unitless)

[sterile pendant]

yeah

[bitter hearth]

yep

[sterile pendant]

but it doesnt all load at once

[bitter hearth]

mine does

[sterile pendant]

the actual sd3 2b transformer only takes up like 4gb vram or so

[bitter hearth]

its a one click install or you can also run it on Tensor Art online generator

platinum that is

[sage burrow]

My next pc us going to be a cloud gpu

[bitter hearth]

yes join cloud gang ☁️

[bitter hearth]

Flux Hyper

[mortal kite]

shark is cool

[bitter hearth]

lol

[sacred jewel]

That was the scout 🤭

[bitter hearth]

normal flux up there

My Flux Hyper checkpoint

Title: Display the title "Flux Hyper" in large playful text at the top or center of the poster.
Main Character: Depict a Space Marine standing next to a Ferrari, dynamic pose,
Background: large explosions with a space marine battle scene, Ensure the background is vibrant and engaging.

[bitter hearth]

JOB DONE. Thats the best way I've found to write a prompt in flux

[mortal kite]

what is flux hyper

[bitter hearth]

https://tensor.art/models/759856135286068673/FLUX-HYPER-TRAINED-DREAM-DIFFUSION-BY-DICE-V-1

Higher trained version of Flux Dev

[bitter hearth]

Title: Display the title "Trustory" in large playful text at the top or center of the poster.
Main Character: A photo of a highly detailed mecha standing next to a human woman, dynamic pose,
Background: Science laboratory, Ensure the background is vibrant and engaging.

[sacred jewel]

oh damn, this is nice!!

oh heeeelllll no!

[mortal kite]

hmmm. Oh screw that, dream just makes everything look the same...

[mortal kite]

YouFunnyGuys I don't get your images, do you just spam random keywords into it lol

[alpine summit]

Title: Display the title " dice.ai" in large playful text at the top or center of the poster.
Main Character: A photo of a highly detailed flying fantasy bio scape, dynamic pose,
Background: cloud, Ensure the background is vibrant and engaging

[bitter hearth]

sweet

😆

Title: Display the title "Dice" in large playful text at the top or center of the poster.
a photo of a cat wearing a Hawaian shirt and colorful surfer shorts , surfing a big wave

Hyper cat rocks

[noble coyote]

I'm guessing he's using MultiLine Prompting?

[bitter hearth]

😆

[noble coyote]

Flux.Dev+LoRAs

[bitter hearth]

well That's one way to prompt things

Needs more balls

[noble coyote]

Amor d'AndroiD?!

[noble coyote]

Are the workflows buried in each png at all?!

[bitter hearth]

ill give you them here Title: Display the title "SpiderMan" in large playful text at the top or center of the poster.
a photo of a SpiderMan , surfing a big wave
Steps: 20, Sampler: Euler, Schedule type: Simple, CFG scale: 1, Distilled CFG Scale: 4, Seed: 1750050164, Size: 1024x576, Model hash: faf2118042, Model: Dream_Diff_Flux_V1, Version: f2.0.1v1.10.1-previous-297-g5fb67f49, Module 1: t5xxl_fp8_e4m3fn

[noble coyote]

My Flux w/f has no place to change the CFG?!

[bitter hearth]

yes your using comfy im assuming

I run in forge UI

faster and i prefer the setting

[noble coyote]

OK, will give it a go in ComfyUI ...

[bitter hearth]

yes show us the results of comfy output bro

with text

[bitter hearth]

webui doesn't make a difference in outputs

[noble coyote]

It will handle noise differently to ComfyUI

[bitter hearth]

you could however apply different techniques but that's apart from the model that gives same results regardless

[bitter hearth]

it does as the setting are different in forge

[bitter hearth]

because how those webui have their own default settings not otherwise

some of the forge settings you wont find nodes for flux in comfy

[noble coyote]

I often re-do my favourite prompts (from ComfyUI) in A1111 - and the results are surprising. It is not the same; but very creatively different!

[bitter hearth]

sure i know, but that's nothing to do with the model itself

[bitter hearth]

I never said anything about the model itself tho

auto111 for example falls behind all the updates features comfyui adds

the webui can do same stuff given the devs put in their features in

[bitter hearth]

dude

and what im trying to say is ... webui doesn't justify what model can do

[noble coyote]

[Is A1111 quietly being retired?]

[bitter hearth]

i dont use a1111 anymore, idk if they are retired

[noble coyote]

Simpler GRadio UIs like Fooocus and Forge ...

[bitter hearth]

im loving comfyui a lot

[noble coyote]

Flux.DD

[bitter hearth]

nice anatomical pose

[bitter hearth]

looking awesome on flux hyper bro

[noble coyote]

All from @bitter hearth spiderman prompt

... going to aqdd some LoRAs ...

[bitter hearth]

heh, i made these two wallpaers out of curiosity for img2img .. but these images are all AI created then put together using a florence2 workflow on comfy

[bitter hearth]

the one thing that i love about comfyui is the tremendous flexibility in it's workflow to do all kinds of interesting stuff

[bitter hearth]

Title: Display the title "SpiderMan" in large playful text at the top or center of the poster.
a photo of a SpiderMan , surfing a big wave, night photography, nocturnal beauty, city lights, starry skies, celestial wonders, moonlit landscapes, urban glow, capturing the essence of darkness, ethereal atmosphere, dramatic shadows, magical ambiance, long exposure techniques, expert use of light sources
Negative prompt: bad anatomy, comics, cropped, cross-eyed, worst quality, low quality, painting, 3D render, drawing, crayon, sketch, graphite, impressionist, cartoon, anime, noisy, blurry, soft, deformed, ugly, lowres, low details, JPEG artifacts, airbrushed, semi-realistic, CGI, render, Blender, digital art, manga, amateur, mutilated, distorted

[bitter hearth]

brillaint work too Dice

[bitter hearth]

cheers bro

[noble coyote]

ComfyUI's modular design has got me interested in "Touch Designer" - python-based and modular as well

[bitter hearth]

nice nice

[noble coyote]

The TM added itself!!!

[bitter hearth]

yes I trained the model with set images

[noble coyote]

"Not" Spiderman!!! 😄

It omitted the text!

[errant dust]

that's what guidance is

[noble coyote]

I have Basic Guider

But no sliders or values to change 🙂

[noble coyote]

Gosh - I have the BG but not the FG - I should add it!

Thanks

[bitter hearth]

lol

[errant dust]

@noble coyote workflow is embedded

If I train a LoRA on black and white comic art, will the model be able to fill in color, or will it just switch to BW?

ex:

[bitter hearth]

lol

dev versions would have guidance effect but not schnell btw

a photo of a SpiderMan , surfing a big wave, night photography,
Steps: 20, Sampler: Euler, Schedule type: Simple, CFG scale: 1, Distilled CFG Scale: 4, Seed: 2222205184, Size: 1024x576, Model hash: faf2118042, Model: Dream_Diff_Flux_V1, dynthres_enabled: True, dynthres_mimic_scale: 7, dynthres_threshold_percentile: 1, dynthres_mimic_mode: Constant, dynthres_mimic_scale_min: 0, dynthres_cfg_mode: Constant, dynthres_cfg_scale_min: 0, dynthres_sched_val: 1, dynthres_separate_feature_channels: enable, dynthres_scaling_startpoint: MEAN, dynthres_variability_measure: AD, dynthres_interpolate_phi: 1, Version: f2.0.1v1.10.1-previous-297-g5fb67f49, Module 1: t5xxl_fp8_e4m3fn

with DynamicThresholding (CFG-Fix) Integrated

get that setting in comfy with flux... NOPE lol

[noble coyote]

Nice

[errant dust]

SO one lesson learned is that Claude is way better at describing images than ChatGPT

as in: Claude can see things ChatGPT is blind to

[sacred jewel]

Sort of... yes and no. I collect LOTS of prompts and dump them into a huge file. Then using a text randomizer, I pick a prompt at random from that list. I then usualy queue up 25 or 100 batch overnight or during the day and hunt for seeds. Some times I run them through LoRAs or not.

At the end of the day, I am NOT INTO generating a Tit or a Dick or an Ass. I am more into art for the sake of art... almost any topic or subject as long as it looks cool/good... that's my ultimate "thing" if it makes sense.

[bitter hearth]

I love messing with there LLMs like removing there admins lol

[errant dust]

Look, I fed this simple, BW imaghe to ChatGPT4

in terms of what is in it, it is fairly straightforward. No tricks

[bitter hearth]

https://youtu.be/KpycTXnRgNg?feature=shared

[errant dust]

You are not getting it

It has nothng to do with censorship

in that image, ChatGPT sees a man standing NEXT to the horse.

I even asked it if it was not sure the man was not sitting on it

nope. he is standing NEXT to it

The image is a black and white illustration depicting a medieval or fantasy scene. In the foreground, there is a figure wearing armor and a cape, standing next to a horse. The background features stone architecture, with stairs leading up to arched doorways and people milling about. There are carts and what appears to be market activity. The scene suggests a bustling town or city environment from an era reminiscent of sword-and-sorcery tales.

That is ChatGPT's final word

I then fed it to Claude 3.5, and it got it all right, in enormous detial it must be said

The image is a detailed black and white illustration in a comic book or graphic novel style. It depicts a bustling street scene in what appears to be an ancient or medieval Middle Eastern or Central Asian city.

In the center of the image, a lone horseman sits atop a horse. The rider is wrapped in a cloak and appears to be the focal point of the scene. The horse is well-drawn, showing muscular definition and a realistic stance.

Around the horseman, there's a lot of activity:

To the left, there's a merchant or vendor sitting behind baskets of what look like fruits or vegetables. He seems to be organizing or sorting his wares.

In the foreground, there are more baskets and containers, suggesting a marketplace setting.

The background shows elaborate architecture with arched doorways, domed structures, and intricate designs on the buildings.

There are other people visible in the scene, mostly in robes or loose clothing typical of the depicted era and region.

The overall composition creates a sense of depth, with detailed foreground elements leading to the central figure and then to the architectural background.

The style of the drawing is highly detailed, with extensive use of cross-hatching and shading to create texture and depth. This gives the image a rich, visual complexity that invites the viewer to explore the various elements of the scene.

[sacred jewel]

poor couple has no idea what's gonna happen to them in 3...2....1... 😛

[errant dust]

So I will repeat. Claude is ten times better at recognizing and describing an image

[noble coyote]

Prompt Injections - frightening!!!

[noble coyote]

Do you have a Claude w/f available? 😄

[sacred jewel]

Let's use this to fine tune Flux in half a second 🤪

https://x.com/elonmusk/status/1827981493924155796?s=46

Elon Musk (@elonmusk) on X

Video of the inside of Cortex today, the giant new AI training supercluster being built at Tesla HQ in Austin to solve real-world AI

Twitter · Aug 26, 2024, 8:08 AM

[errant dust]

I do not. I don't even have a ChatGPT w/f set up. I was planning on just using ChatGPT to create descriptions, but when it bombed on this, I realized it would be useless. Curious, I decided to see if Claude or another might not do better

I didn't even have an account with Claude

[errant dust]

Is that the barbaric cousin of Spock doing a mind meld?

😄

[bitter hearth]

love the graf art

[noble coyote]

Using Flux.DD with prompts crafted by (inter alia) Claude-3, Gemini-1.5

[sacred jewel]

thew last two are #AwesomeSauce

[noble coyote]

The w/f is embedded

I used the term "tensegrity" in my prompt - it does that cage/lattice style

[noble coyote]

Arpillera Stylee

[sacred jewel]

Curious anatomy decisions 😛

[noble coyote]

Metallic LoRA?

[sacred jewel]

Yup, your workflow, not so stock 😛 (modified it to my needs and naming/folder structure etc.)

swapped out the illustration LoRA for Giger

[noble coyote]

Yes, use and reuse is the best policy!

[noble coyote]

Jake-the-Pake LOL

[sacred jewel]

My main workflow is exactly the same except it has a TON OF EXTRA JUNK that I can wire in as needed. Plus it generates 5MB PNGs with all the random prompt collections i have it them lol

[noble coyote]

Yes, your SD Desktop is like a palette - u can just connect up/disconnect - as the mood takes you!

[bitter hearth]

your imge gen styles are generally pretty outstanding, i wonder how you do them

[noble coyote]

YFG might be using multiline prompting - but I won't second-guess him! 😄

[sacred jewel]

Thanks for the kind words...

It is very simple actually... just have a big text node with a bunch of prompts collected from the interwebs... I then pull one from it via a randomizer and generate... As you guys say, prompting is VERY important... agreed.

I do a LOT of seed hunting and random prompting 😛

SD3 was never as satisfactory with first gens and always needed upscaling/refining... Flux has made it such an easy taks of getting great results on first pass...

[sacred jewel]

nope, you're spot on

[bitter hearth]

you make it sound simple but i always thought it required complex workflow .. but yeah i'd like to get an insight with examples maybe

[bitter hearth]

hey bro you might like my iputs if you use that style . tell me what you think over 270 prompt helps https://drive.google.com/file/d/12WJrsnMsVX3qH-fqeZDUxm6bqIEShNZA/view

[bitter hearth]

another killer aesthetics

just add that file to you main webui and name it styles .csv

[bitter hearth]

very nice idea and composition

[noble coyote]

Multiline text also pulls random lines of text and concatenates ...

[bitter hearth]

hey bro you might like my iputs if you use that style . tell me what you think over 270 prompt helps https://drive.google.com/file/d/12WJrsnMsVX3qH-fqeZDUxm6bqIEShNZA/view

[bitter hearth]

so you guys are focusing on prompt detail alone? or are there any other complex workflow involved?

[sacred jewel]

But bulk of real work happens on top right group. I am just LAZY AF to clean up.

Why? because my OCD goes off scale and have to have pixel perfect alignment and groups etc. and waste 100 hours on that instead of creating art... lol

[noble coyote]

I am more and more impressed by prompts which have been worked on by an LLM - is why I like Fooocus so much

[sacred jewel]

Let me create a cleaned up WF and share it... stay tuned.

[noble coyote]

And I like Ollama-based workflows - just need a Clarence-based w/f to complete the set!!!

[bitter hearth]

sure thanks, would love that

[bitter hearth]

interesting, i never thought of integrating LLM into the image gen .. is that like extension you use inside the webui?

[noble coyote]

Never underestimate the huge difference between a prompt (which is a list); and a prompt which uses natural language!

[mortal mesa]

yes join Ollama crew

[noble coyote]

I have some Ollam-based workflows - I shall post one ... wait up!

[bitter hearth]

you can do that with ogaabooga open auto1111 or forge and api your local address and link then you can have you llm reply with images you request

well im fascinated by this concept, so with flux its more feasible to apply the dynamics of LLM im guessing

[kind siren]

Hello guys!!! Is this the right place to ask a question??

[bitter hearth]

going to look into Ollam thing . thanks for mentioning

I use the new vacouna that uses less token and replies with min 1000 words

[noble coyote]

Ollama+CreaPrompt

[bitter hearth]

so you can use that method generally or is that for specific image styles only?

incorp the coqui TTS and your LMM will talk live to you with any voice you sit inside even cloned

im not very fond of using styles tho, but im curios about worfklows that can leverage LLM

[noble coyote]

The Pop Art style is in the prompt - it is essentialy an image-to-image w/f - so the input image will have also been Pop Art stylee

[bitter hearth]

ok, and you are doing that with LLM tools

[bitter hearth]

Pop art illustration style, vibrant and bold colors, uses imagery from popular culture, incorporate elements of mass media and consumerism, celebrates the mundane and everyday objects

i'll be honest i kinda got bored pretty quickly by using styles files with different styles added to it by users, i'd rather prefer the LLM approach to do wild things

[noble coyote]

Ollama is host to a stable of LLMs - just choose the one or ones you intend to use

[bitter hearth]

ok thanks, gong to check em out.

in the middle of gaming atm lol .. but gonna check soon

[noble coyote]

Ollama and SDXL

[mortal mesa]

i like this one, but ive actually taken it out of my workflow and copy paste cuz i want to edit output otften

[noble coyote]

Ollama

[bitter hearth]

https://github.com/ivanfioravanti/chatbot-ollama or https://github.com/ivanfioravanti/chatbot-ollama

[noble coyote]

Yes, that is something like my Ollama w/f as well

[mortal mesa]

all of my images are Ollama with Llama3

[bitter hearth]

https://github.com/oobabooga/text-generation-webui

[noble coyote]

Ollama

[bitter hearth]

Wizard-Vicuna-30B-Uncensored-GPTQ

WizardLM-13B-V1-0-Uncensored-SuperHOT-8K-GPTQ

Those are 2 LLMs I use

[noble coyote]

I'm trying gemini, clarence, llava 2 and 3 etc

[kind siren]

Newbie here!

Just installed ComfyUI and SD3.

Im trying to load an Image and change it (img2img):

What do I have to do to be able to Edit my image????

[bitter hearth]

That is my joker clone voice that i run to forge with flux hyper

[noble coyote]

Edit? Like Photoshop editing?

[kind siren]

I mean like... change the image into a painting by davinci

for example

or like "make a cartoon of the image in the style of so and so

[floral gate]

The issue is that you need to connect those two connectors on the top load image node, but I dont know how to do that. Search up a img2img tutorial on youtube and you should have it working really quickly

[noble coyote]

write make a painting like da vinci - in the text input box

[floral gate]

Cause I can tell you need more nodes just to add that in

[kind siren]

I got this error

Remember: I want it to take my LOADED image... a photo of me I loaded and then change it

[floral gate]

Yeah cause you need more nodes that you'll have to watch a YouTube video for so that it connects to the rest of the nodes, making the img2img actually work

[bitter hearth]

thats with my lightning model checkpoint and my LLM cloned joker TTS

[kind siren]

I want to be able to load any image and change it

[noble coyote]

Flux.DreamDiffusion

[bitter hearth]

here is a basic guide https://youtu.be/Zteta2_JvdA

[noble coyote]

This is so sexy!

[kind siren]

Does that work for sd3?

I have another question: Do I have to always drag and drop a Sample Image for this thing to work?

[bitter hearth]

not the exact worfklow but you get the idea to incorporate it

[kind siren]

My Model only works when I take a sample image and then drop it there and then it can now start working. So I have to do this process everytime I want to use this thing

[kind siren]

it's this the way it's supposed to work ?

[bitter hearth]

for img2img you'd have to load up an image, not sure what you are asking, but yea you could drag and drop or select from folder into node

[sacred jewel]

Here you go. I will also post an image with the workflow embedded... (The above image does not have it embedded but the text file is the JSON of the workflow)

[bitter hearth]

LLMs with checkpoints work great if you set them up perfectly

[bitter hearth]

aaaaa, i dont have comfyui open atm, im gaming

talking in between event breaks

flux upscale workflow if you like it

flux video to video workflow

flux img2img workflow

[kind siren]

Im talking about two subjects here. (Take into consideration I just downloaded this like 30 minutes ago so my knowledge is zero):

Issue 1: From my understanding, to be able to generate my own images, I have downloaded an image I took from a tutorial. The tutor instructed me to take that downloaded image, and drag-and-drop it in my ComfyUI. After doing this, the thing starts to work. I can now create any image I want. Is this drag-and-drop thing something you must do every time you want to use SD3?

Issue 2: How to load my own images into SD3 and use a prompt to change them. I also want to know about this

[bitter hearth]

flux IPAdaptor batch workflow

[sacred jewel]

The below image should have workflow embedded. I made comments in notes to hopefully explain how I use it.

[mortal mesa]

no, some images have meta data of what it used in comfy to make, so comfy will read that and load it

[bitter hearth]

And my fav of all IPIV Morph Img2Vid with animateDiff

with morph

[kind siren]

Anybody knows how to Load and change images with a prompt on SD3 with Comfy?

Here's my Setup:

[bitter hearth]

IPIV Morph Img2Vid with animateDiff

[kind siren]

I'd like to load a photo of mine and change it with a prompt

[bitter hearth]

flux upscaler with 3 upscale options

[mortal mesa]

load image into vae encode into the Ksampler than lower denoise in the sampler

[kind siren]

So First I have to put an image loader right?

[sacred jewel]

Try here first.. .these are usually good enough to start with.

https://comfyanonymous.github.io/ComfyUI_examples/

For img2img whcih is what you call what you want to do, try this:

https://comfyanonymous.github.io/ComfyUI_examples/img2img/

Looks like this:

[mortal mesa]

ya like that

[kind siren]

Ok so I downloaded that image, and I dragged-and-droped it now have the workflow!

[bitter hearth]

1 min to render 3 upscaled images with that workflow for flux

[kind siren]

Ok so I downloaded the image and dropped it and now I have the same workflow....

However I don't know if it's working for sure?

I told it "Painting in the style of Da Vinci"

But Im not sure if the image it generated has anything to do with the image I loaded

[noble coyote]

Load Image - then change node to widget - double-click left top - and an automatic counter should show up - feeding a new image for every prompt in the queue

[bitter hearth]

clarity in that image

[noble coyote]

Male, facing camera, mustachioed - seems that the outcome is quite close?

[kind siren]

Which node should I change to widget?

[sacred jewel]

Play with the Denoise level. 1.00 denoise will completely generate a new image. Start with 0.35 and move up as you generate so you can see the differences

[kind siren]

Ill try that

[noble coyote]

Load Image

[mortal mesa]

where lol, i see blurry

[bitter hearth]

lol ok

[mortal mesa]

looks good small, i wouldnt call it clarity at full size

[sacred jewel]

And GO SLOWLY... make SIMPLE CHANGES and generate and log your results. Making too many changes at once when you are starting out will connfuse you ...

[kind siren]

How do I change it into Widget?

[bitter hearth]

how are you viewing it in full size tho?

[kind siren]

Ok, for now I will just change the denoise

[mortal mesa]

open browser

[sacred jewel]

WAIT before doing so... just focus on denoise for now.

[bitter hearth]

lol that wont show full image thats compressed

[noble coyote]

Convert Widget to Input? (Double-click on the one showing first ... ?)

[sacred jewel]

@bitter hearth thanks for the prompt collection. I am now generating 270 different gens from them 😛 😛 😛

[mortal mesa]

ok you posted it...

[bitter hearth]

lol no probs bro

[kind siren]

Ok, it's working now. I reduced the denoise and it altered the image. So it's working

Like, I can see now that the loaded image has been altered

[sacred jewel]

Awesome...

[noble coyote]

The Marc Chagall effect, aligned with tensegrity and fibonacci

[bitter hearth]

you would need to save the image and open in your pc for a non compressed version. The image is 4k and i can tell by your browser comment your monitor doesnt have 4k cabability

[mortal mesa]

ya i did, at 100% it is not crisp

[kind siren]

Ok, so now that I have this setup working. How can I take a photo of mine and make it look very cool?

[mortal mesa]

you can move on

[kind siren]

like what changes can I do to make it look awesome, what things do i have to change

[noble coyote]

(Put it in the fridge?) 😄

[bitter hearth]

you would need to save the image and open in your pc for a non compressed version. The image is 4k and i can tell by your browser comment your monitor doesnt have 4k cabability

[noble coyote]

Just joking 🙂

[kind siren]

hahaha

[bitter hearth]

your monitor by detection is a 50htz 1080 mate

[kind siren]

I told it to take my image and put medieval clothes on it:

[mortal mesa]

soo you normally an ass or something i guess, move on clown, you know nothing of what i have stop trying

[kind siren]

But the result looks kinda shitty!

[noble coyote]

More 'primeval' than medieval 🙂

[bitter hearth]

you said it was clear . Was educating you to why it wasnt . Your welcome

[kind siren]

hahaha yes! But what things do I have to change to make it do more amazing stuff, to make it more realistic ?

besides changing the denoise thing

[mortal mesa]

more than happy to agree my and your version of clear are diffrent

[noble coyote]

Use words and phrases like photographic, cinematic, 4k, octane render, hyperreal, realistic etc etc

[bitter hearth]

yep and your potatoe pc is very different to mine to

[kind siren]

Ok I will try that

[noble coyote]

(I love Potato PCs - I set mine to high - and by the time I get back - the fries are ready!!!) 😄

[kind siren]

photographic, cinematic, 4k, octane render, hyperreal, realistic, 16th century aristocratic clothes, castle in the background

[bitter hearth]

😆

[kind siren]

Ok now it's working slightly better

[bitter hearth]

Only pulling ya leg bro

[bitter hearth]

LOL this is incredible

[bitter hearth]

I got carried way sorry

[kind siren]

So I have question.

I have these images that I can now use to drag and drop unto comfy and get the worlflows ready.

Is this like the official way to do it?

Do you guys do something like this everytime? Or have you created your own custom workflows and saved them

[noble coyote]

Looks like Madame Tussauds!!!

[bitter hearth]

I haven't tried audio or video ai yet it sounds fun

[kind siren]

hahahaha

[bitter hearth]

I do use the 11labs app to read articles