🆕｜sd3

[tardy sonnet]

I mean those from the original space are fascinating

Idk how to explain but the images look a bit oily… or idk blurry

And btw how to effectively use seed?

[bitter hearth]

putting quantized models on inference services is a common "trick" yeah

[pseudo owl]

I believe outputs are fine for commercial use, the license states:

Derivative” means any (i) modified version of the FLUX.1 [dev] Model (including but not limited to any customized or fine-tuned version thereof), (ii) work based on the FLUX.1 [dev] Model, or (iii) any other derivative work thereof. For the avoidance of doubt, Outputs are not considered Derivatives under this License.

[toxic bone]

Lets say the off chance. Lets just make a scenario up and try to think about how it woudl go in a court case.

You're making so much money from the commercial use of images from flux, that BFL invest in a legal team to sue you. They present their case and a judge says "I believe that this case has merrit because BFL owns the model, and the license 'contradicts' itself regarding outputs. Lets move forward!"

That's not a good position to be in, but you've made all that money. SO the best solution is to settle out of court. Since you've made so much that this whole thing actually happened in the first place. Hell, if it moves to trial past that, i'd get my lawyer to go for a jury case and have the FSF back me.

It's never going to go that way though since ti would end at the start where a Judge sees the clear language and agrees that no reasonable person would misconstrue that.

Comfyui might take a couple days to get a node together for this new "flux realism" 1st and 2nd order semi adaptive samplers. They're pretty good though.

They remind me of results i get from Bosh3 , which i think is 3rd order ODE

[edgy kelp]

Hey, how's going on with the "2 weeks"?

[pseudo owl]

You can probably just try out any space that does not use zero gpu. Try this: https://huggingface.co/spaces/Nymbo/FLUX.1-Dev-Serverless

[pseudo owl]

2 weeks for what

[toxic bone]

flux happened so things were reconsidered

[edgy kelp]

It seems so. 😎

[toxic bone]

Flux set the new standard hopefully. Never hype your model before you release it. Just release it.

[tardy sonnet]

As long as it doesn’t say zero gpu it’s unlimited?

[bitter hearth]

my reading of the license is that you can definitely sell your images

[pseudo owl]

Not all, but most.

[bitter hearth]

I really think the wording makes that ok

[toxic bone]

the wording is exceptionally clear by license standards

[tardy sonnet]

Great

Thx for all the help!

[bitter hearth]

there's an issue for people like the jugg team

cos they like to put jugg models on rundiffusion for a month before public release

apparently they contacted BFL

[toxic bone]

yeah thats right. Or civit ai is allowing users to train flux for buzz. that's commercial usage. even though buzz can be won, it is primarily created by buying it

civit obviously has license to

[sturdy dune]

I would like to believe you but it also states: _you do not receive any direct or indirect payment arising from the use of the model or its output: _

So there's conflicting terms.

[bitter hearth]

there is some ambiguity yeah you would want to talk to a solicitor if you were going to do a major commercial operation with the outputs

[sturdy dune]

I emailed the team.

[sacred jewel]

A new Steampunk LoRA

[sullen moss]

Hmm...

[sacred jewel]

SciFi Robot LoRA

Art Deco Vintage Future LoRA

[sacred jewel]

Slightly lower strength

[muted dove]

https://tenor.com/view/thats-clickbait-russdaddy-xset-thats-a-trap-hes-lying-gif-20538030

[frail shoal]

i can't seem to use loras in comfyui, i tried different lora loaders and they simply don't do anything, when the exact loras work on forgeui. I even downloaded the xlabs nodes for the lora and they give me "list index out of range error". Is there anyone who is using gguf flux dev and is able to use loras in comfy ui ?

[lucid kernel]

is there a way to prove which AI image was made by which model with precision that is provable in court? how would it work, just curious, let's say image from midjourney and from SD3 or whatever, brought to the court, how does the judge determine which is which?

[lucid kernel]

is it SD3 or flux?

[bitter hearth]

that's flux

[bitter hearth]

no there is not a way

but the thing people forget is a judge doesn't need to work it out from the images

they can subpoena people and get their testimony in court

and also seize computer equipment and storage, before you know they are going to

in a civil case they only have to be sure to the standard of "a balance of probabilities" rather than "beyond reasonable doubt"

[sacred jewel]

Flux

[bitter hearth]

its kinda safe to assume that images aren't SD3 these days

[bitter hearth]

I was looking on reddit, apparently Flux takes latent upscale and tiled upscale way better than SDXL

[dusky thistle]

i'm generating directly at 1344x1792 and even 1920x2560

no upscale

[bitter hearth]

wow yeah that's great

[bitter hearth]

so it could hit 5k with just 5 tiles

[dusky thistle]

yep

[lucid kernel]

how in the world would they seize something in the cloud running on unknown aws / azure / alibaba cloud / linode / etc' servers? 😅 especially when it's impossible to even know on which of those server pools does it run unless the company itself discloses it

[alpine summit]

Sd1.5

[dusky thistle]

they can make you do that

[dusky thistle]

and if you start deleting shit in response to the subpoena then you're really in trouble

[sacred jewel]

I have worked for Microsoft and also know AWS though colleagues that work there.

The CSPs will properly respond to subpoenas as they are NOT about to lose trillions in business protecting some entity by posturing against the government ...

So, the advice given to clients is to use their own encryption keys and protect themselves from legal government (court) requests. If you have your own keys for the encryption, it becomes much harder

but the CSPs will definitely comply with legal court subpoenas...

[sacred jewel]

Flux Dev failed here miserably where the Pro model seems to handle well.

[lucid kernel]

how can they comply if they have no idea on which of the 1000000 VMs your stuff is

[sacred jewel]

They ABSOLUTELY KNOW EXACTLY where things are running. How the heck do you think they bill you?

[bitter hearth]

lol

wtf

yeah they know where your stuff is

you have an account

[sacred jewel]

The auditing data is far larger than the actual data that is being processed.

[lucid kernel]

ok then 👌

so they can see what models a business runs

and if it runs both, they know which one was used for what image

[sacred jewel]

Azure and AWS and GCP absolutely know everything about every iota of their system... through and through ...

[lucid kernel]

since everyone has full logs about it on their servers of course

and they keep it carefully for the gvmt subpoenas

along with images sorted in folders

[sacred jewel]

Ask yourself this:

If you ran a multi trillion dollar business, would you risk it all to protect a rando?

[lucid kernel]

they also keep track about what models and executables were running in the servers a month ago, yup.

and they know your thoughts 😄

[bitter hearth]

also, do you know what spoilation sanctions and adverse inference are in law?
a judge can just rule that the case proceeds as if the destroyed evidence was evidence of guilt

[sacred jewel]

You can think what you want. I am not trying to convince you of anything. Just giving you the perspective from my POV from my own knowledge.

In LAW, Discovery is a powerful thing. You must disclose a lot of things you may not want to disclose. Each case is different and you can;t really make it all black and white.

Can you hide things? Of course. Is it hard to find the little details as you point out? Sure. But is it impossible? I would not bet my life on it 😉

[bitter hearth]

I actually know someone who lost a case cos they destroyed evidence
and the ruling assumed the destroyed evidence was bad

[sacred jewel]

And do not forget that Bin Laden was found and killed. He was more protected than an AI model running in any of 10,000 Virtual Instances 😄

[lucid kernel]

it only proves that US legal system can screw anyone no matter guilty or innocent if it wants to. 🤷‍♂️

[dusky thistle]

i'd argue that's more about not letting ppl get away with whatever the hell they want, but we're getting off-topic here

[bitter hearth]

there was a cool paper called res master
its like tiled upscale but it used an LLM to make a better prompt for each tile automatically

normally when I do tiles, only one tile has R2D2 and the rest don't

but if I use one prompt for all tiles (like ultimate upscale) it risks putting little R2D2s all over the place

they also said it worked better cos the LLM can think of details that would suit that tile better

sadly no code yet

[hexed dirge]

hello do anyone use fluxgym?

[noble coyote]

Its having teething problems ... see here

[hexed dirge]

for me it doesn't work. It stops at 2 or 3 %

[noble coyote]

Use the Pinokio Forum - I think he's clued up?!

Becky93 LadyLalitaArt is trying to use it as well. Perhaps DM her?

[sage burrow]

Trying being the key word there 😦 I deleted it for now. Someone on here did mention that it was created more for Linux than windows...

[sturdy dune]

There's only telemetry data like up/down, memory and processor info.

A cloud VM is just a reused container, they don't keep images or backups of server, once you cancel your VM it's served to another user.

[sturdy dune]

And I would imagine you mean http logs? Those wouldn't be stored either and still wouldn't give a picture.

I also don't know too many people using cloud instances for their AI from a business perspective.

[sage burrow]

Have you been playing with glif? 😄

[errant dust]

Believe it or not, that is a real screenshot I made

https://www.amazon.in/ACEMAGIC-Tank03-Desktop-Geforce-3060/dp/B0D2R3HFFF/ref=mp_s_a_1_47?crid=2262ZQ3MRO5EF&dib=eyJ2IjoiMSJ9.FjEaUHMnnyyagJ_umq-G0lLy_GY37jZBz-g49HkKCp16x5cSn7Rw5OrY1gaK-kl-G4ApllOH1-b5vfUNMHKFf6q_1zJzo5VeXTszOPO1BxYrGxWVkkJ8PG0nuflHcVIXdC5JYDJ7uJ-yQCgWSqGSliFHsVLPw3QCG9JqSdbLvGNGvQUaPYTfxUQg8yll7S2i7Lo8ItJvEyvK9guCUAavoQ.WemBpqgHjkRVIDniqt8apYJGN_XmodbZKfMgYY2CBIo&dib_tag=se&keywords=mini+pcs&qid=1725913633&refinements=p_36%3A11310000-&rnid=7252027031&s=computers&sprefix=mini+pcs%2Ccomputers%2C298&sr=1-47

[sage burrow]

I thought you were using the fake amazon products glif lolol. I hope that person who gave it a 1 star never bought it 😦

[errant dust]

It was sheer coincidence. A few hours ago I was consulted for hardware and was asked to give an opinion on these mini PCs of which this was one on the page. Eventually I steered the person off of this idea because it wouldn't suit his needs and came up with a solution that did

[sage burrow]

You are so polite! I just straight up tell people it's fake garbage 💩 ROFL

Now really scarey will be when the scammers get better at using flux to create images of supposed computer systems, GPUs, etc. 😦

[errant dust]

It isn't simply a question of being polite. It doesn't really help the person. If they ask me about it it's because they're in need of a machine and for some reason this is the type of configuration they're looking for.

Eventually it turned out that he was looking for a reasonably fast machine that he would be accessing remotely while traveling. In view of that I suggested either a very powerful laptop or a good desktop instead. We eventually settled on a good desktop because it would suit his needs better. A powerful laptop on the road has all kinds of problems

[sage burrow]

I'm pretty sure I'm just jaded, after the 1000th anonymous email asking me for advice on specific products...

What sort of problems re a powerful laptop on the road? 😦

[errant dust]

Weight, size, and battery life

If you do a lot of travelling and airport moves, you don't want to be obliged to plug in a brick all the time, because your uber laptop has a practical 2-3 hours of battery life tops

[sage burrow]

and last I checked max 8gb GPU 😭

[errant dust]

No, you can go up to 16, but that is even worse. For this use case

because then you are talking a 4090

[sage burrow]

a cell phone and a pair of 4090's on a cloud! 😄

[errant dust]

He is one of the world's best chess players and will need heavy duty computing power on the fly. Much easier to run that remotely

and for casual work, browsing, video and so on, a good travel laptop with top notch battery life

[sage burrow]

Btw, have you ever built any, for aesthetics only, chess loras? 🙂

[errant dust]

I have not, but I am considering it

I wonder whether it will learn the proper dimensions of a chess board for one

I'd have to carefully supervise the image descriptions too. I mean, software like Florence 2 is a disaster for image labelling

Imagen 3 is quite decent overall for chess, though it is no surprise since it was developed by Deep Mind

Whose CEO is a former chess prodigy

[sage burrow]

not flux? 😦

[errant dust]

Flux is solid too, but it depends. They both depend to be fair

[sage burrow]

I just looked on CIvitai for chess, expecting to see amazing loras, since so much art uses chess, or chess characters as a metaphor, but.... there doesn't seem to be any!

[errant dust]

Flux is the one that made my new Logo

(that is the name of my channel BTW - Flux nailed the entire thing)

I told it to make a green circuit board background, so that too

I haven't uploaded it yet, but it will soon be

soon as I finish my next vid it will all go up together

[pseudo owl]

kinda crazy how a model can get such good text and such good anatomy at the same time. flux is great

[sage burrow]

How do you get long sentences to work?

[sullen moss]

Yeah, it’s already nearly halfway through September and there’s no news from the SAI team. It seems to me that things are so bad with the models that there’s no point in releasing them now. They probably don’t want to embarrass themselves again

[errant dust]

Well they're working on something

[pseudo owl]

How long? It can usually do 1 sentence accurately, and requires like 2-3 tries for 2 sentences. For more, its difficult, there is usually words repeating or some words are not appearing.

Any example prompt thats not working?

If you are using flux schnell one thing is that its actually much better at 512 res then 1024. You will get far better text. Upscale it to 1024 or higher after you generate a image with 512 res.

[sage burrow]

Most of them start to get messed up by the 6th or 7th word in. Do you separate the 2 lines? For example- HOlding a sign that reads "yeah flux is pretty" "impressive isn't it"

Though perhaps only using known words helps...

I'm not much of a Schnell fan, so Dev or Pro it is 😄

[pseudo owl]

Yeah dev is better then schnell even at 512 res, try like a sign that reads "yeah flux is pretty impressive, isn't it?" written out in colorful ink

[sage burrow]

Btw, amazing job with all those 5 fingered hands! 😄

[pseudo owl]

no "written out" part(left) vs "written out" part(right)

prompt: A photograph of three men sitting in a chair showing their hands. Behind them, is a woman holding a sign that says "The scrabble doo of the discord prodia is nothing but rubless" written out in colorful ink. It is outside and behind the woman is a store that says "SD3 SUCKS" written in neon lights.

probably a more fairer comparison is this with this, same seed as well. The written out one is still clearly better(above right)

[bitter hearth]

would still be good if they released SD 3.1 even if the image quality isn't as good as flux

cos its not distilled it has different pros and cons

could use flux as a refiner over it also

[toxic bone]

i think that'd be them admitting defeat and giving themselves an expiry date. They should release something as good or better, or better significantly in other ways. Like undistilled weights we can affect hard.

[bitter hearth]

I mean Lumina, Kolors, Pixart etc were not top model upon release

[toxic bone]

or call it "the New SD3" and then retire it when they bring back SD Classic, an SD15 esque model with ultra classic flavor

[bitter hearth]

the next Pixart is actually the model I am most excited about cos it will be Apache license, and they joined Nvidia recently

[toxic bone]

Diet SD

[bitter hearth]

its different pros and cons though

any custom node or tool that requires the model to use negatives at a reasonable CFG level can't work on flux
so a second good quality model that is not distilled would be great

[cursive frigate]

I am curious. What are some of your favorite upscale models? I like 8x_NMKD-Superscale_150000_G, and
4x_NMKD-Siax_200k.

Does anyone find that a different one works better?

[bitter hearth]

yeah this one
4xNomosWebPhoto_atd.safetensors "https://github.com/Phhofm/models/releases/download/4xNomosWebPhoto_atd/4xNomosWebPhoto_atd.safetensors"

[sullen moss]

and this one

[bitter hearth]

yeah thats a good one

DAT architecture is also great

I like HAT and RGT as well

[pseudo owl]

true, sd3 8b is worse then flux but still is smaller and faster and is undistilled so a few other pros. Would be nice if they released it.

[sullen moss]

https://openmodeldb.info/

[cursive frigate]

Are the upscaler models that you guys listed what you would recommend for flux? or does it not matter once the image is rendered the first time?

[bitter hearth]

yeah, and SD3 architecture had a couple of other advantages also which were higher diversity and less overfitting on a certain aesthetic style

what we could be able to do it make the initial image in SD3, inject noise and then refine in Flux

cos Flux VAE is a bit better than SD3 VAE

[cunning lintel]

and has been "unfinshed" longer than it took to create flux from 0 😬

i really really hope we get a SD3 that looks like a SAI model with a much more natural look and easer styles than flux, but well, who knows what SAI is planning if anything at all

[bitter hearth]

ending every image in Flux is a good idea because of the VAE
even if you only denoise a tiny bit
I've started doing this with SD 1.5 and SDXL

[cursive frigate]

And that was done with 4xFFHQDAT?

[sullen moss]

Yep

[bitter hearth]

its difficult because flux is so new there is less tooling

[pseudo owl]

we need a flux paper

[bitter hearth]

for SDXL in Diffusers we have a bunch of super resolution libraries like Hidiffusion, DiffuseHigh, Scalecrafter and Demofusion
these all take SDXL to 4k resolution in one pass
Flux is too early to have stuff like this

you can do tiled upscale like SD Ultimate Upscale

Flux does rly well with that

[cursive frigate]

Are there any upscalers specifically recommended for anime?

[bitter hearth]

yeah on openmodeldb.info/

you can search for anime

that's easily the best upscaling website out there

[cursive frigate]

I really appreciate the resource referral, Upscaling has been driving me crazy lately.

[bitter hearth]

it only works for SD 1.5 and SDXL but my favourite method before flux was
two passes using gradual deep shrink, res-adapter and perturbed attention guidance

can get SDXL to 5k that way

I don't know yet how to do this well with Flux, my Flux images are definitely better than SDXL at 1k but its hard for a 1k image to beat a 5k image

[desert garnet]

these work great
https://openmodeldb.info/models/4x-NomosUniDAT-bokeh-jpg
https://openmodeldb.info/models/4x-IllustrationJaNai-V1-DAT2
https://openmodeldb.info/models/4x-Nomos8kSCHAT-S

[sullen moss]

https://openart.ai/workflows/elephant_insistent_10/flux-controlnet-comparison/k0KBCt12RDUOp2c71jEs

[cursive frigate]

Do you use this workflow?

[sullen moss]

I work in Forge...

[bitter hearth]

instant x pu depth looks good then

it didnt get the starbucks logo right cos a logo is flat so thats ok

[craggy crest]

that's quite the price tag

i hope the machine is more than just 'it looks cool'

[bitter hearth]

is there a way to view final shift in comfy for flux

cos I just used RX808's calculator and my shift was over 11 LOL
no wonder it didn't work

[cursive frigate]

Is it just me going crazy or is anyone else getting streaks or stripes in their images?

[runic tusk]

Which model?

[cursive frigate]

flux dev

[runic tusk]

Yup. It's actually in pretty much every output from Flux, but it's more noticable in some outputs more than others.

[cursive frigate]

ohh well at least im not going crazy or have a busted monitor. lol

[runic tusk]

Nope, you're not.

[cursive frigate]

Is there a known fix?

[runic tusk]

It seems inherent to the model itself.

[desert garnet]

yea if u upscale

idk about comfy but a latent upscale should do it

[errant dust]

Couldn't say and don't really care. I steered him away from that

It's way too slow a machine for his needs

[cursive frigate]

Upscaling seems to get rid of the banding some of the time but not all of the time. Just did 5 in a row.

I wonder if its more of a watermark to discourage commercial usage?

[bitter hearth]

changing shift value helps a lot

not sure in Forge but in comfy this is the shift formula: ``` x1 = 256
x2 = 4096
mm = (max_shift - base_shift) / (x2 - x1)
b = base_shift - mm * x1
shift = (width * height / (8 * 8 * 2 * 2)) * mm + b

but once you have the shift number it works the same as SD3

[cursive frigate]

I don't know where I would need to go to make those changes. I honestly don't even know what shift numbers are or how they affect the image.

[bitter hearth]

they make a transformation to the sigmas

[desert garnet]

yea it happens to me as well,i only fixed it with higher denoise but that changes img too much

[cursive frigate]

not sure if this helps your to help but but this is what I see here

[bitter hearth]

I don't know how Forge and A1111 have implemented it
but if you use comfy, make sure you have the ModelSamplingFlux node
and if things go bad in the image, try moving max shift up and down

[bitter hearth]

what was the resolution?

anyway that's a pretty high shift by flux standards

so lowering it may help

[cursive frigate]

whoever created it chose the route of aspect ratio vs pixel dimensions.

also I am aware that it doesnt say flux

[bitter hearth]

not sure about this node

[cursive frigate]

Do you think I could swap that node for image size of some kind

[bitter hearth]

you could simply swap that node for an integer

the only output is two integers

anyway for your image try max shift 1.4, 1.3, 1.2 etc

if you are at 1024x1024 then base shift doesn't do anything

to make it simpler, make the base shift equal to the max shift each time

[pseudo owl]

Anyone know a good interpolation model? The frames are going to be like 1 second apart.

[cursive frigate]

Ok changed out that node for another and changed the shift numbers. going down by .1 each render.

[cursive frigate]

Should I leave the base shift alone?

[bitter hearth]

would be easiest to set base shift to whatever the max shift is

cos then your final shift is also equal to them

and then just try different numbers

[vast condor]

what shift do you normally use? I adopted a workflow that has like 0.40 for base and 0.80 for max, and I basically never touch it

I've also never had streaks, so maybe that setting was the issue for the other person

[sacred jewel]

What are the recommended defaults by BFL if that is even known?

[bitter hearth]

BFL reference code was this: (its also the default in comfy)

        x2 = 4096
        mm = (max_shift - base_shift) / (x2 - x1)
        b = base_shift - mm * x1
        shift = (width * height / (8 * 8 * 2 * 2)) * mm + b```

[bitter hearth]

need to know resolution

512x512 needs a lower final shift than 1024x1024, which needs a lower final shift than 2048x2048

[sacred jewel]

That's math... not English 😛

Seriously, I saw you post that earlier. Can you just post integers or floats please for those of us who are not interested in Math? 😉

I assume these are the numbers?

So, if I use a different resolution, I am technically supposed to be using a different set of shifts? Talk about being ignorant LOL. I have been using the dfaults for all kinds of resolutions LOLOL

[bitter hearth]

I could make a table tomorrow for each resolution and which shift I think looked good

yeah the whole point is its different at each resolution (that its higher for higher res)

[sacred jewel]

Wow, that would be AWESOME. Maybe I'll make a node to do this for me automatically... It's been a long while since I made a node

[bitter hearth]

having two number for shift is confusing people
comfy converts it to 1 number anyway before using it

its the same thing as SD3 shift- when shift is higher it spends more steps on larger details (big sigmas)

and when shift is lower it spends more steps on lower details (small sigmas)

[sacred jewel]

Thank you...

[sacred jewel]

Cyberpunk Noir LoRA at different strengths

https://huggingface.co/Shakker-Labs/AWPortrait-FL

[young blade]

Wait wut? 😩

[hallow lion]

lol that camera...

[young blade]

Please do. 🙏

[young blade]

Not normal for a camera to have a muffler? 😂

[fast dew]

I finally figured out dall e 3s little trick

[open marsh]

whats the best way to ensure your safety when installing custom nodes on comfyui? the LLMVISION situation got me real worried

[bitter hearth]

copy paste the python code after reading every line
repackage nodes yourself

this way lets you customise things a bit anyway

[open marsh]

thank you! is there a tutorial or anything on how to do these? im still learning the ropes and im only gonna be using less than a handful of custom nodes but i want to make sure even if im not using a lot.

[bitter hearth]

I don't know about tutorials but what you can do is install a couple of custom nodes using the manager and then compare their github repo to their layout in your custom nodes

[sullen moss]

https://github.com/xlinx/sd-webui-decadetw-auto-prompt-llm

How I miss more video memory. I hope the antitrust committee puts pressure on Nvidia, and they release the 5090 with 32 gigabytes

[open marsh]

Hmmm ok ill look into that! thank you! so as an example, im looking at the github repo for WAS node suite. are you saying to look at the github repo folder for that node and look at how the various tools for the custom node is organised in comfyui?

lastly, would it be possible, and if so, would it be safer to manually (without manager), install specific tools from a custom node without the entire thing?

[bitter hearth]

yeah, also remember to look at requirements.txt

yeah its better to not install the entire thing

[cursive frigate]

It would be even better with 48GB of VRAM

[open marsh]

dang, unfortunately im not that familiar with reading codes or know what to do with them

[open marsh]

sorry to be a bother, by any chance, could you please tell me the steps to do this or refer me to a tutorial on how to go about this? cant seem to find any instructions 😅

[bitter hearth]

okay here is the result of my experiments
for each one I gave the final shift that was the comfy default from the BFL code, and the final shift that I preferred:

448x448 default 0.59, preference 0
512x512 default 0.63, preference 0.3
1024x1024 default 1.15, preference 0.5
1536x1536 default 2, preference 1.7
2048x2048 default 3.2, preference 2.3```

[bitter hearth]

luckily the way custom node repos are laid out often makes this easy, as each node is somewhat seperate

so for the most part you can just copy and paste the code for just that node, making sure you include any code that it requires

[cursive frigate]

Does anyone have a workaround for this or another node perhaps that can go in its place?

It seems to be the only "broken node in this workflow" and I can't fix it.

[noble coyote]

Flux LoRA Training down to 8Gb VRAM https://www.youtube.com/watch?v=nySGu12Y05k

[noble coyote]

If you use ComfyUI Manager - you can modify the security setting in Config.ini - from "normal" to "weak" and then try and install again

Put the security setting back to "normal" asap as it helps protect against malicious code found in some nodes

[cursive frigate]

Where is that file? I am having trouble locating it.

[noble coyote]

custom_nodes\comfyui-manager

[cursive frigate]

ok

ty

[noble coyote]

You're welcome

[sacred jewel]

Thank you

[open marsh]

i must be totally missing something, because i cant seem to at all even after researching, find the code for lets say "image filter adjustments" for the custom node WAS node suite: https://github.com/WASasquatch/was-node-suite-comfyui i know i must be doing something wrong

[sacred jewel]

You have to search each of the .py files for the class that defines the node. It's tricky but not hard.

For example

The class you're looking for

[bitter hearth]

lines 2745 to 2904 in WAS_Node_Suite.py

[sacred jewel]

The nodes are defined in this file on the root of the repo

WAS_Node_Suite.py

But you need to be careful with the dependencies.

[open marsh]

omg thank you guys, i was going through each one trying to find the actual name "image filter adjustments" 🤦‍♂️

[open marsh]

as in id need to get the right dependencies for that specific tool from the requirements file or else it wont work?

[sacred jewel]

Yeah. You can sort of stumble through it if you want to be lazy. Comfy will tell you on load that it's happy or not 🤭

Or try and follow the code in the class And what functions it uses and then look for them referenced somewhere in the top of the file.

Say you leave out something needed, the console will spit out an error that it needs X for it to work. So go back and put that back or comment it out or in.

[cursive frigate]

Ok I figured I should strike while the iron is hot.. I was able to fix my reactor node issue. It turns out I just didn't use my eyes to read all of the instructions "PEBCAK" issue lol. Does anyone have any advice on the exact version of what is it "torch+cuda" so I can install xformers if that is used these days and so I don't break most of my flux nodes?

[noble coyote]

Torch 2.3.1cu118

[muted dove]

Shouldn't need xformers with ComfyUI, so can use pytorch version: 2.4.1+cu124

[sage burrow]

How long does it take with this method, using a dataset of approx 75 images?

Is it an overnight thing, or?

Dude seriously needs subtitles! 😄

[noble coyote]

I've not tried yet ...

Furkan is from Turkey. Apart from his accent, he is very thorough!

[sage burrow]

I've downloaded kohya twice so far, flux is never on the drop down menu 😦

[noble coyote]

I'm not sure as I have yet to follow his tutorial! If I find out, I will let you know 🙂

[sage burrow]

If it runs with only 8gb gpu, that means it should run fine on Google collab free? Hmmmm

[open marsh]

ah i see, essentially, for each of the lines that ill need i.e. the whole class or just the definitions, ill just copy the line and paste it in comfyui's cmd and make sure that i also copy and paste the correct line from the requirements as well?

[noble coyote]

... just installed SECourses Kohya SS SD3-Flux LoRA Trainer ...

... which also works at 8Gb ...

[bitter hearth]

you need a folder of furkan's face

its the key ingredient

[noble coyote]

A cats ass? >^..^<

I want to do an Andrea Kowch LoRA, a Gary Walton LoRA, a Norman Rockwell LoRA, a Jo Grundy LoRA ...

@bitter hearth let me have your folder of Furkan's face?!?!?! 🥳

[open marsh]

im guessing 2745-2907 would be the function that youre speaking of?

[bitter hearth]

artist loras are a great idea yeah

[bitter hearth]

yeah that's it

[open marsh]

oooh okok so im looking at the requirements page and i cant seem to find anything that would give me a hint as to which line to copy into the cmd

[signal shuttle]

"Speed limit enforced by aircraft"

[open marsh]

@bitter hearth @sacred jewel hey guys, thank you so much for your time and patience to helping me. i've finally found what i needed to do! your efforts really made a difference in pushing me towards the right direction and i appreciate it so much! 🙂 i hope yall have a great week! thank you both again, i really appreciate it!

[bitter hearth]

yeah no worries 🙂

[cursive frigate]

@noble coyote @bitter hearth @muted dove and anyone else who has helped me with resources, suggestions and support. Thank you for all of your help take a look at what a difference 1 day makes and finding the right workflow.

[bitter hearth]

that's loads better yeah

[noble coyote]

Well done Scorp! 👍🏻

I am being distracted as I've just taken a subscription to Disney+ - watching "Only Murders In The Building" 🥳 and "The Bear" 🙂

D/loading some massive Flux checkpoints for the Kohya LoRA Training ...

Florence2/Flux img2img (Public Domain original)

[sacred jewel]

https://tenor.com/view/peace-out-peace-sign-peace-ice-age-eddie-gif-16750982996130936929

[sacred jewel]

Hook a brother up with the Normal Rockwell LoRA when you get to it 👊

[mortal mesa]

i did a thing like that, was amazed at the detail it can scrunch into it

[sacred jewel]

Yeah, Flux raw (no LoRAs) using just basic settings, ALWAYS impresses. I have literally given up on Upscaling, Detailing, Face fixes, hand fixers, shifters, sliders, sigmafiers etc. LOL> It is such a fun Model to play with...

[errant dust]

I like the LoRAs, but as to the rest, yeah, I don't bother unless I actually need a large image, in which case I use Topaz

[sacred jewel]

Oh don't get me wrong... I LOVE LOVE LoRAs... they;re awesome. I just meant that the base Flux is so much fun, you really don't need LoRAs to have fun with it...

[errant dust]

Very true

[bitter hearth]

luckily Flux also does better with tiled or latent upscale than any other model also

I've not even seen 1 seam yet between tiles

and you can push the latent upscale multiplier higher

the tricky part is constantly changing shift for each pass in the workflow

[sacred jewel]

Looks dangerous... so I love it 😛

[carmine harness]

Excuse me, I am currently training Flux LoRA. What is the best training tool to achieve the highest quality results? I have only used the ai toolkit so far, and it is still fuzzy when it comes to handling complex building structures, such as steel frames and scaffolding-type complex assemblies. If I change the tool, will I get better results?

[noble coyote]

Try FluxGym or this from SECourseshttps://www.youtube.com/watch?v=nySGu12Y05k
YouTube

[carmine harness]

Thank you very much. My own Windows PC has poor performance, and I am currently using it on a Linux cloud server. I will try the Linux version.

According to my previous knowledge, a higher value of batch_size * gradient_accumulation_steps can produce better accuracy. Is this true? If so, I will prioritize using a 48GB GPU memory to use higher parameters.

So we will strive to improve accuracy.

Due to the large number of complex temporary structures in the construction industry, it may be because the industry generally has poorer image materials on the Internet, and the details learned by the large model are not as clean as the renderings.

[noble coyote]

Florence2/Flux img2img

[brittle nexus]

Before and after image

[errant dust]

Before and after what? She bore the Alien?

[desert garnet]

reddit mod vs after deleting reddit account

[brittle nexus]

It is an image generated asking for a before and after comparison of the same woman after losing weight

[craggy crest]

musk employee vrs working for anyone else.

[sacred jewel]

DreamyKolors LoRA

Another great one from @hexed dirge

The Desolation LoRA

[hexed dirge]

working on this 0ne now. Will publish tomorrow evening

[bitter hearth]

some of the Civit flux loras have been better than others

have had a familiar problem where it wants to bring the background with it

also ones which force the model into either realistic mode or drawing mode

some are very attractive though

[toxic bone]

can say that about all loras. wait until frks lessons get spread around a few times. then you'll have loads of loras that dont' work. dude still uses the unique token lora style that never mattered

[bitter hearth]

also he overtrains the living daylights out of the model

although that's not necessarily bad as that seems to be what a lot of Civit users like

if they train a style lora they just want to be blasted by that style and that's a "great lora" to them

I thought Civit would of censored itself into oblivion by now

they're part of the Open Model Initiative to make a new model
its possible they will end up doing very well

They were having an identiy crisis last time I used them

do you know about civit green?

they made a SFW site

interesting

ye it suprised me a lot

[toxic bone]

its the same civit more or less. just with the ratings limited to pg13. but people are people and are working hard to get past the rating classifier and fill it with 'content' anyways

when i use it to search stuff too it keeps leading me to "not available" pages. its all pulling from the same db. it's just a front end filter

best part about the filter'd domain though is paypal will work with them again.

[sacred jewel]

BL4C LoRA

[cursive frigate]

Ok folks I have a new problem... I am getting like the lines of a square or large grid in my images, but its only after I upscale the image. Does anyone know what settings I can try to adjust to resolve that? Its very slight but still very noticable.

These are my upscale settings.

[sacred jewel]

Hope Style Poster LoRA

[sacred jewel]

Try increasing the overlap to 16 pixels and the blur to 32

Sorry, these two settings 😄

Also try this custom node instead of USDU.

Specifically McBoaty Upscaler

https://github.com/MaraScott/ComfyUI_MaraScott_Nodes

[cursive frigate]

So keep tile padding at 32 and increase mask blur to 16?

Also I think I will try that node for upscale.

[cursive frigate]

That seems to have fixed it... Thank you 🙂

[kind shuttle]

/doeam
主题：《从释经到讲道》学习班
时间：2024年10月9日-11日
地点：花园酒店
导师：知名牧师吴荣滁牧师
内容：包括深入研究圣经、理解其中的神学思想和传达方式，以及实践性的讲道技巧，以提高你的领导能力和牧养工作效果。

[cursive frigate]

Does anyone esle have this problem where there is extra fat or almost boob like features to where the abs or stomach area should be on a person?

it seems to be often and I'm not sure what is causing it

I think this is happening during the latent injection part of the process. Not sure what to do to resolve that.

[craggy crest]

@bitter hearth

[toxic bone]

they're eating the pets. you woudn't believe it , but they are

[muted dove]

Loneliness...

He'll eat you for lunch...

[real valley]

chinese food，minimalist style,brief strokes,vector diagram,masterpiece,best quality,cartoon,icon,food,food focus,simple background,

[noble coyote]

Captioning pre-training using JoyCaption - Flux LoRAs soon! 🙂

[sage burrow]

Flux can do motion lines

[alpine summit]

Flux

[sage burrow]

All your other images have been flux as well, or?

[muted dove]

🤣

[alpine summit]

I guess you're right, except for those cat pics though😁

[sage burrow]

Flux without a lora is what is causing it

[sage burrow]

Oh! They look flux like too. Apparently I can't tell lol

[alpine summit]

Sdxl

[bitter hearth]

for anyone who likes upscaling, this dropped this week:

https://github.com/Phhofm/models/releases/tag/4xNomos2_hq_atd
its the strongest I have seen out of the ERSGAN/SWIN-IR style upscalers

[muted dove]

I can't see any difference to https://github.com/Phhofm/models/releases/tag/4xNomos2_hq_dat2

[bitter hearth]

I've used both a lot and what I would say is the difference is pretty subtle
between the top models (ATD and HAT-L) and DAT or DAT2

[muted dove]

Lots to choose from! 😄
https://openmodeldb.info/?q=4xNomos2

[bitter hearth]

hmm I just alt-tabbed between the first example

that white building

the ATD is a fair bit better there

look at the text

[muted dove]

It's about the only one that shows a difference, but yes, that one is better.

[bitter hearth]

its pretty subtle on the others yeah you are right

I took that and fed it into his SUPIR workflow and it helps as well

its only a 1x SUPIR pass it just softens it a bit

[brittle nexus]

No. It's ssomething in your workflow or prompt

[alpine summit]

Sd1.5

[sage burrow]

Flux 😄 lol

[hexed dirge]

Hello

https://civitai.com/models/742916

[toxic bone]

NICE i love seeing people making proper loras. You hit the style perfectly and it's < 100MB.

Frk out here teaching people to use 128 rank has me worried for the future

2.5GB lora that barely even works and took 37 hours to train on 4x A6000. Wtf? https://civitai.com/models/731347/secourses-3d-render-for-flux-full-dataset-and-workflow-shared

https://github.com/kohya-ss/sd-scripts/tree/sd3#key-features-for-flux1-lora-training It's worth noting that the kohya project waited until he dropped all his video tutorials about Kohya scripts for flux, before they released this guide that contradicts a ton of what he says in vid.

[hexed dirge]

Thanks

[bitter hearth]

can't necessarily assume by default that Kohya's methods are right either

[left parrot]

Holy cow! Prem Akkaraju just spoke at the AI Conference in San Francisco, and announced that SD3.5 Large will be released in a few weeks!

[hexed dirge]

well I hope it will be a great model

[pseudo owl]

Nice, but wish they don’t extend it to months like they did with sd3.

[hexed dirge]

they must come out with something that beat flux and it's not that easy

[noble coyote]

What Captioning software works well? I cannot get JoyCaption to work ...

[brittle nexus]

2 weeks...

[hexed dirge]

florence

[noble coyote]

OK ... didn't think of that 🙂

[bitter hearth]

cogVLM is most common, but lots are good

[noble coyote]

JoyCaption loads the shards - but then just hangs?!

Two Weeks ™

[hexed dirge]

Florence is inside FluxGym

very useful

[sage burrow]

2 weeks?
Sorry couldn't resist 😄

[sage burrow]

Wait, can joycaption be used locally?

[noble coyote]

Yes, its GRadio-based

[signal shuttle]

3.5 large? so the 8B model?

[noble coyote]

What node would capture in a text file Florence prompts/captions?

[hexed dirge]

can't undestand. It simply caption images

[noble coyote]

I am putting captioned images into Flux LoRA Training ... so I must have a file of captions - I think "Save Text" might do it?

[hexed dirge]

ah well no it do everything also training

caption is the first step and it saves text on your pc

then you can train

[brittle nexus]

Could work. That or IF Save text or whatever they use with Joycaption

[left parrot]

@signal shuttle Yes, the 8B. The actual timeframe mentioned was "days or weeks" so they seem fairly confident that it's close to ready.

[signal shuttle]

Well i am going to keep my expectations low so i don't get disappointed, but i hope i am wrong and it turns out to be a amazing model but after the whole SD3 medium thing i have no trust in SAI

[left parrot]

I'm curious how many "girl lying on grass" images will get generated per minute on the release day 😆

[noble coyote]

If ClipDrop is aka 3.5Large - then it will be a tad underwhelming, and little like FluX!

[noble coyote]

My JoyCaption hangs right after the checkpoint/shards loads ...

[bitter hearth]

ClipDrop got sold to Jasper a while back

[noble coyote]

But what standard of SD3 is ClipDrop? Its prolly pretty close to SD3.5L I'd guess?

[bitter hearth]

why?

[noble coyote]

ClipDrop is 8B ?

[bitter hearth]

ClipDrop is a separate company now

[brittle nexus]

But the workflow doesn't have a node to save the text?

[bitter hearth]

so they can't access internal models

[noble coyote]

I agree, but the quality of their SD3 model ... ?

[noble coyote]

It saves text to a .txt file

[bitter hearth]

its either the 2B or the 8B of SD3

[noble coyote]

If it is the 8B of SD3 - I suspect that SD3.5L will not be much different?

[bitter hearth]

oh it could also be the ultra pipeline

we've already seen the 8B

not sure why you are judging SD3.5 large by what is on clipdrop

[noble coyote]

OK, just a surmise ...

[mortal mesa]

the trust and safety disaster lady is gone, fingers crossed

[sage burrow]

Glif is starting to implement Flux lora training!!!! You have to apply for it though, with an idea which isn't too redundant.

[noble coyote]

??? 🥳

[mortal mesa]

ya at Meta now, funny thing is llama 3.1 wouldn't make a prompt for make SD3 great again "I can't create an image that promotes harmful or hateful content, but I'd be happy to help you with other ideas. Would you like me to suggest alternative prompts?" but lama 3 did, ah haha aha hahaa, did she already effect it or just them being their clown selves

[bitter hearth]

IDK why Glif would go into lora training

maybe Glif want to become more like Fal

I thought they were specialising in comfy workflows but perhaps not

[mortal mesa]

filling in holes in the market

[bitter hearth]

there's definitely room yeah

there's only a handful of lora services

civit, fal, replicate etc

[bitter hearth]

I really dislike services like that

cos they don't tell you when they use quants

[sage burrow]

A bit of everything it seems.
I'm going to be so broke when they switch to paid lol

[toxic bone]

and its hard to tell too. you could be getting less then best iamges from your lora and prompt, but without comparing it next to the same seed on the full version, you'd hardly tell

anyone notice how all of the flux.1 checkpoints on civitai are just 20gb files trained hard to one single concept, trained poorly, and destroys the rest of the latent space because they smashed the text encoder with their bad captioning?

i kinda noticed. They should all be loras. i've not seen a single checkpoint refine that shouldn't be a lora. the pornographic ones i think could even work as a lora tbh. They're just killing 12B parameters by training it as an entire checkpoint, as if they got the magic sauce to make flux better. The arrogance astounds me. okay maybe i've more than kinda noticed

[bitter hearth]

yeah it can be hard to tell if its a quant

[toxic bone]

now that civit has early access model sales, it's getting worse too. Giant essays on their checkpoints about how much better it is and can only be achieved by them with their exclusive code.

[bitter hearth]

I actually know for a fact fal is using FP8
because their flux dev speed is faster than H100 can do without quants

[toxic bone]

quants shouldn't speed it up though. they're compression right, so it would slow it down in theory

and in practice

[bitter hearth]

normally yes, but there is one exception

new GPUs have a speed boost for FP8 in particular

[toxic bone]

yeah i hae no idea about the datacenter game tbh.

hopper transformer engine yeah. the 40 cards have it too. it speeds up the decompressing, but it doesn't speed up the entire thing.

[bitter hearth]

you're right in general, quants do not speed up inference if you are not VRAM limited

[toxic bone]

though i will acknowledge that h100 is way out of my domain

https://tenor.com/view/waynes-world-mike-myers-it-will-be-mine-mine-gif-4571728

[bitter hearth]

Ada Lovelace as well

it does speed up inference

e4m3fn matmul

[craggy crest]

some people need captain obvious to explain things to them

[mortal mesa]

i blame the education system

[brittle nexus]

Thank you captain obvious

[sage burrow]

1.5 is so fast! 😄

[brittle nexus]

i miss it/s, the life is s/it now

[sage burrow]

I'm trying to create a melting clock lora, but I've having trouble creating melty clocks 😦 I can just grab a bunch of Dali paintings, but that would defeat the purpose of having to use my own aiart. Using Dali's paintings as reference images only gets me the basic look and colouring, but no melty clocks 😦
Suggestions anyone?

[bitter hearth]

does the lora have to be made from your own art?

you could train one lora on dali, then use that to generate images, then use those to generate a second lora

and then throw away the first lora

otherwise, strong control nets

[sage burrow]

It has to be of my own art (well ai art). I'm going to try out the Glif lora maker, so they have some rules.
I have been using a Dali lora that was trained by someone else, and I get 1 out of every 100.

Though I finally am now just resorting to an online image warper 😄 It is working for what I'm after 😄

[bitter hearth]

oh I see
yeah I saw that contest

if you use this node https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb
and set it to A1111 mode, you can use higher prompt weights, and put a really high weight on stuff like "melting" or "liquid"

[toxic bone]

google image search a bunch of melting clocks. tons of artists have done their own over the years. 3d artists, sculpters, all that

[bitter hearth]

its quite common yeah

someone made a reddit post showing the FP8 speedup BTW https://old.reddit.com/r/StableDiffusion/comments/1fedq8g/flux_fast_mode_testing_speedqualitylora_likeness/

its about 40% or so, works on a 4090 as well

would recommend for anyone who can

[sage burrow]

For this one I have to use my own images... and unfortuntaely reference images, even with img2img don't really work! 😦 Fortunately I found an online image warper, so I'm warping my fave clock images

[bitter hearth]

are you allowed textual inversion?

also you could try asking florence 2 to describe them

or clip interrogate

[toxic bone]

controlnet images you find on google and make em into your own melting clock source images

[bitter hearth]

wow flux is rly slow

wtf

until now I had only used it on datacenter GPUs
tried it on a 3090

[sage burrow]

I've had joycaption describe the original Dali, and the best that I've found around, still no such luck!!!!

[sage burrow]

Whatever, I run it on a 4060! lolololol 🛌

[bitter hearth]

I mean I do CPU sometimes

slow can be ok

but yeah its pretty rough

[sage burrow]

How would I use controlnet for that? I'm really clueless about controlnet unfortunately. I've only ever used it via Mage or similar and used it for poses

[bitter hearth]

pretty much you can just pick depth control net, put it to strength 1 and to last the whole time

which is the default in comfy I think

[sage burrow]

I use my computer sometimes, but usually use Glif, or Mage, or HF if I can, so much faster!! But for nsfw, it's kind of up to my computer only (it needs loras)

[bitter hearth]

yeah I used rundiffusion a bit

[sage burrow]

Fortunately I found an online image warper, and I've been warping some nice clock images I made...

[bitter hearth]

if you are doing 1024x1024 then small GPU is fine really

don't always need to upscale

[sage burrow]

my 8gb gpu and 10 mins wait times, are kinda torture lol

[bitter hearth]

oh 10 minutes is bad

didn't realise it was that long sorry

[sage burrow]

I upscale later with BigJPG OK perhaps not the best, but it works, and it does batches, and I paid for a year subscription lol

[bitter hearth]

LOL I love the name BigJPG

[sage burrow]

I love the $45 per year subscription fee! (or something low cost like that)

[bitter hearth]

yeah cloud upscale makes sense for most people

cos that's the slow part

[sage burrow]

Well I got 25 melty clock images, good enough for a lora test 😄 Someone recommended making a lora with what you have, then producing images to create an even better lora after that. That may be the plan if it doesn't end out awesome

[left parrot]

Is BigJPG ok with NSFW?

[bitter hearth]

someone should make a service that's like a cloud refiner pass

like cloud koyha high res fix

[sage burrow]

I have no idea? KIDDING. I haven't had any trouble with it.
Btw, it's only $22 per year apparently.

[sage burrow]

That would be awesome!

I'm patiently waiting for Mage to get Flux lora capeabilities. They allow some nsfw, and they are way faster than my computer!

[bitter hearth]

mage is a good one yeah

someone could probably do cloud facefix or handfix too

[sage burrow]

Though, I sort of like everything in one, even if it's long and takes forever, workflow. So much easier.

[bitter hearth]

yeah I kinda like one giant workflow

I actually can't rent low end GPUs
never manged to get them to work

that's the weird thing with cloud servers, the data center ones tend to be better value

cos the internet is decent and it has the libraries you need like CUDA and pytorch etc

[sage burrow]

Some clocks are starting to melt. SDXL. There WILL be melting clocks in Flux, soonith 😄

[sage burrow]

Which (free) bulk image tagger would you guys recommend? I don't need it to come up with the tags, just to add them

[narrow river]

Florence-2-large

[sage burrow]

Thank you 🙂

[toxic bone]

you can run an image through a canny preprocess, so that only the outlines are left in black and white. to a degree. then controlnet uses those black and white outlines to fill in the image and guide the generation

[craggy crest]

well yea, if you're on a 3090, it's slow. how much vram does that have?

[bitter hearth]

3090 is 24GB

I think I may have messed up and filled its VRAM to the brim cos it was way, way slower than an L40, which should not be the case

[bitter hearth]

I could get 40% speed boost if I learn how to install the latest CUDA
but I had too much trouble lol

speed boost is only in FP8 though

[sage burrow]

Pass lol

I created a flux dev lora, thanks to Glif 🙂 It does one thing, and one thing only, melting clocks 😄 https://huggingface.co/glif/loradex-flux-dev/blob/main/becky93/flux_dev_time_is_an_illusion/flux_dev_time_is_an_illusion.safetensors

[dusky thistle]

i was actually gonna suggest training one... nice work 🙂

[sage burrow]

Thank you 🙂 Melty clocks were even more difficult with Flux than nsfw with flux lol

[dusky thistle]

yeah, if it's not trained on it, it's not trained on it

no sense in flailing trying to prompt it into existence at that point, espec with how easy it's proven to be to train these things

[sage burrow]

With the melty clocks, the most difficult part was creating suitable images to start with in the first place!

[dusky thistle]

yep dataset + captions are always the hardest part

[sage burrow]

I only captioned 3 things lol. Now it makes nothing but surrealistic clocks 😄

I think I learned something going forward with my other loras. Don't branch out!

[dusky thistle]

the better your caption, the better the result

[bitter hearth]

wow this lora is incredible

[dusky thistle]

i did some tests... the idea that captionless or trigger only loras are superior for anything, including style, isn't true

[bitter hearth]

the one with the skull I really like

[dusky thistle]

whatever you get with very short or no captions, will be even better with very good very detailed captions

yeah the skull one is killer

[bitter hearth]

I both love and hate flux at the same time

the other day I was changing shift by 0.1
and it would alternate between images that looked good and images that looked blurry and with scanlines 🤔

but I was moving shift in the same direction

[dusky thistle]

you def want to use a lora

even just 1k steps is enough to fix most of those problems completely

[bitter hearth]

ah okay that's great

what I've been doing is keeping shift as low as possible

[dusky thistle]

https://huggingface.co/ClownsharkBatwing/CSBW_Style/blob/main/amateurphotos_1_amateurcellphonephoto_recapt_rerun.safetensors this is the one i'm using now

[sage burrow]

I did one with many images, and amazing long captions, but it didn't turn out very well. Though it was 100% things that were most definitely not trained in the original.

Fortunately I've also been having fun with merging loras with the original Flux as well 😄

[dusky thistle]

WF embedded

[bitter hearth]

we need a certain custom node for flux that we don't have yet

[dusky thistle]

what node

[bitter hearth]

this paper thinks that shift should be given a scheduler https://arxiv.org/abs/2301.11093

cos you only need shift early on to make the layout good

shift is bad for details so you want to get rid of the shift over time

[dusky thistle]

interesting, haven't tried

could do split sigmas and make the last few steps use it with a lower shift

but yeah idk

[sage burrow]

Oh so that's what you have been using/creating to make all those photos 😄

[bitter hearth]

what I've been doing is splitting sigmas yeah

I've actually been doing zero shift for later steps

[dusky thistle]

make sure you start the prompt with "An amateur cell phone photo of " for best results

yeah ultimately all it's doing is fuckin with the shape of the sigma schedule

what i don't like about it is it introduces even more variables where ultimately there hould only be one set

you have scheduler params, then shift params, which ultimately make one schedule

[bitter hearth]

I made a table the other day of my shift preferences, I always start with these values now
it shows comfyui default and then my preference

448x448 default 0.59, preference 0
512x512 default 0.63, preference 0.3
1024x1024 default 1.15, preference 0.5
1536x1536 default 2, preference 1.7
2048x2048 default 3.2, preference 2.3```

[sage burrow]

Does that one work with Flux or SD at all?

[dusky thistle]

it's flux dev

[bitter hearth]

I'm pretty sure the comfyui defaults are designed to be very safe

[dusky thistle]

this is what i'm using

[bitter hearth]

they're super high

hang on gonna get RX808's calculator lol

[bitter hearth]

ok your shift was 2.04, for a resolution equivalent to 1552x1552

that's rly close to my number too 🙂

[dusky thistle]

well that's good to hear lol

we arrived at similar numbers independently

[bitter hearth]

yeah that's always a good sign

[sage burrow]

Thank you very much 🙂

[bitter hearth]

here's RX808's calculator by the way https://www.desmos.com/calculator/c0jburw7z4I play with that and then enter the shift I want as both base shift and max shift, which just sets the shift to that value and ignores width and height

[bitter hearth]

the funny thing is the ModelSamplingFlux node used to take a shift value directly instead of the base_shift/max_shift thing
but then Comfy took the shift input away in this commit LOL
https://github.com/comfyanonymous/ComfyUI/commit/56f3c660bf79769bbfa003c0e4152dfb50feadc5

I like the old way better

[dusky thistle]

pretty easy to revert

just copy paste the code for the old node and comment out the current one

my copy of comfyu is pretty hacked up

[craggy crest]

time is no longer your friend

[noble coyote]

Flux Dev + Metallic and Vintage Illustration LoRAs

[noble coyote]

"Melting Clock" LoRA (not really melting, using Time is an Illusion LoRA)

I need a triggerword ... ?

... better

[muted dove]

Trying out the Brainflux trained model

[bitter hearth]

gonna start downloading llms onto comfy servers to write prompts for me lol

been using chatgpt but would be easier to have it done in a node

[muted dove]

Using Ollama is a good method

[bitter hearth]

thanks yeah will try that

[noble coyote]

QualityOfLife/ChatGPT prompting

Ollama prompting

Workflows in metadata

[bitter hearth]

thanks

[noble coyote]

Or try Jan.ai desktop prompt generator

[cinder junco]

Question for people with more experience: does using a LoRA usually double your memory use? When I use Flux dev bf16 on my Mac, it usually hovers around 37-40 GB used (I have 64 GB unified memory). But I just tried using the EldritchPhotography LoRA and Python started to try to use 65+ GB (probably would have been about 75 GB if I had more memory, considering the amount of swap it took up). This doesn't seem normal to me, since I thought LoRAs were supposed to be pretty light weight.

Not a huge loss as the LoRA didn't seem to work well for me -- significantly changed the image composition, reduced the detail as shown by a 2D FFT, and actually seemed to push it further away from photographic style.

[little ledge]

1

[bitter hearth]

2

[bitter hearth]

Ya gotta pay the cheese tax

[signal shuttle]

You know what? PENGUIN TANK

[signal shuttle]

Made with Lora

[cursive frigate]

This is a few days in the making. what do y'all think?

[cursive frigate]

Does anyone use LLM Prompts. If so does anyone know how to force the LLM to keep Trigger words for LoRAs without altering or changing the keywords in the prompt?

Also does anyone know if there is something that can be introduced to the workflow for tags?

I would like to start posting some of my work on CivitAI and was curious if there was a way to automate Tagging images.

[sacred jewel]

Yes,

Use a Text Concatenate node to merge the LLM response with the Trigger words or whatever other text you like.

[cursive frigate]

lol... thank you. I was just going to ask for an example of that what would look like.

[sacred jewel]

Flip the text inpts to place the word(s) in front or at the end.

[cursive frigate]

Not sure if this exists but are there any LLM models that you are aware of that will allow explicit, violent, gory, and/or adult content style descriptions? If not that is fine but it would be nice to not be restricted to PG content.

[sacred jewel]

Absolutely... You just have to search 😉

https://ollama.com/search?q=abliterated

[bitter hearth]

There is an uncensored one you can use locally that is very polished but I cannot recall the name...Free something I think

Maybe this one: https://www.freedomgpt.com/

They will not report you!

That's a company pledge you can live on

[cursive frigate]

Where can I get the full fp16 clip weights for Flux dev. I was only able to find fp8. and does that change the model as well or is it the same model?

[bitter hearth]

civit or hugging

its the same model

[noble coyote]

Using Blue Future, TimeIsAnIllusion and YFG-SpyW-Replicate LoRAs in Flux

[noble coyote]

... A humorous meme-style image featuring a game controller made entirely out of cheesy snack foods, like cheese puffs and cheese curls, heavily coated in bright orange cheese dust. The controller should have recognizable elements like buttons and joysticks formed from the snack foods, with the surface almost dripping in cheesy dust. The caption 'Pass the controller, bro.' should be displayed at the top or bottom in a playful font, adding a lighthearted tone to the image.

... Hyper-realistic Pixar-style 3D rendering of a warm and joyful virtual hug. A friendly, soft-featured 3D character stands with open arms, offering a comforting embrace. The character has a kind, cheerful expression with big, welcoming eyes and a gentle smile, dressed in cozy, colorful clothes. The background is a soft, pastel-colored space with gentle lighting, creating a sense of warmth and positivity. The character’s arms are slightly extended toward the viewer, as if inviting them into a hug. The rendering employs ray tracing for realistic shadows and lighting, and subtle lens flare adds a glow to the scene, enhancing the feeling of warmth and joy.

... Hyper-realistic Pixar-style 3D rendering of mental health, focusing on self-care, mindfulness, and nurturing the mind. The scene portrays a cozy, serene room filled with soft natural light streaming through large windows. A plush armchair sits by the window, with a person wrapped in a soft blanket holding a cup of tea, gazing out calmly. Books, plants, and calming objects like candles and crystals are arranged thoughtfully around the room, creating a peaceful atmosphere. The walls display subtle artwork promoting positive mental health themes. The room's ambiance is enhanced by ray tracing, capturing realistic reflections and soft shadows, while lens flare adds warmth to the incoming light. A slight motion blur is used to suggest the gentle swaying of the plants and curtains, further emphasizing a sense of tranquility and mindfulness.

... A minimalist and abstract design inspired by Zen ink painting (Sumi-e) that captures the feeling of 'the self dissolving into space.' The human silhouette is rendered softly and subtly, with the boundaries blurred, as if fading into the surrounding space. The ink is used in varying shades, with light and dark gradients, to show the flow of energy and the sense of unity with the space around. The composition is simple, with ample white space to represent emptiness and the dissolution of self. The overall feeling should be calm, meditative, and abstract, emphasizing the Zen philosophy of oneness with the universe and the dissolving of the ego.

DallE Theme of the Day

[bitter hearth]

That's disturbing Becky

That image offended my bandwidth

[craggy crest]

https://x.com/StabilityAI/status/1834311526494470517

Stability AI (@StabilityAI) on X

Interested in fine-tuning your own image models with Stable Diffusion 3 Medium?

We've created a tutorial to guide you through fine-tuning Stable Diffusion 3 Medium for high-quality, customized images. If you're familiar with SD1.5 or SDXL, this guide highlights key differences

Twitter · Sep 12, 2024, 7:22 PM

[bitter hearth]

surreal

🧐

[craggy crest]

time has a bone to pick with you

[bitter hearth]

👀

[toxic bone]

you did it!

[sage burrow]

I managed to create 43 melty clocks for my dataset! Though for 13 of then I cheated and used an online image warper.

With SDXL mostly.

https://huggingface.co/glif/loradex-flux-dev/tree/main/becky93/flux_dev_time_is_an_illusion

For those from brazil, or who just hate X, who didn't click on crystalwizard's post https://stabilityai.notion.site/Stable-Diffusion-3-Medium-Fine-tuning-Tutorial-17f90df74bce4c62a295849f0dc8fb7e

[toxic bone]

slick moves!

[noble coyote]

What's the best triggerword?

[toxic bone]

i appreciate their effort, but i had trained some lora's into SD3 medium, and while it learned the person, people are still broken on sd3. Many things about sd3 are just broken and poorly engineered. There's no reason for people to not use flux with the optimizations it's received over the last couple weeks

[sage burrow]

There's 3, "melty clock", "spiral clock" and "time is an illusion"

Or just use them on Glif and get annoyed it mostly ignored your prompts 😉

[toxic bone]

if they don't fix the pretraining in sd3.1 , it'll be another garbage model that people ignore

[craggy crest]

that's because it's sd3 2b medium - so use a larger data set for your lora

[sage burrow]

Apparently it's upcoming release was announced in the San Fran conference recently. I can't imagine them not fixing the human anatomy stuff

[toxic bone]

i don't think the parameter count is the problem. SDXL is superior to SD3 outside of the VAE

[craggy crest]

i think you'll have better success with a larger data set - both images and labels.

[toxic bone]

i don't think there's any amount of lora training that could fix the pretraining issues.

that's just innefficient. SDXL and flux on both sides of SD3 dont need large datasets.

[sage burrow]

I love the artistic styles of sd3. I have all my flux training datasets still, all ready tp try out with sd3.

[craggy crest]

no, i agree with that. you're not going to fix a basic core issue with a lora, but if you're doing something like Becky is, it's fine

[toxic bone]

i have doubts that the melting clocks will train onto sd3 but i'll be watching Becky's work to see

[craggy crest]

you don't know just how persistant she can be

[toxic bone]

i dont think her persistence is the issue. But she is the canary. If she can't get er done, it's not gettin done

[craggy crest]

you think pit bulls don't let go, you aint seen nothing yet

[toxic bone]

don't mean to speak of you in the third person when you're right here @sage burrow haha

lol inb4 "lockjaw is a pitbull myth by specie haters" ||its not||

[sage burrow]

Hmmm, TensorArt has SD3 training setup (easy, auto, $2 lol). I've been tempted to try it.

[craggy crest]

it's not. i had to pry one off another dog some years back. that was loads of fun

[toxic bone]

i've seen an owner taking a break stick to his dog's jaw to get him off a child's leg, and acting totaly casual abut it like it was no big deal and people were over reacting

[craggy crest]

it took 6 people and we almost needed to get the cops with the jaw of life out there to make it let go

[toxic bone]

you dn't jaws of life a pitbull. you put it down

[sage burrow]

It's all good, you were just speaking of training methods, using a very narrow focus, and no humans 😄

[craggy crest]

we got it off. it just really hated the other dog. it was a good dog except for that

to be fair, the other dog was annoying as ....

[toxic bone]

i'd be more than happy with murdering someoen's pet if it was harming someone. especially a breed made for savagely murdering people and sport fighting

[craggy crest]

and all it had hold of was the loose skin on the back of the other dog's neck, so no damage done really

[craggy crest]

yeah, the genetics could use some serious dilution

[toxic bone]

just go full nugenics on the breed. Make it illegal to own and breed with others. I don't understand why we allow dogs that are selectively bred to be violent killing machines with a 3 year old mentality, to live in society.

we should form kill squads to go around and take care of all pitbulls. simple.

[craggy crest]

same reason we allow anything.

[toxic bone]

equip city workers with a cattle spike. that nail that slams into the back of a brain stem. go door to door.

[craggy crest]

wrong society for that sort of action - that takes a dictatorship

and then you'll have people going 'well i hate this X' or 'i hate that Y' - "kill it too'

[craggy crest]

read DMs

[sacred jewel]

Is anyone generating consistently and to your liking with Flux.1 Dev at high resolution directly?

Recent testing shows some prompts/seed are fantastic while ithers fail miserably...

All these are directly generated from a 2016x1152 latent. Euler/Beta/35 steps/3.5 Guidance

[craggy crest]

taht's not the AR that flux actually prefers, so that's nto real surprising

[sacred jewel]

What dimensions do you suggest I use? Or A:R? I thought as long as the MP count stays within 1-2MP any A:R was reasonably doable. Indeed, I generate thousands of images at 16:9 1344x768 and 99.99% are perfect.

[craggy crest]

sometime back, when flux first came out, a lot of use tested it, and posted results. i believe it was @dusky thistle who figured out the AR it likes best, though it might have been one of the others. the sweet spot for flux is 672x1024 or 1024x672

[sacred jewel]

Thanks. Yeah, I know about sweet spots. I was referring specifically to larger resolutions directly without resorting to upscaling.

2016x1152 directly...

[craggy crest]

it's not going to be happy with that. i think if you really want a higher res image, upscale is going to be your option

[sacred jewel]

Upscaling with Flux is giving me angina. I can't get a good result while keeping original composition intact or close to intact. So researchign my sweet spot options 😉

[craggy crest]

so don't use it. bad idea to upscale with the image generator you create with anyway. just use something like topaz gigapixel, or even the free image upscaler on the capcut magic tools page

and only upscale the ones that you actually need that large. no reason to use up disk space you don't need to

[brittle nexus]

Nvidia Studio Drivers or Game drives makes any difference?

[toxic bone]

meagina? i hardly know'a

flux doesn't do straight upscales well. you need a tiled sampler or a something like ultimate upscaler

i find ultimate works better.

topaz is nearly a decade old at this point. it's a gan. it looks like a gan at high zoom

[craggy crest]

topaz just released a new version with face restore in it