🆕｜sd3

[bitter hearth]

but similar idea

hmmmmmmmmmm

thats dpmpp 2m on schnell

fairly sure dpmpp won that test

sgm uniform scheduler

yeah thats why i went hmmm

[torn wharf]

this is often the case too. thats how i started using dadaptation myself. reddit advice was mostly just people copy pasting other configgs, and then they'd use the art of bs to explain each setting in a way that made them exude expertise. i went and "read" through the dadapt github and information from the experts who actually made it and found a lot of information that conflicted with the youtuber explanations of things.

[bitter hearth]

yeah these premade scripts that go viral worry me a bit

[torn wharf]

"read" is more like, frantically searching for anything that seems like an infromation nugget, then trying to polish it with spit on my jeans to see if it was good

[bitter hearth]

damn, dpmpp 2m makes a big difference on schnell

I think I will slowly learn how to train checkpoint or lora from either diffusers or raw pytorch/JAX workflow

why are people pushing for euler with schnell

people always push euler lol

but why when dpmpp 2m is much better?

dpmpp 2m has a bunch of improvements
it is second order not first order
it solves the linear part analytically, in order to make the equation less stiff
and it has multistep

could there be some technical reasons that go beyond just the visual?

ok

yeah the DPM papers explain it

well you just laid out technical explanation

this is schnell nf4 v2 btw, could have mistaken for dev maybe, using dpmpp 2m

the order of a solver is a really important thing
so euler is first order
dpmpp 2m is second order
dpmpp 3m SDE is third order etc

dpmpp 3m SDE wouldn't be ODE tho?

yeah that's right

[torn wharf]

i sorta see network dropout like .. okay like water erosion. water goes to the path of least resistance and starts carving out valleys and trenches and thats where you see that artifacts that are obviously water erosion. sorta like training a lora or any network, it'll carve out pathways that work and then reinforce those pathways heavily. that's over fitting if it's done too much. neural pathways that bypass all other pathways.

So dropout, it changes the landscape of the netwokr here and there. forcing it to take other paths for those steps. it forces the process to try different ways now and then and it can lead to a more generalized understanding. I wouldn't put it so high though. A little goes a LONG way. just a dab will do yeh.

When i break down the paper and understand it on a practical level, not just the math it's doing on the network, but what the goal and insights of it are, i don't know how it still doesn't apply to lora training on modern models. It makes a lot of intuitive sense to me. I can think of tons of training analogs when humans learn to do something. a wobble board for balance training for example. they'll wobble on their own so your brain has to learn to compensate and then that training will transfer to your entire concept of balance.

[bitter hearth]

an example for 3rd order ODE solver is Bosh 3

ok so you mentioned ealier we dont need SDE anymore

yeah and SDE will break it also

ahh ok

if you want to go above order 3 it gets more tricky
sometimes 4th order explicit-runge kutta can do okay
but there can be stability issues and sometimes very long convergence times

you can avoid the stability issues using Gauss–Legendre method, but this is very slow

well with what i have now between euler and dpmpp 2m, the latter just improved quality and timing

[torn wharf]

kinda want UI's to just hide all the options that objectively don't work with a given model. sorta why i call .safetensor formats the laser disc format of AI. it's great for now and it brought us a specific tensor file for home use, like laserdiscs was early success for home digital movies. But safetensor don't got no standardization for its metadata like, this model only works with ODE solvers or whatever other technical problems might come up. like why are SDE solvers even showing up when none them will ever work?

[bitter hearth]

i was using dpmpp 2m with flux dev but i didnt realize it would also work with schnell and it does so much better than euler

[torn wharf]

growing pains . always standardization issues. What we need is a new standard!! ||relevant xkcd||

[bitter hearth]

I've seen people make this complaint about UIs before yeah

in some ways its good to let people experiment though

[torn wharf]

Forge made strides with the model preset options at the top. There's opportunity there

it's a rough update, but he did a lot of groundbreaking work on it recently

[bitter hearth]

I kinda feel like stuff like this should be left out of comfy though

[torn wharf]

maybe a toggle on the ksampler - show preffered solvers or something

[bitter hearth]

yeah actually that would be fine

[torn wharf]

i'm also a big believer in keeping ALL functionality open to the user

especially with something like a node graph

[bitter hearth]

I'm switching to just python now so I guess I don't rly mind which direction comfy goes in
they could either keep it experimental or "mainstream" it a bit with guard rails

comfy in its current state has been great for learning though

[mortal mesa]

less than a year old

[torn wharf]

comfy is a great system. swarmui is the perfect front end fusion for it too

[dry wave]

I switched from pure python to comfy cause I was so annoyed by memory leaks and stuff

[torn wharf]

some of the earlier comfyui front end projects were a total bust

[dry wave]

things might got better, but if you work in a Jupyter notebook you experience weird problems from time to time

[torn wharf]

i bought shapez 2 yesterday so i may not actually have much time to train loras . i was about to start dataset building and my brain was like "BRO WAIT. THE GAME!"

Hmm. struggles. i'm going to go touch some grass and ponder this

[torn wharf]

sick album cover

[glossy pilot]

Hey guys I have a dumb question!

Has anyone already tried to create sd 3 LoRA?

[bitter hearth]

check civitai

[glossy pilot]

I would just like to know how they make the Dataset

[bitter hearth]

https://huggingface.co/stabilityai/stable-diffusion-3-medium/discussions/98

people expect sd3 8b model btw, 2b model has it's short comings

[glossy pilot]

otay but what should the data set look like? like what .txt should I write?

https://tenor.com/bTH4D.gif

[cinder junco]

@bitter hearth You seem to know a lot about ODEs. I’ve been using bosh3 by default since the ODE samplers node came out, based mostly on drhead’s declaration that these solvers should be more appropriate than any of the default samplers. However, he never really provided evidence, nor any info for how he identified good tolerance values for each or established the quality ratings he gives on the GitHub page. I did some experimenting over the last two days and noticed that a lot of the custom solvers seem to converge to a very similar result (given appropriate tolerance). The main consideration between those would then be which solver can converge with the least computation (number of steps * computation time per step), though it is also true that convergence and image quality are not really equivalent concepts. Convergence just provides some level of psychological reassurance, given that things won’t continue to change much once you’ve reached it. Anyway, do you have any guidance on whether the custom solvers are really better than the default samplers and which, if any, you would consider to be best for quality and efficiency of generation, particularly for Flux?

Dopri5 seems to converge slower than others, requiring tighter tolerances than the default -2.5/-3.5 for bosh3 and taking longer. I haven’t experimented much with dopri8, because at -2.5/-3.5 it took about an hour and a half to produce a 1MP image, though the result seemed to have converged.

Fehlberg2 produced an output in about 1.5 minutes, much quicker than bosh3, but it was completely borked and obviously needed tighter tolerances. After adjustment, the output was much better but the speed advantage disappeared.

[bitter hearth]

long story short the ODEs for diffusion or rectified flow are too stiff for a lot of solvers to do well

very high order solvers can get stability issues

for the most part one way of dealing with this is to stick to solvers that are not going to be quite as good, e.g. 2-3 orders or below mostly

another way would be to do what DPM does which is to solve the linear part analytically

this makes the remaining ODE less stiff

a third option is to use a more stable solver such as a Gauss–Legendre method

but these do get very slow

also -2.5 tolerance is not tight enough for this sort of thing

its missing the point

the idea of this sort of solver is to get a much more accurate result, not -2.5

[cinder junco]

How do you explain the fact that bosh3 at -2.5/-3.5 produces almost identical results to -5/-5?

[bitter hearth]

it didn't in my tests
you have to bear in mind at this level of local discretisation error (i.e. a much lower level than normal) the benefits to image quality are going to vary a lot between images and depend a lot on the workflow

[cinder junco]

I can’t say I’ve done extensive testing, but this has been consistent across two much different prompts, and between Flux and SD3.

It also seems clear that the different solvers require different tolerances to get similar results.

[bitter hearth]

its not just about prompts its about the whole workflow
a higher tolerance level will be of greater benefit when the trajectories are less straight, so workflows that result in less straight trajectories will see more benefit

[cinder junco]

I’m not sure how anyone using comfy would be able to identify a trajectory as straight or otherwise.

[bitter hearth]

need to be clear that you can't just compare tolerance levels across ODE solvers exactly
because they don't estimate the local discretisation error in the same way

oh you can test for stiffness

there are ways to do that

[cinder junco]

Such as?

And how does stiffness (which I don’t have a deep understanding of) relate to trajectory path?

[bitter hearth]

what you can do is make inferences from the way the ODE solution behaves with different explicit step methods
the main thing that shows stiffness is that it would become unstable, or that it would have areas that require the adaptive step size to get very small
what you can also do is compare the step sizes taken by an explicit method with the step sizes taken by an implicit method
another way is to use a solver that has stiffness detection built in
you can also take the jacobian although that takes a lot of compute
regarding stiffness and trajectory path, the idea is that less straight trajectories could make the ODE more stiff

[bitter hearth]

YEAAAAAAAAAAAAAAAAAAAAAA

dreambooth for flux lets fucking gooooooooooooooooo

[glossy pilot]

mhm mhm now tell me how to use it

[bitter hearth]

aaaaand you need 40gb VRAM -_-

we will have to wait for optimizations to come around my friend. patience is the name of the game

[cinder junco]

I don’t know of any tools for things like this in comfy. The ODE solvers node doesn’t even output the number of steps used by the adaptive solvers.

[bitter hearth]

talkin bout dat dere fp8, dat dere 8-bit-adam

there's not really comprehensive repository of comfy nodes so its hard to know what is out there

[cinder junco]

Well, you made it sound like this is easy to do and something you have done.

[bitter hearth]

5090 will have 28gb VRAM, so I highly suggest start saving up for it now cuz if there are optimizations done for this, it will likely be for 24GB+ users only

this is pretty far from being easy

[cinder junco]

@bitter hearth So, other than avoiding high order methods, do you have any other specific recommendations? Do you think bosh3 is better than dpmpp2m in some way or vice versa? Some other particular solver or sampler? Particular tolerances? Or is the answer just, like too many AI-related things, “there is no answer, just play around and do what works for you”?

I have liked the bosh3 output, but I’d appreciate it to be faster. It also doesn’t adjust its generation time as much as I expected when doing a low denoise (0.2) for img2img.

[bitter hearth]

if you are looking for faster than bosh3 then that's very different from what I would answer if you were just looking for the best image quality

DPM++ 2M or UniPC are fine for just making images fast

[cinder junco]

Well, the point is to reach a certain level of quality with the least computation. You can throw -10/-10 tolerances at anything for “best quality” and it’ll take forever. That’s not practical.

There are diminishing returns, so it is important to identify the location of the dogleg.

[bitter hearth]

practical is a relative term
I've been renting data center GPUs and running slower solvers on them

for your situation I would stick with DPM++ 2M or UniPC
if you want a recommendation

[sterile pendant]

Which is why Euler and dpm++ 2m are the almost unanimously chosen default samplers for most people and even researchers

[sterile pendant]

The potential diminishing returns gains of slower samplers like bosh just don't justify the extra calculation times

[bitter hearth]

yeah one of the DPM papers actually says this
he compares RK4 (which is better than Bosh) and says that the extra time was not worth it

[sterile pendant]

You have to have the precision really high to make them worth it

[cinder junco]

I developed a dislike for UniPC, as its output was decidedly stippled with SDXL. I haven’t tried it with Flux.

[sterile pendant]

So if dpm++ or euler will produce the same image at 99.5% the same quality, in half the time, that's the winner

And why they are the default go to

[bitter hearth]

the DEIS paper independently came to the same conclusion as DPM, that one approach is to solve part of it analytically to get a less stiff resulting ODE to solve:

and numerical simulation (Hochbruck & Ostermann, 2010; Whalen et al., 2015), from which we get
inspirations. The stiff property of the ODEs requires more efficient ODE solvers instead of blackbox solvers that are designed for general ODE problems. In this work, we investigate sovlers for
differential equations in diffusion model and take advantage of the semilinear structure.

[cinder junco]

RK4 is better in what way? Any source for comparison testing? I’ve done a lot of searching, but it is hard to get sources testing the actual methods available for comfy. E.g. most sites I’ve seen talked about bosh5 rather than bosh3.

[bitter hearth]

the thing we are talking about is local discretisation error
since RK4 is fourth order, that can help it to get lower local discretisation error, provided it doesn't stumble on two of the major problems with stiff ODEs- the adaptive solver step size going too small, or instability

also I wanted to explain something from above

the reason why I never bothered to run stiffness testing on these ODEs is that there's no point- because the answer is that they are very stiff

so if you want higher image quality then you need a solver that is robust to stiffness

which is why my actual recommendation for image quality is something like Gauss–Legendre

there's not much point in running stiffness tests because you will just end up at that same conclusion anyway

[cinder junco]

But no solver currently available for AIs uses this method, correct?

[bitter hearth]

well that's what I was saying, there is not central repository of all comfy nodes so it is hard to know everything that is out there

for example you mentioned Dr Head's Comfy ODE node
but there is a significantly more featureful one out there:
https://github.com/wootwootwootwoot/ComfyUI-RK-Sampler

I find all sorts of weird comfy nodes across the internet

[cinder junco]

Hmm, thanks. I’ll check it out.

Explicit methods are not adaptive, correct?

[bitter hearth]

explicit methods can be adaptive

explicit and implicit is more to do with stability and stiffness tolerance

but there is more to it than just those two categories because, for example, gauss legendre is more stable in some ways than implicit adams, even though both are implicit

[cinder junco]

Okay. In a lot of my searches, I was getting info relating more to the actual ODE system. But I don’t know for example, what order ODEs Flux/SD3 use.

[bitter hearth]

well the model and the solver are seperate

[cinder junco]

It seemed like the choice of solver related to what system you’re working on.

In particular the order.

[bitter hearth]

yes but the models themselves don't have a discrete order because they are continuous
its kinda hard to explain

[cinder junco]

Hehe.

[bitter hearth]

this is what discretisation error is referring to
we are trying to make it discrete

but we can't do it perfectly so we get errors

[cinder junco]

I did a double major in physics and math, but never did much in-depth study of differential equations.

[bitter hearth]

oh yeah I didn't study these types of equation much either

[cinder junco]

May I ask how you got your knowledge?

[bitter hearth]

over this summer I just every paper I could find that was about solvers
and then some textbook chapters about ODEs
and then the docs of ODE solvers like in Scipy or some of the old fortran ones

[cinder junco]

So, for you yourself, what do you prefer to use when generating?

[bitter hearth]

I like to use SDXL or SD 1.5 at the moment and I use a sampler called Clownsampler
by someone in this discord (clownshark)

[cinder junco]

Probably an SDE sampler, right?

[bitter hearth]

yeah that's right

it adds a ton more noise than most of the common ones though

for fast sampling with an ODE method I found 15 steps of UniPC with the beta scheduler can be good

and sometimes the beta scheduler with the sigmas multiplied by 0.8

[cinder junco]

What is this beta scheduler?

Haven’t heard of it before.

[bitter hearth]

I think it might be a default in comfy now

it came from a paper where they did heatmaps of when high and low frequency details are produced by the model
and they made the beta scheduler with that analysis in mind

but I tend to like to take their numbers and multiply them by 0.8

because that gets a trade-off that is more in favour of high frequency detail

if you have used Karras, its a relatively similar shape

the actual sampler from the paper the Karras scheduler came from is called "Heunpp" in comfy

for some reason

[cinder junco]

The sigmas basically just determine how much noise the sampler tries to remove at which proportion of the total steps, correct?

[bitter hearth]

yeah that's right

[cinder junco]

And do you think it is true that this schedule does not affect adaptive samplers, since they determine their own step size?

[bitter hearth]

yeah that's right

they choose their own step size

one thing to understand though is that they can also error in their method of choosing the step size

this important if you are using the adaptive step size choices for something such as trying to infer stiffness

just because it chose a certain step size for a certain step, to keep within tolerance, doesn't mean it was 100% correct in that choice

[cinder junco]

Right. And the tolerances operate on an estimate of the solver error, so there is potential inaccuracy there, too.

[bitter hearth]

yeah
they mostly use 2 methods
either they take 2 small steps and compare to 1 step
or they solve it once, and then a second time with a solver with 1 order lower

not exactly the most robust method

if its super unstable due to stiffness then it just goes off the rails both times

[torn wharf]

thanks i din't know about this one

[bitter hearth]

no problem

[torn wharf]

also thanks for the explanation about the mathematical accuracies

[bitter hearth]

the funny thing is
I don't think the test I have been talking about has actually been done
of someone running a solver that is

1. adaptive step
2. pretty high order, at least 4-6
3. very stable against stiffness
4. tolerance of -4 or better

on these diffusion or ret flow models

I can't find any reference of anyone actually trying this

[torn wharf]

https://tenor.com/view/adventure-time-mathematical-adventure-time-finn-gif-8102486

[cinder junco]

Thanks for your time. I’ll try out some more methods in that other node and hopefully it will have some more guidance on good choices.

[bitter hearth]

I was considering renting H100 for a few days and running through the stable ones like higher order gauss legendre

okay sure

the reason I kinda stopped actually running the experiments
is after reading enough papers on diffusion I realised that they are very much pushing in the other direction

their ultimate goal is a model that can be solved with 1 step of euler

this year there were loads of distillation papers I think more distilled models are coming

[sterile pendant]

When in doubt about sigma curves, use the sigma graph node from that one node pack, can't remember which, but you'll see me make examples with it on here afrom time to time

And about beta sigmas: they add a little more importance to the start and end of the schedule

Early=primary forms and overall scene comp, later=small and fine details

[cinder junco]

Not sure about what you are referring to. But I thought Flux (and SD3) were not very tolerant of non-standard schedules.

[sterile pendant]

Beta works with them

It modulates their downhill rounded slope sigma graph

Jesus, I meant convex lol

Va sdxl having a concave curve

But the beta schedule works with the sd3/flux schedule s

[cinder junco]

Is the “simple” schedule just linear? I’ve only used that for Flux.

[sterile pendant]

beta vs simple

comfy automatically picks that curve even if you do simple, its the same as sgm_uniform iirc

normal simple is a flat diagonal line

[bitter hearth]

the beta paper was really funny because they found that almost nothing happens in the middle sigmas
the model just chills in the middle

[sterile pendant]

well primary forms and fine details are really what mostly matters to our perception

[bitter hearth]

yeah

this was the node pack https://github.com/Extraltodeus/sigmas_tools_and_the_golden_scheduler

the actual golden scheduler from the title is not that useful

but I love the graph

[sterile pendant]

and there are other goodies in there as well

[bitter hearth]

ye I always have it installed

it does depend on if i get cloud server with good internet
but if I do I install like 50 custom node packs just in case

also that's why I like unipc, its nice to use 15 steps only on some servers like the 8-16gb vram ones

the fancy ODE samplers I only rly run on the 40GB+ servers I think its not worth it otherwise

even then its only been "worth it" a few times

when there was something confusing in the image

[cinder junco]

Sgm_uniform and simple are the same??? Why would the same thing be listed twice?

[uncut river]

war tank below a waterfall, serene evil aggression, perfect gorgeous chaos, hard contrast, weird

flux

[torn wharf]

is that 4th order ODE solver with -4 tolerances? !? nobodies even trying !

[bitter hearth]

I was wondering that too, turns out comfy explained it here https://github.com/comfyanonymous/ComfyUI/discussions/227

Simple is just one I added because I tried writing the simplest scheduler I could for fun and it actually worked well in some scenarios like second pass of hiresfix so I included it and named it simple.

ddim_uniform is the one that's supposed to be used with ddim if you want it to behave exactly like the ddim sampler in the reference Stable Diffusion implementation.```

[cinder junco]

Not seeing the comparison between sgm_uniform and simple, I’m afraid.

[torn wharf]

but i have no idea how any of this works.

[hallow lion]

If you want it very short and sweet: Karras and Uni PC are film like, Especially Karras. Euler and the rest are like VHS.

[sterile pendant]

for flux they are and i'd assume for sd3 as well, but that has to do with how comfy handles the models under the hood

correction, flux dev

[bitter hearth]

you wanna give it more than 35 max steps

[sterile pendant]

for flux schnell, simple will just be a diagonal line

but if you hook up a model sampling flux node between it, it will give it the dev curve and behaviours

[torn wharf]

it does look nice at 35. i'll allow it 60 max. see what happens. took 123 seconds before and finished at 23 steps

[sterile pendant]

(it's technically a bug that i should make an issue for on comfyui)

[bitter hearth]

its ok if you want to do that but bare in mind
it might basically be worse than a normal sampler
if you give a big sampler less steps than it needs to converge

[torn wharf]

max set to 60 steps. took a little bit longer but only used 22 steps

i like the look of images from the ode nodes dr head gave us. i don't know enough to say why i find them more pleasing though.

[bitter hearth]

bit confused as to why it was so fast
but it is the case that not every generation has the same shape

[torn wharf]

using dev nf4. dopri5 is fifth order if i understand it right

[bitter hearth]

oh this is flux

[torn wharf]

oh yeh true maybe i should try on sd3

[bitter hearth]

maybe its super straight trajectories and is less stiff

[torn wharf]

tahts what rectified flow wants right? straight paths?

[bitter hearth]

hopefully yes

sometimes the papers talk about single step of euler although we didn't get that far

but they should be straighter than diffusion

[sterile pendant]

I miss being able to use SDE samplers with these flow models, it was my favorite sdxl sampler

[bitter hearth]

yeah its a shame

[sterile pendant]

Like dpm++ 2m sde was my go to

[bitter hearth]

I liked clown and shark

yeah dpm++ 2m/3m were nice

[sterile pendant]

Sometimes I'd use 3m adaptive

From that sampler pack

[bitter hearth]

I actually started liking euler a too

for super soft look

but only the cfg++ version

an unintentional benefit of cfg++ is that it looks nicer if it fails to converge

its not really designed for unconverged generation but it helps that

[sterile pendant]

some people also do the trick of setting denoise to like 95% instead of 100, to essentially skip the final step and leave things a little softer

well not quite like that with the regular ksampler node since the % kind of just scales the sigma graph and you still do all N steps, but with advanced ksampler, you can set it to like 30 steps, then end it at 28 or 29

[torn wharf]

sd3 with same dopri5 at 60 max steps -4 and -4. did 31 of the steps and took 60 seconds. after i warmed it up with one that took longer. same prompt.

this one should ahve meta data. not sure about last one

i'm trying dopri8 now and it reached 60 steps and is going still

counting new steps

it's just like "fk your max steps i got my own" its up to 85 now

[hallow lion]

Isn't anything above 30-ish redundant...

[torn wharf]

dopri8 says no

[cinder junco]

Max steps is really just a max. It should by design crash out if the sampler exceeds that value.

[cinder junco]

This must be a bug!

[torn wharf]

it's at 305 steps now and the image is still coming in as it ticks away

[sage burrow]

My review of Civitai's Flux lora making.....
1 out of 10

Civitai's SDXL lora making 7 out of 10

Maybe it's good for SFW, but that's what Dalle is for 😄

[bitter hearth]

@sage burrow

[sage burrow]

Perhaps I should try with a few hundred images instead of only 65

[hallow lion]

Or at least 69.

[torn wharf]

it finally finished and its pretty good detail

[hallow lion]

How many steps? XD

[torn wharf]

made it all the way to 529 steps in 1945 seconds

[hallow lion]

omg

[sage burrow]

I didn't even think of... oh nm lol

next lora...

[sterile pendant]

Lol...

No matter how anal the sampler gets in latent space, it's still only going to decode into rgb8 per pixel

But damn you let that run for a minute

God it would actually still be bottlenecked by the vae at that rate. The quality of the decode can only ever be as good as the vae

[sacred jewel]

What's the prompt and seed?

[torn wharf]

there should be metadata. it's a dopri8 solve so i dont think seed would matter for other samplers. prompt is llm copy paste. In the depths of the Black Forest, Germany, a haunting landscape unfolds. A thick, emerald-green moss bed blankets the forest floor, softening the terrain. Hundreds of fir trees, once majestic, now stand as skeletal sentinels, their branches etched against the sky like bony fingers. Broken twig branches litter the ground, a testament to the forest's decline. The air is heavy with the scent of damp earth and decay, and the misty lighting casts an eerie gloom over the scene. The trees' trunks, once sturdy, now bear the scars of time and neglect, their bark rough and weathered. The forest floor, once teeming with life, now lies silent and still, a haunting reminder of nature's fragility.

not sure what seed is i dont have it open atm

[young blade]

can't wait for flux to have a negate and a solid depth control net

[sterile pendant]

They'll have to put out another model that isn't a distilled model that actually uses cfg. For now, any of the faux cfg posts and workflows you see are a hack and result in lower quality generations, unless you're doing a multistage resampling workflow to clean it up

[sacred jewel]

I find Flux good enough not to need negative prompting. 🤷‍♂️

[sacred jewel]

Thanks bro... no worries. Just wanted to play around and compare... 👊

I mean, this is inasanity. I literally prompted EXACTLY this... this thing just keeps delivering!

[young blade]

these are dope

[sterile pendant]

da fuq am i gunna do with all this overgrown accidental nirnroot

[zealous ore]

A serene and futuristic illustration of a person sitting in front of an old, yet perfectly maintained laptop, with a warm desk lamp shining down. The laptop is covered in stickers and has a few nostalgic gadgets plugged into it, such as a vintage keyboard and a floppy disk drive. In the background, there's a subtle glimpse of a home office setup, with books and plants surrounding the space.

The person in the foreground is engrossed in their work, typing away on the laptop with a look of deep satisfaction on their face. The AI they're working with is depicted as a gentle, pulsing glow emanating from the screen, representing the seamless integration of technology and human creativity.

In the corner of the illustration, there's a faint image of a person deleting a cookie notification or a tracking pixel from their browser history, symbolizing the complete control and freedom that comes with self-hosting an AI on an old laptop. The atmosphere is peaceful, yet empowering, conveying the message that embracing digital independence is not only possible but also liberating.

Style: Mix of minimalist futurism with nostalgic retro-futurism. Colors: Soothing pastels like light blue, pale green, and creamy white, with hints of warm beige and earthy brown to evoke a sense of comfort and familiarity. Incorporate fine details such as circuit boards, wireframes, or coding symbols to add depth and visual interest.

Dimensions: 16:9 (1920*1080) size illustration, suitable for print media or digital display."

[sterile pendant]

here's the novel you requested

[young blade]

does anyone have a decent image 2 image workflow on flux yet?

and mind sharing?

[errant dust]

So what exactly is this?

[errant dust]

What are these image pairs supposed to be showing? I mean from a technical or creative POV

[sacred jewel]

They show what I like

[errant dust]

sure, but why two of each, and the second being slightly different

[errant dust]

are they modified prompts? Or are they using a new model or setting?

[sacred jewel]

Nope, base workflow.

Flux.1 Dev ... all of them. random seeds, straight CLIP Text Encode

[errant dust]

so the second image is the same prompt but differnt seed only?

[sacred jewel]

correct

[errant dust]

huh. are the prompts so very detailed they bring out nearly twin results?

[sacred jewel]

I guess... I am not much into the technical details as I am abotu the output I like.

The sampler is EULER and the scheduler is BETA if it matters

[errant dust]

That's cool. Could you shre the prompt for one of them?

[sacred jewel]

In some cases, the changes are more obvious... some times, not so much

[sacred jewel]

Sure, here is the one for the last two above

expressionist Dune movie poster, gigantic sandworm in the background, surreal, black and gold bicolor, extremely detailed, pixelated, random style

[errant dust]

Heh. I like it. starts with expressionist, and then you throw the poor thing a curve ball with "random style"

[errant dust]

Anyhow, it is still interesting it deviates so little

[sacred jewel]

Matteo mentioned in his last video how the model is VERY strict... that it does not have as broad a path through the network as SD3 ... if the prompt does not carve a fairly straight path, the model will not pull in from outside the path too much.

[errant dust]

One thing I noticed with Flux is that euler and ddim give identical results

[errant dust]

Might be worth trying to lower the cfg and see how it responds. Nothing catastrophic, but enough to randomize more

like 2.5

[sacred jewel]

I have not been using guidance at all. I did hear that some say a low guidance gives more artistic and higher gives more realism

[errant dust]

Mmmm.... if you lower it too much it begins to produce plain bad images

hence the suggestion of lowerin a bit

and not too much

you can test with a fixed seed

[errant dust]

below a certain point it begins to mess up text a lot

[sacred jewel]

This particular one was very different seed to seed. It certainly isn't always the case of twinsies

[errant dust]

What was the prompt?

[young blade]

@sacred jewel are you using a random prompt picker node with batch set to 2?

[sage burrow]

How do I add a lora node to my flux workflow?

all the premade workflows I download always have so many missing custom nodes....

[bitter hearth]

I haven't started using flux yet but it very much looks like a model that needs a bunch of passes

so on L2 discord they found that you can generate in graphics or artistic style for the first pass, which gives you that deep depth of field

then do a second pass as photograph

and it will keep the deep depth of field of the art work

another person there first generated an image which was just background with no person/subject
and the model made the depth of field deep
then they composited a person/subject and ran pass 2

and the model blended the subject in but didn't reduce the depth of field

[sage burrow]

finally found a workflow that adds loras to flux... guess I'll go out for dinner, or a hike or something now while I wait for it to finish the 1 image lol

[bitter hearth]

I went back to SD 1.5 and its rly fast

the model is aging now but its so much easier for realistic photos of people than SDXL

with SDXL you always have to fight the model to not give the airbrushed skin texture

[sage burrow]

I made a few flux loras today, so I want to tast them out. So much for instant gratification lol

[signal iron]

Hold on. I'll post a basic one in a a minute.

[sage burrow]

Fortunately I finally found one that works 🙂

Works without even having to hunt down a sozen different custom nodes which don't auto install!

https://comfyworkflows.com/workflows/f589b78e-ab81-4f84-8fc2-048a0422d216

[bitter hearth]

I started making characters
they have lots of R2D2s in the IP adapter so their clothes are white and blue
and the background is always a bit like the rococo palace that R2D2 apparently lives in

sorry for slightly NSFW, these models are over-trained on that 🙃

does anyone know what the most censored SDXL model is LOL

[sage burrow]

just go to civitai, set your viewing preferences for sfw, not able to see any of the n sfw stuff, then you will find some

also your prompting matters

what is this BS, I"m not liking flux at all today

[bitter hearth]

ok thanks

yeah I am probably the worst prompter in this community

that prompt was practically just "woman" with a handful more tokens

and most of my image are just "photo of R2D2 in a setting"

[torn wharf]

save image with metadata node f'd something up.

[bitter hearth]

I have a chatGPT thread with a prompt that RX808 gave at some point

and sometimes the Florence 2 node

those do ok

[noble coyote]

XLabsAI has a whole stable of Flux goodies https://discord.com/channels/1271080914692341801/1271086905743638591

https://huggingface.co/XLabs-AI/flux-lora-collection

[sage burrow]

I'm confused though, when people create loras, they are supposed to work though right?

[torn wharf]

i think so

[noble coyote]

Flux.Dev is the only model I can get to work with Flux LoRAs ... and the results are good after waiting a month of Sundays!!! 😄

I'm unaware of any nf4 LoRAs?!?!?!?

Error occurred when executing CheckpointLoaderSimple:

ERROR: Could not detect model type of: X:\ComfyUI\models\checkpoints\AuraFlow-v0.3\text_encoder\Aura-Flow-3-model.fp16.safetensors

???

[bitter hearth]

I'm skipping flux and the new comfy versions for a couple months
to let it "settle down"

[sage burrow]

So I created a lora. The epoch images look so so, but then I ran the lora on civitai and got cartoon like fat people, or really crappy looking furries, all mostly sfw. I'll try again with a couple thousand images or something.

[bitter hearth]

is this considered a good or bad result?

[sage burrow]

considering the amazing images I used to create it I'd say, horrible result 😦

also it's not nsfw at all! Which is pretty amazing considering my training data lol

and they have that faded flux look 😦

SDXL, Pony, and 1.5 are all so much easier to train!

[bitter hearth]

that faded flux look yeah

[junior dune]

How do you train a lora for sd3? I’ve been wanting to do that with some paintings

[sacred jewel]

Nope. Those are all identical settings with two random seeds.

[bitter hearth]

You can make them on Tensor Art if your not parfe with coding

[junior dune]

Could you send me a link on where exactly on tensor art I can input the images to train for the lora?

Nvm found it

[bitter hearth]

lol

[sage burrow]

The full proper onfo isn't out yet to do it decently, best to wait a bit more for that.

[next willow]

👍

[bitter hearth]

I can

https://tensor.art/models/751649208697700869/SD3-Text-Fonts-v1

runs with my Hyper Trained SD3 checkpoint https://tensor.art/models/753956516668680191/SD3-PLATINUM-Dream-Diffusion-By-Dice-V1

four bottles lined up on a table. from left to right, they are numbered "4" then "3" then "1" then "2". from left to right, they are red, blue, green, and orange. the background is a nightclub with a neon sign that says "Dream Diffusion".

[junior dune]

whats this?

[sage burrow]

Does it come out better than just awesome prompting with claud's help?

Is there an online place that does it? 😄 I don't have much vram.

[bitter hearth]

You can run my flux on Tensor Art . they give you 17 uses of flux a day

but to behonest they arnt great renders with there set up

[junior dune]

this is on tensor.art in the training tab

[bitter hearth]

I havent fully tested tensor art SD3 lora maker I just know people have tried it. I use my own programme to train

I'm trying to learn pytorch workflow for fine tuning
its gonna take me a while though

for the VAE?

[uncut river]

yes! I've got flux working local 😄

[bitter hearth]

Comfy Workflow : https://openart.ai/workflows/maitruclam/comfyui-workflow-for-flux-simple/iuRdGnfzmTbOOzONIiVV

Vae : https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main/vae

Clip download clip_l.safetensors and t5xxl_fp8_e4m3fn.safetensors from :

https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main

[uncut river]

using https://github.com/city96/ComfyUI-GGUF and the smallest possible flux model flux1-schnell-Q4_0.gguf

[bitter hearth]

any good?

I run swarm ui with comfy ui as the backend

I even use a simple workflow

But I mostly run all my renders in Forge UI as it now has flux and I run my SD3 Gold checkpoint in forge and auto1111

[noble coyote]

I'm using AuraFlow3 - initially prompt coherence is good; but as a series of images develops, the later images are more creative than the specified prompt. Prompt = a giraffe wearing a red hat. on the right side of the hat is a yellow snail. The giraffe is reading a blue book. There is a green frog on the left side of the book. There is a waterfall in the background

[uncut river]

my first comfyui flux workflow

[bitter hearth]

Thats a cool prompt

[junior dune]

What program if it’s not private? I have a 4070 so I think I could train stuff locally pretty well

[bitter hearth]

img 2 img for flux

[noble coyote]

The naive and elementary prompt style is simply to see how close to the instruction the final image emerges

[bitter hearth]

My set up is still private for training models and loras. But just going through the processes with the leywers atm before i release it as open source

[junior dune]

Gotcha,I’ll be on the lookout for it

[bitter hearth]

vid2vid workflow for FLUX

Should be dropping in a few weeks time

ipiv morph img2vid workflow for FLUX

[sage burrow]

Does anyone have a Flux workflow where one can add a lora? I've yet to find a wirjfliw that works

[bitter hearth]

yes 2 secs

[noble coyote]

Yes, if you use Flux.Dev

[sage burrow]

That's the one I need

Trying to test my crapoy flux lora

[bitter hearth]

dev is the base I hyper trained

[noble coyote]

Flux.Dev and Landscape LoRA - just swap out the landscape LoRA and d/load new ones at Civitai

[bitter hearth]

there is the flux upscaler to

[sage burrow]

Thank you 🙂

[sage burrow]

Thank you 🙂

[bitter hearth]

no worries

https://www.shakker.ai/modelinfo/bbe99aa7e86540a082f1e18b3af9131c?from=personal_page

[noble coyote]

Flux.Dev and Art Nouveau LoRA

[bitter hearth]

this a article i put together on hyper flux https://tensor.art/articles/760080809324596639

[sage burrow]

I trained a flux lora on civitai, but it sucks 😦 Is it likely that my lack of images is the problem, or my using auto tagging, then adding only 5 of my own to each image?

The epochs came out so so, but when using it from civitai front end, it sucks!

[bitter hearth]

I dont use loras anymore bro I have every style all pre prompted https://drive.google.com/file/d/12WJrsnMsVX3qH-fqeZDUxm6bqIEShNZA/view

[lunar rivet]

has anybody managed to convert the kohya loras to something working in comfyui? 😖
the keys are all different

[sage burrow]

Loras are needed for nsfw with Flux tho

[lunar rivet]

huh what

is this some spambot

[bitter hearth]

If you really want to bother with a flux lora . you will need a minimum of 100 4k images to get anywhere near what the base model can already produce without a lora

[uncut river]

it can do some basic nsfw nude stuff

[bitter hearth]

my FLux is NSFW uncensored

So are my SD3 checkpoints

[lunar rivet]

people are training with 512 res images and seeing good effects...

[bitter hearth]

they are untrained waste of time

[lunar rivet]

riiiiight

[bitter hearth]

the base model is trained on 576 x 1024

[sage burrow]

For males, no.

[bitter hearth]

The only real time a lora can be handy with flux is if you have a custom charater

[sage burrow]

Thank you. Fortunately I have a few hundred 🙂

[bitter hearth]

short of that the base will create what you want from a half decent written prompt

[sage burrow]

So, does anyone have a Flux (not nf4) workflow I can use, that includes adding a lora? :>

I need to test my loras, but don't want to use up all my civitai buzz lol

[bitter hearth]

I used to love making loras for SDXL checkpoints I stopped when i found SDXL turbo. I trained that model so hard that dont even need a lora now

[junior dune]

id need to pay 10 bucks to process my training images on the tensor site,bummer i can do it for free or locally

[bitter hearth]

use the img 2 img it has the lora node in it

[junior dune]

i dont mind paying but i literally cant since the payment platform doesnt work

[bitter hearth]

always a catch hey

[bitter hearth]

the lora node is in that workflow Becky

[junior dune]

in any case

since i cant do sd3

how could i train a lora locally for sdxl

i have around 111 training images

[bitter hearth]

Hold tight bro i have 1000 on 1000s of credits, ill see if i can spin you over what you need

[junior dune]

sickk,can we take this to dms then?

[bitter hearth]

has to be done on the tensor discord server .

[junior dune]

alright,1 sec

[bitter hearth]

how much do you need

[bitter hearth]

is that 1800

[junior dune]

just joined the server,1 sec

[junior dune]

yeah

[bitter hearth]

on my alt account i have 4000 so ill just drain that one i dont use that anyway

[junior dune]

:o

[bitter hearth]

my main account has 80k

they give you credit the more people use your models on there

[sage burrow]

Er, where? I figure it's not called lora, but I still can't figure out where to add it 😦

[icy drift]

Testing the effect of prompting Clip_L and T5xxl separately.
Two prompts:
Clip_L
anime girl, red fox ears, holding sign that says "PROMPT", wearing blue kimono with gold stars, with red ribbon in hair, in center of image, on right of image a large white wolf, on left of image a fantasy monster mimic, open treasure chest with teeth on lid and teeth and tongue inside, background stone hallway, crumbling ruins, eerie lighting
T5xxl
An anime girl with red fox ears is holding a sign that says "PROMPT". She is wearing blue kimono with gold stars on it. She has a red ribbon in her hair. She is standing in the center of the image. On the right of the image is a large white wolf. On the left of the image is a treasure chest that is open, and around the rim of the chest's lid are razor sharp teeth. Inside the chest is a large tongue and more monster teeth. The chest is a mimic, a fantasy monster. In the background is a crumbling stonework hallway in an eerily lit dungeon.

Renders are seeds 1000 to 1003, all settings held constant. Only prompts vary.
Seed 1000 and 1003 are perfect in every way.
1001 is missing the hair ribbon.
1002 has too many fingers.

[junior dune]

so,where do we go from here?

[icy drift]

Next, both Clip_L and T5xxl are prompted with full english sentences, using the T5xxl prompt given above.
Prompt adherence drops, but more importantly, seeds 1001 and 1003 are now utterly mangled abominations.

Conclusion: Giving Clip_L full english sentences will result in at least a 50% drop in quality across general knowledge. 😐

[bitter hearth]

call my user name in tensor discord chat

[icy drift]

Next, both Clip_L and T5xxl are prompted using only comma-separated descriptors.
Prompt adherence drops to 25%, and mangled abominations still emerge.

Conclusion: Giving T5xxl comma-separated descriptors will cause Flux prompt adherence to fail.

[junior dune]

Am I in the right server?

[noble coyote]

LoRA normally feed-in from Checkpoint MODEL-out; and LoRA-out to KSampler MODEL-in

[bitter hearth]

what version t5xxl are you using... 8 or 16

[icy drift]

16

[bitter hearth]

damm i cant find it in my list now bro

[bitter hearth]

and what vae are you using?

[icy drift]

It literally doesn't matter. I'm using the same settings across all these tests. I'm only varying the prompt.

[junior dune]

Where do we talk then?

[bitter hearth]

I think i may have been booted from the tensor server as i cant get in ther for some reason

says try later

lol wtf

[junior dune]

:/

[bitter hearth]

ill keep trying bro and when i do ill shout you up

[junior dune]

Sounds good,thanks man

[icy drift]

Finally, Clip_L and T5xxl are given the same prompt as a concatenation of full English sentences followed by comma-separated descriptors.
Prompt adherence reemergers, and seed 1003 is perfect.
But seeds 1000 to 1002 again feature unusably mangled forms.

[noble coyote]

Nice to see the word "beauty" accepting a male image - AI's getting diversified (AuraFlow3) - prompt = beauty style of michael parkes eric ravilious rob gonsalves remedios varo arunas rutkus rafal olbinski venice carnival grisfest

[bitter hearth]

I still cant get my head around how today the best model still uses Euler sampler lol

ill run your prompt on my flux hyper

just read your prompt lol I cant make any of it out lol

[icy drift]

Conclusion:

Using a unified prompt for both Clip_L and T5xxl reduces Flux's overall quality by 50% to 75%, while mangling forms. This happens regardless of the format of that prompt.
😨

Realized this might be happening when I tried using Forge for the first time yesterday, which forces you to use a unified prompt. I was like, "Why is Flux suddenly awful?" Tested now to figure out if it was just my imagination.
It's totally real.
And Comfy still doesn't work on my phone... Only Forge.

[bitter hearth]

its running now

doing 2 renders

[sage burrow]

So add a flux lora node (doesn't come up in search), then add a ksampler?

[noble coyote]

Doesn't my png contain the workflow?

[bitter hearth]

thats what i got from your prompt

Good morning sd3 channel

[noble coyote]

Yes, it's not often the word beauty gives a male image (except if you use the Hindi word for beauty - then you get 50% men and 50% women!!!)

[bitter hearth]

hi 4GB vram lol

[sage burrow]

So nf4 works with Dev loras? (why did I not try this already?!)

[noble coyote]

No, nf4 no LoRAs yet. My landscape image (based on Dev) uses a Flux.Dev LoRA - the workflow shd be contained in the png

[bitter hearth]

its right there

[sage burrow]

Your ladscae image workflow

[bitter hearth]

img2img flux workflow

[noble coyote]

Anyone have a Florence-based w/flow I can have? 🙂

[bitter hearth]

Florence-based

[noble coyote]

Have you a png with the Florence w/f embedded?

afk

[bitter hearth]

I have the Json workflow

download that to your workflow folder

also if you want the Vid2VId flow download that to

that is mega handy save all the prompt data to image so you can upload your image to site with all the nodes and settings

as you know its issues when you gen on comfy and enter a site comp without the data your image gets regected , use that and your golden

[sage burrow]

Put it in where it says florence? My list of loras doesn't come up (only florence ones). So would I replace that node with a Flux lora loader, or would I just put my loras into the florence folder?

[bitter hearth]

(tried to imitate those Facebook gifs lmao)

[bitter hearth]

switch your loras to where the florance loras are saved to

[sage burrow]

Ty 🙂

and where do you guys all find the flux lora node? What is it called? I keep trying to build my own workflow but regular lora nodes don't work, and when I search for flux ones... nothing??????

[bitter hearth]

the text adherance in my hyper flux is crazy

[sage burrow]

I'm skeptical that you used SD3 to make that 😉

[bitter hearth]

render 4 images and 1 will be slightly off

[noble coyote]

Totally Rad!!! Ideogram - eat your socks!!!

[bitter hearth]

lol nom nom

[bitter hearth]

Took 2 tries, one had "again" at the end twice, like the model was mocking me

Does shakker have sd3 8b

A certain site is limiting me at 20 gens per day

Thats my hyper trained SD3 Gold checkpoint https://tensor.art/models/751454663859189150/SD3-GOLD-Dream-Diffusion-By-DICE-v1

[noble coyote]

Some of mine - wouldn't say was as easy as Ideogram - some took over 10 tries!

[sage burrow]

That's only bc you haven't applied for the glif creatore role. You get 200 per day that way 😉

[bitter hearth]

shaker charge 140 for 1 sd3 render

and 180 for 2 flux renders

https://tenor.com/view/they-kinda-pricey-dante-dangelo-theyre-kinda-pricey-theyre-a-bit-expensive-too-pricey-gif-26721910

[noble coyote]

I have a yearly sub for both MJ and ClipDrop - but desktop SD is so good now - I hardly use them!

[icy drift]

These two images say it all.
Can you tell which one tried feeding the same prompt to both text encoders?
Both images used identical seed and render settings.

[bitter hearth]

I'm useless I do nothing but gen balls half the time

[icy drift]

I posted the results to my forge issue on Github.
https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1182#discussioncomment-10357729

[noble coyote]

The first one got the hound to stick its head up its own a**!!!! LOL

[bitter hearth]

Lmao

[icy drift]

General domain knowledge drops by 50% to 75% if you use the same prompt for both text encoders in flux. 😕

[sage burrow]

I don't think there's any ball glifs yet. All you have to do is find a glif that you like, hit the remix icon at the top, then type in your balls prompt instead of their boaring whatever prompt. Then voila, everyone who uses your glif, gets balls.

[noble coyote]

OK, I changed the prompt slightly in AuraFlow3 = ravishing female beautiful woman style of michael parkes, eric ravilious, rob gonsalves, remedios varo, arunas rutkus, rafal olbinski, venice carnival, grisfest

[bitter hearth]

lol i try that prompt

[noble coyote]

Its good to see diversity in language working itself into the output

[bitter hearth]

ravishing female beautiful woman style of michael parkes, eric ravilious, rob gonsalves, remedios varo, arunas rutkus, rafal olbinski, venice carnival, grisfest <<< are they artist names in that prompt?

[noble coyote]

Yes, living artists

[bitter hearth]

as they are very stylised

[noble coyote]

Some places (SD3 included) will strip-oiut living artists names

[noble coyote]

If you take out Venice Carnival - mebbe less stylised

[bitter hearth]

Yes some stabilty ai models were set to remove famous people to protect the deep fake shananigons lol

trying now

[noble coyote]

The lady with yellow gloves has 8 fingers on her right hand?!

[bitter hearth]

lol she has spider virus

[noble coyote]

Are your images AuraFlow or Flux?

[bitter hearth]

without the venice carnval bit

[noble coyote]

Less brocade, possibly

[bitter hearth]

mine are all my trained hyper flux

[noble coyote]

OK

[bitter hearth]

running in forge UI

thats also with the t5xxl fp8

[noble coyote]

Arunas Rutkus has a very beautiful and delicate look - mebbe drop him from the prompt for something with more va va voom?!

[bitter hearth]

its just so easy to create what you want now . Finially SD got to where it needed to be

this is with my flux model

[noble coyote]

Can we d/load a copy of Flux.Hyper?

[bitter hearth]

yes ofc bro

https://www.shakker.ai/modelinfo/bbe99aa7e86540a082f1e18b3af9131c?from=personal_page right there

[noble coyote]

Thanks

[bitter hearth]

check the gallery as some have prompts since i started using forge ui

[mortal kite]

hehehe

[bitter hearth]

love a good car render

[mortal kite]

snu-snu

"2d flat vector pixel art of a full set of 52 playing cards laid out in a grid pattern,"

decently smart

[bitter hearth]

flux hyper and unreal engine 5

[mortal kite]

Flux can't do ASCII art, though. At all.

[bitter hearth]

Cool prompt but

@sage burrow 👍runs away

[mortal kite]

hehe didn't mean fo rit

[noble coyote]

I went to the library and asked the lady where the books on paranoia were?!

Behind you! Behind you!!! she said

[sage burrow]

What did you use to make it? (hopes that wasn't my lora lol)

[mortal kite]

Chuthulu softens the rift to the void

[bitter hearth]

ASCII art, lol handy if your stuck in the 80s bro lol

[mortal kite]

vector logo 😄

[bitter hearth]

it will do vector logos

[mortal kite]

Flux at least knows what Nikola Tesla looks like. SD3 was a massive fail for that

[mortal kite]

ya bro

[bitter hearth]

ill give you a simple vector logo prompt

[mortal kite]

I redid some of my SD3 prompts in Flux to see how they would turn out. In most cases, 200% better

[bitter hearth]

Vector logo of a bear, vector logo art style, on solid color background, vector logo symbol, stylized logo, logo icon, minimalist, clean, clean and scalable designs, precise lines, professional-grade execution, emphasizes simplicity and versatility
Negative prompt: signature, shading, gradient, painting, watermark, photo, photorealistic, realism, ugly, off-center, deformed, 35mm film, dslr, cropped, frame, worst quality, low quality, lowres, JPEG artifacts
Steps: 30, Sampler: Euler, Schedule type: Simple, CFG scale: 1, Distilled CFG Scale: 4, Seed: 3907803808, Size: 576x1024, Model hash: faf2118042, Model: Dream_Diff_Flux_V1, Version: f2.0.1v1.10.1-previous-297-g5fb67f49, Module 1: t5xxl_fp8_e4m3fn

[mortal kite]

why would you give that to me, couldn't you do it yourself?

[bitter hearth]

oh mate flux makes sd3 look like sd1.5

thought you wanted to try it lol

[mortal kite]

Shoggoth

[mortal kite]

Flux doesn't really need all that comma separated stuff..

[bitter hearth]

also dont need negative prompts

[bitter hearth]

Nai 👍

but my vector was from a copy paste from a sdxl model

[mortal kite]

if you turn the Flux "guidance scale" down some, it can result a bit more "realism" look

things will look less plastic/shiny

[bitter hearth]

that red truck looks more real yeah

[mortal kite]

except for the creepy log arm that is coming out from under it

always have to try the hamburger

[bitter hearth]

I wonder if Perturbed Attention Guidance would work with flux if ported right

cos it would fix small stuff like that

[mortal kite]

straight up SD3 prompts right into flux, usually perfect result the very first render I have an even better one than this but its more nsfw...

[bitter hearth]

lol

[mortal kite]

hubba

[bitter hearth]

lol

[mortal kite]

most women choose the bear dude

[bitter hearth]

lol

[mortal kite]

jussst a little snack

[bitter hearth]

sharp asf

[mortal kite]

Who needs a crescent wrench when you can have a crescent wench?

[bitter hearth]

Do you like making real scene movies into anime style ?

[mortal kite]

meh not necessarily

seamless tile

[bitter hearth]

I made a lightning checkpoint that smashes it , really funny results

[sage burrow]

I didn't add enouch images to my lora when I made it (at least I'm hoping that's the problem anyways), so it spits out crap like this 😦

[bitter hearth]

lol

bet the result you want will be without the lora

[noble coyote]

I need a bigger rig - AuraFlow 16Gb, Flux 21Gb - sheesh!!! 8Gb VRAM is go-karting@Monza!!! 😦

[bitter hearth]

install my version of flux its 15GB

[noble coyote]

Will do, just trying out AuraFlow3

These 4 are AuraFlow3

[bitter hearth]

yes its a nice model aura, how do you find the text adherence ?

[errant dust]

It has rough edges. It is not so much the adherence that is questionable, just its ability to do some things. But to be fair it is in beta, and it has improved a LOT since 0.1 six weeks ago. My 2 cents

[errant dust]

so here are a few images of same settings except with a lowered CFG. The first is default CFG 3.5, the second is CFG 2.5 and the third is CFG 2.0.

[noble coyote]

At the start of a sequence of images, prompt coherence seems good; but as the images output increases, prompt coherence seems to fade ... ?!

[rain current]

nf4 <> gguf

[noble coyote]

AuraFlow3 using 1.73 Model Sampling

[errant dust]

guessing prompt was: "My train trip on acid"

[noble coyote]

dancing beauty peruvian arpillera and a painting of the train under a sky, in the style of psychedelic dreamscapes, illuminated visions, i can't believe how beautiful this is, moody color schemes, chicano art, whistlerian, trace monotone , extreme cosmic texture, nebulous Oort Cloud, galaxy stars neptune Saturn mercury Jupiter mars moon venus that glow psychedelic galaxstars Fantasy, Mystical: Nikon D6 with Nikon AF-S NIKKOR 24mm f/1.4G ED lens

1280x720

[crimson salmon]

a dog

[errant dust]

Could have summarized it to just "word salad". 😄

[noble coyote]

Its one of my favourite prompts - even looks majestic in SD3 Medium!!!

With the advent of t5xxl_fp16 I have upped my prompting game away from 'lists' to a more natural language

And even if not using t5xxl my prompts keep a level ...

But word salad sometimes works 😉

[mortal kite]

word salad works for Kamala 😄

[bitter hearth]

you can fix that if you use forge

[rain current]

gguf-q8 (+celeb lora) - 24steps

[noble coyote]

I have ComfyUI, A1111, SD.Next, SwarmUI, Fooocus and Omost - I think Forge would bring 'overload'

😄

[noble coyote]

AuraFlow3

[mortal kite]

that just looks nasty eww lool

[rain current]

[mortal kite]

random?

[bitter hearth]

might aswell get the full set lol

[rain current]

Several different prompts together, a madness

[mortal kite]

ah another day in Chicago

[mortal kite]

lol

[alpine summit]

Sd1.5 Rpg4

[mortal kite]

nice movie still

[noble coyote]

SD3 Gold checkpoint via SwarmUI

[alpine summit]

Sd3

Sd1.5

[noble coyote]

SD3 Gold safetensor

[alpine summit]

Sd1.5

[alpine summit]

Sd1.5

Sdxl

Sdxl

[bitter hearth]

noooice what check point you on for that bud

[alpine summit]

Sd1.5

[noble coyote]

Looks like something by Clownshark!

[bitter hearth]

sd3 likeness check point https://tensor.art/models/751438781070140461/SD3-LIKENESS-Dream-Diffusion-By-Dice-v1

all that lot from a hyper trained SDXL turbo base checkpoint

and still SD3 cant match them renders

[sage burrow]

My flux lora with 64 images didn't work out very well, so now, I'm using 758 images!!!!!

[alpine summit]

Sd 1.5, simplemix

[bitter hearth]

what you training your lora on ?

[bitter hearth]

i remember the simple mix sd models , the word mix was everywhere in the day of 1,5 lol

[sage burrow]

The only thing Flux can't do: 🥒

It's only $2 per try with civitai 😄

[bitter hearth]

lol cucumber lol

[sage burrow]

Trying to keep it PG for discord lololol

Well futa furries to be precise 😉

[bitter hearth]

oh civitai flux lora maker . lol thats not worth the paper its written on. epoch is so off its bonkers

do you have a webui like auto or forge?

[sage burrow]

the epochs don't turn out too terribly, but they don't look like the resulting images. So I"m seeing if adding MORE images works

[bitter hearth]

if you try what im gonna suggest you notice how off civitai is

but you need a webui

[sage burrow]

I'm using comfy locally, but I only have 8gb gpu. I'm tempted to just use a rented cloud gpu service, but worried it'll take me days to figure out how to setup , or I'll forget to turn it off LOL. Google loud is ridiculously easy BUT I've read too much about never being able to land a decent GPU even on paid

[hallow lion]

Lmao. This is the Internet in a nutshell.

[sage burrow]

CIvitai claims to use Kohya...

[hallow lion]

Actually its the image above with the cat in the kitchen but im tired of battling discord reply system

[bitter hearth]

if you have auto or forge you can install the extension called dream booth. and train the lora in there, takes approx 15 to 20 mins and way better than civitai lora training

[sage burrow]

You mean train using somehting other than my own gpu? I did use everydreamtrainer before and liked it, and my computer nearly died with a 16 imag checkpoint training ROFL

[alpine summit]

Sdxl, leonardo

[bitter hearth]

if you install it id suggest watching a youtube video tutorial for the first run as it looks very dawnting at fiirst. but its very easy

[sage burrow]

Which is better/easer/closer to comfy?

[sage burrow]

Do you mean using a cloud service to install forge, then run dreambooth? My computer only dreams of running such things.

[bitter hearth]

i only use 2 webui now days thats swarm with a comfy ui backend and forge , forge update now runs sd3 and flux

forge local

[sage burrow]

Can dreambooth run via comfy? 😄

but I'm guessing I need to run dreambooth externally

[bitter hearth]

ive honesty never looked in to that as was happy witht he forge way

[alpine summit]

Sd 1.5, dreamshaper8

[bitter hearth]

https://dreambooth.github.io/ https://jimclydemonge.medium.com/how-to-run-dreambooth-locally-a-step-by-step-gyu-88c028ab01a4

i tested it from interest and was impressed as it free and user friendly

[sage burrow]

8gb GPU here....

[bitter hearth]

where as training in civitai they store all your training images

[bitter hearth]

you might just squeeze in there

[sage burrow]

Oh, requires only 10gb vram, close enough

[bitter hearth]

im surprised its not as popular as it should be

[sage burrow]

I'll start small 😄

[bitter hearth]

thats what she said

[sage burrow]

Last article i read said 16gb GPU....

for dreambooth

[bitter hearth]

i never checked i just added as an extension in auto and forge

[sage burrow]

Need to find a paywall bypasser grumble

[bitter hearth]

for what?

[sterile pendant]

For those having a hard time understanding Flux's base/max_shift values and how they actually interact with the calculations, I threw together a quick interactive Desmos calculator for it. Basically, if you stay within 1024^2 pixels, base shift will do absolutely nothing and it only applies when the width * height product is not equal to 1024^2. So for instance, if you picked resolutions like 768x768 or 1280x1280 or some other combo that the product is not equal to 1024^2.
https://www.desmos.com/calculator/c0jburw7z4

[bitter hearth]

hi RX808, flux base on both models are trianed with 576 x 1024 if that helps you

[alpine summit]

Sd1.5, dreamshaper8

[sage burrow]

Your link, is paywalled

[bitter hearth]

(why hope it's not your lora?)

[sage burrow]

Because it's missing some key features of my lora!!!!!!!

[bitter hearth]

I won't even ask what

runs

[uncut river]

running up becky's leaning curve

[alpine summit]

Sd1.5, simplemix

[sage burrow]

Now I get to learn how to run comfy and auto111 on the same computer..... Apparently I should have read more articles before it decided to reinstall torch (and prob going to do python and cuda next :P) LOL

[noble coyote]

Symlink your custom nodes and models and loras folders

So that A1111 and ComfyUI share the same folders

[sage burrow]

Oh ty

[uncut river]

yeah, i do that

[sage burrow]

maybe I'll delete it and start over again lol

[noble coyote]

mklink /D "Empty File" "File to be used"

[sterile pendant]

honestly, i wouldn't even bother with symlinks. both have yaml files where you can specify locations

[sage burrow]

Do you happen to have a link? GOogle is useless, it's giving me comfy vs auto111 links 😛

[noble coyote]

I can never get SMUDGE to work - as I have all the models I need alreayd d/loaded - yet a new installation will insist on d/loading its own models!!!

This is how to do a Symlink - open a CMD terminal

mklink /D "Empty-folder-any-name-you-choose" "Models folder (full path"

[sage burrow]

should I uninstall auto111 first? Or can I just make them move and delete as needed?

[noble coyote]

All my models are stored under ComfyUI - so I get my A1111 folders to point to the A1111 models folder via a symlink

[alpine summit]

Sd1.5, rpg4

[sage burrow]

I only installed auto111 so I could run dreambooth via it

I did look up comf and dreambooth, but I don't know how to make that comfy mode start/work/etc.

[uncut river]

flux1-schnell-Q4_0.gguf running local with 12gb vram and 16gb ram, quite old pc AMD Ryzen 5 1500X

[bitter hearth]

oh i did look i just grabbed it. its not when running as an extension tho

[alpine summit]

Sdxl, leonardo

[sage burrow]

my backup plan now that I have 750 images: 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

[sacred jewel]

I would run the portable versions of each... save yourself a lot of grief.

[sage burrow]

I am running the portable version of comfy.... oops didn't kmmow there was a portable 111!

[noble coyote]

I think A111 is always portable - it has a venv

[uncut river]

though you can run it without venv, which is what I did

now running comfyui and never go back to a1111

[sage burrow]

I only installed 111 for one reason, dreambooth 😄

[uncut river]

oh, no dreambooth nodes for comfyui?

[sage burrow]

There is a custom node, but I have noooo idea how to make it run

[sacred jewel]

There was someone who created fork of the official one and made it PORTABLE... but yeah, the default one is a VENV ... however it still uses your global Python install I believe

[sage burrow]

so I'm just using the how to for 111

ooooh, a dreambooth training version optimized for low vram! 🙂 🙂 🙂

[bitter hearth]

lol

[noble coyote]

I often run my more successful ComfyUI Prompts thru A1111 - the results are often amazing as A1111 processes noise differently

[uncut river]

just random chance

indeed there is a difference between a1111 and comfyui where they get they're noise

[sage burrow]

more so than changing the seed?

[sacred jewel]

CPU vs GPU... noise is noise and it is always random... not sure why anyone would think one random number is better than the other [all the time]

LUCK has more to do with it and the actual processing pipeline, not the seed source itself

Yes, A1111 and Forge differ in how they process the pipeline from ComfyUI... very different results with the same prompt

Some nodes in Comfy try to approximate A1111 but it is never 100%

[bitter hearth]

the noise was made completely different with SD3 tho

[sacred jewel]

Yes, SD3 latent is different

However, you can use a SDXL latent with SD3 but not the other way around...

[bitter hearth]

Noise Predictor estimates the amount of noise in the latent space and subtracts the from the image. This process is repeated for a specific number of times, reducing noise according to user-specific steps. Older models like 1 and 2 use the U-Net Noise predictor architecture. On the other hand, SD3 uses a repeating stack of Diffusion Transformers meaning it uses multiple transformers for the diffusion process when compared to previous SD models.

[sacred jewel]

Upscaled using McBoaty Large Refiner

[sage burrow]

@bitter hearth that how to link you gave me is the best and easist how to I have ever used! 😄

[sacred jewel]

original

[bitter hearth]

lol are you loving it lol

[sage burrow]

Flux isn't listed under model type 😦

[bitter hearth]

your having a right mare hey, is sd3 in there

[sage burrow]

It is not either

looks for update things buttons

[sage burrow]

No nodes and no spaghetti, this looks so easy!!!!!

[bitter hearth]

worms and nodes, the bane of ever coders life

[sage burrow]

Why did I not start with 111 months ago? LOL

I prob put the checkpoints in the wrong folder, seems to not show my flux one

[bitter hearth]

flux is a Unet

not a stable diffusion folder

[torn wharf]

soem of the comfyui nodes want check points in /models/stable diffusion, and some want checkpoints in a specific comfyui folder outside of swarms called /models/checkpoints, and then some want them in a folder called /models/unets

at this point i'm just symlinking everything to one folder. i dont' understand how they couldn't have standardized this one part of comfyui across many nodes. and when a node is hard coded to look in a specific folder, there's NO indication of where that folder is in the UI or how to get there.

you have to recode the node to change the folder it wants

text encoders too. some want them all in clip. some look for t5 elsewhere. i think that's gotten better but i remember things were hairer in that regard once before. nodes looking for diff folders. it's a catastrophe

[sage burrow]

Hole up, is 111 10x faster for flux images than comfy? I'm only doing 512X512, but still....

[bitter hearth]

or just drop the json in the folder its current path is set to

i find forge renders flux images quicker than comfy

[torn wharf]

new version of forge is snappy fast

[bitter hearth]

the first render of flux always takes a bit longer anyway

[torn wharf]

always do a warmup render so that memory can prime. any model will suffer that

[bitter hearth]

thats my flux hyper running in forge

[torn wharf]

the one thing i dont' like about forge is that it has no way to prompt the tencs individually. less of a problem with SDXL but it was still hurting it back then too. i would've figured that webui would have something to feed different prompts to different tencs by now

[bitter hearth]

the update to forge lost a few good extensions tho which was a shame

[torn wharf]

the forge update shouldn't have been a standard release and kept as a dev branch for the time being. it broke so much

i think he wanted to clean sweep all the other branches that were "taking over" though

[torn wharf]

made a right mare

[sage burrow]

It definitely isn't really running flux, it can't to txt

[bitter hearth]

it seems as it stands at the moment the forge latest will not port any form of video generation extensions

[torn wharf]

yeah i don't get that. why do comfy nodes look for flux in a Unet folder? why do none of the previous unet models need to look there? what's with the split standardization

[sage burrow]

So I was supposed to create a unet folder and put flux in there?

[torn wharf]

the forge version of animatediff is busted too anyways. it isn't kept up to date with the main branch and can't do things like sparsectrl or the frame consistency passes

animatediff for forge has been busted and left to decay for many months

[bitter hearth]

in forge you put the flux checkpoint in the stable diffision folder in models webui_forge_cu121_torch21\webui\models\Stable-diffusion

in comfy it goes in unet

the flux Vae just goes in the Vae folder along with the T5xxl

[torn wharf]

i feel like i'm being gas lit the past couple weeks. i keep hearing things like "flux is a unet" but .. i don't understand where a unet fits into a DiT block

i just trust that people who make UI's know that flux is a unet better than me i guess

[bitter hearth]

I use the turbo samplers for all my trained SDXL checkpoints to find auto 1111 did have them. so i got to work on re scripting auto 1111 and found out i could add them so now all my turbo samplers are in my places lol

[torn wharf]

all my models are in teh one stable diffusion folder, in sub folders to organize them. then if a ui wants another destination , like comfy nodes needing a very specific folder name that doesn't seem to relate, i just symlink them. explorer extensions help this process greatly. this old gooder. basically as essential as a zip utility on your system these days imo. https://schinagl.priv.at/nt/hardlinkshellext/linkshellextension.html

[bitter hearth]

if you want to add them here is how https://youtu.be/zu3x2DUyng0?feature=shared

[sage burrow]

Like this? I wasn't sure if the " " are needed or not, so I left them in for first try. I also made something up for the unet based on the others, prob wrong LOL
.@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=

--ckpt-dir "C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models"

--vae-dir "C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\vae_approx"

--lora-dir "C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\loras"

--unet dir "C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\unet"

call webui.bat

[bitter hearth]

yes thats a sync

[sage burrow]

I should add freeu dir as well 😄

[bitter hearth]

but you may find the checkpoint needs to be moved to the normal place for forge to action it

[torn wharf]

Everything i dig up about DiT models say that the DiT block replaces the Unet. UGHGHGUh i'm going crazy!! why did comfyanon require a specific node for flux that looks in a unet folder?

why is flux a unet!?!

[bitter hearth]

i trained a DiT checkpoint https://www.shakker.ai/modelinfo/c0a9f61d7b0e495ca3b14148e6098b04?from=personal_page

[sage burrow]

THere's this super long scholarly article on why unet is better... ? I don't remember the link now tho

[torn wharf]

quotte from the DiT paper. "With this work, we aim to demystify the significance of
architectural choices in diffusion models and offer empiri-
cal baselines for future generative modeling research. We
show that the U-Net inductive bias is not crucial to the per-
formance of diffusion models, and they can be readily re-
placed with standard designs such as transformers"

REPLACE THE UNET. why is comfyui insisting that the flux weights go in the unet folder!? i'm breaking down into a mental panic here. the world is inherently untrustable. (dont worry i'm not really. i'mjust being dramatic for effect)

[bitter hearth]

and Ripping SD3 down to run on auto 1111 was great run https://youtu.be/FrVITtD0q_Y?feature=shared

[torn wharf]

i think i was once in here talking about how unet models feel dated compared to new transformer models, and someone here was telling me that transformers are unets and unets use transformers. This whole discussion literealy has me questioning my sanity and thinking i have no research skills and am just way out of my depth

[sage burrow]

"'--ckpt-dir' is not recognized as an internal or external command,
operable program or batch file.
'--vae-dir' is not recognized as an internal or external command,
operable program or batch file.
'--lora-dir' is not recognized as an internal or external command,
operable program or batch file.
'--unet' is not recognized as an internal or external command,
operable program or batch file.
venv "C:\Users\oipte\Downloads\stable-diffusion-webui\venv\Scripts\Python.exe"

INCOMPATIBLE PYTHON VERSION
"

[torn wharf]

I don't bother with sd3 in a1111 or forge. both of them want to use more vram than i got and can't efficiently operate. so i use comfyui instead

[torn wharf]

looks like the forge project killed all that command line configuration

[sterile pendant]

Unet just became the go to term for the model portion of things since it's all there really was until dit models came out. For dit models, it's called a transformer.

[torn wharf]

i know what a unet is. i'm just going crazy trying to find out why comfyui workflows require people put flux models into a unet folder. in the past i've asked about that here, i forget who but i was told that unets are transformers and the new DiT models still have unets.

i feel really gaslit because i was confident that unets and DiT models are entirely different architectures

Then flux came out and comfyui required it to be saved into a wholly different /unet/ folder that i've never seen used before

[bitter hearth]

as a new comfy install you have to add the unet folder yourself to lol

[torn wharf]

oh ok

i just created a symbolic link to the /models/stable diffusion/ folder instead

my confusion still persists