🆕｜sd3

[noble coyote]

Dev.Bnb.Nf4 is good at 30 steps

[sterile pendant]

i know, i just didn't switch back earlier after making examples for here

[noble coyote]

Flux.Dev is good at 20+ steps - but takes a month of Sundays!

Is Schnell at 35 steps really sharp?

At 4 steps, Schnell is good, but a tad soft!

NF4

[sterile pendant]

yeah a little bit of haloing was happening as well, almost like having too high of a cfg, but they aren't terrible

like here was one of the accidents with schnell at 35 steps

[noble coyote]

When I use Schnell, I add an SDXL-out for the sharpness

[urban totem]

is flux NF4 20 steps?

[noble coyote]

Good realism

[noble coyote]

I'm using NF4 at 30 steps - which is good quality. You can try NF4 at 20 and see what gives ...

NF4@30steps

[sterile pendant]

i recommend 25+ since it's working with less precision

personally, i use 35 steps with dev nf4

[noble coyote]

I will up my steps to 35 and see what happens

[sterile pendant]

probably not a whole lot, it will shift things around on an image if you're trying to compare it to a current seed or something

so i wouldnt use a 30 vs 35 with the same seed as some kind of litmus test

[noble coyote]

Just to see if it gets that extra sharp edge

[sterile pendant]

maybe try a prompt like "sword against a plain gray background"

something that you can guarantee a straight edge on to look for sharpness

[noble coyote]

For the separation?

OK...

[sterile pendant]

if you're interested in edge sharpness

[noble coyote]

There are some ControlNets for Edge Sharpness ...

[sterile pendant]

meh, i do all that shit in post if i really care

[noble coyote]

NF4@30steps

[sterile pendant]

99% of what i do with diffusion is simply brainstorming for game assets, so i'm not too picky about having absolute 100% quality

[noble coyote]

As an artist, not having 99% quality is an artistic asset!

[sterile pendant]

i'm an actual artist and sculpter as well, both digitally and with physical media

[noble coyote]

Way to go

I have stuff at fineartamerica, society6, printler ...

[sterile pendant]

i don't view ai generated content as important to me, but it makes for awesome brainstorming

30 vs 35 steps btw

[noble coyote]

In 16 years of 'selling' my art, I think I have collected £300 😄

[sage burrow]

So let me get this right, i can pay Google collab for faster gpus so I can make loras, BUT, I'll probably never get an a100, without trying everyday for weeks?

How is this even something anyone even uses?

[noble coyote]

The 35 steps has a tad more filigree

I think this one is 35 steps NF4

[sterile pendant]

the reflections near the tip of the sword, on the spine, are more accurate on 35 steps as well

but you could probably split hairs between a 1000 images and would likely see a marginal preferential difference of maybe 2% leaning toward 35 steps

[noble coyote]

It's the difference between fine-art and photographic realism ... and always quite subjective. Both versions have their uses.

If you're going to expand and blow images up, then 35 steps is best

35 Steps - there is better detail

[sterile pendant]

Well in theory, more steps is almost always going to be better. If you look at the curve fitting gif I showed earlier, it lets the sampler wiggle tighter and tighter around the data, using smaller steps rather than chunkier shifts. However, if there are flaws in the quality of the data in the actual model, it can make them more obvious. Kind of like if you overfit a lora when training. Things like artifacts in the actual training set can become more visible. Like bad jpg compression and stuff

[noble coyote]

Glad to have you onboard to explain - I'm "colour-blind" when it comes to maths!!!

35

[sterile pendant]

Thanks, well my actual formal education was in engineering, so I breathe math

[noble coyote]

My father graduated 2 maths degrees, and went into civil-engineering to build nuclear power.

I can do long-division! 'nuff said 😄

[icy drift]

All settings constant, just using different steps:
25, 30, 35, 50, and 100.
There are minor changes, even from 50 to 100, but not worth the time. You can get better results by just using that time for rerolling.

[noble coyote]

The more steps, the deeper the d-o-f

[bitter hearth]

if you use something like adaptive Bosh3 then you will see benefits from 100 steps

high steps requires a different type of sampler to see gains

[noble coyote]

Oh! I've turned Flux nf4 to humour ...

35 steps

Prompt = astronauts playing bagpipes, bath tub, light house, jonas peterson, andrea kowch, umbrella, andrew wyeth, harbour, ship, sea shells, crab, lobster, wheels, hat, rain clouds, watercolor, victo ngai, matisse, monet, catrin welz-stein, vladimir kush, henri rousseau

[icy drift]

You might be right about that.

[noble coyote]

Proper d-o-f requires clockmaker precision in detail

35 steps, 16:9

[bitter hearth]

generally at very high steps its stuff like architecture, furniture, doors and windows that get fixed

at least in my generations

I've seen stuff get fixed beyond 300 steps

[bitter hearth]

Now have them breakdancing on a bathtub

X3

[noble coyote]

Pin-tack sharpness, what every pro-photographer raves about!

[noble coyote]

Great idea - just adding ... 😄

[noble coyote]

But the output was usable?

[bitter hearth]

if you use SDE samplers with low they can need a really high number of steps to converge
with euler A at CFG 0.1 (using CFG++) I found that the image wasn't useable at all until after 150 steps

but that's a quirk of SDE

[noble coyote]

I've never really swapped in/swapped out samplers and schedulers - unless I was really convinced.

Ancestral Samplers respond to steps/iterations. Others to CFG ...

35 Steps NF4 16:9

[sterile pendant]

double majored in electrical and mechanical engineering. was going to go for a phd in mechatronics or some other robotics field, but life shit happened. thankfully, there's a ton of overlap with game development. if i can code the logic into a robot arm for something like inverse kinematics, i can do it for a game lol

[noble coyote]

I did a part-time degree in English Literature - figured if I was lying on the beach reading all the greats ...

[noble coyote]

He's not playing the bagpipes; nor grating cheese!!! 🥳

"Can you play the bagpipes in space?!"

[bitter hearth]

I actually started doing quite a low step method these days

12-15 steps of UniPC

with beta sched

[noble coyote]

Flux.Schnell at 4 steps is cool. But it is also soft

[bitter hearth]

I don't really want to run super long generations all the time

[noble coyote]

So I use and sdxl-out to increase sharpness

[bitter hearth]

yeah refine is good

[sterile pendant]

just expect from and treat schnell like you would a lightning model

[noble coyote]

My cousin was a mech-eng - he made pacemakers 😄

[bitter hearth]

there's a Lora made by the Leosam HelloWorld XL creator
that combines Turbo with LCM but not quite as strong
I like that one for 8 steps SDXL

[noble coyote]

NF4 - does it work with LoRAs?

[sterile pendant]

hell yeah

[bitter hearth]

not by defualt

there might be a way to convert loras though

not sure

doing this to flux was kind experimental even according to the guys who did it

(its in diffusers as well)

[noble coyote]

I think that Flux.Dev works with LoRAs - but not dev.nf4

[bitter hearth]

if you want to ever try a really fancy sampler combo

clownshark's stuff is really good

clownsampler and sharksampler

they work together

[noble coyote]

ClownShark's stuff is memorably unique - it takes 11 minutes/image on my 8Gb RTX2070 - but every image is worth it!

[sterile pendant]

yeah i talked about it over in the comfyui section. my guess is due to the weights of the nf4 models being variable like U8/bfp16/f32/etc and the lora needs to be up or down casted every block to modify the model weights

[bitter hearth]

its very slow yeah

Supreme Sampler by Clybius is good too

but this won't work on flow model

like flux/SD3/auraflow

[noble coyote]

I added 'breakdancing' - nf4 seems tame on that!!! 🥳

[sterile pendant]

yeah clown makes some really cool shit

[noble coyote]

The lobster seems to want to though!!!

[icy drift]

30 gens and not a single one with correct anatomy. 😕
Oh well. Time to go to work.

[bitter hearth]

IDK if they will actually pull this off
but according to some of the flow papers
the goal of flow architecture is to make it solvable by a single step of euler
that might require a truly massive model though

the SD3 paper did note that the flow paths got more straight as the model got bigger

[noble coyote]

We've come to a tipping-point: bigger, better, larger, quicker hardware?

Or smaller, more efficient software?

[bitter hearth]

we still need more hardware rly

distillation has all sorts of side effects

turbo / lightning stuff too

I haven't seen a "free lunch"

apart from the node that is literally called free lunch (FreeU)

[noble coyote]

35 iterations@nf4

[sterile pendant]

sometimes it helps if you use the fp16 t5. just add a dual loader, set it to flux and load the fp16 t5 and clipL, connect that to the prompt instead of the one from the checkpoint loader

doesn't always fix issues, but sometimes it helps

[noble coyote]

I'm using @errant dust 's w/f

[sterile pendant]

i experiment too much to bother with using actual consistent workflows, i just keep my shit like a messy desk with a ton of muted node groups everywhere lol. thats why i rarely share images with the workflows attached

[noble coyote]

Flux Guidance: what does it do?

ClownSharks stuff is like that - littered with unused nodes 😄

[sterile pendant]

basically, it just adjusts some of the "signal to noise" ratios of the prompt in the model vs the noise. higher guidance tends to look more burned, similar to CFG, but also adheres better to the prompt usually

[noble coyote]

Maybe that's why text is so coherent?

[sterile pendant]

the flux distills that we can use locally do not have cfgs though, so no real negative prompting. people do some hacky stuff to kind of make faux negative prompts, but it's not functionally the same and has a lot of drawbacks

[sterile pendant]

higher guidance probably makes it do a better job of adhering, but i dunno, probably has more to do with the training and the architecture

we still dont have a paper, so it's hard to say what they did to pull it off

[noble coyote]

Kilted bagpiping cosmonauts!!! 35steps@nf4

afk - lunch

[errant dust]

Vast.Ai FTW

https://decrypt.co/244307/flux-ai-generator-less-powerful-fp8-np4

[bitter hearth]

vast ai is good yeah

or their 20 similar competitors, offering pretty much the same thing

better value than colab pro

[errant dust]

it's now mainstream news in India

[dry wave]

in general yes, but not as Comfy implemented it

[cunning lintel]

Ugh, this all could have been SAI with their 8b model, if they only bothered to release it and not quickly make a crippled 2b one for release :/

[dry wave]

Comfyui injects the Lora weights directly into the quantised model, instead of applying the lora on the fly during inference

the latter (the way diffusers is doing it by default) would allow to use the lora at any quantisation without problems

the way comfyui is doing it basically ends up in a lot of information loss within the lora. A workaround would be to first apply the Lora, THEN apply quantisation, but that would be slow as hell

[bitter hearth]

I'm trying to learn diffusers cos
some stuff is easier in comfy and some stuff is easier in diffusers

[dry wave]

yeah, it's a shame

diffusers fails as an API in my opinion

they abstract the low level components like unets, transformers and so on, but not the high level components like inpainting, cfg, diffusion

so implementing anything in diffusers mean starting from 0 almost

on the other hand many complicated things like lora or quantisation are just one-liners in diffusers which is really neat. Basically, you can use the whole huggingface ecosystem easily within diffusers

currently my workaround is to run Flux in bf16 in Comfyui, so that I can still use loras. But it's not viable option, as it crashs as soon as any other process (e.g. browser) takes more vram

[alpine summit]

Sd3

[cunning lintel]

Diffusers doesn't really do API, they re-implement everything even if it's the same functionality, make it more fragmented than it needs to. But for an easy backend, i prefer diffusers, comfy is pretty quirky too (and i really don't like the extension/custom nodes ecosystem with all their own pythdon deps that break things all the time)

[bitter hearth]

I really love the syntax of crowsonkb/k-diffusion
which is basically what comfy is based on
its so much nicer than the diffusers sampling code

k-diffusion is basically someone taking the Karras et al. (2022) paper
and writing a pytorch implementation
this is the same paper that Karras schedule and Heunpp came from

I have two big struggles with comfy

1. the way it is structured makes it hard to predict clashes if you don't know the library well
2. custom nodes are poorly documented and sometimes get stuff slightly wrong

[sterile pendant]

Because those vary on a per pipeline basis. Not every architecture is going to be compatible with every feature. But they usually eventually get adapted and added to major models. Like someone worked in PAG to sd3 recently, even though sd3 doesn't use a unet

[dry wave]

most stuff works 1:1 in every model

the funny thing is: this is not the case for the low level components

like unets, transformers and so on can be implemented in very different ways

that's why Flux transformer implementation is a few lines of clean and easy to read code

while diffusers transformer implementation is a mess with thousands of options and configs

because it has to match any implementation from flux, pixart, sdxl, kadinsky and so on

but the high level components like inpainting, img2img and so on, which are almost identical in each framework, they are reimplemented for every single model

[sterile pendant]

Right, they get the architecture creators to help and make major decisions. Most of the time they want 1:1 parity with the creator's code

[noble coyote]

35 iterations@nf4

[dry wave]

also that is not a valid argument. In a good API you would make everything modular and easy to adapt. If a component is different between two models, it would have to override this tiny part of the code

[sterile pendant]

It's not a true API though in the sense of everything being modular and interchangeable

[dry wave]

that's my point. It fails as API

it's a nice endpoint system

[sterile pendant]

But I use diffusers a lot, I wrote two apps standalone apps for hunyuan and pixart and there are definitely major differences in how they handle things like embeds

[bitter hearth]

do you think it is viable to do diffusion projects in pure pytorch?

or is that gonna be too much work

[sterile pendant]

my portable ratchet set has both metric and imperial sockets in it. i wouldn't call it a poor set just because i don't have a metric size that fit's a common imperial size, i'd use the imperial size for it. but the ratchet, extender and case are compatible with all of them.

the problem with the whole AI world in general is that there are no actual standards or unified ways of doing things. you have a million different models doing different and similar things, but there's no standard for how they are supposed to go about doing them. most of them do try to use some of the similar backends like pytorch, python, some similar libraries, etc etc, but there's no real consistency in it.

[noble coyote]

Since Flux - the PiXart room has gone 'horribly quiet'!

[dry wave]

again, I don't say diffusers is bad in itself. It's just bad as an replacement for ComfyUI. Not because an UI is missing, but because a proper API is missing

[bitter hearth]

yeah its a big shame because the pixart guys gave true open licenses

[dry wave]

of course it should be possible to just create a proper API library on top of diffusers. Maybe that would be a cool project

[sterile pendant]

diffusers aren't meant to replace modular apps like comfyui. diffusers are mostly meant for just using the vanilla workflow with a few features. like a basic checkpoint-encoder-ksampler-decoder workflow in comfy

[noble coyote]

That amateur, volunteering, public-spirited world of open licence software - and it is incredible - will necessarily invite allcomers, and so diverge and splinter as all and sundry contribute

And the best open licence software will rise to the top ... where unified standards can start?

[bitter hearth]

yeah I was surprised by this
diffusers is only really about replicating the basic workflow
not experimental stuff

before I looked into it, I had assumed that diffusers was lower level than it is

I thought it was a very thin pytorch wrapper but its not

[sterile pendant]

it's meant for people to easily create a diffuser that can easily be interfaced with things like gradio, if you want and it literally takes like 60 seconds to set up a basic app

[noble coyote]

AllYourTechAI room says that there are some Flux Pro LoRAs ... ?

[sacred jewel]

https://tenor.com/view/do-it-shia-la-beouf-flame-gif-4445204

[noble coyote]

Its normally a blue flame!!! 😄

[sterile pendant]

`import torch
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16)
pipe.enable_model_cpu_offload()

prompt = "a tiny astronaut hatching from an egg on the moon"
out = pipe(
prompt=prompt,
guidance_scale=3.5,
height=768,
width=1360,
num_inference_steps=50,
).images[0]
out.save("image.png")`

[bitter hearth]

nice

[sterile pendant]

from their example on HF docs lol

[noble coyote]

Very good - Little Moon Ice-cream?!

[sterile pendant]

but pretty much every diffuser will look like this

[bitter hearth]

ye this isn't what I am looking for
its a nice library though
I'm just gonna use raw pytorch

even if it takes me a while

I thought diffusers was just some small utilities and abstractions
to aide doing diffusion in pytorch but that's not what diffusers is

[sacred jewel]

Is Pro being considered for release or always gonna be API/Paywall?

[bitter hearth]

they might release it in like a year

[sterile pendant]

and then this is what a simple GUI can look like using diffusers

[bitter hearth]

that's a really nice UI

[sterile pendant]

(i wrote it a while back)

[bitter hearth]

I like how the silders are audio sliders

[noble coyote]

AllYourTechAI is minting it via PhotoDojo

So it remains to be seen who will slay the goose which lays the golden eggs?

[sterile pendant]

damn the app was over 600 lines of code, didn't realize it got that fat

[bitter hearth]

I have to admit though
integration / similarity with huggingface transformers is a big plus

I will probably still try to learn it because of that

[sterile pendant]

oh right right, it was all the tk labels and fields and grid placement shit

[bitter hearth]

ah yeah tk

[sterile pendant]

and a bunch of incremental steps to save memory and vram

[bitter hearth]

I always forget the name but I used pyside 2 or pyside 6

[sterile pendant]

like i precalculate the prompt embeds, then unload the tencs

then load the tformer and diffuse

"Max memory allocated: 7.484710693359375 GB"

and that's with the fp16 t5 for this

[sacred jewel]

Is the T5 encoder loaded in VRAM or RAM? Or both?

[bitter hearth]

ah yeah loading and unloading can help

[noble coyote]

I love the wait in Flux as t5xxl_fp16 analyses the prompt - it's like it's saying "I'm really paying attention here!"

[dry wave]

don't underestimate all the annoying things like cpu offloading, memory handling, proper model loading and casting

diffusers does simplify all this stuff

[sterile pendant]

if they have it set up correctly in the library for the model, yeah

but a lot dont

[bitter hearth]

yeah I am predicting its gonna take a lot of work

[noble coyote]

(Is it just me, or am I alone in being the one not going back to SD3?)

[cunning lintel]

all your tech is full of 💩 , their "pro"is dev, the sad thing is they don't need these lies, they're one of the few sites actually offering unlimited dev, not just schnell.

[sterile pendant]

it will look something like this "model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->transformer->vae"" but doesn't always work, especially if something is screwy with the dtypes where they won't allow calculations with cpu and cuda:0 at the same time. i'm sure you've seen this error before in comfy working with stuff from time to time

[cunning lintel]

once something like the sd3 8b api is released, i'll be back in no time, i like its ability for different styles and even its photo's look more real

[noble coyote]

Mebbe, but I regularly use 8b @ClipDrop - and even then I'm not impressed

I guess that reality in SD is not my forté - art survives its own realité

[cunning lintel]

for different style it's good, but i also tried some api prompts from artisan chan, and it was a bit painfull to compare (flux was way ahead)

[sage burrow]

I'm trying to see if I can figure out how to convert a Google workbook to vast. Vast doesn't have any (useful) just click buttons and it works already made workflows 😦

It took me a few days to install everydream2 trainer on my computer, with help from their discord, so I worry that trying to create a Flux lora on vast might take as long and cost a small fortune 😦

[noble coyote]

Yes, I still think that 8b needs refining and polishing

35 steps nf4

Monopeds

[bitter hearth]

apparently only 2B had had full aesthetic fine tuning

[noble coyote]

Jellyfishalicious

[hallow lion]

phenomenal model

im my image generation jeourney flux was the first model that rly blows me away

and we can run it locally damn, what a time to be alive

[bitter hearth]

flux is amazing yeah

[hallow lion]

two papers down the line well all be uploading our brains to the matrix

[hallow lion]

you know what will blow me away hard like this next?

Real consistency

[alpine summit]

Flux

[noble coyote]

Thought it was ClownShark for a moment!!!

nf4

[hollow swift]

how long it takes you to render nf4?

[bitter hearth]

She needs more clothes

[radiant quiver]

Can stable diffusion

Do img2img

And make a nice skybox

[sterile pendant]

You'd need some kind of controlnet to handle the wrap points. I know there are cnets to do stuff like making tiling textures, so it's definitely possible.

Though in games, we don't use that style of cube mapping anymore really and if there's one in that format, we convert it to a normal panorama style

[radiant quiver]

For skybox, thats the only template

[radiant quiver]

Where can i find such controlnets

[bitter hearth]

there is some company out there that makes vr skyboxes with diffusion

closed source though

but it shows its possible

[noble coyote]

1024x1024 about 2 minutes and 10 seconds

nf4@35 steps

[hallow lion]

So the realism lora for flux is 20 megs...

[noble coyote]

None of my Flux LoRAs is larger than 445,804 Kb

[noble coyote]

They don't seem to work that well - unless I'm using them wrong?

nf4@35 steps

[hallow lion]

They don;t at all in forge. I think they work in Comfyu. Will try in a bit, it seemed to do soemthign yesterday.

[brittle nexus]

Loras don't work with the NF4 version on Forge

[noble coyote]

Lurv!

Red Knows Day

[errant dust]

So Flux seems to have strong guardrails against art styles of any kind, whether comic art or classic art such as impressionist, surrealist, and so on. At a sufficiently low CFG it breaks free of them very slightly, at a real cost in adherence too, but not enough to actually show it can distinguish between styles. That said, if the ability to truly train a checkpoint or fully supported LoRAs come out, people will be able to build their own preferences. Since that was a defining feature of SD in the past, one can only presume and hope this will be true of Flux moving forward too

[rain current]

https://civitai.com/models/640459/impressionist-landscape-lora-for-flux?modelVersionId=716306

[errant dust]

I'm totally aware, but sadly unless I accept 45 minute renders, I need to wait for some kind of support with NF4

[noble coyote]

I find it hard to make Flux LoRAs work at all - do they need Dev or Schnell? NF4?

[errant dust]

Plus as I understand, the LoRAs don't really have full effect like classic SD LoRAs.

[noble coyote]

NF4 35 iterations

[errant dust]

I suspect the Flux team will eventually open the doors to more understanding of what is needed at which point we will witness a tsunami of such

[errant dust]

Dev only. NF4 and Schnell won't work, though for different reasons

[noble coyote]

It'll take me forever using a LoRA in Dev 😦

[errant dust]

LoRAs require guidance control which Schnell does not have, and NF4 due to the way it is made, is a small nightmare to make LoRAs

So it is Dev alone

[rain current]

Ok, I went back to fp8, because in my case it takes the same time as nf4

[errant dust]

I assume you have a lot more VRAM than I do though

I am the pauper of paupers for this use case. Meaning 8GB of VRAM

and even then I have to regularly restart Comfy to flush out the memory

[noble coyote]

I too am an 8Gb VRAMmer

[rain current]

In my case, it consumes the same. I have 16GB

[errant dust]

right, a different use case

with fp8 you can no doubt fit it all onto the GPU memory

[noble coyote]

Where does the Flux LoRA go? After model ...?

[errant dust]

Load the checkpoint, connect the Model node output to the LoRA and then on to the rest

[noble coyote]

Unet works as well as Checkpoint version?

[errant dust]

Yes

[real terrace]

Are you using Comfy? What workflow?

[errant dust]

Just grab the image, no?

[real terrace]

And which model specifically? if anybody knows, I wanted to try it

[noble coyote]

I'm using one supplied by @errant dust

This is the first Art Nouveau Flux.Dev/LoRA image I have made - quite good

20 minutes in the making

[sage burrow]

I haven't been able to make my flux images look like the style of any particular painter yet

[errant dust]

I strongly suspect you won't be able to at all. Let me show you what "impressionist oil painting of...." yields:

[noble coyote]

Apparently you need trigger words with that LoRA ...

[sage burrow]

What about glif with their comfy feature? Or civitai?

[noble coyote]

Flux.Dev image using Art Nouveau Flux LoRA - quite impressive!

[errant dust]

GLif with Flux Pro is the same

Yes, LoRAs will be the solution.

[sage burrow]

You should see my by Pieter Bruegel the Elder images 😭😱

[dull star]

??

[real terrace]

which one is it? On Comfy?

[dull star]

its absolutely not perfect for humans, but it does work

you need to use the trigger words and take down the strength to about 0.6 or lower

these are all with the lora

[sage burrow]

Flux has never heard of Salvador Dali 😦

[noble coyote]

Flux.Dev using Flux Art Nouveau LoRA

[noble coyote]

Use any png I have posted it contains the workflow

[bitter hearth]

wow art nouveau era

[tough oriole]

Have any of yall trained on tags only? So far my NLP loras have been great but i wanna know if i can use my huge tag dataset. Edit: for flux loras

[bitter hearth]

do you know about pony

[tough oriole]

sorry i meant for flux

[tough oriole]

should have added that

[bitter hearth]

I think that's kinda an unknown question at the moment

[young blade]

flux with style prompt isn't terrible

[bitter hearth]

what is style prompt?

[real terrace]

ok ty, and is this one model?

[bitter hearth]

ty

[young blade]

not as good as a lora for sure tho

[bitter hearth]

I need to learn the good prompting nodes

[real terrace]

for some reason (after doing updates) I cannot get this node

[real terrace]

I git pull Comfy and also Update All from manager

Also restarted Comfy

[sage burrow]

Install directly from comfy. Got pull didn't work for me. Also do that python stiff the git explains

[real terrace]

what do you mean install directly

like re-install Comfy?

[mortal mesa]

most of the time its a bitsandbytes not installed or old version issue

[real terrace]

I think I didnt install that

[real terrace]

I re-installed Comfy and it was the same

[sage burrow]

I mean from within confy," install from git url"

You definitely need the bits and bytes

[real terrace]

what is that exactly?

[real terrace]

what is the url for that loader, or what?

I'm on Ubuntu

[sage burrow]

Here's a guide for nf4 installation if you read from here down. The patient folks helped me get it installed #🆕｜sd3 message

That's probably a bit different, but I figure install via git url from within comfy should still work.
Also the whole add the python stiff should be similar

Just ignore the windows directory structure in the above convo

[real terrace]

I don't understand how to install this https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4

oh I think I found it

[real terrace]

oh boi

[sullen moss]

https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Canny

[real terrace]

I'll try to git pull manually

[errant dust]

Just go to the folder of Custom_nodes in CMD, and git clone it

make sure you have bitsandbytes already installed first

[torn wharf]

the 1024 version nice. hmmm. do i want nf4 comfyui or 1024 controlnet comfyui? is it update day today?

https://github.com/comfyanonymous/ComfyUI/pull/4260 woohoo its mainbranch now

[real terrace]

I'm trying to install that

git clone what?

[noble coyote]

Flux.Dev and Flux Impressionist Landscape LoRA

[torn wharf]

just some advice. you might want to do some basic youtube courses on how to manage your comfyui. it's quite intuitive what he means. git cloning the custom node project into the custom node folder

[sage burrow]

Both! 🙂

[torn wharf]

i've stopped helping people install stuff like that since more cereal experts get mad at me when i try to clarify. so i'll just zip it again. i'm getting into deep water

[errant dust]

The Comfy NF4 node you were trying to install via the Manager

[real terrace]

I did pull, not clone

[noble coyote]

Cezanne inspired impressionist landscape Flux.Dev and Impressionist LoRA

[real terrace]

how to install bitsandbytes?

I need to install it on the venv?

mm I tried but it is installing torch, I don't want to touch that

Also I see it is installing much cuda-nvidia things, I'm on AMD

and it crashed

[noble coyote]

Final Flux.Dev Impressionist LoRA image for the day

[real terrace]

It seems it doesn't work with AMD

[mortal mesa]

i think you are right, i didnt think of that

[sage burrow]

the install via git url button

[sage burrow]

What are your system specs?

[torn wharf]

wait. wait wait wait. that's the bitsandbytes project. not the python compiled module

https://pypi.org/project/bitsandbytes/ this is the more appropriate version to install as a python project dependency

[mortal mesa]

ya needs cuda

[sage burrow]

INstall via git url

[torn wharf]

portable comfy needs a bat file. run it, and it asks you "what pip packages would you like installed" then you type bitsandbytes or whatever and bblamo. ez pip install.

I've figured out how to run the embedded python's pip by now, but i see this advice being repeated and distorted so often that i think a simplified tool like a bat file being available in the root project folder would benefit 10000 newbs

theres the install pip command in comfy manager but it never works i've noticed

security clearance or some

[hollow swift]

where should i pip this

[mortal mesa]

in venv

[hollow swift]

which in comfy

[torn wharf]

you have portable comfyui? open the python_embedded folder in powershell and type ./python.exe -m pip install -U bitsandbytes

[torn wharf]

to which i say, we need a better way of advising the newbs. some people have a venv. some people have an embedded python environment

[hollow swift]

i have portable one, ah thanks for the explanation!

[torn wharf]

dont think 'm using newb condescendingly btw. it's from the heart. we're all newbs

[mortal mesa]

im just a dabbler, i try to help on basic things ive seen myself

but when problem 2 comes along im out 🙂

[sage burrow]

?

[hollow swift]

do you notice any different nf4 on forge and comfy?

[hollow swift]

type on address bar 'powershell'

[sage burrow]

that's diff than regular command prompt?

found it, arg, then I gotta cd 100 times

[mortal mesa]

brows to the folder in windows and copy it from the address bar and paste it in CMD window

[sage burrow]

It was working perfectly yestrday grumble

[sage burrow]

did, but then this

so found powershell, but I put my comfy in way too deep of a directory lol

[mortal mesa]

i dunno i just use the plain CMD window, not powershell

[sage burrow]

nm not important enough, I played with it yesterday, so got my fun

not sure what happened to it

[mortal mesa]

take the slash out?

[radiant ledge]

or .\python.exe

[torn wharf]

thats the python folder yup.

[torn wharf]

just one cd lol. though typing absolute file paths is something i happen to be great at. as you're typing it, tab will auto complete the folder names

also, windows 11 hacks. right click in hte folder in explorer and "open terminal here"

[sage burrow]

I tried it in the correct path, seems you can put cd and pase and entire path these days! (last I used DOS was decades ago lol)

anyways, I got the terminal for it, then got the fancy powershell version, either worked

[torn wharf]

yeah cmd.exe works different. you have to go to the drive letter first

the next step is to add a custom transparent background to your powershell and make it look like matrix terminal letters

[sage burrow]

so I'll just wait until Comfy upgrades to add it. I already installed it yesterday, and it worked fine, but not today

[torn wharf]

also in cmd.exe you don't need to do ./ before executable commands

[sage burrow]

I had it installing but...

[mortal mesa]

how many characters was it than the ~

[sage burrow]

nf4 so not worth it!

[mortal mesa]

i refute this for me

[sage burrow]

it still takes 3-4 mins per image tho! 😦

whereas I can just use glif or HF and only wait 1 min with flux dev or pro

and Flux sucks at nsfw, so I don't even get that advantage on my own system

LOL (from yesterday)

[sage burrow]

Btw, is making flux loras worth it? I'm tempted to try out that Flux google collab lora making. $18 isn't too bad for a whole bunch of loras...

(if I ever luck out and get an A100 that is)

[mortal mesa]

i haven't been that impressed yet, my images are all no lora

[sage burrow]

All the loras I see on civitae all have that slightly faded, slightly blurry look 😦 I don't know if that's a flux thing, or a model maker bad images thing

[torn wharf]

python dependency issues aren't inherent to nf4 node. while i'ts unfortunate that the rollout of this comfyui node has taken the wind out of people's sails, its still really good. I haven't found any situation where i can't get what i want out of nf4. and its faster and lighter weight

[sage burrow]

Can you manage a decent one with male anatomy? If so I'll try harder (no pun intended lol).

[torn wharf]

i think the style loras aren't worth it. i'm not going to bother making a ball lora for flux. i think a prompt template is so much better. 1KB vs GB for the same effect.

the loras that add people though work the best i've ever seen

[sage burrow]

hmm, so character consisency you mean? Or specific actresses, or?

[torn wharf]

lol well, i should qualify that things i might try may be different from things others might try. can you get "male anatomy" out of stock dev?

[torn wharf]

the actors people have bene posting loras for. arnold and the other girl, queens gambit

[sage burrow]

No lol. But also not with stock SDXL, so you never know 😄

[torn wharf]

gotta head out for a couple. be back later. i think today is more importantly 1024 controlnet day instead of play wiht nf4 more day

can't wait to get my hands on ip adapters

[sage burrow]

Is it just me, or does it not know any Renaissance or Medieval painters? <pout>

I'm defi intely going to try that one first 😄

[torn wharf]

i think it does but the distillation has collapsed the "painting" class down to a couple styles. prompt aggressively, fidget with guidance and step count. try ddim or something. fiddle. it's in there. just , how to get it out?

remember. 12B params

bye for now

[sage burrow]

I'm suprirsed though, since SD3 has so many painers!!!!

[torn wharf]

one more thought before i'm out the door. does pro have the painters?

[real terrace]

I crashed in the installation, and maybe I should try to install it in the venv. But I guess it would be same. It seems it won't work on AMD

I don't want to mess my Automatic1111 venv

[mortal mesa]

any particular, ive done some

[sage burrow]

Not last I checked. But I'll try with some more mainstream ones

[sage burrow]

Oooh I want to see pics! 😄
Pietre Bruegel the Elder is one of my faves (SD3 def has him)

[bitter hearth]

looks around

No balls

bad channel

Ball girl

im using dev nf4

making good timing on my gpu

and image qualities are apparently better than schnell

[bitter hearth]

aww shit i clicked one of the spoiler images

thought you posted something sexy but it's all gore

[sullen moss]

I tested upscaling and inpainting with FLUX in Forge today. I don't feel like going back to the SDXL model anymore.

[bitter hearth]

im fast losing interest in pony / sdxl too

[sullen moss]

It's time to free up some space on my SSD. 😁

[icy drift]

Going to a better place.

[bitter hearth]

Powerful cookie

Looks very clean

[bitter hearth]

Make them play ping pong

[mortal mesa]

doesn't count, does it? upgraded to VR

[bitter hearth]

Those are some fancy headsets

Cx

[bitter hearth]

They are playing so fast there are multiple balls being captured in a frame

[mortal mesa]

to make an image you must first train the AI to play ping pong

[sacred jewel]

Cool

[tough oriole]

Damn Depth and HED Controlnets Xlabs is cooking

[sacred jewel]

Guidance 80

[mortal kite]

I've moved on, guys, to Flux. It's everything SD3 wanted to be..

[bitter hearth]

except cool art

[mortal kite]

Flux can actually put drivers in the vehicle

[mortal kite]

nah it's awesome. It does pixel art really well..

and gets hands right almost always

and is amazing with text..

it's literally what SD3 was supposed to be

I mean this is cool isn't it?

[bitter hearth]

I haven't seen any pixel art from flux yet I think

but most stuff looks very generic

like dalle kinda

its consistent at things thats good

[torn wharf]

Pixel art and jpeg artifacts are the one thing I find sd3 slays flux at

[mortal kite]

it works really well for it, I have a couple examples :

I mean..it seems to look decent IMO. I don't know if it's "true" pixel art

[bitter hearth]

jpeg artifacts ? lmao

[torn wharf]

I struggle to get flux to make anything align to a grid like pixel art

[mortal kite]

just getting hands right is worth it

No AI image gen makes "actual" pixel art

[mortal mesa]

searches google for the pixel art foundation rules

[torn wharf]

Yeh like, security cam footage. Some of the results had me checking my export settings on my UI

[bitter hearth]

jesus comes to tell you

[torn wharf]

Even pixel artists are telling each other they're not making real pixel art

[torn wharf]

It's a very opinionated art form

[mortal kite]

it works super good IMO

I just say "an 8-bit pixel art image"

[torn wharf]

Read a twitter thread the other day a guy was saying pixel art was invented after LCD flat panels because we weren't making pixel art on CRT screens. I honestly just think he was being a dolt though

[mortal kite]

well he's technically right. CRT was phosphors not pixels

[bitter hearth]

thats some "ackshually" statement

lmao

[torn wharf]

Monitors had clear pixels

Tvs didn't but my PC 1280x1024 screen had ultra sharp pixels

I don't even want to think of 640x480 monitors or 1024x768. How the fuck did we even manage

Not only that, but video game designers back then made their art with graphing paper and used expensive workstations that were essentially pixel editors.

I remember when pixel art models for sd15 first started coming out. There was an extension to cook the image out in a way that made it pixelated. Pixel artists came out of their caves then too to tell us all that those weren't actually pixelart

[bitter hearth]

really incredible

sometimes I feel like there is little point in using anything other than flux

when I see ones like this

[mortal mesa]

i think about refining them with other stuff than dont do it and make more pictures

[bitter hearth]

when full SD3 comes out it will be more viable

[pseudo owl]

Yeah 8b one will be nice to use, flux will still be better but 8b should be close.

[bitter hearth]

nah flux in the trash bro fr fr

I don't think flux will be better

flux is distilled so much less tricks work with it

and its overtrained on a certain style

even just the 2B is better for realism

[mortal mesa]

sometimes its better every time

[errant dust]

Sounds like the honeymoon is over

[bitter hearth]

whats that prompt ? Theres some meme potential there

I'm really sensitive to looks

like midjourney/instagram look

realistic stock photo is my preferred style

or analogue film

[errant dust]

Look, my honest take? Flux is fantastic. Saying otherwise would be ridiculous. It is simply not the 'second coming' as many wanted. Like all the top image AIs of the moment, the very best ones, it has its strengths and weaknesses

Personally I learn what those are, and use the best tool for the job, and Flux won't always be it. Tis life.

[sacred jewel]

(DARPA,Lockheed Martin,Northrop Grumman,Reaction Engines Limited,DSTL,BAE Systems,Airbus Defence and Space,DGA,Arianespace,OHB SE,BAAINBw,MTU Aero Engines,Energia,Russian Space Forces,Sukhoi,Antrix Corporation,DRDO,HAL,UAESA,CASC,CASIC,PLA Strategic Support Force,Mitsubishi Heavy Industries,IHI Corporation,ATLA:1.5), (NASA,CSA,AEB,UKSA,ESA,CNES,DLR,ISRA,Roscosmos,UAESA,ISRO,ASRI,KARI,JAXA,CNSA:0.5)

[bitter hearth]

yeah this take is fine

[bitter hearth]

https://media.discordapp.net/attachments/194772750957674497/1039982173618438164/1034723981825159228-1.gif

a what now

[sacred jewel]

Hey, I don't make the rules 😛

[errant dust]

https://www.reddit.com/r/StableDiffusion/comments/1erj8a1/flux_full_fine_tuning_achieved_with_24gb_gpu/

[bitter hearth]

I like how trash prompts still make great images

[bitter hearth]

hey I might... just might have as weird prompts don't call them trash

[mortal mesa]

Scientific laboratory setup, high-tech equipment, advanced technology, precision instruments, futuristic design, scientists working together, highly focused attention, neon green GenStone placed at center stage, surrounded by complex arrangement of laser beams, splitting mirrors, lenses, spectrometers, fluorescent tubes, data collection devices, digital displays, lab coats, goggles, gloves, safety protocols in place, cutting-edge innovation, groundbreaking discovery on the horizon.

[mortal mesa]

i just wanted to shoot lasers at something

[bitter hearth]

"giant robot" definitely not a transformer

lol

you could use perpneg to get rid of transformer

might not work without perpneg

[bitter hearth]

I don't blame the model

Giant robots are transformers in my mind anyway cx

Also I'm just using api, none of the noodles for me

[sacred jewel]

Flatulent ballz

[bitter hearth]

wow i love these

this is flux dev?

[vital crag]

Wild. This will give stable diffusion a run for its money now

[sacred jewel]

Yup... I have been exclusively using Dev since it was released...

[torn wharf]

flux is way better at death stars than sd3. sd3 keeps giving them big glass balls where the delfector goes. neither are quite great at it, but they're a lot more creative on flux and can be made all sorts of ways

sd3 can do a lego death star sorta but wants to pixel art it and always that black lens

[torn wharf]

foiled by diffusers again. the new instantx 1024 resolution controlnet is diffusers file. might not even load in comfy if it were a regular safetensor file

[fossil vector]

What are you concerns or pros about SD3? Asking for a friend

[sage burrow]

No pro?

[torn wharf]

very nice

[vital crag]

flux or sd3?

[bitter hearth]

that's flux

[noble coyote]

Flux@nf4

[noble coyote]

When you prompt - be like a Hollywood Casting Director - prompt against type!

If Flux is too refined, use prompt words like distressed, grunge, dystopia, chaos, entropy, depressed etc etc which counter its rather glamorous and glossy look

[rain current]

https://x.com/bfl_ml/status/1823608452096123120

Black Forest Labs (@bfl_ml) on X

FLUX.1 is now part of Grok-2!

Twitter · Aug 14, 2024, 6:31 AM

[noble coyote]

Flux.dev.bnb.nf4

[sacred jewel]

Just dev.

[sterile pendant]

He put out a V2 version of the flux dev nf4 model:

"Update:
Always use V2 by default.

V2 is quantized in a better way to turn off the second stage of double quant.

V2 is 0.5 GB larger than the previous version, since the chunk 64 norm is now stored in full precision float32, making it much more precise than the previous version. Also, since V2 does not have second compression stage, it now has less computation overhead for on-the-fly decompression, making the inference a bit faster.

The only drawback of V2 is being 0.5 GB larger."
https://huggingface.co/lllyasviel/flux1-dev-bnb-nf4/tree/main

From my understanding, what he did is kind of similar to llm quants where you'll have something like a q4_0 vs a q4_k_m or k_s where parts of it are quantized differently

[frigid mango]

Flux is just from another planet. Incredible model with that "extra" that we all were waiting for. Getting interesting, how SD will handle with that.

[sullen moss]

I think SAI needs to start considering a fundamentally new model because by the time the 8B model is released, there will already be a huge ecosystem built around FLUX, with LoRAs, ControlNets, fine-tuned models, and so on. And with new compression formats, FLUX is now accessible to a larger number of users. In general, SD is in an unenviable position.

[sterile pendant]

testing the V2 nf4 dev model, yeah the precision is a hair higher. the overall quality is still very similar, but random artifacts in things like backgrounds seem much less common. the speed is a hair faster as well, but not but a massive amount or anything. maybe 5% faster in my case

[hollow swift]

nf4v2

[bitter hearth]

Kinda horror

[bitter hearth]

No more pizza man, they kinda creepy sometimes

[errant dust]

Prompt?

[bitter hearth]

A cow like horror creature with extremely long legs, white background, game art , digital drawing, black and white cow print

I couldn't remember "concept art"

It's what I wanted to use lmao

[hollow swift]

which flux version do you use?

[errant dust]

My guess is this is actually SD3

[bitter hearth]

SD3

[errant dust]

No surprise. Flux cannot produce that kind of comic art natively.

[errant dust]

Honestly, I don't think the new compression formats will have much of an impact on the overall access by users. The number of users of open source models like Flux and SD who actually have the technical wherewithal and hardware to install and run it locally is microscopic compared to the overall userbase. The overwhelming number of users who do use the open source models will be doing so from service providers who have Pro or Dev and offer it for free or fee. And those service providers won't be concerned with Nf4 or the like. Don't get me wrong, I am incredibly grateful for the NF4 builds, as I am a direct and ideal beneficiary, but I also know what is what in the larger scheme of things.

[rain current]

They are going to give you a Nobel Prize

[brittle nexus]

flux, the original image is Cascade

[errant dust]

Frankly, SD3 8GB made no sense financially when it first came out, nevermind with competition by Flux now. Via API (and I think their subscription Assistant with credits yields the same thing today) it amounted to 190 images total for $20. For that amount a month at Ideogram I get that many renders (with 4 images each) per day, same for Dall-E 3, and MJ gives you literally an infinite number of renders per month for $30/month ($24 if you buy the yearly plan). Is SD3 so great and overwhelmingly superior to warrant their asking price? Never was.

[bitter hearth]

what do you mean

[rain current]

For trying to keep SD3 alive.

[bitter hearth]

its fun to use, flux doesnt do the stuff I like well most of the time so its not worth using >.>

[errant dust]

The biggest irony about Stability's Assistant subscription service is that all the other services, or rather techs, embedded are actually free to download and install. The image generator is the one they keep under wraps as if this were the real magic sauce and must not be 'given away'. The truth is I also use online sites for.... Flux! Despite having it running and working here. Why? Because they have hardware I don't and can render in seconds and without hogging my machine. Free renders in Flux Pro.... what's not to like? I give up some of the control I get locally, so there is a price, but one that I am sometimes willing to trade off.

[errant dust]

Amazing

[bitter hearth]

the ultra model pipeline has something special

as far as I can tell

significantly better than vanilla SD3 8B

[errant dust]

there is only one single 8GB offering, not two afaik

[bitter hearth]

I mean just whatever comes out of the ultra endpoint

they said its a pipeline

[errant dust]

nomenclature won't change a thing. It is one model. The Ultra being peddled as a name is to differentiate from the 'Medium'

And yes, it has zero relation to Medium. They are trained on entirely different data sets

[bitter hearth]

what I mean is
what's the difference between SD3 8B and Ultra
not the medium, which is 2B

[errant dust]

And I am saying 8GB = Ultra

[bitter hearth]

but it has a separate endpoint and the results are better

[errant dust]

Ah. And you have evidence of this?

THat there are two different 8GBs available and one is better than the other?

[bitter hearth]

its not that its a different model its that its a pipeline apparently

one of the comfy people said that its a comfy workflow

[errant dust]

I think we are not communicating properly. What are the two different SD3 8GB available?

[bitter hearth]

there is only one SD3 8GB model
but there are two endpoints that use the same SD3 8GB model
the regular one just serves the model
the second, called Ultra, does something in addition to that

[errant dust]

Link?

[bitter hearth]

its just on the main stability site

[sacred jewel]

https://discord.com/channels/1002292111942635562/1237461679286128730

[sage burrow]

So what are the benefits of SD3, and the benefits of Flux, over Midjourney? I mean aside from price point 😄

[errant dust]

I found it. And looked at the API details too. Can I be brutally honest? I think the only difference between the 'Ultra' at 8 credits per image' and the vanilla at 6.5 credits per image is.... more steps.

[sacred jewel]

Midjourney is an always learning, always evolving platform vs. a point in time checkpoint. IMO MJ and the like will always be more adaptive, dynamic and able to produce whatever images the developers train it to produce at any time...

We could have the same from SD3 and Flux is there were continuous releases with different training focus ... but with the limitations of consumer hardware... OR they could do the same as MJ and have an API platform that does the same thign that MJ does.

I may be wrong but SD3 API is likely able to do this or is doing this already?

[sage burrow]

Maybe the checkpoints and loras do that extra stuff for SD models

[errant dust]

If you are talking about SD3 8GB (any flavor), the benefit is simply.... variety. It is not better, just different, and that means it might hit that perfect render none of the others gets, which is likewise true of them all

[sacred jewel]

Absolutely... but, again, not very dynamically. Always a point in time locked in set of weights. MJ can technically do this on the fly... but what do I know LOL

[sage burrow]

how about the benefits of flux vs MJ?

[sage burrow]

MJ tends to update about once every 6 months from what I remember.

[sacred jewel]

I think flux brings a new level of quality over previous models. And a seemingly much better text encoder implementation.

[errant dust]

let me give you an example of an output by Dall-E 3 that NONE of them can do even close. Does it make it the king? No, not at all. I can show outputs by all of them, that none of the others can do also. It simply illustrates why there is no universal do it all. Prompt: mug of cappuccino with a world map made from the milk foam and coffee swirls.

[torn marsh]

Ecosystem of custom LoRAs, controlNet's and other extensions. Open weights are just like Open Source. Almost

[sage burrow]

Dalle definitely is the best image wise imo. THough I've heard that they have downgraded lately 😦

[errant dust]

The others all produce something along the lines of: (here is Flux)

[zenith hemlock]

RIP SD3 man 😭

flux does a really great job

[sage burrow]

LOL

But SD3 does ancient painnters so well!

[zenith hemlock]

There are still some things SD3 can do well (I’m DEFINITELY not talking about 2B), but in terms of human anatomy, Flux is almost flawless. Also, RIP Midjourney. That was the first thing it killed anyway 😄

[sage burrow]

I'm hoping the next SD model will also have many flux qualities

[zenith hemlock]

I still have some sympathy for StabilityAI, but they haven't been doing very good work lately

[sage burrow]

Yeah MJ should lower their prices, and offer free tiers now 😉

[zenith hemlock]

Is this flux schnell?

[errant dust]

But don't think this is me say DE3 is pure best. I can show sample output by MJ none of them can match, for a variety of reasons (probably legal ones... lol): I asked MJ (and others) for an image of a large fantasy warrior with a scantily clad princess painted in the style of Frank Frazetta (this was with MJ 6.0, which is now on 6.1): This too none of the others can come close to. It literally looks like a painting made by the former great artist

[zenith hemlock]

I'll try that prompt on my own now

on flux dev* I mean coffee prompt

[errant dust]

And then you have Ideogram which can produce insane things with text and art, hitting it out of the ballpark over and over again in ways none can match. Its typography and variety is unmatched, even by Flux Pro. But that's cool. It is not perfect either, just has different strengths.

[sage burrow]

No luck with SD3 8b on that one? I love what SD3 does with mimicking artist's works!

[zenith hemlock]

my output

[errant dust]

I honestly never tried it on Frazetta. The cost barrier made it prohibitive to really explore much

[errant dust]

That's nice, but as you can see, the map sort of looks like a map texturized with a coffee color. Dall-E's actually looks like it was made from milk foam and coffee

No diss to Flux, since the rivals are no better

[zenith hemlock]

DALL-E is not a standalone model, but a service where many different processes run in the background. Therefore, comparing DALL-E with a single model file might be a bit unfair. However, based on my tests yesterday, flux dev seems to do a better job at understanding prompts than DALL-E

[errant dust]

It is irrelevant. And I am interested in results, not rationalizations or justifications. If one alone gives me the result, that is that.

[zenith hemlock]

Yeahh youre right

Dalle's output looks more artistic

but let me give it a try with realism lora 😄

[errant dust]

I use this as a bit of a litmus test, but let me be super clear: I have PLENTY of horrible fails by Dall-E 3 too. This is simply one that hit all its strengths.

[sage burrow]

Flux pro (via glif) version

[zenith hemlock]

And I believe that FLUX will become much better once fine-tunes are available for it; we must remember that this is a base model. When I tried SDXL last year, I experienced a slight disappointment until the fine-tunes came out

[errant dust]

And since no one else has come close, it has become a useful measuring point

[errant dust]

Yes, i can get great images of the concept, no arguments, but nothing that says: Frank Frazetta

[sage burrow]

flux dev (via glif) version

[zenith hemlock]

man

flux pro is insane

I wish I can use it on comfy 😭

[errant dust]

That is original Frank

for comparison

[sacred jewel]

FLUX DEV first seed

[sage burrow]

SD3 via my glif 😉 a large fantasy warrior with a scantily clad princess painted in the style of Frank Frazetta

[bitter hearth]

my coffee be like

[errant dust]

I got tons of those. Really really tried. 🙂

[bitter hearth]

Now this is beauty

[errant dust]

Interestingly, the closest I got was SD XL with a Frank Frazetta LoRA (yes, there is one):

It actually has the slight painterly style to lend it authenticity

[sage burrow]

brb grabbing a new lora 😄

there's a few! 🙂

[bitter hearth]

[sage burrow]

you been snooping my civitae images again? 😄

[bitter hearth]

I haven't been to Civitai since then

I don't use local so there's nothing there for me

[sage burrow]

how come you don't use SD locally?

[bitter hearth]

He's stuck step bro

[errant dust]

Look, be careful as there are two completely different sets of Dall-E 3. One that is completely censored and crippled by OpenAI, tied to ChatGPT 4, and the other that is offered by MS and has far fewer guardrails. If I ask for OpenAI's version make an illustration of a tree filled with whimsical cats in the style of Keith Haring it will flat out refuse. MS will do it in an instant, and perfectly too.

[bitter hearth]

4gb vram, old amd card, my PC goes boom

[errant dust]

You can try MS's version of Dall-E 3 for free 15-20 times a day (4 images per render) here: https://www.bing.com/images/create/

[sage burrow]

I do kow someoe who does all their images using 6gb vram, but 4 sounds nearly impossible 😦

[sage burrow]

Why on earth would it refuse that prompt? I'm used to more adult ones being refused!

[errant dust]

Because you are asking it to do it in the style of someone's name

[sage burrow]

LOL oh LOL

[errant dust]

And the second you add that name it will tell you forget it

but not MS, which is why I brought up the difference. It has an impact on all the outputs though, regardless of the anme used or not

I tested them both and compared to death

[errant dust]

in the end, I concluded MS's Dall-E 3 had a clear edge over OpenAI's output

[bitter hearth]

Damn

[errant dust]

Try telling OpenAI's version to make an image from: "an Impressionist Cartoon of a tree covered in whimsical cats on the base and branches all drawn in a variety of colors and facial expressions in the style of Andy Kehoe and Skottie Young." It won't. It will choke all the way. MS's version does this:

In this particular aspect, DE3 has no peers IMHO. So this is definitely one of its great strengths

It can be crap for some other stuff though

Original link BTW (and you can see the three other renders from the same prompt): https://www.bing.com/images/create/an-impressionist-cartoon-of-a-tree-covered-in-whim/1-66b5746985304d2e92633339eda9287d?id=1VAi4%2FW3pnBPFKIongJkSQ%3D%3D&view=detailv2&idpp=genimg&thId=OIG3.wh5X.pzQKdk8R4kYXWgc&skey=mUoZ5_u07GI2VYaMOZRhB4Qz3Adqgmo9xkzpORdU9ag&FORM=GCRIDP&mode=overlay

[bitter hearth]

they are stuck in space now look what you have done

[errant dust]

lol

[bitter hearth]

[errant dust]

Cthulhu would be proud

[bitter hearth]

[errant dust]

I really want to see mom

[bitter hearth]

I tried

Now time to eat

[errant dust]

Well, since it was a bird's nest, I had in mind some large cat with flapping wings to feed its brood. 🙂

[bitter hearth]

Have this delicious pizza with broccoli soup and metal bearings

I'm proud of myself

[sage burrow]

I love glif for everything, however, it bens most of my skulls! 😦

[cunning lintel]

[errant dust]

Hmmm. That gave me an idea. First tried on DE3, but will test on others. Might be less difficult for other AIs:

a large pizza in which a world map can be seen made from the various toppings

[sage burrow]

SD3 via glif (it finally let a skull through)

[sage burrow]

SD3

SD3

[urban arch]

Interesting. I tried glif using Flux and any time I used the word "Scantily", it had a fit. Bikini, it would do.

[bitter hearth]

[fossil vector]

What is the sentiment on SD3 and commercialized models versus open source?

[worldly fog]

Hi everyone! New to SD. What are you guys seeing as the pros and cons of SD3?

[noble coyote]

Whatever happened to Sytan at all?

[noble coyote]

SD3 Medium is bad with human anatomy. It is good at photorealism, but distinctly poor with fingers and limbs

SD3 Medium dataset only 2 billion items ... Flux is 12 billion.

SD3 should release their 8 billion dataset into the community ...

But SD3 has been crushed by in-fighting and politics - and dare I say it? - money!

[mortal mesa]

fell into a trust and safety hole

[errant dust]

Hate to say it, but the v2 version of the NFT is a fail for me. While I can squeeze v1 into my laptop 4060's 8GB of VRAM for relatively quick renders, v2 will not fit in no matter how much I try.

[errant dust]

Keep in mind that 2 billion is the number of parameters in the NN, not the number of items in the dataset

[sullen moss]

Dog World 😅

[craggy crest]

the version that was open sourced is a small, unfinished model - so this question really doesn't have an answer

[craggy crest]

you know this how? other than rumors?

[sterile pendant]

me 99% of the time... (flux dev nf4 v2, 30 steps, guidance 3.5)

[sterile pendant]

you can revert it back to the legacy view

and there's a setting for the search box

and it will pretty much be identical to how it was before. been using the new frontend for a while now

[bitter hearth]

the difference is more than just steps, although Ultra very likely does have more total steps
its got a higher amount of high frequency detail and its compositions are often a lot better

it implies some sort of noise injection or multiple passes

[sterile pendant]

I guess twitter/X is using groq with flux now, this could either be good or bad depending on things

[bitter hearth]

the second spaceman one is really good

[sage burrow]

It's amazing at backgrounds, alsl the styles of ancient painters 🙂

[sage burrow]

Which model?

[sacred jewel]

Flux.1 Dev

[sage burrow]

I was actually surprised when it kept doing this prompt: a large fantasy warrior with a scantily clad princess painted in the style of Frank Frazetta. Perhaps it's based on the image output? I did have one go blurry.

SD3 via glif

[bitter hearth]

seems ok

the white point is super low

[bitter hearth]

the lion looks worse in the discord preview if you download and zoom in then it looks okay

[errant dust]

Ok, so like the coffee example shared earlier, the concept of a pizza in which the world map could be seen made from the toppings was won by Dall-E 3. It clearly works with its strenghs so not a big suprise. Here are in order: SD3 Medium, SD3 Large (through Glif), Flux Pro, Ideogram, and Dall-E 3:

SD3 and Flux are fails in my book since I see nothing in the map to suggest toppings

[sterile pendant]

flux can actually handle some abstract shit pretty well

been hammering it with vague prompts

[hallow lion]

So how is Wednesday treating everyone?

[bitter hearth]

Dalle is so incredible sometimes

[hallow lion]

Wednesday is a no man's land. Not as cool as Friday and not as soul crushing as Monday. Wednesday just is. It's kinda OK.

[young blade]

kolors

[hallow lion]

That is indeed colorful.

[sage burrow]

lunch anyone? Flux Dev

[young blade]

overall a pretty bad model, but why not it's wednesday

[bitter hearth]

wow nice its transparent

[young blade]

seems to do ok for the most part, if the image includes a shadow it doesn't do so well
can't wait for Flux to have a negate

[bitter hearth]

is this rembg

[young blade]

yeah

just plugged into flux

[errant dust]

https://www.hollywoodreporter.com/business/business-news/artists-score-major-win-copyright-case-against-ai-art-generators-1235973601/

[errant dust]

“ the product operates necessarily invokes copies or protected elements of those works.”

You have to wonder just how much the judge understands

That's a quote from the judge not from one of the lawyers

[bitter hearth]

ah its just advancing to discovery though

[errant dust]

Of course, and it's interesting to see all of the tech sites claiming that stability is screwed in view of this.

[bitter hearth]

I kinda feel like
with this Supreme Court
there is very little chance of them siding with artists over capital

[errant dust]

Well this is not Supreme Court so let's not exaggerate.

[bitter hearth]

yeah true but this stuff will get there

[errant dust]

It might or it might not. It would take a long time though to get that far. The Supreme Court is not a litigation court at all. It is a purely Appeals court

And the Supreme Court isn't obliged to accept a case at all. If it judges that a case was properly evaluated and has no real appeals value it can reject to weigh in on it

[bitter hearth]

it can reject to weigh in yeah I just don't think it will

I think there will at some point be a strict ruling against AI by one of the lower courts
and then it will bubble up to supreme court and get ruled unconstitutional

they'll just claim Originalism as the reason

[errant dust]

The funny thing about the articles coming out about the refusal to dismiss is that the news sites act as if this were final judgement, and not simply: we will weigh in on it and not dismiss it out of hand.

[bitter hearth]

yeah definitely

they write the article with the title that makes it sounds like a ruling has happened

I was kinda worried at first

and then its just nothing

[errant dust]

exactly, and it is not a little odd. I can uderstand a site such as ... Hollywood Reporter, or even Fox News getting this wrong, but actual Tech sites who understand tech better or should?

[mortal mesa]

they are not all equal

[errant dust]

True, and some are more than little biased on the topic of AI in general

[bitter hearth]

sometimes they get it wrong deliberately

[errant dust]

Yes, Wired for example wld be one

they are vocally against the Image AIs in general

[bitter hearth]

I haven't checked what New York Times view on this is
but I suspect they are anti AI

[sage burrow]

or the enlish language?

[hallow lion]

I like how these models get R2 right but magle C3PO.

[errant dust]

and The Verge once had an editor report about Nvidia's keynote and their ambition to help power AI development to the hilt, and his reply was... not a little hilarious. I kid you not, he wrote he said he had cold sweats as he thought of the end of the world, and all our energy consumption to do this wiping out the ecosystem and how could the Nvidia CEO not see this?

[young blade]

the flux/grok partnership has X saturated with images today, jeesh

[hallow lion]

Man good manners get you nowhere you need to assert yourself like R2. He so hardcore all his lines were beeped out.

[mortal mesa]

i think artists shouldn't show their work in a public setting, someone may remember what it looks like and reproduce it or capture the essence of it

[bitter hearth]

its relying on the word "invokes" a lot
but its not very clear what invokes mean
because famously these models cannot actually regurgitate their training data without an ablation
aside from a few small examples, such as a certain Sony image that got into SD 1.5 over 2,000 times

[sage burrow]

that sells more clicks/advertising

[errant dust]

Maybe, but it was so absurd and over the top it literally left me gobsmacked for a good many seconds

[uncut river]

This doesnt look very good though, from the article: "In a thread on Discord, the platform where Midjourney operates, chief executive David Holz posted the names of roughly 4,700 artists he said that its AI tool can replicate. This followed Stability chief executive Prem Akkaraju saying that the company downloaded from the internet troves of images and compressed it in a way that “recreate” any of those images."

[mortal mesa]

well susceptible people are trained like this

[bitter hearth]

every now and then I take a look at discussions by the other side of the AI art debate
and I see such crazy stuff

[uncut river]

based on this logic, they should kill pop music and pop culture, at least 95% of it

[hallow lion]

🤣

AI will kill us all!

[errant dust]

The funny thing is this: let us suppose indeed it can reproduce ... Woody from Toy Story to perfection, ok? It is STILL not illegal! I cannot use that image for commerical reasons or othger, which is wher3e the copyright kicks in, but it is no more illegal than my hiring an actual flesh and blood artist to make a perfect picture using Woody

[mortal mesa]

tresspass the eyes and ears

[uncut river]

if it's all free and no money involved, you can do a lot of stuff

once money is involved, freedom looses much of its meaning

[hallow lion]

Money is dead.

[errant dust]

I cannot use that artiss's picture of Woody to sell things, or promote anything, and so on, but I can absolutely hire an artist to make pics of Woody, Spiderman, and more all I want, and neither of us is breaking any laws

[hallow lion]

They're just keeping it alive coz everyone is used to it. But it died long ago.

[errant dust]

and it won't matter if he had pictures of Woody and Spiderman, the real pics, right next to him to model

[hallow lion]

by the time youll make your first million itll be worth 300k

[rain current]

Darth maul

[errant dust]

If Darth Maul were the love child of Jessica Rabbit

[hallow lion]

Darth Maul should not have been on Solo

Maul ruined Solo

its the scene that labeled the movie as a dumbass action comic book movie

The day star wars was reduced the marvel level garbage

[mortal mesa]

Jar Jar was the canary in the coal mine

[hallow lion]

He was the key to all of it.

If they could have gotten him right...

[mortal mesa]

i think that girl had a horrible part also i understand she is a good actress, just not in that movie, maybe they wrote the part crap

Portman

[hallow lion]

Natalie?

shes good

with or without star wars

Poor Christian got fked bad tho

[errant dust]

She can be great even, and has proven it more than once

[hallow lion]

Also the kid who playe dthe little boy

he got so fked he stopped acting

bullied to pieces

thats how much people hated the prequals kids

you think they wer elove dno

before disney the rpequels were torn to bits and pieces by everyone

now folks have their nostalgia googles on