🆕｜sd3

[sterile pendant]

on this 2080

seems to be doing alright with quality and artifacts, and is still reliably saying "puddin!" in every image. though it lacks the guidance control with the dev model

[bitter hearth]

"cat howling at the moon"

Just like a cat would, it gave no crap about my prompt

[sterile pendant]

the cat is kind of an optical illusion though, looks like hes turned toward the camera

and howling

but cat and howl aren't likely very close to each other in concept space

just say meow instead

[hollow swift]

nf4 taking so long to render compared to fp8

[sterile pendant]

are you using the right checkpoint loader?

[hollow swift]

yep!

[sterile pendant]

well then something else weird is going on

[hollow swift]

on fp8 it only took 30 seconds for 1 step

[sterile pendant]

what GPU do you have?

[hollow swift]

3050 laptop

[sterile pendant]

and have you updated comfy today?

[bitter hearth]

i set my comfyui channel to Dev then fetched update but i dont see any nf4 loader

[sterile pendant]

install it by clicking the install from git in the bottom right corner

https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4.git

paste that name into it when it asks

[bitter hearth]

Turned back to stare at me look stupid

[sterile pendant]

schnell 4 step nf4 "cat meowing at the moon"

[bitter hearth]

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

[sterile pendant]

lol

"vibrant impressionistic painting. bright vivid rainbow of colors" added to the prompt

[bitter hearth]

oddly when i enter the link in git url via manger it says cant allow for security reason.. so i have to manually git pull https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4.git ?

ofc from D:\ComfyUI\ComfyUI\custom_nodes

[sterile pendant]

oh someone else had this issue earlier. i guess you can edit a config file to allow it. they put the security stuff in there after that asshat slipped in malware to that vlm node months back

lemme see if i can find it real quick

[bitter hearth]

ok ty

Fluf

[sterile pendant]

security_level = normal within your ComfyUI\custom_nodes\ComfyUI-Manager folder https://github.com/ltdrdata/ComfyUI-Manager?tab=readme-ov-file#security-policy here are the options

edit the normal to something else and reboot comfy

like set it to weak for now

and then after you're done, set it back to normal so you don't accidentally forget

it's the config.ini file in that folder

[sterile pendant]

no meow, but still wow

[sterile pendant]

this cat reminds me of the yelling cowboy song meme

[bitter hearth]

taking too long to download this small file.... i wonder if the server is getting very high traffic Downloading bitsandbytes-0.43.3-py3-none-win_amd64.whl (136.5 MB)

[sterile pendant]

yeah a lot of python hosts get slammed with people downloading their millionth pytorch library lol

[hallow lion]

That Flux Boring Lora is pretty good.

[sterile pendant]

i just saw that, i need to download it. i love the boring loras

[bitter hearth]

thanksfully it's only about a hundred mb

[hallow lion]

yeah its amazingly tiny

=0

[bitter hearth]

lol yeah but still downloading

[hallow lion]

it makes even the schnell images more relaistic

[bitter hearth]

oooh

nice

[hallow lion]

they dotn look voerbaked anymore

[bitter hearth]

cant wait to see

[hallow lion]

they look like dev stuff

[bitter hearth]

post some

[sterile pendant]

that's the one thing i dont like about schnell is the lack of guidance. i normally use guidance around 2.0-2.5 because it makes things like pictures not feel so contrasty. the default contrast/saturation of schnell is a little too high for my taste

but it depends on what you're making

[bitter hearth]

yeah i liked the guidance around 2 -2.5 as well the image tone looks nicer than 3.5

[hollow swift]

did you solved this issue?

[drifting oak]

Before vs after lightroom

[noble coyote]

Not yet - I will try a re-boot of my PC shortly ...

[hallow lion]

ah yeah how do i set cfg on flux? that fluxguidance node is very elusive and its ot there in the nodes list with the manager

[bitter hearth]

depending on how you set up the nodes but in general there is no need for cfg as flux has no use for negative prompt

[noble coyote]

Flux.Dev_FP8 - 9 minutes 1024x1024 - 20 iterations - 8Gb VRAM, 64Gb RAM - five times as fast as Flux.Dev

[bitter hearth]

but if you wanna try cfg ... here is a workflow example i put together..

[sterile pendant]

Yeah there are some faux cfg workflows you can use to kind of have a ghetto negative prompt, but it's still not a real cfg

[hallow lion]

When your model is so good fk negative prompts. XD

[sterile pendant]

It has ups and downs. But you can always resample afterwards

[bitter hearth]

thats the idea, there is no function for it with flux

[hallow lion]

Well the need for them has been decreasing...

[bitter hearth]

i tested it tho out of curiosity hence that workflow

[hallow lion]

Remember the short novels for negatives we had to enter for SD15

😄

[bitter hearth]

yeah and also flux doesn't require any quality prompts like high res, best quality etc

its so liberating lol

[sterile pendant]

I always hated db style tag prompting with a passion

[hallow lion]

lol

"masterpiece"

[bitter hearth]

i used to make prompt templates lol.. i still maintain some for pony models but i've cut them short to bare minimum

[hallow lion]

I had a 5000 token one for negative body stuff for sd15. i dont even know if all of it when in due to token limitss

it seemed to make a difference....

[bitter hearth]

do you still use sd1.5 btw?

[hallow lion]

hell ye

but onyl as a fine fine refiner

[bitter hearth]

i see lol ... fingers ..

yes some SD 1.5 models are the best at their niche

i only use sdxl pony and flux now

[sterile pendant]

i havent touched sd1.5 since sdxl came out. wait maybe when that ic-light thing came out, but just to experiment

[bitter hearth]

I not learnt pony yet

I'm using SDXL at the moment

[hallow lion]

some tetxures are extremely highly rfined in the newest sd15 models

like skin detailes

[bitter hearth]

yeah some of the sd1.5 models are pretty good in texture but fingers are not so much

anime is an area where there are some good sd 1.5 ones

[hallow lion]

absolutely

even sdxl hands are sad

[bitter hearth]

oh yeah i love the anime with sd15 a lot

[hallow lion]

<--- there is a reason for this icon

[bitter hearth]

but i've come across some good anime pony models that are pretty similar

[sterile pendant]

the problem with the 1.5 finetunes is they all lack variety in what they can do. i make some completely random shit and don't really work much with humans or characters. so for me, they are mostly useless and i find myself using base models a lot more often than i should

[bitter hearth]

@hallow lion if you are interested

https://civitai.com/models/316174/copycat?modelVersionId=490480
https://civitai.com/models/402492/hadrian-delicexl?modelVersionId=530985

sdxl hands are ok now if you are willing to add SEGS detailer and meshgraphormer to workflows

[sterile pendant]

the gooner crowd always seem to ruin models and it creates a vicious circle of further and further refining them to be better and better at making waifus and shit

and that's what sd1.5 became basically, that and anime pron

[bitter hearth]

there is a second issue which isn't as bad
which is models getting more cinematic and contrasty
but too far

[sterile pendant]

sdxl has basically became the same thing as well, but there are still some good models that aren't overcooked on concepts

[bitter hearth]

@noble coyote this is my working workflow for bnb nf4 ... if you wanna load it up and see if you can update anything missing

if that doesnt work, rtx 2x might not be supported

[noble coyote]

I will re-boot my PC, and give it a go!

[bitter hearth]

ok

16s for an image

[sterile pendant]

das geht

[bitter hearth]

im happy

used to get 40ish seconds before with regular schnell

[hallow lion]

Whats the bare minimum steps for DEv?

[sterile pendant]

you can go as low as like 15 and still get decent results

i showed a two stage approach the other day where it was 12 steps and then 4 steps refine 0.2 denoise same seed, and it was better looking than at 25 steps in one sampler

but your mileage will vary

[hallow lion]

LCM Flux when? XD

[bitter hearth]

hmm https://huggingface.co/silveroxides/flux1-nf4-weights

[bitter hearth]

schnell is already lightning at 4 steps

[errant dust]

You're not wrong, but some put out some cool LoRas or models. I had one for Frank Frazetta for XL which was fun. Within the limits of XL's abilities, but still

[bitter hearth]

👀

anyone tried the 6gb checkpoint?

[dull star]

dad enjoying his first cup of lean (he doesn't know what's awaiting him)

[errant dust]

@sterile pendant ahhhh. That guy built a Dev-Schnell for NF4. You can't tell from the confusing names he put, but the community comments and replies make it clear. The Bitte model is basically Dev-Schnell NF4. "Bitte is schnell with double blocks from dev." https://huggingface.co/silveroxides/flux1-nf4-weights/discussions/1

[dull star]

please tell me what is stands for

[errant dust]

Dev-Schnell was an improved version of Schnell based on an idea promoted by Comfy when he offered support for Flux. It was what he said, though Schnell was still horribly Schlow on my laptop regardless. Now someone has compiled a quantized version of it for NF4

[errant dust]

That was withy Flux? Wow. Fantastic. I thought only MJ was able to do that

[dull star]

Flux-Dev with Boreal lora (Boring Reality for Flux)

my favorite sdxl lora is now in flux

[dull star]

this too was made with it

[errant dust]

Well, I cannot run normal Dev. I mean, the VRAM it needs means my laptop slows to 65s per it. Really. So 45 mins for a render is not unheard of. That is why I will burn candles and sacrifice cows to the creator of NF4

Blessed be his name (said in reverential tone while waving hands to sky vaguely.)

🙂

[dull star]

45 mins for a render :(

and how much with nf4

[hallow lion]

Bitte means please

[errant dust]

2 mins

[dull star]

nice

[errant dust]

for 30 steps

[dull star]

that is especially nice

I only do 20 steps

[errant dust]

there is a mini catch I have come to terms with. Must be 1k. Larger and the greater VRAM demands start to impact again

So 1024 x 1024 fine, 1280 x 1280, NOT so fine

I may test some in between but I don't see any real benefit other than to test and know

[dull star]

yeah I always do 1024x1024 res with different aspect ratios

such as 1152x896 or something

[errant dust]

Still, the fact is I can actually render with Dev realistically and happily on my laptop. There may be plenty of free options to run Pro, which is totally cool, but this is still an amazing boon

[noble coyote]

This works in dev_fp8

About 5 minutes/image

[rain current]

15 steps dev-nf4

[sterile pendant]

It can also mean a request as well. Like if I said "ich habe eine bitte" it means I have a request

[bitter hearth]

Dev-Schnell NF4 might be the best combo

[bitter hearth]

yeah both dev nf4 and schnell nf4 should work in that workflow

[bitter hearth]

how many steps do you use for that combo?

no idea

I use adaptive steps with everything so I never know

[bitter hearth]

dpm adaptive?

how long does that take tho

not dpm adaptive but a similar idea

the way adaptive steps works is you can still change how long it takes
but instead of setting the steps you set relative precision and absolute precision

and you're applying that on flux models?

[noble coyote]

Only Schnell - Dev doesn't like

[bitter hearth]

there is only really one rule with samplers
and that rule is don't use SDE samplers with flow models
so long as you don't break that one rule any sampler works on any model, for the most part

[bitter hearth]

in most cases i just use euler a.. that seems to work across all sdxl, pony, and flux

except for flux im using just euler

euler is ok but its literally the worst sampler

in terms of what

local discretisation error

idk what that means but i get good speed and decent image quality with euler

its basically how accurate it is

euler is ok I'm not saying euler is bad

but any other choice would have been better

[hollow swift]

any tips for forge flux?

[brittle nexus]

With NF4 maybe we will have even larger than 12b parameters models in the near future.

[dull star]

im so happy for you 8GB and lower users being able to run these, I expected the future models to just ignore lower vram users, but the community keeps finding better ways to run these on lower end.

I wonder if we are entering an era of running larger models with quantization (like how LLMs are used nowadays, you barely see people running small models at fp16, everyone uses 2 to 6 bit quantization with larger models instead, depending on the size)

[dusky thistle]

i look forward to 1 bit quantization 😛

[dull star]

bitnet moment

has that ever been proven to be good or viable

I remember 1 instance for a DiT-like model I think, but it was a small test and had like a trade off

[noble coyote]

Flux LoRA - like before - from Model -> LoRA. And from LoRA -> VAE Encoder?

[strange grotto]

i think forge flux inpaint is amazing。it's completely different from the comfyui version

[noble coyote]

Isn't Forge based on A1111 at all?

[dusky thistle]

yeah, it is

[noble coyote]

A1111 is less and less supported these days! Where is the folder which contains the schedulers and the sampler models in A1111 at all?

[dusky thistle]

no idea

i haven't bothered with it in a long time... comfyui and stableswarm are where it's at these days

[sage burrow]

10 mins and 4 mins on mine for dev nf4

[noble coyote]

I love the rich deep warm colours and contrast of Fooocus - and it can really give some motor (via its LLM) to your prompts. And it is the simplest to use - not a tweakers toy!!! 😄

[noble coyote]

Swarm is OK - I get confused between "in-house " ComfyUI workflow; and "imported ComfyUI workflow!" 😄

[dusky thistle]

yeah i'm just not aware of anything a1111 can do that comfyui can't

but i know of a TON of shit comfyui can do that a1111 can't

[noble coyote]

If I run all my best ComfyUI prompts via A1111 - I get a whole different set of images - and a lot of them keepers!

I like Latent Mirroring Extension in A1111

[errant dust]

well, the Bitte NF4, AKA Dev Schnell, was a fail. It works, but it botched simple text so badly it left scars.

[sage burrow]

Anyone else see the Aurora last night?

[bitter hearth]

nice ball

[bitter hearth]

going out ?!

[noble coyote]

It bored me alis!!! 😄

The Perseids

[sage burrow]

Those also last night 🙂

I was up late trying to talk comfyui into letting me install nf4, so I got to see them

[noble coyote]

Flux Dev.FP8 - LoRA weren't good!

[noble coyote]

My RTX 2070 8Gb VRAM GPU does not like nf4 - so I'm making do with dev.fp8

[dusky thistle]

no need to go outside anymore when we can just generate the outdoors locally

[bitter hearth]

what the hell is that lora trying to do

[noble coyote]

I used the wrong Sampler - Euler A hates LoRAs (I think?!)

[sage burrow]

doesn't nf4 make it easier? It's 4 mins for me instad of 10. 8gb nvidia gput, 16GB RAM RTX 4060

[noble coyote]

nVidia series 2XXXX GPU's don't always work with nf4 - mine is a case in point; although fp8 is 5 times faster!!!

[sage burrow]

try my aurora skull workflow just in case?

[noble coyote]

'k

[sage burrow]

So now I just need SDXL at the end to "enhance" things 😉

[sage burrow]

Oh, perhaps I should have mentioned my 180gb swap file.....

[noble coyote]

Poodle size!!! 😄

Flux.Dev - 20 iterations

[sage burrow]

I've been deleting unecessary files to try to get that up to 400gb lolol

[alpine summit]

Sd 1.5

[sage burrow]

I still love 1.5!

[noble coyote]

As an artist - and not too bothered with "real" images - 1.4 and 1.5, 2.0 qnd 2.1 are exquisite - as are SDXL 0.9 and 1.0

[dull star]

https://civitai.com/models/640459/impressionist-landscape-lora-for-flux

I swear to fricking god I'm being spoonfed

[noble coyote]

I cannot get the Flux LoRA to do anything?!

[dull star]

first there's boring reality, which I really fricking loved, and now impressionist paintings

[noble coyote]

Pop-art. Flux-style. Art-nouveau - nothing, nix, nada

Flux.Dev with Flux Art-Nouveau LoRA - as you can see - nothing worked (art-nouveau in prompt)

[sage burrow]

hmm, I'm going to try a bunch of Renaissance and Medieval artists prompts! 🙂

[errant dust]

Ah if only the NF4s were compatible with LoRas... No doubt will be in no time.

[alpine summit]

Sd 1.5

[dry wave]

do you use quantisation?

[noble coyote]

Let me know how?

[errant dust]

It's fine (where are they?) and I am patient (what is taking so long??). I am a zen person in sync with the universe (I want it now! Now! NOW!). You would do well to learn from my example (Gaaaahhhh!).

[noble coyote]

Buddhist monk to McDonald's employee: make me one with everything!!!

[noble coyote]

Quantisation?

[dry wave]

like fp8

[dry wave]

or NF4

[noble coyote]

I'm using fp8 as nf4 does not like my nVidia RTX2070

[alpine summit]

Sd 1.5

[errant dust]

Doesn't like? How so?

[dry wave]

this will mess up with Loras probably

[sage burrow]

My review of Flux nsfw loras and checkpoints: "tell me you have never seen a nude person, in person, without telling me you have never seen one"

[noble coyote]

NF4 throws up error messages like Error occurred when executing UNETLoader:

Error(s) in loading state_dict for Flux:
size mismatch for img_in.weight: copying a param with shape torch.Size([98304, 1]) from checkpoint, the shape in current model is torch.Size([3072, 64]).
size mismatch for time_in.in_layer.weight: copying a param with shape torch.Size([393216, 1]) from checkpoint, the shape in current model is torch.Size([3072, 256]).
size mismatch for time_in.out_layer.weight: copying a param with shape torch.Size([4718592, 1]) from checkpoint, the shape in current model is torch.Size([3072, 3072]).

[errant dust]

Add "tell me you have no mirrors in your house or bathroom."

[alpine summit]

Sd 1.5

[errant dust]

This is because you don't know how to load them. You cannot load them that way and need a speical node in Comfy as well as Python package

[noble coyote]

OK, is it the ComfyUI Add-on Node?

CheckPointLoaderNF4?

[sage burrow]

@noble coyote have you installed the nf4 bits and bytes via comfy? (after changing the config tile so you can bypass the safety)?

[errant dust]

This grows almost tiresome, and I am not criticizing you Torcello, because you'd think cler instructions would be left for others to use and not require endless good faith from the rest

[alpine summit]

Sd 1.5

[noble coyote]

I have the bits and bytes yes - did not change config file to bypass safety ...!

[errant dust]

You need the special node that supports it

[noble coyote]

I appreciate your effort, keep going as I'm taking it all in 🙂

[errant dust]

Are you comfortable Git cloning?

[noble coyote]

Yes

[sage burrow]

It didn't install for me properly until I changed the safety and did it via comfy. My self install flopped (even with help lol)

[errant dust]

https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4

You need to git clone that into the Custom_nodes folder

[dull star]

I have to take down the strength to like 0.6

[errant dust]

if you need help with that, tell me, but that is the next step.

[dull star]

HOLY CRAP

[errant dust]

ok, you need to add the node into your workflow

[sage burrow]

there's a final step of installing yet another instance of python something or another

[errant dust]

Like so:

[noble coyote]

I have installed via git clone ComfyUI_bitsandbytes_NF4

[errant dust]

there can be a tricky moment if you are using the embedded Python

[noble coyote]

OK, wait up ...

I'm using system Python

[errant dust]

then in theory, this workflow should show no errors

[hollow swift]

Which is better settings on forge?

[errant dust]

ok, the checkpoint of NF4 goes to the checkpoint folder, not unet

[dull star]

indeed, because it contains the T5 encoder and clip and the VAE too

[noble coyote]

Yep, there so far ...

[errant dust]

You need to load it from that node in my workflow, though you can use the fp16 Clip as I did, which is what I suggest. Not forced to though, so feel free to ignore

[sage burrow]

Starting from here, they (esp rx808) walked me through it step by step. Don't mind my still re-learning windows in it too lol) @noble coyote

#🆕｜sd3 message

[noble coyote]

Error at KSampler stage - Error occurred when executing SamplerCustomAdvanced:

'ForgeParams4bit' object has no attribute 'quant_storage'

File "X:\ComfyUI\execution.py", line 152, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
File "X:\ComfyUI\execution.py", line 82, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
File "X:\ComfyUI\execution.py", line 75, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
File "X:\ComfyUI\comfy_extras\nodes_custom_sampler.py", line 612, in sample
samples = guider.sample(noise.generate_noise(latent), latent_image, sampler, sigmas, denoise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=noise.seed)
File "X:\ComfyUI\comfy\samplers.py", line 706, in sample
self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds)
File "X:\ComfyUI\comfy\sampler_helpers.py", line 66, in prepare_sampling
comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required, minimum_memory_required=minimum_memory_required)
File "X:\ComfyUI\comfy\model_management.py", line 526, in load_models_gpu
cur_loaded_model = loaded_model.model_load(lowvram_model_memory, force_patch_weights=force_patch_weights)
File "X:\ComfyUI\comfy\model_management.py", line 323, in model_load
self.model.unpatch_model(self.model.offload_device)
File "X:\ComfyUI\comfy\model_patcher.py", line 614, in unpatch_model
self.model.to(device_to)
etc

[errant dust]

how are you starting up Comfy?

[noble coyote]

Via .bat

FP8 works OK, but NF4 errors-out!

[errant dust]

ok, so you may have python, but are likely using the embedded python. The .bat uses it by default unless you actively edit it

[sage burrow]

Wait what, is there a way without using a web browser? (hopeful)

[alpine summit]

Sd 1.5

[noble coyote]

I don't have embedded Python - I am using 3.10 which is system

In my ComfyUI_Windows_Portable - the emebeded Python is 3.11

[errant dust]

ok, so try updating comfy and python first witht he update function

[noble coyote]

OK ...

[errant dust]

this .bat file is the one that comes with Comfy? unedited? or one you made or edited?

[sage burrow]

When I "installed" it via git, it didn't really install. So I had to install it via comfy manager.

[errant dust]

The issue is that the update .bat files are all designed to update the embedded python dependencies, not your system's Python

[noble coyote]

Its the ComfyUI main.py file

All I have is a script which says ComfyUI/python main.py

[errant dust]

ok

Try updating all the dependencies Comfy uses

and Comfy itself

[sage burrow]

Hehehehehe all this bc none of us want to wait a day or 2 for comfy to add the proper nf4 node 😄

[dull star]

pink

[noble coyote]

You mean go to the ComfyUI folder and git pull?

[errant dust]

if that is how you do it, yes

[dull star]

that's just simply because bitsandbytes is outdated

[mortal mesa]

yup

[dull star]

go to the python_embedded folder and type in the terminal/powershell: "python.exe -m pip install -U bitsandbytes"

[mortal mesa]

an answer

[dull star]

that's all

[errant dust]

and update bitsandbytes if you are using an older version

(that never crossed my mind)

[dull star]

I had the exact same issue

[mortal mesa]

same

[dull star]

and it was literally answered in the issue tab lol

[errant dust]

he isn't using the embedded python

[noble coyote]

OK, well ComfyUI updated; bits and bytes was up-2-date

[errant dust]

Nor am I for that matter

[mortal mesa]

pip show bitsandbytes, what version comes up

[errant dust]

update TOrch too. BTW, there is a newer CUDA out you can install independently: v12.6

[noble coyote]

bitsandbytes 0.43.1

[mortal mesa]

that one doesnt work, needs to be 0.43.3, pip install bitsandbytes --upgrade

[errant dust]

If you don't have CUDA installed in fact, at least for me, the Torch it installs is not compiled for CUDA

[noble coyote]

I have torch 2.3.1+cu118

[errant dust]

I have 2.4.2+cu124

you still need CUDA to get a Torch compiled ro suport it

[noble coyote]

OK, bitsandbytes upgrading!

[errant dust]

in Python, try this:

import torch
torch.cuda.is_available()

It will say False or True

just a sanity check

[noble coyote]

Still have this erro Error occurred when executing SamplerCustomAdvanced:

'ForgeParams4bit' object has no attribute 'quant_storag

[noble coyote]

True

[errant dust]

ok

[noble coyote]

Go to pytorch.org and get torch update?

[errant dust]

cannot hurt

[noble coyote]

OK, will do ...

[mortal mesa]

probably what ide do

[bitter hearth]

you should not be loading unets

[mortal mesa]

oh ya, you load the linked checkpoint from that repo into that node, not others

[errant dust]

ok, look, let's just cover our bases with a simple, and presumably unneeded step. in CMD, go to the ComfyUI folder, and run: ..\python_embeded\python.exe -s -m pip install -U bitsandbytes

[bitter hearth]

once again, here is the complete workflow for nf4 checkpoints

[torn wharf]

i think a bunch of comfy code refers to these transformer networks as unets, since it's legacy code that still works and they just repurposed it

i see unets in the logs a lot while using flux and sd3

[noble coyote]

I'm not using ComfyUI_Windows_Portable - but just so there's no confusion ...

[sage burrow]

and if all that doesn't work, just install it via comfyui manager (now that you have the bat etc. files and python all ready

[errant dust]

I get it. I perosnally upgraded every single dependency in th requirements list

even my CUDA

and I am not running from embedded either

[errant dust]

The error you are reporting is actually the very first one reported in the issues of the Comfy node: https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4/issues/1

and was due to BNB not properly installed or updated

[noble coyote]

Done! Just waiting for pytorch.org to finish

[errant dust]

but if you are still getting it... time to see of other ancillary dependencies are not compromising the setup

[torn wharf]

its the same old story. people dont' read the readme, or they assume that requirements installed correctly when it didn't.

[bitter hearth]

just to emphasize the use case of loras ... left one w/o lora and the right one w/ realism lora ...
there is obvious difference

[torn wharf]

the 80s' arnold lora is the most impressive imo

[errant dust]

In Portuguese 'torcer' can mean 'to twist' or 'to root for'

[noble coyote]

Ta daah! it's working!!! Yippee 😄

[sage burrow]

I read the readme on that one, didn't help one bit lol

[noble coyote]

Just getting that cup-of-coffee going here ..

[sage burrow]

Don't worry, there'll be new release of something in a couple of days for us to all figure out 😉

[noble coyote]

I've had such troubles with Xformers - it says it needs torch 2.3.1+cu118 - so I was loathe to change the version!!!

[mortal mesa]

so like uninstall xformers, unless you have something that specifically needs it, not a comfy requirement anymore

[torn wharf]

i wouldn't be so quick to join subtleone in trying to frame it as a project issue. this is classic user error in most cases. the readme file covers most people's problems, like not having b&b installed

[bitter hearth]

so with this nf4 i notice every time i change a word in the prompt it takes time to cache everything in from the start all over again

but other than that render time is pretty good

[torn wharf]

do you have nvidia 30 or 40?

[bitter hearth]

3060

[torn wharf]

thats why. you don't have the hopper transformer engine on your silicon to speed up these operations. it's doing it on cpu instead

[bitter hearth]

well once it caches in after the first render the 2nd time it goes way faster

unless i make changes

[errant dust]

Wooohooo!!!!!

[torn wharf]

its the number format casting that slows it down

[sage burrow]

the older versions seem to blazingly fast now!

[mortal mesa]

the speedup is worth the slow down if that makes sense

[bitter hearth]

the difference between that is like from 1 min down to 17 seconds

[sage burrow]

Giant swap files can help!

[torn wharf]

feel like comfyui got this out as fast as possible and over looked things like user error. i'm not sure why it didn't roll out as a project node. If they didn't release this official extra node, then all week long forge woulda been getting praise for having nf4 support. there's kind of a perception that this is somehting achieved by the comfyui team because they ported his code so quick

[noble coyote]

So the new Flux LoRAs just go in between the Model and the VAE Encoder?

[errant dust]

Yes, keep in mind the default workflow I gave you has it using the fp16 Clip text parser. Ample testing has shown this to be superior, but it does come at a cost, as you described. You can use the embedded fp8 if you want and test it. My output had very irregular text output and mistakes that the fp16 never had. And that is likely affecting all output, so bear that in mind whatever your choice.

[torn wharf]

40 seconds to cast is reasonable sounding for 30 series. mine in about 15 seconds or less

[bitter hearth]

not using any default workflow that uses encoders, im using single checkpoint with custom nodes

[errant dust]

Yeah, I mixed you up with torcello, sorry

[noble coyote]

My Flux.Dev at 20 iterations often takes 40 minutes at 1024x1024 - it'll be interesting to see what NF4 can do?

[mortal mesa]

prepare to be astounded

[errant dust]

so you share the same issue. Just be sure to close anything that might be borrowing vram

and even use MS Edge for Comfy

[torn wharf]

if you can afford to use the fp16 t5, use it. this is one of those side by side tests that when shown with same seed and long prompts, fp16 t5 is superior undoubtedly.

[errant dust]

(not an MS fan, but it does use the least memory of all browsers

[mortal mesa]

fast but slow

[sage burrow]

oh?

[torn wharf]

edge is just chrome. if it's using less than chrome it's because you've installed extensions that make chrome innefficient

[errant dust]

Its well docuemtned that whatever you may think of Edge, and I don't use it as my main driver, it uses the least memory

[noble coyote]

Flux.Schnell takes 100 seconds for 1024x1024 - but can be a tad soft. Dev is definitely the crisper look - but I added an SDXL-Out to Schnell, and can get a crisper and sharper look way sooner than waiting for Dev

[bitter hearth]

out of the big browsers yeah

schnell nf4 with lora

[noble coyote]

Somebody remind me how to "plumb-in" the LoRAs with NF4?

[errant dust]

they are incompatible

no LoRA support for NF4s for now

[torn wharf]

edge is just chrome under the hood. microsoft ties it into a ton of their services and has some smart precaching with it, but effectively, just chrome. if it's using less memory than chrome for you, it's because of your use cases for each browser

[noble coyote]

The Flux LoRAs "aren't Flux LoRAs?!?!?!?"

[sage burrow]

The 100 tabs I keep open I know never helps either!

[bitter hearth]

are you sure, cause im getting some visible differences

[noble coyote]

Flux.Dev or Flux.Schnell LoRAs at all?

[torn wharf]

having a 100 tabs open is why my chrome eats 20gb of system emmory yeah

[errant dust]

Unless they changed. because I don't get any 'differences'. I get a straight out crash

[bitter hearth]

i tried w/o lora and i dont get that realistic feel i get with realism lora

[noble coyote]

I get grainy and horrible with the LoRAs

[torn wharf]

theres also this myth that using less memory means it's a better program. but i'm a keen believer in the machine philosophy, if you got the ram then use it

[bitter hearth]

are you trying the dev ?

[noble coyote]

NF4 has pushed generation time for 1024 x 1024 to about 9 minutes - approximately 5 times quicker

[torn wharf]

you'd think your lora would fail out if it wasn't working. i've heard from others that it doesn't work too, but i'm not quick to believe people. a lot of people rush testing and reach their conclusion TOO quickly

[bitter hearth]

no technical explanation on my part but im getting consistent results that are better with lora

[torn wharf]

did you guys know loading weights as nf4 works for sdxl too? i haven't tried it yet but it should work 100% in forge ui.

[noble coyote]

I am just trying out @errant dust workflow - and here are the first fruits (9 minutes)

[errant dust]

Should still be faster than that. Really, close every bloody program, app, or other. Restart Comfy and try again. I may be overestimating the 2070, but still

[bitter hearth]

looks good but 9 mins is painful

[noble coyote]

I'm just glad it worked 😄

[sage burrow]

so which main Flux model are you all using for those super fast times? I've been using this one:

[torn wharf]

NF4 and the new memory management and preset options in forge UI. llly gone and changed the game all over again. that dude spreads himself thin but still manages to knock out game changing bangers

[errant dust]

That isn't NF4

[torn wharf]

version 10 ??

i can has version 10?

[torn wharf]

sneaky

[bitter hearth]

dev is possibly 3x slower than schnell but also that's not a nf4 single checkpoint workflow

[sage burrow]

Anyone have a straight up Nf4 workflow?

[errant dust]

That is NF4 though.

workflow is embedded

Nothing fancy

[noble coyote]

Here is another image - and its only taken 3 minutes and 50 seconds in NF4

[bitter hearth]

yeah but you'd first have to update a few things https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4

[errant dust]

@sage burrow As I suggested to Torcello: close EVERYTHING you don't need. Anything that might steal even 0.2GB of VRAM. All of it.

[bitter hearth]

gotta update the nf4 node

[noble coyote]

That 9 minutes seems to have been an anomaly for the Coffee picture!

[bitter hearth]

actually i mean install the nf4 node

[errant dust]

And even then, on some occasions I need to ''reset" and that means just restarting Comfy via .bat

[torn wharf]

nf4 sounds cool too. llike, need for 4

i got a need for 4 and nothing more!

16 weights? naw make em 4!

[errant dust]

This is on a 'squeaking by' setup, as you, with a laptop with 16GB of Ram, and a laptop 4060

[mortal mesa]

it really should be faster on what you have, you ran a marathon, if you wanted to run another get rid of xformers and bump up pytorch and cuda, just my opinion, but glad it works

[noble coyote]

I got it down from 9 minutes to 3 minutes 50 seconds ...

[errant dust]

That may be what to expect and that is still AMAZING, considering the previous reference was 45 mins

Be warned the times will skyrocket again if you try larger than 1k images

That said, it cannot hurt to update CUDA to 12.6 and then Torch too, supporting 12.4

[sage burrow]

179 seconds for the coffee workflow compared with 164 for the workflow

[errant dust]

https://developer.nvidia.com/cuda-downloads

[tough oriole]

Do higher parameter models train quicker? tried 2 concept loras for flux's and it trains so fast. 700 steps feels like 2000 steps in SDXL

[errant dust]

lol

[noble coyote]

3 minutes for this picture

Less than 3 minutes ...

3 minutes - seems to have troughed/plateau'd

[sage burrow]

Flux doesn't support dynaimic prompts does it?

[uncut river]

what is a dynamic prompt?

[noble coyote]

You bracket certain terms together (colours/textures/sizes/shapes) and then ComfyUI makes up a prompt by selecting a word from each category - dynamically

[sage burrow]

A cup of coffee which reads: "my poor GPU needs this", "tea is better", "I want orange juice". Forgot the proper format now, but then you come back to 3 different images

[uncut river]

oh, that stuff yeah i know

not sure if flux can do it

[sage burrow]

So you go to sleep at night, and you wake up with your drive full of images if 1.5 and about 20 if flux 😉

[uncut river]

there might be some nodes to create that feature from scratch

[noble coyote]

That text in the coffee cup is really coherent. Is NF4 that good at text?

[errant dust]

Yes

[noble coyote]

Cool

[noble coyote]

The Hakaa

[errant dust]

( famous rugby team)

[sage burrow]

every flux expt schnell or schnell/dev merge

[uncut river]

https://github.com/adieyal/comfyui-dynamicprompts might help in a comfyui workflow with flux

[noble coyote]

Am I that b*tch?

[sage burrow]

Glif already has flux dev and flux pro, perhaps I can request flux nf4 as a feature 🙂

Though flux is so fast, prob doesn't matter

[errant dust]

https://glif.app/@AmericanCarioca/glifs/clzoyqqv00000jjj58x96kqd3

Flux Pro BTW

[torn wharf]

some nf4 results look better than default fp16 dev. it's interesting. i wonder if the normalization of weights led to something

[sage burrow]

I made one for flux dev and flux pro! <-- glif addict lol

[errant dust]

Yeah, it was you who led me to the promised land

[torn wharf]

kind of like how perturbing the weights has been claimed to improve things

[noble coyote]

I'm still in the dark about the safety check at Flux?

[uncut river]

yeah, most "tricks" do work but they tend to narrow results

[sage burrow]

With flux as well? I did perturb sd3, but only a very slight difference.

[uncut river]

so, it always seems to come at a cost

[errant dust]

I made this for my complete neophyte friends, and especially Brazilians

[torn wharf]

i haven't seen flux weight perturbing studies at all yet but i was thinking of the sd3 ones

[uncut river]

She didnt had to hold up the sign to know ...

[torn wharf]

yeah. i'm not a believer in perturbed weights as a "solve" quite either, but it does something at least. thats kind of how i feel about nf4. it seems to improve things slightly but i'm left wondering "at what cost"

[uncut river]

it's the general principle, expert things are much better in a given subject at the cost of variation

[noble coyote]

My first fluxual text!!! Cough, splutter 😄

[uncut river]

add model size to the mix and you can choose: small size, highly specialized, high variation. you can only pick 1 or 2 at maximum

[errant dust]

from my Gliph with Flux Pro (took seconds.... sigh!):

[uncut river]

real photo of a cup of coffee with text in the cream "FLUx bla dh"

[errant dust]

Really good and precise

[noble coyote]

Well all be asking the Starbucks baristas if they've ever heard of Flux?!?!?!?!?

[errant dust]

I will be making that last one of mine into a coffee mug for myself.

[hollow swift]

How does the API work? (fluxpro) do you still need to download the model? how it works on comfy?

[errant dust]

Pro cannot be downloaded

[uncut river]

i dont think you can use fluxpro in comfy

[noble coyote]

I like coffee - but after a bout of Covid - my taste changed.

I like coffee less so

[uncut river]

real photo of a cup of coffee with text in the cream "SD3 bla dh"

SD3

[errant dust]

You blame the coffee?

Really good! I can make out the letter D!!

[uncut river]

sorry, I stirred it ..

[errant dust]

I love my cappuccino, but only indulge twice a day. Albeit more a Flat White, meaning double espresso shot

home made

[torn wharf]

i think those principles might yet be shaken up by large models still. i'm not sure if nf4 is losing anything yet. that's my big concern. what was lost by normalizing weights the way they do. its a double quantization if i understand it right. i intend to look at how the nf4 format has performed in LLM arena today. i'm so out of touch with LLMs.

[noble coyote]

Latte with either almond syrup or hazelnut

[uncut river]

well, even flux is not the best in real photos

it smooths and polished out about everything, it's a generalist

flux is not a specialized model

[errant dust]

my relatives look at my minuscule amount of sugar and comment: "why bother if you are only putting that much?"

[noble coyote]

Splenda sweetener

[uncut river]

so while it can do the popular styles and many compositions, poses, etc, it's not the best in more specific styles

[torn wharf]

i used to be all snobby about coffee but then i got to experience kona coffee in kona hawaii itself. the balance of acidity changed my world. now anything i buy from local stores is basicall all gas station coffee in comparison. there's no reason to be snotty about what gas station brand i'm buying.

[uncut river]

I rather have some smaller models specialized in a certain style, so they can be smaller

[errant dust]

Sorry, tried all the sweeteners. Can't stand them. Besides, they are not far wrong. I hate it remotely sweet. I just like to reduce fractionally the bitterness

[torn wharf]

i've actually stoped with grinds and i drink instant nescafe now because it's consistent and good. it's all gas station swill after i had proper kona coffee

[uncut river]

if you have a bunch of very specific style models, without much overlap, they are a great resource for model merging

[mortal mesa]

that also cause death and DNA damage it seems check WHO among others

[torn wharf]

if you don't like sweet but want to cut the acidity of coffee, little salt. old cowboy trick

[uncut river]

sunlight and baked meat also cause dna damage

oh, and oxygen

[torn wharf]

yeah oxygen is a harsh destroyer of the dna molecule. don't fuck with o2

oh healing to often too. don't do that. mitosis is completely mutilates dna at times

[mortal mesa]

ya the salt trick works nice, makes it a whole new thing with a tiny amount

[torn wharf]

emphasis on tiny amount

[uncut river]

hey, dont get me wrong, I still like sunlight baked meat and to heal

[mortal mesa]

ya sure there is just a diffrence of taking a known poison or not

[torn wharf]

i look at my family lineage and see no cancer so i figure i've got lobster genes

poison is actually a really broad term. any substance that can produce an adverse effect in a given circustance. so dihydrogen monoxide is a poison.

[torn wharf]

venomous too humans are venoumous because our saliva is part of our digestion process

[noble coyote]

A rude bhoy!!!! NF4

[uncut river]

yeah hope the crap and lobster will get him!

[mortal mesa]

got to ban DHMO for sure causes all kinds of deaths

[uncut river]

sd3 was looking through some holiday picture of your mom, yeah it's in the datasets now ...

[errant dust]

Judging by the image, I am guessing DMHO is a boy band? And the deaths are people listening to it realize they cannot unhear it.

[torn wharf]

DHMO sounds so ominous

[noble coyote]

NF4

[uncut river]

can you maybe put a lobster in the bathtub with him ?!

[hollow swift]

how do i make a cctv footage image? is there any keyword?

[noble coyote]

D/load the workflow png and put some lobsters in there yourself!!! 😄

[uncut river]

something like "grainy video, bad quality, cctv footage" ??

[torn wharf]

cctv works but it's a bit of a dated term. modern systems nothing is closed circuit anymore, its all encryption on the same network

[hollow swift]

i already put the cctv footage but it still look great. thank you for the keyword!

[uncut river]

flux is still too smooth for this ...

grainy video, bad quality, cctv footage, isometric street view, dark lighting

[errant dust]

If so, it is a sentiment I can sympathize with

[torn wharf]

yeah sd3 is way better at getting granular pixels and jpeg artifacts

[uncut river]

lemme try

[mortal mesa]

nah someone said rude boy so thats a brit rude boy

[torn wharf]

i've had gens on sd3 and it had me convinced my ui settings were fkd. thought it swapped to low quality jpg or something

[noble coyote]

DHMO - dihydrogen monoxide aka H20 aka water!!! LOL

[errant dust]

😄 I was having fun

[uncut river]

hmm, somehow sd3 is switching to anime-ish style here

[noble coyote]

🤩 #

Has anybody used the bitte NF4 safetensors yet?

[errant dust]

Crap

[noble coyote]

OK

[torn wharf]

i'd wait till a guided hand produced a file to judge

[noble coyote]

My SSDs are getting overfull - guess I will not d/load any!!!

[torn wharf]

a lot of people have just been cranking out nf4 quants of weights, but i'm thinking that llly didn't put one out for it because he hasn't done the proper code for it yet

[errant dust]

This was the first output of my coffee test prompt, and then the rerun:

[noble coyote]

Cool

But not so good

[errant dust]

All text came out wrong or garbled, so Bitte was deleted soon thereafter

[hollow swift]

banana character, eating a banana, sitting on a banana shaped couch, under a banana tree.
FluxDevNF4

[mortal mesa]

i smooshed fp16 T5xxl back into the Dev BNB NF4, still satisfied with the speed/quality, dont really have a compelling reason for the combo or schnell version

[torn wharf]

i'm sticking to dev nf4 right now.

so what are qloras? this nf4 stuff is the first i've heard of them

[errant dust]

As in you have a single checkpoint file?

[torn wharf]

i might remember them being announced, but it blew past me becasue i felt like q was becoming the new "throw it on anything to make it sound modern" gimmick. q* qanon startrek q, i'ts all q. its qetting annoying

[uncut river]

damn you flux!

[mortal mesa]

me? yup

[uncut river]

she also knows urmom

[torn wharf]

thats what she said

[noble coyote]

That we take all this time and effort with flux just to get an iphone look!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 😄

[uncut river]

portrait photos with iphone already have filters right? for how long now?

[uncut river]

spooky lady

guess flux had some eyes on them

[torn wharf]

i think the nexus phones were the first to do it. might've been the nexus 3 had default filter enhancements on the photo app

could be wrong. that was a fast paced time in smart phone development

[uncut river]

filters went from simple colors and contrast adjustment fluently into ai assisted enhancement

[torn wharf]

i loved the nexus phones. i've not felt the same about the pixel phones. the motorola nexus was a beast. one the best cameras

[uncut river]

meanwhile from SD3, since idk on-topic or something

[uncut river]

ok, perfect fingers. but we need to talk about the thumb

[hollow swift]

can i have your prompt? it looks amazing!

[noble coyote]

Its trying to worm its way out of the picture!!!

OK, its been a great day staying out of the fabulously hot English sun.

[uncut river]

it's nice to discuss this, it's a nice change from "fear of godlike AI powers taking over the world" to "this thing cannot even do hands! wait, it cannot do thumbs or real stuff!"

[noble coyote]

One more and I'm gone! Kudos to all who helped me to install NF4

[uncut river]

"Now drink it."

i find it highly annoying the handle is on the wrong side

[torn wharf]

https://tenor.com/view/elf-coffee-santa-christmas-congratulations-gif-5704070

[bitter hearth]

the prompt was this:
Epic cinematic movie still of R2D2 in a rococo palace. There is intricate and detailed ornate furnishings. There is fancy jewellery and gemstones. There are expensive ornate cups and saucers. Soft lighting, sharp focus
but to get it looking that soft required FreeU, CADS and a CFG++ SDE sampler

[uncut river]

meanwhile sd3 spits out results which reminds of my polish aunt. there must have been some snooping around in my google drive archives. where to check the dataset.

[errant dust]

It is a prompting issue. Just add "...with handle on the other side"

[errant dust]

Any place to download them?

[proven pecan]

schnell

[mortal mesa]

uh don't know, make your own super easy, what i did

[uncut river]

also flux-schnell

[proven pecan]

schnell

[torn wharf]

ugh. comfy. i guess last night before i went to bed i was updating nodes and getting some other workflows setup. one of them i guess killed bits and bytes. downgraded it to an ancient version that wasn't compatible. i was confused because i knew ealrier last night i had bnb installed and the nf4 node was working but it kept erroring out just now like it wasn't installed right

python dependencies are stupid. every project stepping all over each other

[torn wharf]

a standard library of nodes is sorely needed instead of every new workflow having a dozen unique custom nodes

[dusky thistle]

at a minimum, a standard set of dependencies and their versions

[torn wharf]

i'm not sure which custom node update downgraded bnb to 0.41 from 0.43. i dont think that's necessary at all

[uncut river]

same prompt as above, but with sd3 instead of flux-schnell

[dusky thistle]

it usually is'nt

tons have random deps they don't even use

or versions without any reason for it

[proven pecan]

more schnell

[hollow swift]

how do i save jpg in forge?

[bitter hearth]

comfy ecosystem has issues yeah

[uncut river]

for me sd3 still beats flux in the context of real photo style

[bitter hearth]

100% yes

[torn wharf]

stdio.comfy

[torn wharf]

c3d2

[bitter hearth]

lol yeah

it was a control net on a flux generation funnily enough

[torn wharf]

controlnet works so good. do you got it working on the newest comfy update? i saw that the experiment branch had been hanging for a few days so i abandoned it

[bitter hearth]

I'm a few versions back I think

[torn wharf]

yeah i jumped off at the nf2 point

[bitter hearth]

there was a very good commit for me recently
that reduced the minimum PAG/SAG level

I need it cos I often have 0.1 CFG

[bitter hearth]

haha nice

its nice when the models put some gold on R2D2 like that

not every model does it

[errant dust]

That did the trick, thanks!

[errant dust]

Let me guess: woman wearing cocktail umbrella as hat surrounded by poisonous mushrooms

[torn wharf]

No those are power up mushrooms the poison ones have little 💀 s on them. Trust me life long Mario Bros player

[rain current]

w/Lora Emma

[errant dust]

Interesting sidenote: the giant walkers were all filmed with frame-by-frame motion capture for months, which was modeled on the walk of an elephant they brought in and filmed on a parking lot to copy.

[bitter hearth]

wow that really is an interesting sidenote

[errant dust]

https://youtu.be/h_Cp9xGzPWs?si=_M1iKSGZDVihhQ7b&t=1077

Starts at the 18:00 mark about the snow walkers

[cunning mesa]

I mean, could be, but both of you SD3 examples don't really look like photos (or subjectively good) and the flux example is one of the worst realistic gens I have seen from Flux.

[uncut river]

yeah all is subjectively

[cunning mesa]

Flux is interesting because it can gen both fairly convincing lowdef images and also too highdef.

[bitter hearth]

all of these with lora

and im using schnell

[bitter hearth]

[errant dust]

Nice! What was the prompt? To get that cartoon style that is

[bitter hearth]

Here I go disappoint another person who asks for my genius prompts
My prompt for that cartoon style was: Cartoon style, the rest

👍

[errant dust]

and it is randomly choosing that variety of cartoon styles?

well, good to know it knows them. Now to figure out how to trigger them specifically. Thanks though.

[bitter hearth]

All 3 images I posted it cats were "Cartoon style"

Oh

And it had "detailed background" on negative

Not sure if that did much

[errant dust]

you are using negs on Flux? I know it works to a degree, however slow, but wouldn't have thought it would be desirable here

[bitter hearth]

I am using sd3

I am one of the few alive in this channel

wtf

[errant dust]

It did a nice job on the text

[bitter hearth]

you can get more effective negatives if you add perpneg or clipnegpip

[errant dust]

in flux, yes, but it is insanely slow

[bitter hearth]

depends on hardwar

[errant dust]

doubles or triples the time

[bitter hearth]

I was actually talking about SD3 anyway

but yes

it works on flux a bit too

I'm personally skipping flux in hope of a strong non-distilled model coming

[icy drift]

Does anyone know of a Clip-Vision model that works with Flux?

[bitter hearth]

I need more ideas

Otherwise

Balls

[icy drift]

Have you considered a hat trick?

[torn wharf]

you see if you put a traffic control person in the scene, now it 100% explains why everyone is so calm where godzillas chillin. its so smart im telling us

[bitter hearth]

[bitter hearth]

is that Albert Einstein

[torn wharf]

no longer its like "why everyone so calm around a disaster zone" now its clearly shown why people are so chill

[torn wharf]

thats just how smart it is

[bitter hearth]

Make him sign a tornado to stop

[mortal mesa]

tornado waiting at a traffic light

[torn wharf]

thats preposterous you can't traffic control a tornado. i'm going for realism here. this is to create a believable photor ealistic image

[bitter hearth]

The model is probably so confused by this prompt

[icy drift]

Just a bit of vacuuming. Not getting very good suction for some reason.

[mortal mesa]

Rat Capone's cigar tactic is new to me

[icy drift]

When you're not that rich, but you still want to say you have a swimming pool with cash in it for exercise.

[torn wharf]

did you try banging it hard?

[uncut river]

this (anime-sd15) --> (sd3) workflow gets a bit creepy ... uncanny stuff

[torn wharf]

i woudln't call that uncanny. its well within the bratz doll stylez that it just looks like harley quinn bratz edition

[torn wharf]

that one though oh god

[bitter hearth]

What's this? Hollywood actresses when they stop being famous?

[uncut river]

idk rich mom trying anime cosplay overdone hyper style

this one came out decent, a bit harley quinn as you noticed

[sage burrow]

Is this Flux, or?

[bitter hearth]

https://tenor.com/view/spongebob-no-nope-nah-naw-gif-16159366

[torn wharf]

classic

[real terrace]

Does someone has a workflow of the Schnell model with img2img added? I'm trying to do it but I'm dubious

[old plaza]

how do I send a prompt to SD instead of interacting with forum members?

[dusky thistle]

Cascade, the real sd3 lol

[limpid thunder]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[dull star]

beerus!

[sage burrow]

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

[rain current]

https://x.com/i/status/1822456967236821062 🤣

Blaine Brown  (@blizaine) on X

She must feel so silly

Twitter · Aug 11, 2024, 2:16 AM

[rain current]

Well, but that also happens to us when we're drunk. I've ended up drinking a cigarette and smoking a beer

[torn wharf]

love the part where the ai implodes her mouth and grows a new one

[fleet meteor]

cute 🥺

[alpine summit]

Sdxl

[torn wharf]

https://x.com/thibaudz/status/1823095294994632979

Thibaud Zamora ∞ (@thibaudz) on X

@mrflosunday don't know, i'm training it. i'm just waiting for it

Twitter · Aug 12, 2024, 8:32 PM

[sage burrow]

What makes you call it the real sd3? Also does it have nearly as much knowlege as SD3? I'll try it out.

[dusky thistle]

imo it is currently the best quality model we have from sai

[dull star]

I wonder what auraflow v0.3 will be like

[alpine summit]

Flux

[alpine summit]

Flux

[alpine summit]

Flux+sd 1.5

[torn wharf]

oh no look what i dropped. https://x.com/angrypenguinPNG/status/1823064101528260772

AP (@angrypenguinPNG) on X

FLUX [DEV] LoRA Training in Google Colab ❗️❗️

Notebook and code by the 🐐 @ostrisai

Twitter · Aug 12, 2024, 6:28 PM

hope no one takes advantage of that

collab keeps only giving me 15gb gpus lately though so i'll just keep checkin for the large being available

oh no the free high tier gpus are gone shhuu nm. that's just a link for people that wanna pay fr collab

[sage burrow]

THis is what it told me:

Gemini advanced tells me I won't be charged actual $ on google collab without me realizing it, so seeing how far I can get with my free comes with my yearly drive subscription collab

[compact forge]

Flux FS4 native upscaling workflow: https://civitai.com/models/642062

[sage burrow]

I had to try lol

[mortal kite]

is Flux allowed here

[torn wharf]

They seem to tolerate it. But not images of cylinders between balls

Imo it showcases what could be achieved with sd3 yet but rn it's more sd3 than medium is. M still has some strengths

[sage burrow]

It has kinda turned into 99% flux here oops lol

[odd basalt]

Yes shhh 🤫 😂

[sage burrow]

Glif is back to 100 per day, and has Flux Pro and Dev
(Schnell also, but who cares lol)

[sacred jewel]

Stolen prompt... very nice... thank you

[errant dust]

so interestingly enough, there may be guardrails against the most basic things such as "impressionist oil painting". Someone in Reddit suggested that lowering the CFG to 1.2 would bypass this, but a question I had was whether it was simply producing such a bad image that it 'looked like' an impressionist oil painting. I therefore set it to 1.2 and asked it first for an oil painting of two young men playing a game of chess in a park sitting on stone stools and playing on a stone table. Then I asked it for the same, but replaced 'impressionist oil painting of' to 'photo of'.

Both are CFG 1.2. So there may indeed be something to the idea of a very low CFG to access some not readily available choices. At least for art.

[bitter hearth]

not sure if this will work for flux
but for SDXL by far the best painterly styles I made came using the CFG++ versions of SDE samplers that were deliberately not given enough steps

[errant dust]

I used a massive 40 steps each here

[bitter hearth]

CFG++ is crucial because with CFG++ the image comes in much softer at lower steps

[errant dust]

so agan, this is not the result of it producing a lesser image than it planned

and as you can see, the second image still looks ultra sharp like a photo, even with CFG 1.2

[bitter hearth]

which sampler was it?

[errant dust]

euler

or ddim

let me look

they yield the same results

[bitter hearth]

its possible that euler was not solving well at 40 steps

[errant dust]

no, that is not the case

Euler produces stellar images at 40 steps

[bitter hearth]

it depends on the image sometimes
(because trajectories are sometimes more or less straight)

[errant dust]

Ok, so maybe me and the other guys are just having randomly bad results with Euler that happen to look like impressionist oil paintings

[bitter hearth]

generally to really look like a certain painting style, that would have to be in the prompt
but what I am saying is that euler tends towards a softer and more painterly look in general until it has had a ton of steps

because when the trajectory has wiggles, euler just goes straight over them instead of going into the dips

[errant dust]

Which it was

as I explained

[bitter hearth]

looking at the two images
it does look like the model has deliberately gone for "impressionist oil painting" style
rather than merely euler doing its thing and giving a soft painterly style
so I think you are right

[errant dust]

Regardless, it is something worth exploring

[bitter hearth]

is there any chance you could test it with DPM++ 2M at 60 steps or something like that

cos that would remove the possibility of euler adding painterly effects

would be a good way to counter the argument that its just the solver making the image that way

[errant dust]

Remember I told it to make a photo and it did

that did not look like a painting

also at CFG 1.2

[bitter hearth]

its ok if you don't want to
I just meant that that would eliminate the possibility of it being euler painterly

[errant dust]

Yes, same. If I can ask it to make a photo instead, all else being the same, and it does, then it is not randomly making it into a painting. In one it respects my choice of an oil painting, and in the second it respected by choice of a photo

same seed, size, and CFG of 1.2

If I increase to 3.5 then 'impressionist oil painting' looks like this:

[bitter hearth]

ok this is interesting

you may have found a jailbreak

[errant dust]

Not me, but I confirmed their results

[alpine summit]

Sd 3

Sd3

[alpine summit]

Flux

[hallow lion]

So best sampler and scheduler for schnell and dev. Did we settle on one yet?

[bitter hearth]

there's not really gonna be one best method cos people have wildly different preferences

[hallow lion]

ok

for realism

[bitter hearth]

I don't mean different styles of image
I mean that some people want fixed steps and some people want variable steps
and some people want a sampler optimised for 4 steps, some people want 20 steps, some people want 60 steps and some people want hundreds of steps

but your sampler choice is different for these different preferences

[bitter hearth]

I should add if you just want a general recommendation that will work well with a wide variety of different architectures then DPM++ 2M 20-40 steps beta scheduler

[noble coyote]

NF4 in Flux - 10th attempt gets the text right! Prompt = Times Square NYC a large neon sign high up on a building which says "Where were you going to, when I saw you coming back?"

Near Miss 1

Near Miss 2

[noble coyote]

Olivio Sarkas uses uni_pc_bh2/sgm_uniform - works OK

[hallow lion]

ok thanks guys

Hehe I am rocking Forge now. So simple compared for comfy... just to bang out some images fast lol

[sterile pendant]

Flux dev will make its own sd3-like sigma curve even on simple (which should just be a straight linear line), but it will work with sgmuniform and beta as well

Schnell will follow the curve of whatever noise you use though and won't use the built in sd3-like sigma curve like dev does with some of the schedulers

[sterile pendant]

here's a better visual of it. note on normal the last step, this usually bricks quality these flow models during the last step (makes them grainy or causes artifacts sometimes)

so that's why they use simple/sgm_uniform/beta

and then here's schnell:

(changed it to show 4 steps since it's centered around that)

[hollow swift]

what this graph means (grainyness)?

[sterile pendant]

it's the sigma graph of how much noise the model adds and expects at each step of solving

[hollow swift]

so it is better when there is curvature or linear?

[sterile pendant]

for these flow models, they expect that downward hill graph look of the three on the right in the first image

that's why you cant use karras with sd3 or flux

[hollow swift]

aaahh i see

[sterile pendant]

normal almost works with them, but note that last step

that i circled in red

[hollow swift]

thanks for the explaination, new knowledge for me!

[sterile pendant]

no problem, it's a super confusing aspect of image diffusion. but everything revolves around noise. the samplers are what then take expect noise levels and run calculations to fit the data

which eventually makes your image

this is curve fitting, now picture this with like a 30,000 dimension graph

that's what the samplers do

[noble coyote]

NF4

[bitter hearth]

Flux dev will make its own sd3-like sigma curve even on simplewill depend on nodes

this is partly why comfy gets confusing cos sigmas can both be changed by nodes that have green scheduler dots, and by nodes have have purple model dots

[sterile pendant]

The purple node pulls from the model to get the min/max ranges from it and the model type. Like if it's sdxl, max sigmas will be 14, if it's playground 2.5, they'll be like 120, sd3 or flux will be 1.0 and so on and it also preps some model related branches of code for the ksampler

But the curve calculations that are used are like what I showed. If you're manually screwing with the green sigma pins after that, then anything goes

Internally, with a basic loader prompt ksampler workflow, the sigma graphs will be what I showed

Things like shift will adjust how front or back heavy the curve is

Or how flat or curved they are

[urban totem]

In ComfyUI ,how do you show preview of image being outputted?

I had it once ,now its gone

[sterile pendant]

if you have comfy manager, it's in the settings on the right of the window

[noble coyote]

Go to manager and select TAESD preview method

[urban totem]

Whats the setting called?

[sterile pendant]

you can use taesd previews as well, but they are a hair slower

[noble coyote]

...changed to Latent2RGB!

[urban totem]

Oh ok I was looking in the gear settings

[sterile pendant]

oh and the setting will apply on your next queued image

if you're currently in the middle of diffusing one

[urban totem]

Got it! thanks good people

[sterile pendant]

damn look at my poor gpu over there roasting at 71C after a batch

[noble coyote]

I have a fan on top of PC as the internal fans are blasting a bit in this weather!!!

... but in cooler weather, my PC literally heats the room 😄

NF4

[sterile pendant]

well 71C isn't really that warm, just don't ever see temps like that during gaming or anything else. the hotspot temp is probably 80C and i checked to see if it's thermal throttling at all and it isnt

this is a 2080 FE that has seen heavy use since the month it came out and has never been repasted/repadded lol

i keep telling myself that i'll get around to it one of these days or that i'll buy a new gpu, but i keep procrastinating about either option

[noble coyote]

I am hoping to get a "5090" - or at least 24Gb VRAM; and 128Gb RAM

[sterile pendant]

china is getting some new version of the 4090 with like 48gb vram or some shit like that

[noble coyote]

Now 48Gb VRAM would be nice - but Beijing would see/hear everything I do!!! 😄

[sterile pendant]

nah, nvidia wouldn't do that for china. they are from taiwan

[noble coyote]

The detail from NF4! The hair on her arm, the goosebumps ...

[sterile pendant]

oh and it looks like there will also be a 4080 super version with 32gb vram

[noble coyote]

When BIG China eventually wipes-out Taiwan there'll be a GPU chip crisis.

In fact - an "everything-with-a-chip" crisis!!!

[noble coyote]

Me want!

[sterile pendant]

the 4080 super with 32gb sounds hella enticing. i'll hold off till i know what the 5070 models will look like though

watch them only bump it up to 16 or 18gb lol

[noble coyote]

Mebbe an A100

[sterile pendant]

i game

and am a game dev

so a100s, quadros, etc are a no go

[noble coyote]

My experience of games: Pac-Man, Asteroids, Space-Invaders, Solitaire ...

[sterile pendant]

well i'm in the latter half of my 30s, so i've done a lot of gaming in my life and still do when i'm not busy or dealing with family things

[noble coyote]

Waaay back in 1982 - a game for the BBC Microcomputer (32k memory) = Elite. The data was loaded by cassette tape!!! 😄

[urban totem]

Why don't they pack gpus with a lot more vram?

[noble coyote]

Price-point

BBC Microcomputers finally grew up and became ARM!

[sterile pendant]

cost and there's no actual consumer level demand for it

[noble coyote]

nf4

[sterile pendant]

contrary to clickbait youtube videos, gaming doesn't really need more than 16gb vram right now and if a game does, i'll tell you as a game dev, they have shit optimization

plus, with everything using nvme drives, there's no reason to keep massive texture pools when you can load gigs per second

wow F me... i've been using schnell for dozens of generations at 35 steps because i made that example earlier and didn't switch it back lol... the quality is still good though

[noble coyote]

Schnell is good at 4 steps ...

[sterile pendant]

yeah, but im using it for a full 35 steps lol

ooops